Two years ago, I found that the Gen AI value chain was inverted: the compute layer captured ~83% of all revenue and ~87% of all gross profit. The application layer, despite being closest to end customers, earned almost nothing. I predicted this would flip over time, following the pattern of every prior platform shift. Two years ago, I said:

Two years since, the AI ecosystem has grown roughly 5x, from ~$90B to ~$435B in annualized revenue. But what’s remarkable is how little the shape of economics has changed.

The bottom line upfront: Semi is a one-player game. Apps is a two-player game. Infra is the only competitive layer. The most profitable strategy in AI is still selling the shovels. :)

Let’s dig in!

#1 The value chain is still inverted

Here’s how each layer breaks down:

• Semi (~$300B): This is overwhelmingly NVIDIA. Their data center business did $62B last quarter, annualizing to ~$250B. Broadcom’s AI semi business (custom accelerators for Google, Meta, ByteDance) adds ~$34B. An estimated ~$25B in high bandwidth memory purchased directly by hyperscalers for custom chip programs rounds it out. The concentration is extreme: NVIDIA is ~80% of the layer.

• Infra (~$75B): AI-specific cloud infrastructure. Azure, AWS, GCP, Oracle each contribute $10-20B in AI-attributable revenue. CoreWeave adds ~$6B. Baseten, Together, Modal and other inference providers collectively make up the rest. Unlike the other layers, this one is relatively evenly distributed across the major clouds.

• Apps (~$60B): Also extremely concentrated, but in two companies. OpenAI and Anthropic together make up ~45B annualized revenue and are ~75% of the layer. A distant third are the coding AI players like Cursor, followed by the fast growing agent companies like ElevenLabs, Glean, Sierra, Perplexity, Replit, Lovable, Harvey, Abridge, etc..
  

Bottom line: semi layer still captures ~70% of all AI revenues. In the cloud stack, semis capture 1/10th that share: ~8%. That gap is the core distortion between this supercycle and the prior one.

#2 AI grew 5x but NVIDIA captured most of the value

Despite 5x growth in the overall ecosystem, the shape hasn’t changed much. The app layer grew fastest in percentage terms (12x in two years). But on absolute dollars, the semi layer added ~$225B vs ~$55B for apps. NVIDIA alone added $175B of incremental revenue. That’s roughly 3x the size of the entire app layer today.

The profitability picture is even more skewed. Semis run at ~73% gross margins. The infra layer runs at ~55%. Apps run at ~33% based on my estimates.

Multiplying through: the semi layer earns ~$225B in gross profit. Infra earns ~$40B. Apps earn ~$20B. Semis capture 79% of all gross profit dollars in the AI ecosystem. For comparison, in the cloud stack apps capture 70% and semis capture 6%. The AI stack is almost exactly the mirror image.

Over two years, infra and apps each gained about 4 points of profit share. Semis went from 87% to 79%. At that pace, it would take well over a decade for the app layer to reach the kind of share that apps enjoy in cloud.

#3 The capex question: is it worth it?

The top five hyperscalers spent ~$443B in capex in 2025, up 73% from $256B in 2024. That number is projected to exceed $600B in 2026, with roughly 75% (~$450B) directed at AI infrastructure. Which raises the obvious question: is all this capex generating positive ROI?

The CEOs are betting it’s worth it:

• Andy Jassy, Amazon (Q4 2025 earnings call, Feb 2026): “As fast as we’re adding capacity right now, we’re monetizing it.”

• Sundar Pichai, Alphabet (Q4 2025 earnings call, Feb 2026): “We are seeing our AI investments and infrastructure drive revenue and growth across the board.” Though he also acknowledged “elements of irrationality” in the current scale of AI investing.

• Mark Zuckerberg, Meta (Q4 2025 earnings call, Jan 2026): “I think it’s the right strategy to aggressively front load building capacity. In the worst case, we would just slow building new infrastructure for some period while we grow into what we build.”

Here’s what they are buying with all that money. The chart below shows cumulative AI compute capacity by chip type. The answer is overwhelmingly NVIDIA, though non-NVIDIA silicon is starting to show up at the margins:

Source: Epoch AI. Cumulative compute capacity measured in H100-equivalents.

Every major hyperscaler is hedging by developing custom silicon. Sorted by maturity:

• Google TPU (most mature): 7th-gen Ironwood now generally available. Anthropic ordered up to 1 million TPU chips, worth tens of billions. Google is now selling TPUs as merchant hardware. Competitive pressure from TPUs has reportedly forced NVIDIA to cut pricing ~30% for some customers.

• Amazon Trainium (scaling fast): 1.4 million Trainium2 chips deployed, powering majority of Bedrock inference. According to Amazon, custom chips business crossed $10B annual run rate, growing triple digits. Trainium3 (3nm) in production, nearly all supply committed through mid-2026.

• OpenAI custom ASICs (new entrant) + AMD: Signed multiyear deal with Broadcom for 10GW of custom accelerators starting 2026. Separately signed 6GW deal with AMD for Instinct MI450 GPUs.

• Microsoft Maia + AMD: Deployed Maia 200 in Azure (Jan 2026). Claims 3x performance of Trainium3 on inference. Now powering a portion of ChatGPT workloads. Nadella said Microsoft will still buy from NVIDIA and AMD alongside Maia.

• Meta MTIA (internal only): MTIA v3 deployed for internal inference. Acquired chip startup Rivos (Sept 2025) for RISC-V designs targeting 2026.

Jensen Huang has dismissed custom ASICs as “noncompetitive,” noting “a lot of ASICs get canceled.” He’s not wrong historically :)

#4 The stack will eventually flip. It may take longer than 10 years.

I still believe the stack flips eventually. I’m just less sure about when.

For the shape to invert, either the app layer needs to keep exploding or the semi layer needs to get cheaper. Both are happening, but neither fast enough to change the shape within this decade.

The cloud stack took roughly 15 years to evolve from hardware-dominated to software-dominated. The AI stack may follow a similar timeline. The custom silicon efforts are the key variable. If TPUs, Trainium and the other ASICs succeed at scale, they could compress NVIDIA’s margins and shift profits up the stack. But outside Google’s TPUs, no custom silicon program has proven it can compete with NVIDIA at scale for training workloads.

The biggest open opportunity remains in the application layer. That’s where I spend most of my time. But investors and builders should be clear-eyed about where the economics sit today. The most profitable company in AI is still the one selling the shovels.

For the underlying data behind this analysis, ping me. Happy to share the full datasheet.

The information presented in this newsletter is the opinion of the author and does not reflect the view of any other person or entity, including Altimeter Capital Management, LP (”Altimeter”). The information provided is believed to be from reliable sources but no liability is accepted for any inaccuracies. This is for informational purposes and should not be construed as investment advice or an investment recommendation. Past performance is no guarantee of future performance. Altimeter is an investment adviser registered with the U.S. Securities and Exchange Commission. Registration does not imply a certain level of skill or training. Altimeter and its clients trade in public securities and have made and/or may make investments in or investment decisions relating to the companies referenced herein. The views expressed herein are those of the author and not of Altimeter or its clients, which reserve the right to make investment decisions or engage in trading activity that would be (or could be construed as) consistent and/or inconsistent with the views expressed herein.

This post and the information presented are intended for informational purposes only. The views expressed herein are the author’s alone and do not constitute an offer to sell, or a recommendation to purchase, or a solicitation of an offer to buy, any security, nor a recommendation for any investment product or service. While certain information contained herein has been obtained from sources believed to be reliable, neither the author nor any of his employers or their affiliates have independently verified this information, and its accuracy and completeness cannot be guaranteed. Accordingly, no representation or warranty, express or implied, is made as to, and no reliance should be placed on the fairness, accuracy, timeliness or completeness of this information. The author and all employers and their affiliated persons assume no liability for this information and no obligation to update the information or analysis contained herein in the future.

Those two posts were about Q in the Revenue = Price × Quantity equation. How many users, how often they come back, and whether the habit is real. This post is about P. What can you actually charge?

Time spent is the bridge between the two. In consumer technology, time is the raw material of monetization. Subscription businesses convert time into perceived value and willingness to pay. Advertising businesses convert time into ad inventory. Both start with the same input: how much of a user’s day does your product command?

Bottom line up front: I think the advertising revenue opportunity for leading consumer AI apps may be larger than the subscription opportunity. The reason is simple: consumer AI is accumulating the same raw material that powers the largest internet businesses, time and attention. The ad revenue equation is straightforward: Ad revenue = Total Time × Ad Volume × Price of Ads. Across these three variables, the data suggests:

1. Total time spent in AI apps is exploding. AI mindshare follows the same power law as the number of users, even adjusted for how much time each user spends.

2. Time per user is rising → large and growing ad volumes. AI apps trail consumer benchmarks but are starting to rival enterprise apps. ChatGPT is behaving more like a work and productivity tool than a social feed. That is a strong signal for the ability to increase ad loads over time.

3. ChatGPT queries have stronger intent signals than search → compelling ad pricing. More on this in Section 3 below.

Let’s dig in.

1. Total Time: ChatGPT owns 68% of Consumer AI’s attention

Total time spent in generative AI apps grew approximately 10x over the past two years and 3.6x in 2025 alone. No app category has scaled this fast.

A few things jump out. First, the inflection around January 2025 is striking. Total time spent roughly doubled in the first half of 2025, driven by ChatGPT’s expansion into voice, image generation, and search. Second, Gemini emerged around mid-2024 and has grown meaningfully, but remains a distant second.

ChatGPT commands 68% of total AI time spent. Gemini is at 16%. Everyone else combines for roughly 16%. This concentration makes ChatGPT the most plausible place for the first scaled AI-native advertising franchise to emerge. It also helps explain why OpenAI is experimenting earlier and more aggressively with monetization than peers with smaller shares of attention. This is also important to note because you can’t really do advertising in a sub-scale platform.

Where users spend time is the real estate available to advertisers. And 68% of that inventory sits inside one product - ChatGPT. For advertisers evaluating AI-native ad spend, the concentration of attention in a single product is hard to ignore.

2. Time per user is rising. That means more surface area for ads.

Every AI app on this chart is trending up. ChatGPT has roughly tripled its time per user since early 2023. Claude, Gemini, and Grok have all climbed steeply over the past year. The trend is clear: people are spending more time in AI apps, not just downloading them and leaving.

But how much time, relative to what we already know about consumer and enterprise apps?

Versus consumer: still a lot lower. ChatGPT at 16 minutes per day is well below TikTok, YouTube, Instagram, and the rest. But this gap isn’t really a fair comparison because ChatGPT lacks the two things that drive massive time spent in consumer apps.

First, it has no social network effects. TikTok and Instagram are sticky in part because your friends, creators, and communities are there. The content is personalized to you and generated by people you follow. That creates a pull to keep coming back and keep scrolling. ChatGPT has none of that. There is no feed, no followers, no social graph.

Second, it has no dopamine loop. You don’t open ChatGPT to scroll through cat videos or see what your ex is up to. Consumer social apps are engineered for variable-reward engagement: you never know if the next swipe will be boring or hilarious, and that unpredictability keeps you glued. AI assistants are the opposite. You come with a specific task, you get an answer, you leave.

Versus enterprise: higher than most. The enterprise comparison is more useful, with one important caveat: this is mobile-only data from SensorTower, so desktop-heavy products like Slack, Gmail, and Google Docs are understated here. Even so, the signal is notable. On mobile alone, ChatGPT already looks like a high-frequency productivity utility. That matters because productivity products monetize well even at much lower time-spent levels than consumer entertainment apps.

Slack charges $7-12 per user per month. If AI assistants are already capturing this level of daily attention while operating at consumer scale, the monetization surface area is meaningful.

Rising time per user means two things for the revenue equation: more perceived value to support willingness to pay for subscriptions, and more surface area to serve ads. Both point in the right direction.

3. The revenue opportunity

3a. Why ads can outscale subscriptions

Now to the price side: what is all of this attention actually worth? At consumer scale, the two monetization models that matter most are subscriptions and advertising. The important point is not that subscriptions are weak. It is that the largest consumer internet businesses have historically monetized far more revenue through ads than through subscriptions.

The magnitude gap is the point. Google’s advertising business alone is roughly 5x Netflix’s revenue. Meta’s advertising business is roughly 4x. Even Amazon’s ad business, which barely existed a decade ago, is larger than Netflix today. The natural question is whether higher subscription ARPU makes up for the smaller paying base. For some businesses it does. But at scale, the larger monetized base from ads tends to win.

Alphabet and Amazon are particularly interesting because they have both models. In both cases, the ad business is larger than the subscription business, and growing faster. Netflix, Disney, and Spotify are almost entirely subscription. Meta and TikTok are almost entirely ads. OpenAI today is almost entirely subscription revenue, with some ad revenue from the recent launch across a fraction of their 95% free users in the US. That is an enormous pool of attention that is currently generating almost zero revenue. OpenAI is the first mover here, but the same free-user monetization question applies to every AI app with a large unpaid base.

3b. How AI attention gets priced

Ad revenue = Total Time × Ad Volume × Price of Ads

• Total Time: We already know Total Time from Sections 1 and 2. ChatGPT has ~900M WAU, DAU:MAU of 45%, ~16 minutes per day on mobile.

• Ad volume (or ad load): is a product decision. How many ads do you show per session? Google shows 3-4 ads per search results page. Meta shows an ad every 3-5 posts in the feed. ChatGPT currently shows at most one ad per conversation, and only for ~5% of mobile users. That restraint is smart for preserving trust, but it means the ad load variable is currently very low.

• Price of ads (CPM) is the price advertisers will pay per thousand impressions. This is where it gets interesting, because not all attention is priced equally. CPM is ultimately a function of one question: will this user buy something? That breaks down into three things: intent (is the user actively making a decision?), attribution (can the advertiser trace the ad to a purchase?), and audience quality (does this user have spending power?).

The biggest ad franchises each lean on different strengths here. Google Search has a strong intent signal because when someone types “best mortgage rates 2026,” they are declaring a commercial intention in real time. CPMs of $15-200+ depending on category. Revenue per user of ~$84/year globally. Meta has weaker intent but massive time spent. Users scroll 30-90 minutes per day. Meta compensates with extraordinary targeting precision, inferring intent from behavior and people graphs. Revenue per user of ~$57/year. YouTube sits in between: mid CPMs, long sessions, video creative.

To summarize, Google sells intent. Meta sells attention. YouTube sells watch time.

3c. Where AI assistants fit - a ChatGPT case study

Let’s use ChatGPT as the test case since it has the largest free user base and the most ad data to work with. ChatGPT ad pricing will likely land closer to Google than Meta, and potentially advantaged in categories where conversational context improves commercial intent.

When someone opens ChatGPT and asks for a laptop recommendation, compares insurance plans, or plans a family trip, the interaction resembles search, but with richer context. The user often provides budget, preferences, constraints, and intent in a single prompt. That can make the commercial signal more legible to an advertiser, even if it does not automatically make every AI query more valuable than a search query.

I’d expect ChatGPT’s realized CPMs to be at least comparable to Google Search and potentially higher for certain categories. The early data supports this. OpenAI’s premium placements are priced at roughly $60 CPM, well above display advertising and in the range of high-intent search.

Today ChatGPT has ~800-900M free users (95% of WAU). If ChatGPT can generate $30 in annual ad revenue per free user, that implies $25B in ad revenue at current scale. For reference, Meta generates $57/user and Google generates $84/user, so $30 for a high-intent logged-in product is not aggressive.

The early data shows no impact on trust metrics, but the test is still early. Scaling ads 20x without degrading the experience that built the habit is the real execution challenge. The main reason this opportunity remains unproven is that not all AI time is commercial. A meaningful share of ChatGPT usage is informational, creative, or productivity-oriented rather than transactional. And unlike a feed or search results page, a conversational interface gives you fewer obvious places to inject ads without harming trust. So the upside is real, but the execution constraint is equally real: OpenAI has to monetize the habit without degrading the product that created it.

There is also a more optimistic possibility. AI does not just create ad inventory. It can create entirely new ad formats. Conversational ads, where the product recommendation is woven into the dialogue rather than bolted onto the side of it, could actually improve the user experience rather than degrade it. Imagine asking ChatGPT to plan a weekend trip and having it surface a relevant hotel deal within the conversation, informed by your unique preferences and memory. That is not an interruption. That is a feature. If AI enables hyper-personalization, agentic behavior, and truly conversational moments with brands, the advertising opportunity may not just be large. It may be meaningfully different from any ad experience that exists today.

3d. Why Google can wait

Google’s strategy is notably different from OpenAI’s. Google has repeatedly said it has no plans to run ads in Gemini. At Davos in January, DeepMind CEO Demis Hassabis said he was “surprised” that OpenAI was rushing forward with ads in ChatGPT. Google Ads VP Dan Taylor stated on X in December 2025: “There are no ads in the Gemini app and there are no current plans to change that.”

This is a strategic luxury only Google can afford. Google already has a $295B advertising cash machine in Search. It can subsidize Gemini as a loss leader, using the ad-free experience to grow users and deepen engagement while monetizing AI through its existing Search infrastructure (AI Overviews and AI Mode already carry ads). OpenAI does not have that luxury. With no separate cash cow to lean on, it must monetize the chat interface directly.

For now, Gemini monetizes only through subscriptions. That is a much smaller revenue opportunity per user than what OpenAI is building with ads + subscriptions. But Google is playing a different game. It is protecting the assistant to retain users while aggressively monetizing the search results page. Whether that strategy holds as free-user inference costs rise and Gemini’s user base scales past 750M MAU is an open question. At some point, the economics of running a massive free AI assistant may force even Google’s hand.

Putting it all together

This series asked a simple question: is consumer AI just a lot of usage, or is it becoming a real business? Part 1 showed the reach. Part 2 showed the habit. Part 3 suggests the monetization opportunity may be larger than many people appreciate.

The underappreciated point is that leading AI assistants, ChatGPT foremost among them, already have properties that the largest consumer internet businesses monetize best: scaled, recurring attention. Roughly 95% of its weekly active users are still free, which means most of that attention is barely monetized today.

That does not guarantee an ad business on the scale of Google or Meta. Conversational interfaces are harder to monetize cleanly, and trust is the product’s most valuable asset. But if OpenAI can prove that ads can live inside a high-intent assistant without breaking the user experience, the long-term advertising opportunity could end up larger than the subscription business. And if that happens, the real strategic question for Google will no longer be whether Gemini should stay ad-free. It will be how long it can afford to.

Thanks to Sarah Friar and Fidji Simo for reviewing drafts of this article and series.

Sources used in this post include Sensortower.

The information presented in this newsletter is the opinion of the author and does not necessarily reflect the view of any other person or entity, including Altimeter Capital Management, LP (”Altimeter”). The information provided is believed to be from reliable sources but no liability is accepted for any inaccuracies. This is for information purposes and should not be construed as an investment recommendation. Past performance is no guarantee of future performance. Altimeter is an investment adviser registered with the U.S. Securities and Exchange Commission. Registration does not imply a certain level of skill or training. Altimeter and its clients trade in public securities and have made and/or may make investments in or investment decisions relating to the companies referenced herein. The views expressed herein are those of the author and not of Altimeter or its clients, which reserve the right to make investment decisions or engage in trading activity that would be (or could be construed as) consistent and/or inconsistent with the views expressed herein. This post and the information presented are intended for informational purposes only. The views expressed herein are the author’s alone and do not constitute an offer to sell, or a recommendation to purchase, or a solicitation of an offer to buy, any security, nor a recommendation for any investment product or service. While certain information contained herein has been obtained from sources believed to be reliable, neither the author nor any of his employers or their affiliates have independently verified this information, and its accuracy and completeness cannot be guaranteed. Accordingly, no representation or warranty, express or implied, is made as to, and no reliance should be placed on, the fairness, accuracy, timeliness or completeness of this information. The author and all employers and their affiliated persons assume no liability for this information and no obligation to update the information or analysis contained herein in the future.

As we showed in Part 1, consumer AI is following the same script. ChatGPT holds ~70% market share. Gemini sits at ~20%. Everyone else is fighting over scraps.

But market share alone doesn’t tell you whether a winner is durable. The history of consumer technology is littered with products that spiked in usage and then faded. Remember Google Wave? Clubhouse? BeReal? Clubhouse hit 10M downloads in its first month. BeReal was the #1 app in the App Store. Neither built a habit. Not all usage is created equal. A user who opens an app once a month out of curiosity is fundamentally different from one who opens it every day out of habit. The real question isn’t who has the most users. It’s who has the most habitual users. And that’s where engagement and retention come in.

Part 2 goes a layer deeper. Usage tells you who showed up. Engagement and retention tell you who stayed, and whether they’re forming a habit or just browsing. How often are people actually using these products? Are they sticky enough to survive a better or cheaper competitor? These are the questions that separate a market leader from a durable franchise.

Bottom line up front: ChatGPT is not just the largest AI app. It is also the stickiest. And it is getting stickier. Looking at three years of engagement and retention data, three findings stood out:

1. ChatGPT’s engagement lead is wider than its market-share lead among AI peers.

2. Zooming out, ChatGPT has better engagement than most enterprise apps (Slack, Gmail, Outlook) and matches the engagement levels of consumer staples like Facebook, Spotify, and X.

3. On retention, the same pattern holds. ChatGPT has significantly better retention than AI peers and top enterprise apps, and is steadily converging with the best consumer platforms.

1. ChatGPT’s engagement lead is wider than its market-share lead

In Part 1, I showed that ChatGPT commands ~70% of AI WAU and Gemini is second at ~20%. The engagement data tells an even more concentrated story.

On daily engagement (DAU:MAU), ChatGPT is at 45%. Gemini is at 22%. That is more than a 2x gap. Gemini’s reported MAU is impressive (they reported 750M on the last earnings call), but DAU:MAU tells you more about depth of usage. MAU tells you how many people tried the product in a month. DAU:MAU tells you how many of them are actually using it as part of their day. On that metric, ChatGPT is in a different league.

2. ChatGPT has entered consumer-grade engagement territory

Let’s zoom out beyond AI and compare the new AI app category with established consumer and enterprise applications. And let’s use WAU:MAU because it captures the rhythm of knowledge-work better: people use enterprise apps a few days a week, not every single day.

To put these numbers in context: WhatsApp is at 96%. Instagram is at 92%. Gmail is at 78%. Spotify is at 79%. ChatGPT’s engagement is better than Gmail / Spotify and nearing Instagram, which is remarkable for a product that is three years old and doesn’t have a social graph, a notification-driven feed, or an inbox that fills up on its own. And it’s improving:

Three years ago, ChatGPT’s WAU:MAU was around 50%. I argued at the time that apps need 80%+ WAU:MAU to reach a billion users. Historical data across decades of consumer and enterprise apps showed that engagement levels are largely set by a product’s category and core value proposition.

ChatGPT broke that pattern. Its WAU:MAU has gone from ~50% in mid-2023 to 82% today. That is a 30+ percentage point improvement. No other AI app comes close except for Claude which seems to be catching up (a quiet standout and a huge leap over the last month).

The implication: ChatGPT is no longer engaging like an AI curiosity. It is engaging like a tool people rely on every week. The engagement profile matches my Part 1 thesis: ChatGPT is behaving like a utility, not a social app. People don’t scroll it for dopamine. They open it because they need it.

3. The smile curve: ChatGPT is getting stickier over time

Engagement tells you how often people come back in a given window. Retention tells you whether they stick around at all. And the rarest signal in consumer products is whether a product can win back the users who left.

There are three shapes a retention curve can take after the initial drop-off: declining, flattening, or smiling. I covered these in my original analysis two years ago. Most apps flatten or decline. A smile curve, where retention dips and then recovers, is the rarest. That likely reflects product improvements pulling users back in.

In our dataset, very few products show a true smile curve in retention, with rare examples including Gmail, ChatGPT, and Chrome. That’s it. WhatsApp, Instagram, TikTok, Spotify, Slack, Outlook, YouTube: all flat or declining curves. They hold users well, but they don’t win them back over time. ChatGPT does. And it’s doing it at 900M WAUs, which means OpenAI’s pace of product iteration (voice, vision, canvas, search, memory) is not just attracting new users but reactivating lapsed ones.

The floor is already high

The smile curve is even more notable given where retention starts. At Week 4, ChatGPT retains ~66% of its users. Two out of three people who download it are still active four weeks later.

Among AI apps, the gap is material: Character AI is at 48%, Gemini 44%, Claude 41%, Perplexity 24%. At 66%, ChatGPT is building an installed base. At 24%, Perplexity is still struggling to build an installed base on mobile.

Against established apps, ChatGPT’s Week 4 retention beats every enterprise app in our dataset, including Slack. It trails the top consumer utilities (WhatsApp, Instagram, Chrome) but sits in the same range as TikTok. Two years ago, I plotted adoption against retention and found the same clustering: consumer apps top-right, enterprise apps in the middle, AI apps bottom-left. ChatGPT has since migrated out of the AI cluster and into enterprise territory, approaching consumer.

Retention that improves at scale

Three years ago, ChatGPT’s Week 4 retention was closer to 40%. It has climbed to 66% today, a roughly 25 percentage point gain. Most apps see retention set early and stay flat or decline as the user base matures and more casual users come in. ChatGPT has done the opposite: as its user base has grown 10x, retention has gotten better. The most likely explanation is product improvement. Each major release has given users a new reason to come back.

That trajectory is what makes the smile curve credible. ChatGPT isn’t recovering from a weak starting point. It’s recovering from an already-strong floor, and doing so while adding hundreds of millions of users. That combination is genuinely rare.

Conclusion

ChatGPT is no longer just the biggest AI app. It is becoming one of the stickiest products on the internet. High weekly engagement, strong early retention, and improving stickiness even at massive scale. That is not novelty. That is habit formation.

And in consumer technology, habit is what separates reach from revenue. A product with 900 million weekly users and improving retention is not just winning attention. It is building a franchise.

For competitors, the window to displace ChatGPT is narrowing. Not because the product is unbeatable, but because the habit is compounding. Every week that engagement and retention improve at this scale, switching costs go up and the gravitational pull gets stronger.

Part 3 next week will look at time spent and what that tells us about monetization.

Sources used in this post include Sensortower.

The information presented in this newsletter is the opinion of the author and does not necessarily reflect the view of any other person or entity, including Altimeter Capital Management, LP (”Altimeter”). The information provided is believed to be from reliable sources but no liability is accepted for any inaccuracies. This is for information purposes and should not be construed as an investment recommendation. Past performance is no guarantee of future performance. Altimeter is an investment adviser registered with the U.S. Securities and Exchange Commission. Registration does not imply a certain level of skill or training. Altimeter and its clients trade in public securities and have made and/or may make investments in or investment decisions relating to the companies referenced herein. The views expressed herein are those of the author and not of Altimeter or its clients, which reserve the right to make investment decisions or engage in trading activity that would be (or could be construed as) consistent and/or inconsistent with the views expressed herein.

This post and the information presented are intended for informational purposes only. The views expressed herein are the author’s alone and do not constitute an offer to sell, or a recommendation to purchase, or a solicitation of an offer to buy, any security, nor a recommendation for any investment product or service. While certain information contained herein has been obtained from sources believed to be reliable, neither the author nor any of his employers or their affiliates have independently verified this information, and its accuracy and completeness cannot be guaranteed. Accordingly, no representation or warranty, express or implied, is made as to, and no reliance should be placed on, the fairness, accuracy, timeliness or completeness of this information. The author and all employers and their affiliated persons assume no liability for this information and no obligation to update the information or analysis contained herein in the future.

That debate has played out. I was wrong.

Today, AI apps have crossed 1 billion weekly active users and ChatGPT alone accounts for 900M of that. It built that user base from scratch, without an existing distribution platform. Below I analyze usage data across apps to make the case that ChatGPT is the only AI app with both the installed base and new user momentum to become a core utility, like WhatsApp and Chrome are today. Every other challenger wave has faded. ChatGPT keeps compounding.

This is Part 1 of a multi-part series focused on usage. Engagement, retention, and time spent are next. I analyze usage across four dimensions:

1. ChatGPT pulls away in a two-horse AI race

2. ChatGPT vs. the Consumer giants: closer than you think

3. Four seasons of the year: OpenAI, Google, XAI, China

4. Among AI apps, only ChatGPT looks like a plausible new utility

1. ChatGPT pulls away in a two-horse AI race

Source: Sensortower

Methodology note: Source is Sensortower, so this analysis is mobile-only and likely understates web-heavy AI usage. I use weekly active users because it is a useful midpoint between DAU and MAU and a reasonable way to compare emerging AI products against established consumer apps.

Total AI app WAUs have gone from ~100M in January 2024 to over 1.2 billion in February 2026. That is a roughly 20x increase in two years. No app category in history has scaled this fast.

ChatGPT at 900 million WAUs is larger than Spotify at ~600 million, in the same neighborhood as TikTok (~1.5B) and Instagram (~1.3B), and approaching WhatsApp and Chrome (~2.5B each) which I consider essential utilities. When I wrote the original post, ChatGPT was a rounding error next to consumer giants. Now it’s in the same league. Most people, myself included, thought this would take much longer.

Source: Sensortower

ChatGPT commands approximately 70% of total AI WAU. It has held that share consistently for the past year. Gemini accounts for roughly 15-20%. Everyone else (DeepSeek, Grok, Perplexity, Claude, Character AI) combines for ~10%.

The AI category grew 20x, but almost all of that growth accrued to two apps. ChatGPT built its user base from scratch, without an existing distribution platform. Gemini is riding Google’s 4-billion user distribution.

2. ChatGPT vs. the Consumer giants: closer than you think

To understand where AI apps fit, it helps to look at how consumer apps have stratified over the past decade.

Source: Sensortower

Consumer app usage has settled into three tiers:

Core Utility (~2-3B WAU): YouTube, Chrome, WhatsApp. These are infrastructure. Apps so embedded in daily life that they function more like operating system features than standalone products. Growth is slow and steady, driven by smartphone penetration rather than viral loops.

Social platforms (~1-1.5B WAU): Facebook, Instagram, TikTok. Habitual-use apps driven by content graphs and social reinforcement. They took years to climb into this tier and growth has largely plateaued.

Niche apps (~300-600M WAU): Spotify, Amazon, X. Category leaders with enormous user bases that serve more specific needs. Usage is frequent but not universal.

AI is no longer just a category to benchmark against itself. It now belongs on the same chart as the biggest consumer products ever built:

ChatGPT, at 800-900M WAU, cleared the niche tier and is now moving toward the social tier. It did so in roughly three years, a timeline that took Instagram five and TikTok four. Gemini, at roughly 200M WAU, is entering the “Niche” tier from below.

3. Four seasons of the year: OpenAI, Google, XAI, China

The WAU chart tells a story of concentration. The download chart tells a story of chaos.

Source: Sensortower

Over the past year, each season brought a different challenger into the spotlight, each riding a wave of hype, product launches, or distribution muscle. Think of it as four seasons of AI downloads:

Winter 2025: China. DeepSeek exploded past 25 million weekly downloads in January 2025, briefly rivaling ChatGPT’s own numbers. The open-source model’s performance caught the industry off guard and triggered a wave of “is OpenAI in trouble?” commentary. Within weeks, downloads fell back to a fraction of their peak.

Spring 2025: XAI. Grok rode a surge through Q1 and into Q2, fueled by Elon Musk’s distribution via X and a string of model updates. Downloads spiked, press coverage followed. But sustained usage never materialized at the same scale.

Summer 2025: Google. Gemini’s biggest moment. Weekly downloads surged above 20 million, driven largely by the nano banana model — Google’s viral image generation release around Google I/O. The spike was almost entirely image-gen related, which also explains the subsequent drop: it was a model moment, not a sustained behavior shift. Google’s distribution engine likely amplified the reach, but the catalyst was having a genuinely good image model that created viral sharing.

Every season: OpenAI. Through all of it, ChatGPT kept compounding. While each challenger grabbed headlines for a quarter, ChatGPT’s download numbers remained consistently at or near the top, week after week, with none of the volatility. The steady line on the chart is the one that matters.

Honorable mention: Anthropic. As of this writing, Claude hit #1 free app in the App Store. Whether this becomes another seasonal spike or something more durable is worth watching. But the pattern so far has been clear: attention is easier to capture, daily habits are hard to build.

The pattern is clear. Each quarter produces a new “threat to ChatGPT” narrative. Each one fades. Downloads spike on hype. Usage accrues to the apps that earn daily habits.

4. Among AI apps, only ChatGPT looks like a plausible new utility

I think about consumer apps through the lens of stock and flow. Stock is your current user base. Flow is new downloads. Plot both and you get a clear picture of momentum:

Source: Sensortower

This scatter plot captures the end result. ChatGPT occupies a position that, among consumer apps, only the core utility apps (WhatsApp, YouTube, and Chrome) and the dominant social apps (TikTok, Instagram, and Facebook) share. A massive existing user base and high new downloads. Gemini is pulling meaningful downloads but hasn’t converted them into a comparable installed base yet. The rest of the AI field is clustered in the bottom-left: small base, modest downloads. Unless something dramatic changes, ChatGPT is the only AI app on a path to a “Core Utility”.

Conclusion

AI apps have crossed a billion weekly active users. ChatGPT owns 70% of that and is the only AI app operating at true consumer internet scale, reaching 800-900M WAUs in roughly three years without an existing distribution platform.

Google is a real but distant second. Gemini at 200-250M WAUs is the only other AI app that has converted downloads into durable usage, largely by leveraging Android, Search, and 4 billion existing users. No one else is close.

But not all usage is equal. In Part 2, we’ll look at engagement, where ChatGPT’s lead is even more pronounced across DAU:MAU ratios, retention, and time spent.

In consumer markets, once one product owns both the installed base and the flow of new users, the default outcome is further consolidation. ChatGPT has both. The question now is whether that usage is deepening into genuine habit or still behaving like a utility people visit briefly and leave. That’s the real battleground from here.

* * *

Sources used in this post include Sensortower.

The information presented in this newsletter is the opinion of the author and does not necessarily reflect the view of any other person or entity, including Altimeter Capital Management, LP (”Altimeter”). The information provided is believed to be from reliable sources but no liability is accepted for any inaccuracies. This is for information purposes and should not be construed as an investment recommendation. Past performance is no guarantee of future performance. Altimeter is an investment adviser registered with the U.S. Securities and Exchange Commission. Registration does not imply a certain level of skill or training. Altimeter and its clients trade in public securities and have made and/or may make investments in or investment decisions relating to the companies referenced herein. The views expressed herein are those of the author and not of Altimeter or its clients, which reserve the right to make investment decisions or engage in trading activity that would be (or could be construed as) consistent and/or inconsistent with the views expressed herein.

This post and the information presented are intended for informational purposes only. The views expressed herein are the author’s alone and do not constitute an offer to sell, or a recommendation to purchase, or a solicitation of an offer to buy, any security, nor a recommendation for any investment product or service. While certain information contained herein has been obtained from sources believed to be reliable, neither the author nor any of his employers or their affiliates have independently verified this information, and its accuracy and completeness cannot be guaranteed. Accordingly, no representation or warranty, express or implied, is made as to, and no reliance should be placed on, the fairness, accuracy, timeliness or completeness of this information. The author and all employers and their affiliated persons assume no liability for this information and no obligation to update the information or analysis contained herein in the future.

“[Inference] is about to go up by a billion times.” Jensen Huang said it plainly on BG2, a perspective gaining traction across the industry. The implication is straightforward: as AI shifts from demos to default, the bottleneck moves from training to production inference.

Early AI products looked like single requests and single responses. The next generation looks like loops: retrieve, reason, call tools, rerank, synthesize, and often do it multiple times per user interaction. That structure multiplies inference events per task, and it raises the bar on latency, reliability, and cost.

At the same time, more of the AI stack is becoming “always on.” If a product’s core workflow depends on a model call, downtime and jitter are no longer acceptable. Teams need a platform that can meet production expectations while they stay focused on their differentiated product work.

The data backs this up. Google processed 9.7 trillion tokens in the month of April 2024. By April 2025, that number hit 480 trillion - a 50x increase in twelve months. By October 2025, it crossed 1.3 quadrillion per month. Another 3x in 3 months! And we’re still in the early innings.

Open weights expand the surface area

A second tailwind is the rapid improvement of open-weight models. As open models get more capable, more teams choose to run a portfolio of models: frontier APIs for certain tasks, open models for high volume workloads, and fine-tuned models for domain-specific performance. The world trends toward a multi-model future, not a one-model monoculture.

That choice creates a new kind of complexity. Running open models well requires expertise across runtimes, quantization, compilation, caching, hardware selection, and capacity management. Most application teams do not want to build an internal “GPU platform team” just to ship their product.

Source: Artificial Analysis (Dec 2025)

Baseten’s core value: abstraction

We view Baseten as doing for inference what Stripe did for payments: abstracting away the operational and engineering burden so product teams can move faster. Stripe made payments programmable by hiding the messy primitives behind a clean interface. Baseten does the same for model serving by abstracting deployment, scaling, routing, and performance tuning across models and environments.

There is also a Snowflake and Databricks parallel. The hyperscalers provide the raw infrastructure. The winning platforms make it simple to use, easy to scale, and predictable to operate. Baseten sits on top of cloud primitives and turns inference into a higher-level product.

Yes, a cloud provider could build more of this over time. We underwrite that reality. Our bet is that a focused platform can deliver the best mix of performance and total cost of ownership, with a product and support model designed specifically for production inference.

Baseten’s wedge is not just “GPUs as an API.” It is an inference stack and workflow that looks closer to a production system than a managed endpoint:

• Performance engineering: systematic optimization work across runtimes and serving techniques that improves latency and throughput for real workloads.

• Multi-cloud capacity management: the ability to source and manage capacity across environments, and to design for resiliency when regions or providers have issues.

• Deployment flexibility: options that fit how teams actually operate, including dedicated deployments and hybrid approaches for stricter security requirements.

• Developer experience: a clean interface that helps teams ship, debug, and iterate without turning infrastructure into their full-time job.

• Deep customer partnership: an “own the outcome” posture that shows up in how quickly issues get resolved and how much help customers get in production. This posture comes from the team. Co-founders Tuhin Srivastava and Amir built Baseten around customer co-design, and institutionalized it with a forward deployed engineering model led by Joey Zwicker, where engineers embed with customers and own production outcomes.

Source: Baseten

Baseten = an index on AI agents

In the same spirit that Stripe is an index of the internet economy, Baseten is an index on the AI economy. Baseten scales with the success of the application layer. When an app grows, inference volume grows with it. You get exposure to the compounding of AI-native usage, not just seat expansion.

Baseten’s publicly disclosed customers include: Cursor, Notion, OpenEvidence, Abridge, Clay, World Labs, Gamma, Writer, Bland AI, Descript, Ambience, Hex, Wispr Flow, Latent Health, etc. some of the fastest growing businesses! This helps to explain the explosive momentum : Baseten’s revenue 10x-ed in 2025.

Source: Baseten

Customer proof: OpenEvidence

A good way to understand Baseten is through the teams that ship high-stakes products. OpenEvidence serves clinicians who need fast, trustworthy information at the point of care. As OpenEvidence scaled, running inference in-house consumed more time and attention, pulling focus away from the product. They moved to Baseten to offload infrastructure work and improve performance, while keeping their team concentrated on building tools for physicians.

The story is common: customers start with in-house infrastructure or generic cloud building blocks, then hit a wall as production requirements harden. Baseten is designed to be the point where teams can scale inference without becoming an infrastructure company.

The thesis

We believe inference is becoming the dominant workload of AI. Open-weight models and multi-model architectures are expanding what gets deployed. Baseten is building the abstraction layer that makes production inference simple to adopt, efficient to run, and reliable at scale.

We invested because we believe Baseten can become a foundational platform in the AI application economy: the infrastructure choice for teams that care deeply about performance, reliability, and shipping.

Disclaimers: https://www.altimeter.com/terms

This post and the information presented are intended for informational purposes only. The views expressed herein are the author’s alone and do not constitute an offer to sell, or a recommendation to purchase, or a solicitation of an offer to buy, any security, nor a recommendation for any investment product or service. While certain information contained herein has been obtained from sources believed to be reliable, neither the author nor any of his employers or their affiliates have independently verified this information, and its accuracy and completeness cannot be guaranteed. Accordingly, no representation or warranty, express or implied, is made as to, and no reliance should be placed on the fairness, accuracy, timeliness or completeness of this information. The author and all employers and their affiliated persons assume no liability for this information and no obligation to update the information or analysis contained herein in the future.

At Altimeter, we ask ourselves: where are the killer AI applications that will transform industries? Customer experience is one of the largest and most consequential.

We first partnered with Parloa in February 2024, leading their Series B. We doubled down in March 2025, co-leading their Series C. Today, Parloa has raised a $350M Series D led by General Catalyst at a $3B valuation—and we couldn’t be more excited about their trajectory.

The Problem

The call center agent job is brutal. 63% of agents report high burnout. Annual turnover runs high around ~35%—more than double other occupations. For employers, turnover costs $10,000-$21,000 per employee—a vicious cycle. This isn’t just a labor problem—it’s a customer experience problem. Burned-out agents can’t provide empathetic service, and customers expect personalized, immediate help across every channel.

AI offers a fundamental reimagining. An AI agent doesn’t get frustrated after its 50th difficult call. It remains patient and consistent—whether it’s the first interaction or the ten-thousandth. This isn’t about eliminating humans; it’s about deploying AI where it excels and freeing human agents for high-value conversations. Global spend on contact centers is $500 billion annually—enterprises leading this shift are fundamentally rebuilding their customer interaction stacks around AI.

Designed for Large Enterprise

Parloa is not building for SMBs. This is a large enterprise solution—purpose-built for Fortune 200 companies with complex contact center environments and millions of customer interactions. Enterprise contact centers involve integrations with legacy telephony (Genesys, Avaya, Amazon Connect), CRM platforms, payment gateways, and dozens of backend systems across multiple geographies and languages.

This is not a self-serve motion. Parloa deploys Forward Deployed Engineers (FDEs)—technical experts who embed directly with customer teams to implement and operationalize the platform. When deploying mission-critical AI into complex environments, you need engineers on the front lines. Once Parloa is embedded in a Fortune 200 contact center, switching costs are substantial.

Built for Global Complexity

Parloa’s success in Europe demonstrates its ability to win in the world’s most demanding market—a proof point for global scalability. Europe is uniquely complex: 24 official EU languages, strict regulatory requirements, and challenging labor economics. English-speaking offshore agents cost $5-6/hour; German-speaking agents cost $15-17/hour—roughly 3x more. A pan-European enterprise serving customers in Germany, France, Italy, and the Nordics faces a staffing nightmare.

AI changes this equation. Parloa’s AI Agent Management Platform is generative AI-native—not a legacy system with LLMs bolted on. AI agents speak 100+ languages, and when human agents take the call, they can too with real-time translations. A single agent handles German, French, Dutch, and Swedish—no separate teams required. The platform handles natural language briefings, enterprise integrations, and large-scale simulation testing before deployment. If you can win in Europe’s regulatory, linguistic, and labor complexity, you can win anywhere.

The customers speak for themselves. Multiple Fortune 200 companies including Booking.com, Allianz, Swiss Life, and HealthEquity. Swiss Life routes calls 60% faster with 96% accuracy. Berlin Brandenburg Airport’s multilingual voice agent achieved 85% CSAT. These aren’t pilots—they’re production deployments handling millions of conversations.

The Path Forward

Having proven the platform in Europe’s most demanding environments, Parloa is now scaling aggressively in North America—the largest market for enterprise software. The team is crushing it. Chris Silver joined as CRO from Five9. Latané Conant joined as CMO from 6sense, where she helped scale the company from 75 to 1,200+ employees. Together with Csaba Tamas (CPO, ex-AWS), they round out an executive team built to win in enterprise. With offices in Berlin, Munich, and Manhattan, Parloa has surpassed $50M ARR and quadrupled revenue since our Series B.

While customer service is the beachhead, Parloa’s vision extends across the entire customer experience. Every touchpoint becomes conversational—not just support calls, but apps, websites, and proactive outreach. AI agents that proactively identify problems, guide onboarding, and surface upsell opportunities—turning the contact center from a cost center into a growth engine. The opportunity is not just to automate customer service—it’s to fundamentally improve how businesses connect with their customers. That’s the vision Malte and Stefan have pursued since founding Parloa, and why we’ve partnered with them across three funding rounds.

Disclaimer

The information presented in this newsletter is the opinion of the author and does not necessarily reflect the view of any other person or entity, including Altimeter Capital Management, LP (”Altimeter”). The information provided is believed to be from reliable sources but no liability is accepted for any inaccuracies. This is for information purposes and should not be construed as an investment recommendation. Past performance is no guarantee of future performance. Altimeter is an investment adviser registered with the U.S. Securities and Exchange Commission. Registration does not imply a certain level of skill or training. Altimeter and its clients trade in public securities and have made and/or may make investments in or investment decisions relating to the companies referenced herein. The views expressed herein are those of the author and not of Altimeter or its clients, which reserve the right to make investment decisions or engage in trading activity that would be (or could be construed as) consistent and/or inconsistent with the views expressed herein.

This post and the information presented are intended for informational purposes only. The views expressed herein are the author’s alone and do not constitute an offer to sell, or a recommendation to purchase, or a solicitation of an offer to buy, any security, nor a recommendation for any investment product or service. While certain information contained herein has been obtained from sources believed to be reliable, neither the author nor any of his employers or their affiliates have independently verified this information, and its accuracy and completeness cannot be guaranteed. Accordingly, no representation or warranty, express or implied, is made as to, and no reliance should be placed on, the fairness, accuracy, timeliness or completeness of this information. The author and all employers and their affiliated persons assume no liability for this information and no obligation to update the information or analysis contained herein in the future.

1. The Year of Liquidity. $1T+ in new public market cap from AI/tech IPOs and direct listings. The private-to-public logjam breaks. Potential candidates: SpaceX, Anthropic, OpenAI, Databricks.

2. Google Gains Ground. Said it last year “Don’t forget Google.” 2026: Gemini may or may not be the best model... doesn’t matter. It wins because of smart deals (Apple, Reliance Jio) + distribution (Gmail, Maps, Search). 2 billion users don’t pick models... they use defaults.

3. The great AI sorting. Users actively compare. ChatGPT for X, Claude for Y, Gemini for Z. They pick favorites based on vibes and lived experience. Memory compounds. Switching costs rise. The AI landscape ends 2026 more fragmented than it started.

4. AI learns to speak first. Not a chatbot waiting in a box. Ambient. Contextual. It notices, nudges, checks in. The relationship inverts.

5. Open source keeps climbing. Agents learn to mix and match... frontier for the heavy lifting, fine-tuned OSS for everything else. The hybrid stack becomes default. The cash register rings at the model serving platforms.

6. The Meta question. Nothing major since Llama 4 in Q1 2025. Key departures. No long-term research bets. 2026 is time to deliver.

7. The bar rises. Labs spending $5-25B on training this year need to show $5-25B in gross margin over next 12-24 months. R&D payback becomes the metric as they get ready for public markets.

8. Consumer AI experiences growing pains. 40+ countries hold elections including US midterms. AI tries to stay neutral. The “AI is swaying elections” narrative takes hold anyway.

   8b) Consumer AI easy money zone is over. 8B people on the planet of which 1B already using AI up 3x from last year. Tripling from here means becoming the #1 consumer app globally. For context, the most used consumer apps: Facebook, Whatsapp, Youtube etc. are around the ~3B user scale.

What am I missing? Which one ages worst? What’s your most wild prediction for AI in 2026? I’d love to hear from you.

2025 Scorecard

A year ago, I made 6 predictions. Here’s how they held up.

1) AI shakeout intensifies. Frontier labs face consolidation or acquihires.

✅ Nailed it. Big Tech deployed $40B+ through “license & acquihire” structures: Microsoft + Inflection ($650M), Amazon + Adept ($330M), Google + Character.AI ($2.7B), Google + Windsurf ($2.4B), Meta + Scale AI ($14.3B), Nvidia + Groq ($20B). The playbook: license the IP, hire the team, leave the shell behind.

2) Enterprise AI blossoms. SWE orgs transform. Compute vs payroll tradeoff becomes real.

✅ Hit. Software Engineering, Customer Support, Legal, and Healthcare being some areas where it saw 10x adoption in 2025.

3) New hardware form factors with AI as killer app.

❌ Miss. Ray-Ban Meta glasses were modest. The AI killer app lives in your phone and browser, not new hardware.

4) Personalization improves 10x. Usage concentrates in a super app.

🟡 Partial. Memory features shipped (ChatGPT, Claude, Gemini all have it). Context windows hit 10M+ tokens. But the super app didn’t emerge. ChatGPT hit 800M+ weekly users but isn’t replacing your phone apps. Yet…

5) Startups collide with frontier models. Consolidation accelerates.

✅ Hit. Frontier models ate “AI wrapper” territory. The value chain compressed to model providers or deeply vertical apps. AI acquihires totaled $40B+. ZIRP-era startups faced down rounds or shutdowns. The IPO dam held... but 2026 is when it breaks.

6) Big tech stays dominant. Don’t forget Google.

✅ Nailed it. Hyperscalers spent $200B+ on AI infrastructure and absorbed the best talent. Google recovered: Gemini improved, deep integration into Gmail/Maps/Search, strategic deals with Apple and Reliance Jio. Distribution + deals + improved models = momentum.

Final Score

4 out of 6 hit. 1 partial. 1 miss.

Disclaimer

The information presented in this newsletter is the opinion of the author and does not necessarily reflect the view of any other person or entity, including Altimeter Capital Management, LP (”Altimeter”). The information provided is believed to be from reliable sources but no liability is accepted for any inaccuracies. This is for information purposes and should not be construed as an investment recommendation. Past performance is no guarantee of future performance. Altimeter is an investment adviser registered with the U.S. Securities and Exchange Commission. Registration does not imply a certain level of skill or training. Altimeter and its clients trade in public securities and have made and/or may make investments in or investment decisions relating to the companies referenced herein. The views expressed herein are those of the author and not of Altimeter or its clients, which reserve the right to make investment decisions or engage in trading activity that would be (or could be construed as) consistent and/or inconsistent with the views expressed herein.

This post and the information presented are intended for informational purposes only. The views expressed herein are the author’s alone and do not constitute an offer to sell, or a recommendation to purchase, or a solicitation of an offer to buy, any security, nor a recommendation for any investment product or service. While certain information contained herein has been obtained from sources believed to be reliable, neither the author nor any of his employers or their affiliates have independently verified this information, and its accuracy and completeness cannot be guaranteed. Accordingly, no representation or warranty, express or implied, is made as to, and no reliance should be placed on, the fairness, accuracy, timeliness or completeness of this information. The author and all employers and their affiliated persons assume no liability for this information and no obligation to update the information or analysis contained herein in the future.

How It Started

In the mid-2010s, consumer banking was on the verge of a transformation. Smartphones had become ubiquitous, trust in traditional banks was shaky post-2008, and a new wave of “neobanks” and fintech startups emerged to reimagine banking for the digital age. The neobank thesis was elegantly simple: traditional banks had terrible apps, charged hidden fees, and moved at the speed of bureaucracy. A new generation of mobile-first challengers could do better.

Around 2015, the starting gun fired almost simultaneously across Europe. N26 launched in Berlin. Monzo and Starling emerged in London. Chime started acquiring users in San Francisco. TransferWise (now Wise) tackled cross-border transfers. Each had its own wedge. Monzo built a social, mobile-first checking account. Chime partnered with banks to offer fee-free accounts with early wage access. N26 focused on elegant European design. TransferWise made international transfers cheap. The fintech playbook was evolving rapidly, but most players were betting on winning one market or one product category at a time.

Revolut entered this crowded field in July 2015 with a different thesis entirely. Revolut’s founding insight was that banking could be approached like a global software product, full-stack and borderless from day one. Nikolay Storonsky (ex-Lehman, Credit Suisse) and Vlad Yatsenko launched with a simple mobile app linked to a prepaid card for travelers. The card let users hold multiple currencies and get interbank FX rates, a radical convenience compared to banks quietly extracting 2-3% on foreign transactions.

Fast forward to today, Revolut has pulled away from its peer group with a staggering 65M customers globally.

Why Revolut Broke Out

The typical fintech advice was to win one market or one product category at a time. Revolut decided to do both from the very early days, simultaneously adding new products each year while entering new countries. This was bold and controversial. But it worked.

Speed as strategy. Culturally, the team became known for lightning-fast execution and relentless product launches. Storonsky’s view: shipping faster and iterating more gives you more chances to win. Within a few years, Revolut rolled out not just spending and FX, but crypto trading (2017), stock trading, savings vaults, budgeting tools, insurance, P2P payments, business accounts, and more, effectively molding itself into a financial super-app while competitors stuck to narrower playbooks.

Full-stack ownership. While competitors built on banking-as-a-service partners or relied on middleware, Revolut invested obsessively in proprietary infrastructure. They now hold dozens of regulatory licenses globally. When you control the stack, you control the user experience, you can ship faster, and you can enter new markets without renegotiating with partners.

Geographic ambition with geographic capability. Many fintechs “launch” internationally but Revolut actually wins markets. Monzo’s US expansion stalled. N26 withdrew from the UK entirely post-Brexit. Chime stayed US-only. Revolut now operates in 48 countries and is the #1 finance app by downloads in over a dozen EU countries.

Revenue Diversification —> Business Resilience

Here’s what I find most compelling about Revolut’s model: the revenue diversification.

Most fintechs have significant concentration on one business model. Nubank derives ~85% of revenue from lending. Chime is ~70% interchange. Stripe is 90%+ payments. Ramp is ~90% interchange. When your dominant revenue stream faces headwinds, such as regulatory caps, rate changes, competitive pressure, the whole business feels it.

Revolut has built something structurally different. They now have 8+ revenue streams:

• Card payments and interchange

• Subscriptions (Plus, Premium, Metal, Ultra tiers)

• Foreign exchange and cross-border transfers

• Wealth products (stocks, crypto, commodities, money market funds)

• Credit (personal loans and credit cards)

• Business accounts and B2B services

• Merchant acquiring (Revolut Pay)

• Net interest income on deposits

No single product or segment accounts for more than ~30% of revenue. The value is resilience. When crypto volumes crashed, subscriptions kept growing. When interchange gets squeezed by regulation, credit products scale. When rates eventually fall, payments and FX compensate. The diversification smooths cyclicality and creates more durable compounding than single-product businesses can achieve.

It also means Revolut’s marginal economics improve with scale. Many of these revenue streams are pure software margins, subscriptions, trading fees, FX spreads. Each incremental customer adds revenue at low marginal cost. Coupled with disciplined cost control, this is why they’re running 25%+ net margins at scale.

Revolut quarterly revenue across segments:

Thinking In Bets

What makes Revolut’s diversification sustainable is how they think about new products.

The company runs itself like a portfolio of venture-like businesses inside the firm. Dozens of bets outside the core business lines are active at any time. Only a fraction of them will graduate to real business lines. And there are dozens more in the hopper that haven’t even started pilots yet. This is a fundamentally different operating model than most financial services companies. Traditional banks committee-approve three-year product roadmaps. Revolut runs experiments, measures ruthlessly, kills what doesn’t work, and doubles down on what does.

The 8+ revenue streams didn’t emerge from a strategic planning exercise. They emerged from this bets culture, crypto trading started as a bet in 2017, business accounts started as a bet, credit started as a bet. The ones that worked became real businesses. The ones that didn’t got killed quickly. It’s the same logic that venture portfolios use: accept that most experiments won’t work, but make sure you’re running enough of them that the winners more than compensate.

What makes Revolut interesting today isn’t just past execution, it’s the optionality in their current portfolio. A few emerging growth levers that stand out are 1) Credit: still nascent but growing fast. They’re live with credit products in 9 countries and have signaled mortgages are on the roadmap. Credit is high-margin but high-risk; if they execute here without blowing up on losses, it transforms the unit economics further. 2) Business banking is working. Revolut now has a remarkable B2B franchise built on top of the consumer platform. 3) Geographic expansion continues. They’re now pursuing licenses in Brazil, India, Mexico, Colombia, and across Asia-Pacific. Their “mature” European markets have 40+ million users serving ~560 million population. Expansion markets represent billions more.

The stated goal is audacious: 100 countries by 2040. A truly global digital bank. If the last decade is any indication, they’ll get there.

A decade in, the compounding is just getting started.

This post and the information presented are intended for informational purposes only. The views expressed herein are the author’s alone and do not constitute an offer to sell, or a recommendation to purchase, or a solicitation of an offer to buy, any security, nor a recommendation for any investment product or service. While certain information contained herein has been obtained from sources believed to be reliable, neither the author nor any of his employers or their affiliates have independently verified this information, and its accuracy and completeness cannot be guaranteed. Accordingly, no representation or warranty, express or implied, is made as to, and no reliance should be placed on, the fairness, accuracy, timeliness or completeness of this information. The author and all employers and their affiliated persons assume no liability for this information and no obligation to update the information or analysis contained herein in the future.

When a sufficiently advanced technology can transform multiple industries, a curious pattern emerges: the technology vendors that win aren’t just the ones with the best products, highest IQ models, or even the slickest demos. They’re the ones who figure out how to make the technology actually work in the messy reality of enterprise operations.

Enter Forward Deployed Engineering (FDE).

The concept is simple: take engineers who are experts in the core technology and embed them directly with industry experts. Work in tight loops. Understand the domain deeply. Build solutions that work for that specific context—then extract the generalizable pieces that can be reused across similar problems.

As Shyam Sankar, CTO of Palantir, puts it: FDEs “metabolize pain and excrete product.”

Learning from OpenAI’s FDE Playbook

This week I had a conversation with Colin Jarvis who leads Forward Deployed Engineering at OpenAI. Alongside Palantir, OpenAI is doing FDE work at remarkable scale – engagements that target tens of millions to hundreds of millions in customer value, sometimes reaching into the low billions. Colin shared several principles that have guided their approach:

Start with genuinely high-stakes problems. The companies that succeed with generative AI don’t pick edge cases on the periphery of their business. Morgan Stanley deployed GPT-4 across their wealth management practice—one of their largest business units. The semiconductor company OpenAI works with asked them to tackle the biggest sources of waste across their entire value chain.

Build trust through iteration, not just technical excellence. With Morgan Stanley, the core technical pipeline was working within six to eight weeks. But it took six months of pilots, evals, and iteration before the wealth advisors trusted it enough to use it with their clients. The outcome: 98% adoption and a 3x increase in research report usage.

Use eval-driven development. Every piece of LLM-written code isn’t done until you have a set of evals that verify its efficacy. This creates the foundation for trust and for handing off to internal teams once the FDE engagement ends.

Trade off determinism and probabilism deliberately. Use LLMs for what they’re best at—handling nuance and complexity. But wrap them with deterministic guardrails for the things that must be right 100% of the time.

Advice for CEOs Considering FDE

Colin’s sharpest advice: be ruthlessly clear about purpose.

“Be very clear as to what your FDE team is going to accomplish,” he said. “Are you going to be relying on services revenue as a key revenue stream? That’s a very different motion than betting on the increase in revenue you’re going to get from successful product bets that come out of these engagements.”

The failure mode he’s seen repeatedly: companies that have a vision of becoming a product company, but get dragged toward services revenue because it’s easier to capture in the short term. “Suddenly you lose the strategic view.”

The discipline required is saying no at difficult times, even when someone offers significant money for work that isn’t strategic.

The north star: you must leave with product. FDE is a zero-to-one motion. The first engagement might be 20% reusable. Do it two or three more times and you get to 50%. Then push it into the scaled part of the business. OpenAI’s Agent SDK and Agent Kit emerged exactly this way – from custom work at Klarna, extended at T-Mobile, then productized for everyone.

If you’re not building toward something reusable, you’re just consulting with extra steps.

Hope you enjoy my conversation with Colin Jarvis.

Every innovation in music technology has eaten music revenues - consistently. CDs replaced vinyl. Downloads challenged CDs. Streaming conquered downloads.

Each iteration of music technology has cannibalized the prior. The script: new technology arrives, threatens to upend everything, lawsuits fly, and then (maybe) everyone figures out how to make money together.

Most notably with internet: music revenues peaked in the late 1990s at $20B+ all thanks to CDs. However internet allowed for p2p file sharing (Napster, uTorrent et al) and those revenues cratered.

Streaming rescued the industry a decade later. Fun fact: Daniel Ek was the CEO of uTorrent before starting Spotify that rescued the industry! Today, platforms like Spotify and AppleMusic pay roughly ~70% of revenue to labels (Universal, Sony, Warner) who in turn pay ~15% to the artists. By 2021, U.S. music revenue was back to historical peaks $20B+ 😅

But this time is different?

AI music generators like Suno and Udio don't just distribute (incredible!) music, they create it. By training on existing songs, they can produce entirely new tracks that sound professionally made. This raises fundamental questions:

-- If an AI creates a hit song after training on Taylor Swift's catalog, who owns it?
-- How do you track which songs influenced an AI-generated track?
-- What happens when AI-created music becomes indistinguishable from human-made music?


Labels are in early talks with AI startups to build a “Content ID for AI,” so every time an AI model uses a catalog, royalties flow back. They’re hashing out fingerprinting, attribution, and fair pay. These negotiations will decide whether AI becomes music’s next golden era... or its next big disruption

Both systems rest on the same four-layer autonomy stack: Perception → Reasoning → Action → Data Flywheel. The basic plumbing at the “Action” layer is the biggest hold-up for AI agents today. Cars have a simpler “Action” step: throttle, brake, steer, on roads whose rules rarely change. Agents have vastly more complexity. Their 'Action' layer operates in a fast-changing digital environment with orders-of-magnitude more degrees of freedom: where to click, what to type, what tools to use. Meanwhile, they must manage brittle APIs, authentication scopes, and data privacy challenges.

I predict this world order will flip. Why? First, the data flywheels for AI agents are much faster, cheaper, better than cars. The cost of error can be fatal with cars. This data flywheel will lead to compounding advantages in other parts of the stack. Second, AI agents just got much better plumbing with a standard protocol like MCP that helps agents collaborate. This significantly improved the “Action” layer and together with a strong data flywheel puts AI agents on a much steeper improvement curve.

We’re living in a brief moment where robots roam the streets while desktop bots are tethered to human intervention but that balance won’t last. Prediction: AI agents will reach mass adoption way before Autonomous Vehicles do. More precisely, I predict that a personal AI secretary will clear a 100M paid user mark years before a 100M users ride a Level-4 autonomous vehicle.

To understand why this reversal will happen, we need to break down what makes any autonomous system work. Both self-driving cars and AI agents share the same fundamental architecture, but they perform very differently on each component. In this article I dive deeper into:

1. The principal components of autonomy

2. Deep dive into AI Agents

3. Takeaways for builders and investors

Part 1 Principal components of autonomy

Under the hood, every autonomous stack can be broken down into four fundamental layers: Perception → Reasoning → Action → Data Flywheel.

The autonomous systems share some attributes in common (e.g. the Transformer architecture turbo-charged the Reasoning layer for both physical and digital world), their performance on these four layers is quite varying. For instance, the promise of humanoid robots was re-ignited with the ChatGPT launch. However I suspect that they are both farther behind and slower in getting to L4-equivalent autonomy. For two main reasons: 1) lack of large diversified data for pretraining and 2) data flywheel is quite slow / expensive for post training improvements as well. With this framework in mind, I evaluate how different autonomous systems stack up across these four layers:

Data flywheel is emerging as a key variable. It compounds with implicit and explicit human feedback - essentially anything that can provide ground truth for the model to improve its behavior. For instance, this is how a toddler learns to walk via reinforcement learning. Falling is failure :) The fall reinforces to the toddler “don’t fall”. Similarly, OpenAI's products have a massive audience of 500m Weekly Active Users (WAUs) providing feedback at billions of feedback points per day. This has been the secret of why OpenAI products are improving fast and are likable. This is not too dissimilar from other versions of Reinforcement Learning (RL) we have seen in consumer products before.

On X/Twitter overdoing targeting makes content feel like “clickbait”. Similarly, overdoing RL in ChatGPT feels like sycophancy! In the case of self-driving cars, the feedback does exist but is less frequent, harder to integrate and more expensive.

This framework reveals why cars currently lead the autonomy race but it also shows why that lead won't last. The data flywheel advantage that AI agents possess is just beginning to compound. To see how this plays out, let's examine digital autonomy in detail and trace its rapid evolution.

Part 2 Deep dive into digital autonomy (AI Agents)

Software agents have leapt from “chatbots that answer” to “co-workers that do.” Unlike embodied robots, they benefit from (1) cheap, abundant feedback loops and (2) instant distribution over the Web, letting them learn orders-of-magnitude faster than their hardware cousins. Digital agents can see and think incredibly well, and they get smarter with every user interaction. However, they still struggle to do things on our behalf with the consistency, safety, and speed that real-world workflows demand. Closing that last mile is the biggest remaining unlock.

Where are we today? The table below benchmarks today’s best-known agent classes on the four autonomy pillars:

This snapshot shows digital agents' current strengths and weaknesses, but the picture is changing fast. Three major trends are reshaping the landscape

#1 From single agent to swarms of agents: In just twenty-four months we’ve watched digital agents sprint through three waves: (1) single-turn chatbots that only answer questions, (2) single-agent “copilots” that can decide what sub-steps to run like code-gen, SQL, classification, etc. (case in point: GitHub Copilot has ~15 million users who let the tool auto-write “46 % of their code” ) and now (3) multi-agent “crews” and agentic workflows that delegate tasks among specialised workers (planner, researcher, executor, evaluator). Frameworks like crewAI, AutoGen and LangGraph have made it trivial to spin up these swarms, turning what used to be brittle prompt-chaining into reusable, event-driven microservices.

#2 “Action” is the brittle tier of the stack but a few great efforts to improve it such as MCP: Model Context Protocol (MCP) is like a USB-C that joins Reasoning to Action. It standardised how agents talk to external tools and data. Instead of bespoke wrappers for GitHub, Slack or a private DB, a Host agent sends an MCP request to a lightweight Server that owns auth, auditing and schema versioning. The Server returns a prompt-ready response the model can reason over, then loops back when the model issues the next call. By moving the “what credential, what endpoint, what retry” logic out of the prompt and into the protocol, MCP lets the reasoning layer plan confidently while the action layer executes safely and deterministically. The prize: agents that improve with every feedback datapoint and ship real transactions without humans hovering over the “Undo” button. MCP has tradeoffs—ultimately, we're trading flexibility for cost. A MCP call is more expensive than an API call.

#3 Identity is another key part of the agentic stack that needs to mature. The moment an agent graduates from read-only to write access, it inherits all the messy realities of enterprise auth. Each API call must run under the right SSO session, respect per-object ACLs, and leave an audit trail. A single over-broad token can leak customer data or nuke a prod table. A few great efforts in the space are 1) Descope (offers drag-and-drop authentication management workflows so agents can borrow a user’s identity without developers hand-rolling OAuth flows) and 2) Arcade.dev (provides pre-authenticated “tool-calling” connectors that let LLMs act on Gmail, Calendars, or internal APIs while keeping scopes tight and revocable)

Part 3 Takeaways for builders & investors

Autonomous systems win by improving on the slowest link in a four-layer chain (Perception → Reasoning → Action → Flywheel). Transformers solved “good-enough” reasoning for most tasks; now the race is about plumbing reliable inputs, executing deterministically, and compounding data faster than peers. Here are a few best practices I picked up surveying practitioners:

1. Figure out the bottleneck by auditing each layer separately. A flashy demo often masks a weak action layer (PC agents) or a constrained flywheel (HD-map AVs). Transformers aren’t the end of the story. Reasoning is solved enough for most use-cases; returns now come from better perception (plumbing to connectors, sensors), precise action or faster flywheels (data rights, edge deployment).

   • For instance, in humanoid robotics the biggest bottleneck is access to diversified data at scale for pre-training (Perception)!

   • In customer service agents, the biggest bottleneck right now is instruction following while minimizing latency (Perception <> Action). In a few months, hopefully that bottleneck is something different.

   • In the case of Waymo you can see the evolution of Waymo’s performance and the gains were driven by different parts of the stack at different times. See chart:

How to read this data:

• Big drop 2015-20 – steady software & sensor gains on easier suburban routes.

• Spike in 2021 – most CA testing shifted to dense San Francisco; tougher edge-cases meant safety drivers intervened more often.

• Recovery 2022-23 – model, policy & mapping updates pushed rates back down.

• ‘24 uptick – fleet scaled 4x and Waymo mixed commercial robotaxi duty with drivered test miles; test miles skew toward the hardest, still-unsolved scenarios.

2. Find ways to accelerate the Flywheel! Capture explicit and implicit user feedback from day one; automate labeling and retraining to shorten the loop and drive marginal-cost-efficient quality gains. A few examples:

   • Glean learns an incredible amount of user activity that helps improve personalization but also gathers post training data.

   • In voice customer service, where instruction following in live audio is tough, creating data on what tool calls are not working with precision (e.g. call routing, reservations, cancellations etc.), with real world examples (even if anonymized with masked PII) can be valuable. This creates valuable post-training data that can help improve underlying model performance.

3. Embed auto-evaluations. Stand up an automated eval harness that 1) houses deterministic task suites, 2) executes every new model/agent build (CI-style), and 3) blocks deployment while surfacing granular traces. A few real-life examples include Zapier (they block PRs that fail tool-call and task-success checks) and Airtable (their AI Assistant streams all prompt logs to Braintrust, tags failures, and retrains prompts/models only after the test suite is green).

4. Harden execution & security. Adopt structured tool-calling (MCP is the leading protocol), least-privilege tokens (not doing so could create tough times), and replayable audit logs. Target a minimum task success rate before scaling to customers.

Thanks to Alexander Matthey, Sud Bhatija, Jamie Beaton, Cobi Gantz, Sahil Aggarwal, Axel Dittman, Angad Singh, Matt Elkherj, and Will McCarthy for their feedback on this article.

The information presented in this newsletter is the opinion of the author and does not necessarily reflect the view of any other person or entity, including Altimeter Capital Management, LP ("Altimeter"). The information provided is believed to be from reliable sources but no liability is accepted for any inaccuracies. This is for information purposes and should not be construed as an investment recommendation. Past performance is no guarantee of future performance. Altimeter is an investment adviser registered with the U.S. Securities and Exchange Commission. Registration does not imply a certain level of skill or training. Altimeter and its clients trade in public securities and have made and/or may make investments in or investment decisions relating to the companies referenced herein. The views expressed herein are those of the author and not of Altimeter or its clients, which reserve the right to make investment decisions or engage in trading activity that would be (or could be construed as) consistent and/or inconsistent with the views expressed herein.

This post and the information presented are intended for informational purposes only. The views expressed herein are the author’s alone and do not constitute an offer to sell, or a recommendation to purchase, or a solicitation of an offer to buy, any security, nor a recommendation for any investment product or service. While certain information contained herein has been obtained from sources believed to be reliable, neither the author nor any of his employers or their affiliates have independently verified this information, and its accuracy and completeness cannot be guaranteed. Accordingly, no representation or warranty, express or implied, is made as to, and no reliance should be placed on, the fairness, accuracy, timeliness or completeness of this information. The author and all employers and their affiliated persons assume no liability for this information and no obligation to update the information or analysis contained herein in the future.

#1 Where are we today?

Over the past century, the Brain Machine Interface (BMI) space has evolved from early breakthroughs such as the introduction of electroencephalography (EEG) in 1924 to groundbreaking animal trials like monkeys playing Pong in 2021. This year Neuralink’s N1 implant was successfully implanted in multiple humans residing in their homes.

This is a breakthrough for people with quadriplegia, restoring autonomy by converting neural signals from the motor cortex into digital commands. This allows users to engage with the digital world (move cursors and type through thought alone), communicate via text, email, speech and engage with creative outlets like art and music - ultimately delivering life changing human value.

Noland Arbaugh, the first participant in Neuralink’s clinical trial, now plays video games, live streams, and navigates his laptop—all through the power of his mind. Similarly, Alex, another participant, is designing complex 3D objects using CAD software, unlocking creative potential once thought inaccessible. These stories are not isolated; they are the first chapters in a revolution to restore independence to millions globally. Reviews from Noland Arbaugh, Neuralink’s first participant:

• “(Neuralink) has helped me reconnect with the world, my friends, and my family. It's given me the ability to do things on my own again without needing my family at all hours of the day and night.”

• “The biggest thing with comfort is that I can lie in my bed and use (Neuralink). Any other assistive technology had to have someone else help or have me sit up. Sitting causes stress mentally and on my body which would give me pressure sores or spasms. It lets me live on my own time, not needing to have someone adjust me, etc. throughout the day.”

• “I think it should give a lot of people a lot of hope for what this thing can do for them”

#2 How does it work?

The N1 is the size of a coin, containing a microprocessor, Bluetooth transmitter, rechargeable battery, and ultra-thin wires with 1,024 electrodes distributed across 64 threads. These highly-flexible, ultra-thin threads are key to minimize damage during implantation. The threads of our implant are so fine that they can't be inserted by the human hand. A surgical robot has been designed to reliably and efficiently insert these threads exactly where they need to be.

Once inserted, the electrodes detect neural signals and translate them into digital commands, transmitted via Bluetooth to external devices. Thinking about moving a cursor triggers corresponding neurons in the motor cortex, enabling precise cursor movement on a screen.

#3 Why Now?

This year Neuralink’s patients achieved bandwidth that’s ~50% of a normal human via the N1 implant. It’s poised to meet human level bandwidth next year and the team aims to exceed it by the end of the decade. A quick breakdown on why bandwidth matters. Today’s intelligence operates on three levels:

1. Limbic Intelligence (“thinking fast” – primal needs like safety and survival)

2. Higher-Order Reasoning (“thinking slow” – complex thought and planning)

3. Digital Intelligence (enabled by phones, internet, software)

Our reliance on digital intelligence is undeniable. Forgetting your phone feels like losing a limb. Yet, the bottleneck lies in the bandwidth—the rate at which information flows between biology and technology. Bandwidth is quantified in Bits Per Second (BPS), as popularized by the late Professor Krishna Shenoy:

1. Average human output (no devices): 1 BPS

2. Typing/speaking: ~20 BPS

3. Peak human performance: ~40 BPS

With BMIs (Brain-Machine Interfaces):

• 2017: Achieved ~2-4 BPS

• 2021: Reached 6.56 BPS

• 2024: Neuralink’s N1 reached 9.5 BPS

• 2025: Neuralink targets 40 BPS

• By 2030: Neuralink targets 100 BPS (edit: Elon forecasts 1Mbps of bit rate and >1M humans by 2030)

The advances next year will likely come from two main levers: 1) More electrodes (going from 1024 to 3000) and increased utilization of electrodes (currently N1 utilizes ~10% of the 1024 electrodes and is expected to go as high as 40% next year). Achieving near-human bandwidth will help dramatically improve the quality of life for people with special abilities.

#4 What's next?

1. Scaling N1: Neuralink aims for 27 procedures in 2025 and 79 in 2026, expanding the number of users.

2. Vision Restoration: FDA-approved BMI for vision restoration (“Blindsight”) interacts with the visual cortex to restore basic light and shape perception.

3. Speech Restoration: Expanding to auditory cortex applications to restore speech capabilities.

4. Beyond Restoration: BMIs could replace smartphones. BMIs could surpass traditional devices, becoming high-bandwidth interfaces for AI-human interaction. They could enhance human cognition, memory, and productivity for 100M+ users

From restoring autonomy for quadriplegics to enhancing human potential, Neuralink will revolutionize how we interact with machines, each other, and the world. It’s not just about catching up to technology, but about defining the next frontier for humanity!

Thanks to Cobi Blumenfeld-Gantz and Shreya Bhargava for their inputs on this article.

We wanted to dig into what makes Palantir special through the lens of company building and helping other founders apply best practices. A bit about us: Apoorv was a Forward Deployed Engineer at Palantir from 2013 to 2017 working primarily on the commercial side of the business (working with Palantir’s non-government customers). Cobi was at Palantir from 2014 to 2020 working predominantly on the US Government side of the business. Given our different functional and customer focus areas, we saw a broad spectrum of what makes Palantir special and below we distill some of the common elements that stand out. 

Palantir: a misunderstood company

Palantir has been -- and in many ways still is -- a misunderstood company. Over the years, Palantir has been called everything from a secretive spy organization that steals citizens’ data to a services business to a consulting shop.

Today however that tide has largely turned. The perception change has to do in part with defense tech becoming in vogue and with Palantir's strong public market performance (nearing $100b in market cap, and now a part of the S&P 500). While those are lagging indicators, the leading indicators lie in the special environment that Palantir creates and the associated training Palantirians receive. 

Here are the key elements that create the magic of Palantir’s culture:

1. Uncompromising Talent Bar

2. Mission Orientation

3. Forward Deployed Engineering: Deliver Outcomes, Not Software

4. Seek Truth: Solve Real Problems, Not Shiny Ones

5. Define a Shared Enemy: United The Team

6. Flat Structure: Direct Access To Leadership

7. Radical Transparency: Learn From Everything

#1 Uncompromising Talent Bar

Palantir maintains a high bar for talent. For as long as we were there, every candidate was interviewed by one of the founders (typically Alex Karp or Stephen Cohen) as their final interview. This created a tax on the organization given the founders' schedules were limited, but it also kept the bar incredibly high. Palantir screens for people with T-shaped skills—deep in certain fields while being able to collaborate across disciplines with experts in other areas. Palantir CTO Shyam Sankar writes: 

“When forming a team, I don't want to assemble a polite roster of cross-functional professionals. I want the X-Men: a medley of mutants united for good.”

One piece of training Palantir prescribed to interviewers was to always ask ourselves: “What would make us think the opposite of what we currently think about a candidate?” If a candidate was doing really well in the interviews, we’d design their next set of interviews to see what would make us change our decision to not hire. Similarly, for candidates with sharp T-shaped skills but who were struggling in the process, we’d also ask the counterfactual: what could the candidate demonstrate to change our minds and receive an offer?

Finally, retention of your strong performers matters. One of the hardest parts about having a talented pool of employees is that it’s hard to keep them together. Humans go through phases and so do companies. As Palantir’s Head of Commercial, Ted Mabrey, describes it:

“(At) Palantir, a person would have to stay committed through phases that included: products that were closer to pipedreams than viable for years, raising hyped private rounds as a sexy unicorn, being the most maligned company in the valley with literal protesters at your door, having your options underwater, a life-changing event in going public, an ~80% drop in stock price, and deciding to climb those mountains all over again with yet a new platform.”

But Palantir did an exceptional job of keeping this group together, in large part driven by the strong mission orientation.

Apoorv at a Palantir recruiting booth on a college campus

#2 Mission Orientation

In our first weeks at Palantir, we quickly learned that sleep is overrated and coffee is underrated. Shyam covers it well in his post "The Case Against Work-Life Balance: Owning Your Future" as to why working 70+ hour weeks was not heroic, it was standard. Palantir consistently outworked its competitors, but the reason behind this relentless drive was far more compelling. 

Palantir employees were driven by a sense of purpose far beyond monetary rewards. The mission was everything—it became a personal, deeply ingrained commitment to achieve a shared goal that was larger than each of our individual existence. 

Teams were deeply mission-aligned, with every project outcome evaluated as either a WIN or LOSS. That is an incredibly powerful way to fuel your fire! This culture of winning was ingrained and our team operated with singular focus on achieving mission success. Shyam summarizes it well in “The Primacy of Winning”. Drawing from its roots in defense technology, Palantir extended this mission-driven mindset beyond defense and into commercial applications. Even commercial teams were structured like military units, with Commanding Officers (COs) leading Forward Deployed Engineers (FDEs) referred to as Deltas and less technical forward deployed individuals known as Echoes. 

Setting the mission to be larger than one’s own identity is a winning formula. Teams who are able to rally around that shared mission are in possession of a superpower. Some teams we see that have a similar level of alignment towards a shared goal are SpaceX (mission: making humanity multiplanetary), OpenAI (mission: developing AGI to benefit humanity), and Anduril (mission: equip US and allies with defense tech to secure sovereignty).

#3 Forward Deployed Engineering: Deliver Outcomes, Not Software

Palantir pioneered a new breed of engineer—FDEs—who didn't just deliver software, they delivered outcomes. These were engineers embedded directly with customers to solve on-the-ground challenges, ensuring the tech created real-world value, no matter the complexity.

When delivering tangible value is the mission, FDEs don’t stop at delivering software. They go above and beyond to build the supporting ecosystem. This often includes the less glamorous “plumbing” work, like data integrations, or solving non-technical challenges like incentive alignment within an organization (e.g., what if Palantir’s implementation causes certain jobs to become redundant?). Operating as an FDE is rarely glamorous, but it is arguably the most critical role for ensuring success. The focus is on results, not just technology.

You may wonder: isn’t it incredibly expensive to serve customers with these FDEs? The answer is yes, but Palantir’s median Annual Contract Value (ACV) is $5 million. When a customer is paying that much, they are not just purchasing technology; they are purchasing outcomes, and FDEs are the ones ensuring those outcomes are delivered.

Source: Meritech Comps

The need for such a role is not unique to Palantir. The same pattern is evident in today’s AI landscape. As with any new technology, the existing tech stack is often not mature enough for something like generative AI (today) or Spark (a decade ago) to “just work” and deliver value. FDEs make these new technologies viable and ensure they deliver results, which is why their skillset is highly sought after today.

FDEs have come a long way—from being criticized as mere “consultants” to being revered as the frontline “AI engineers.” 

#4 Seek Truth: Solve Real Problems, Not Shiny Ones

A core principle at Palantir—especially among FDEs—was the relentless pursuit of solving real user pain points, not just shiny problems. This often required 'tasting the soup' by spending time with users in their environment—whether that was down the street or on the other side of the globe.

For instance, in our work with the military, we had servers deployed on ships. We didn’t initially consider how the physical rocking of ships at sea would affect these servers—like becoming unplugged—which had implications for caching frequency, downtime procedures, and a host of other technical and process issues. If we hadn’t been physically present, we wouldn’t have known to solve these problems. The challenges weren’t always strictly software.

Similarly, at a forward operating base in the Middle East, the military unit had to take inventory of every hard drive on every server each week. This meant physically pulling out hard drives from servers we had installed. We had to develop clear shutdown procedures and a reinstallation sequence to make this possible without interrupting operations.

Our work often meant solving less glamorous, yet crucial, issues, like cleaning data or building software to make data easier to use. It wasn’t always the sexiest or hardest technical challenge, but it was often the most critical to achieving user outcomes.

Cobi on a military base training US Marines

#5 Define a Shared Enemy: Unite The Team

One key cultural element at Palantir was defining a shared enemy. Mandatory reading included "The Looming Tower" by Lawrence Wright, which highlighted the importance of understanding your adversary. Every team, including those not focused on government work, managed to define their own shared enemy.

For government projects, the shared enemy was often the customer itself—the government—with all of its inefficiencies and bureaucracy. In commercial projects, the enemy sometimes was consulting giants like BCG or McKinsey, and other times it was the drudgery of spreadsheets and siloed information.

Palantir took on entrenched bureaucracies so seriously that they even sued the U.S. Army to overcome procurement roadblocks—and won. This was in the spirit of overcoming inefficiencies, similar to SpaceX’s lawsuit years before.

In one instance, we were working with a civilian US Government agency, competing with Booz Allen Hamilton for a large contract. It wasn’t a question whether Palantir’s offering or Booz Allen’s cadre of dashboard makers was a better choice - Palantir delivered better performance at a better price point. It was a question of justice -- would the Government choose the right answer (Palantir) over the legacy approach (Booz). Using the shared enemy of ‘large government contractors trying to solve foundational data challenges with flimsy dashboards’ allowed our team to rally together to work exceedingly hard for this agency.

With the combination of mission orientation and a shared enemy, we were motivated to try to solve any problem. This shared enemy was also valuable for long term orientation and motivation -- taking down the largest evils in the world takes time and can’t be accomplished overnight. To beat Raytheon or Booz Allen -- or to beat technological ineptitude out of the US Government -- takes consistent, durable work.

Cobi at a DHS deployment in support of a National Special Security Event

#6 Flat Structure: Direct Access To Leadership

Palantir tried to maintain a flat organizational structure. Alex Karp and Shyam Sankar often reached out directly to individual contributors, cutting through layers of management and delivering “Gamma radiation”. Shyam writes:

“Mentorship is likewise critical when directing mutant powers towards the greatest possible good. The X-Men would not have become X-Men without Professor X’s School for Gifted Youngsters. But again, the standard model doesn't apply. To begin with, you need mutants to mentor mutants, and in many cases, to provide the initial dose of radiation. Otherwise, even the best institution of higher learning will predictably devolve into a lemming academy. Once mutation is in process, one of the greatest aspects of mentorship is, paradoxically, autonomy. This is especially important because extreme growth doesn't happen on schedule, but is subject to periods of intense activity. As a mentor of mutants, you need to be attuned to these periods, and when they come, confer even more autonomy. Above all, fight your instinct to handhold (hard to do when both hands are always clenched in a fist anyway!)” 

This "gamma radiation" approach—similar to the leadership styles of Elon Musk and Jeff Bezos—allowed direct communication and rapid decision-making. It also had a side effect of providing great opportunities to everyone in the company to learn.

 #7 Radical Transparency: Learn From Everything

Radical transparency at Palantir wasn't just a practice—it was foundational for building alignment across the company. Every email, decision, and retrospective analysis was available to all in the company, breaking down silos and creating a hive mind that thrived on collective learning

Several practices exemplified this transparency:

• Almost all emails were accessible to everyone in the company.

• Meeting reports were written and shared with the entire company after every important meeting.

• When something didn’t go according to plan, retrospective analyses (e.g., the Five Whys) were documented and made available in the internal company wiki for all teams to learn from.

• Most team wikis and notes were open for others to access, breaking down silos and encouraging cross-team learning.

This transparency helped Palantir quickly gather and spread lessons across teams, ensuring that everyone learned from each other's experiences and drove collective outcomes.

Palantir’s True Legacy

Palantir’s true legacy isn’t just its software or contracts—it’s the people it has trained and launched into the world. The leaders, engineers and founders that have emerged from Palantir carry with them a relentless focus on mission, a commitment to delivering results, and the ability to lead teams in tackling big, audacious challenges.

For any company or startup, the lessons from Palantir are clear: Set an uncompromising talent bar, rally around a shared mission, and always seek to solve real problems.

Thanks to Cobi Blumenfeld-Gantz, Sud Bhatija and Shreya Bhargava for their inputs on this article. 

At Altimeter, we have been studying the AI landscape across the stack, from semiconductors to applications, and asking ourselves: Where will value accrue? As we dug deeper into the ecosystem, we have come to believe that Glean is on a path to becoming not just an enterprise search tool, but the enterprise AI platform. Today, we are excited to announce Altimeter’s partnership with Glean as a co-lead for their Series E fundraise.

The problem

The vast majority of company knowledge is scattered in silos of unstructured data. It is not just documents, but also code, screenshots, emails, presentations, diagrams, and more. Consequently, knowledge workers today spend more time than ever looking for answers spread across more apps than ever, which hurts organization productivity and competitive edge. Generative AI promises to bring more efficiency to the workplace, but operationalizing it at scale while leveraging an organization’s proprietary datasets securely is a significant technical challenge. 

Glean’s solution

CEO Arvind Jain and the founding team worked on Google Search in the early days, making them particularly suited to solve one of the biggest challenges in operationalizing enterprise generative AI: seamless search across heterogeneous enterprise datasets. Glean delivers a turnkey enterprise search and generative AI solution with enterprise-grade security, role-based access control logic, and fully referenceable source documents, all on petabytes of data with constant real-time changes. Some Glean customers connect up to 50 disparate applications across email, messaging, calendar, document storage, ticketing systems, code repositories, CRM tools, project management tools, and more¹. Glean combines its powerful search with generative AI capabilities to produce answers that are precisely tailored to the user’s query with the proper enterprise context, driving improved employee productivity. Glean is leading the charge in enterprise AI, trusted to deliver the latest and greatest advancements— from data retrieval to analysis, summarization, content generation, and task automation— and continuously evolving to meet future needs.

Glean Product Overview:

Customers report strong engagement and penetration… 

Engagement and penetration are two important metrics we look at for how much utility enterprises get from software products. Glean stands out in both usage frequency (engagement) and diversity of use cases (penetration). At Glean, 70-80% of monthly users engage on a weekly basis. This is just shy of essential enterprise tools like calendar (80%), email (95%), and browsers (95%)² (full analysis here). This high engagement is not limited to a small subsect of employees within a customer organization— we observe it across a wide range of personas in engineering, sales, support, business development, operations, IT, marketing, finance, and more. In some of the most mature customer deployments, over 80% of employees are monthly active users. In contrast, other AI solutions tend to benefit a small subset of enterprise employees (e.g. coding AI for junior developers or image generation for marketing teams). 

…delivering tangible ROI with customer love

The ROI of AI investments is a hot topic of debate. Customers are quantifying the ROI on Glean spend in terms of productivity gains, dollars saved, and employee satisfaction. One CIO noted that they procured Glean at a time when they conducted layoffs because Glean helped them consolidate spend¹. This demonstrates that the ROI makes sense even in tough times. Not only are employees spending time on Glean, but they also love it. Our customer survey shows that Glean has a 47% net promoter score, compared to 43% for Slack and 33% for Microsoft products. Another CIO shared that employees would revolt if he took Glean away¹. 

The path forward

We believe we are still in the early innings of the generative AI revolution, and there is much more innovation to come. Glean has the potential to be the platform that brings AI innovation to the enterprise—in search, chat, agents, and beyond. In its next generation of products focused on effective prompting, Glean is driving AI-centric behavior among users, a critical step in advancing AI at work. 

We are excited to support Arvind and the Glean team as they redefine how work happens in the enterprise.

Disclosures

Reflects Altimeter's views and beliefs unless otherwise noted. There is no assurance that Altimeter will continue to see similar levels or types of investment opportunities, execute on any particular investment opportunity in the future, or that any such investment would be profitable. Third party views are not necessarily representative of Altimeter and content has not been reviewed for completeness or accuracy

No Updates and No Warranties: Unless otherwise specifically indicated in a particular instance, everything in this Document speaks only as of the date of this document.  This Document is necessarily incomplete and necessarily reflects only a small, non-exclusive, and out-of-context sample of Altimeter’s opinions, assumptions, projections, processes, valuation metrics, considerations, and available information (including information provided by third-parties).  While Altimeter currently believes that this Document is accurate in isolation, Altimeter makes no express or implied warranty as to this Document current or future accuracy, relevance, or completeness, and Altimeter cannot accept responsibility for any errors or omissions appearing in this Document.  This Document subject matter (including each forward-looking statement) is constantly evolving, and is constantly under scrutiny, re-evaluation, and revision by Altimeter, and always remains subject to change, modification, termination, supplementation, or reversal at any time in Altimeter’s sole and absolute discretion or without Altimeter’s knowledge.  Altimeter has no obligation to update, supplement, or revise this Document, or to otherwise notify you, if or when (a) any information in this Document (including each forward-looking statement) changes or reverses or if any information (including any forward-looking statement) proves to be inaccurate; (b) new, changed, or different events, circumstances, developments, or other factors materially change or reverse the conclusions, opinions, assumptions, forward-looking statements, or other information contained in this Document; or (c) Altimeter modifies, revokes, terminates, supplements, reverses, or otherwise changes any Altimeter opinions, conclusions, assumptions, projections, personnel, or investment strategies, restrictions, parameters, procedures, or policies, or any other information.

Sources

1. Glean Customer Survey. There is no assurance that this data is an accurate representation of all Glean customers. 

2. External Sources:  Certain information contained herein has been obtained from third-party sources. Although Altimeter believes the information from such sources to be reliable, Altimeter makes no representation as to its accuracy and completeness.

1. Engagement: How Entrenched Are Users?  

2. Retention: ChatGPT starts to smile!

3. Time Spent: Attention is not all you need!

4. Putting it all together

In the AI gold rush, with trillions poured into capex and sky-high valuations, ChatGPT has emerged as the killer AI app, practically a household name today. However, will it be the most adopted AI app by the end of the decade? 

My analysis reveals a twist. Unless OpenAI radically evolves, the dark horses (Meta and Google) are poised to dominate consumer AI by decade's end.

Forget the hype. Ignore the flashy demos. Real-world adoption and usage tell the true story of AI's future. As the former Product Manager at Palantir, I've learned to trust product metrics. I dissect a panel of usage metrics for Gen AI apps versus best-in-class consumer and enterprise apps. The results? 

Take ChatGPT. Its weekly-to-monthly active user ratio (a measure of engagement) is 40%. Impressive for a newcomer, but a far cry from the 80% mark of billion-user apps. Meanwhile, Meta and Google's suite - Instagram, WhatsApp, Gmail, Chrome - have sky-high engagement and massive user bases. These aren't just apps. They're digital empires, primed for AI domination.

In this deep dive, we'll explore AI's true potential through concrete, quantifiable metrics. For entrepreneurs and investors eyeing the AI prize, consider this your reality check. The AI race isn't just about demos, it’ll be about user adoption, engagement, and stickiness. And in that game, the tech giants already have a massive head start.

#1 Engagement: How Entrenched Are Your Users?  

Engagement metrics reveal how deeply an app is integrated into users' lives -- a critical indicator of long-term success and growth potential. While DAU:MAU (Daily Active Users to Monthly Active Users) is the go-to metric for consumer applications, I use a modified version: WAU:MAU (Weekly Active Users to MAUs). This metric better aligns with the work week, accommodates intensive but sporadic use of AI apps within knowledge work, and enables fairer comparison with enterprise software applications.

High WAU:MAU ratios signal that users find consistent value week after week -- crucial for products aiming to become indispensable tools. When we look at the data, a striking pattern emerges:

This chart plots MAU (y-axis) against Engagement (WAU:MAU on x-axis). Takeaways:

1. The Billion-User Threshold: Apps need to achieve 80%+ WAU:MAU to reach a billion users. This exclusive club includes Instagram, WhatsApp, GMail, GCalendar, and Chrome -- all owned by Meta and Google.

2. Enterprise apps: Enterprise apps like Outlook, Teams, and Slack form the next tier at 60-70% engagement.

3. AI's Current Standing: Most Gen AI apps, including ChatGPT, hover around 40-50% engagement -- significantly behind both consumer and enterprise leaders. The notable exception is Character AI, which has broken out of this range.

4. The Character AI Exception: Notably, Character AI (recently acquired by Google) shows higher engagement, breaking out of the typical Gen AI cluster.

5. Meta + Google’s Distribution Advantage: Meta owns two of the most engaging, billion-user apps (Instagram and WhatsApp), while Google dominates with Chrome and Gmail. This positions both companies strongly for AI integration and workflows that let users access AI from within this existing real esatate.

So how can ChatGPT get to a billion users?

Data suggests that for ChatGPT or any AI app to reach billion-user status, it needs to dramatically increase their WAU:MAU ratio to 80%+. However, historical data shows this is a formidable challenging because WAU:MAU ratios stay remarkably stable over time. See chart:

In this chart above, only Duolingo and Google Calendar show significant improvements over a decade, and even these are modest (dotted lines). This stability implies that an app's engagement level is largely determined by its category and core value proposition, not easily altered by feature additions or marketing efforts.

Implications for Consumer AI:

1. ChatGPT's Uphill Battle: To reach a billion users, ChatGPT needs to more than double its current engagement levels - a feat rarely achieved in tech history.

2. Meta's Edge: With Instagram and WhatsApp already in the billion-user, high-engagement quadrant, Meta has a significant head start in deploying AI to a massive, engaged user base.

3. Google's Dual Strategy:

   • High-Engagement Products: Google can leverage its highly engaged user base in Chrome and Gmail to introduce AI features.

   • AI Specialization: The acquisition of Character AI, combined with their own Gemini, gives Google both a well-engaged AI product and the potential to improve their lower-engaged offering.

4. The Integration Advantage: Rather than standalone AI apps, the future might belong to AI deeply integrated into existing high-engagement platforms - a scenario that favors established tech giants like Meta and Google.

This data suggests that the race for AI dominance may not be won by AI-first companies alone, but by those who can most effectively integrate AI into already-indispensable applications, especially those with network effects. The established tech giants, with their massive, highly-engaged user bases, appear to have a significant advantage in this regard.

#2 Retention: ChatGPT starts to smile!

While engagement metrics show how often users interact with an app, retention reveals whether they find lasting value. Let's examine how AI apps compare to established enterprise and consumer applications in terms of retention.

This chart comparing 1-month retention rates reveals some intriguing patterns:

1. Big Tech Dominates Consumer: Meta and Google's consumer apps lead the pack. WhatsApp (90%), Instagram (90%), and Chrome (89%) show impressive Month 1 retention rates, underlining these tech giants' ability to create sticky products.

2. Enterprise Apps Steady: Google's Gmail (75%) maintains a solid retention rate, outperforming many other enterprise tools.

3. AI Apps Show Promise: Gen AI apps are performing relatively well, with ChatGPT (66%) showing retention rates comparable to some enterprise apps. This is particularly impressive given its novelty.

4. The Gap Persists: While promising, Gen AI apps still lag behind the best consumer apps in retention, mirroring what we saw with engagement metrics.

However, the story doesn't end with first-month retention. Let's look at how retention trends over time. But first a primer of what to expect in retention curves.  There are three kinds of retention curves: smiling, flattening, declining. 

Sequoia does a good breakdown on these: 

• Flattening: Suggests a core group of users found ongoing value.

• Declining: Indicates continuous user loss over time.

• Smiling: Rare and highly desirable, showing users returning after initially churning.

Analyzing retention from Month 0 to Month 10 reveals some fascinating insights:

1. Google's Dominance: Chrome and Gmail (both Google properties) display rare and valuable "smiling" retention curves, indicating increasing value to users over time.

2. Meta's Strong Performance: WhatsApp and Instagram show exceptionally stable, high retention curves, demonstrating Meta's ability to create lasting user value.

3. ChatGPT's Surprising Trend: Intriguingly, ChatGPT is the third app in our dataset showing a "smiling" curve. While not as pronounced as Google's apps, this trend suggests ChatGPT is providing increasing value to its users over time.

4. The AI Challenge: Most other AI apps show steeper drops in retention, highlighting the difficulty of maintaining a user base in this competitive space.

Implications for the Consumer AI Race:

1. Google and Meta's Advantage: With multiple products showing industry-leading retention, these tech giants have a solid foundation for introducing and scaling AI features to a loyal user base.

2. ChatGPT's Potential: The "smiling" retention curve for ChatGPT is a positive sign, indicating it might be on a path to creating lasting user value. However, it still has ground to cover to match the retention levels of top consumer apps.

3. Integration vs. Standalone: The strong retention of existing popular apps from Google and Meta suggests that integrating AI into these platforms could be more effective than creating standalone AI apps.

4. The Long Game: While AI-first companies like OpenAI (with ChatGPT) show promise, the consistently high retention of Meta and Google's products underscores the challenge they face in creating habitual, long-term use.

In conclusion, retention data further supports the notion that tech giants like Meta and Google are well-positioned to dominate consumer AI. Their existing products already demonstrate the ability to create lasting user value – a crucial factor in the AI race. However, ChatGPT's positive retention trend suggests that AI-first companies shouldn't be counted out just yet. The key to success may lie in bridging the gap between promising technology and ingrained user habits.

#3 Time Spent: Attention is not all you need

After examining engagement and retention, let's look at another crucial metric: time spent. This metric gives us insight into how deeply users are engaging with these apps and the potential for monetization.

Takeaways:

1. Social Media Dominance: TikTok leads, with Meta's apps (Instagram, Facebook) also showing strong performance.

2. Google's Ecosystem: While not topping the chart, Google's suite (YouTube, Chrome, Gmail) collectively captures significant user time.

3. AI Apps' Promise: Character AI (now owned by Google) stands out among AI apps, with usage times comparable to some social media platforms.

4. ChatGPT's Position: Shows respectable time spent, considering its utility focus versus entertainment-centric apps. It’s also showing some improvement over time up ~100% y/y from 13 mins / week to 26. 

It's crucial to note that time spent metrics are heavily influenced by business models:

• Ad-based models (like social media platforms) naturally incentivize more time spent.

• Paid or transactional apps often focus on efficiently completing tasks or transactions, potentially resulting in less time spent.

Implications for Consumer AI:

1. Meta and Google's Advantage: Their apps already command significant user time, providing a captive audience for AI feature integration.

2. Integration Potential: Incorporating AI into existing, heavily-used apps may be more effective than standalone AI applications in capturing sustained attention.

3. AI-First Challenge: Companies like OpenAI must create compelling experiences that carve out time in users' busy digital lives.

4. Business Model Considerations: As AI apps evolve, their approach to monetization will likely influence design choices that impact time spent metrics.

The time spent data reinforces Meta and Google's potential advantage in consumer AI. Their existing apps not only boast high engagement and retention but also command a significant share of users' digital time. This provides ample opportunities to introduce AI features within contexts where users already spend considerable time.

For AI-first companies, the path to widespread adoption may require either exceptional standalone experiences or strategic partnerships with platforms that already capture significant user attention.

#4 Putting it all together 

The data suggests Meta and Google are strongly positioned to lead in consumer AI, potentially outpacing current AI-first leaders and even ChatGPT - the killer app. 

Why? It's simple:

1. Engagement: Their apps are already daily habits for billions.

2. Retention: Users stick around, providing continuous data and opportunities.

3. Time Share: They own the lion's share of our digital attention.

4. Distribution: Unparalleled ability to deploy AI features at scale.

5. Network Effects: Massive user bases accelerate AI improvement.

Meta's secret weapon? Zuckerberg's "move fast and copy" ethos. His track record of rapid iteration and feature replication positions Meta to swiftly integrate and scale AI.

Companies like OpenAI continue to push technological boundaries, and ChatGPT has shown impressive traction. While their current incarnation may face challenges in reaching billions of daily users, I remain optimistic about their potential for significant product upgrades that could shift the landscape.

The AI race is far from over, but the current advantage lies with the giants who already dominate our digital lives. They're not just playing the game; they have home field advantage.

Thanks to Brad Gerstner, Thomas Reiner and Shreya Bhargava for their input on this article. 

The information presented in this newsletter is the opinion of the author and does not necessarily reflect the view of any other person or entity, including Altimeter Capital Management, LP ("Altimeter"). The information provided is believed to be from reliable sources but no liability is accepted for any inaccuracies. This is for information purposes and should not be construed as an investment recommendation. Past performance is no guarantee of future performance. Altimeter is an investment adviser registered with the U.S. Securities and Exchange Commission. Registration does not imply a certain level of skill or training. This post and the information presented are intended for informational purposes only. The views expressed herein are the author’s alone and do not constitute an offer to sell, or a recommendation to purchase, or a solicitation of an offer to buy, any security, nor a recommendation for any investment product or service. While certain information contained herein has been obtained from sources believed to be reliable, neither the author nor any of his employers or their affiliates have independently verified this information, and its accuracy and completeness cannot be guaranteed. Accordingly, no representation or warranty, express or implied, is made as to, and no reliance should be placed on, the fairness, accuracy, timeliness or completeness of this information. The author and all employers and their affiliated persons assume no liability for this information and no obligation to update the information or analysis contained herein in the future

It’s been 18 months since the “iPhone moment of AI”1 but the pace of developments has not slowed down. One of the biggest questions I’ve been thinking about is where does value accrue in Gen AI - today and in the future. 

Today, the Gen AI stack seems to be A-shaped vs V-shaped for cloud2: 

I split the stack into three layers - semis, infra and applications. Here’s a tally of the Gen AI revenues:

• Semi: Nvidia earned ~$18B of data center revenues in the last reported quarter (ending Jan 2024). Given they have 95%+ market share, let’s estimate ~$75B in annualized revenue for this part of the stack

• Infra: This layer would have hyperscalers (AWS, GCP, Azure) and the prominent inference clouds (Coreweave, Lambda, etc.). Generously estimate this layer at ~$10B in annualized revenue 

• Applications: Large language models (OpenAI, Anthropic, xAI, et al), image models (Midjourney et al), and other pure play generative AI applications. Some Gen AI use-cases might earn revenue disguised as “software” revenues, so for now I estimate this layer generously at ~$5B of annualized revenues 

In contrast, the cloud economy exhibits a much “intuitive” value distribution where the applications closest to the more end-customer earn the most value.

Bottom Line: The semis layer has captured ~83% ($75B of $90B revenues) of the Gen AI stack today. This is much higher than the ~10% that semis capture in the cloud stack today! 

#2 Where have profits accrued in Gen AI so far?

We get an inverted stack for profitability as well, with semis extracting the highest share today3:

Here’s the tally:

• Apps: Anthropic is estimated to earn ~50-55% gross margins. I assume the same for the overall layer

• Infra: I estimate the infra players earn ~65% GMs (without including GPU depreciation). If you include depreciation, that number drops to 25-30%

• Semis: NVIDIA is estimated to earn upwards of ~85% gross margins on their gen AI datacenter products. 

• The cloud stack is well studied with the exception of hyperscale gross margins. Azure is estimated to earn ~63% GMs and we assume that for all the infra layer 

Putting it together, the gross profits are significantly concentrated at the semis layer extracting $64B (out of $73B total). Visualizing this below (revenue x gross margin %) : 

It’s worth visualizing this on a bar chart to intuit the staggering relative proportions. If 100% represents the total gross profits in the system.

#3 Where do we go from here? 

We are in early phases of a platform shift where semis capture most of the value. I don’t expect the current structure of Gen AI revenues (inverted pyramid) to stay this way forever. I expect the application layer to capture a similarly high proportion of the value chain in due time. 

Here is a case study of how the value accrued in the mobile wave. Over a the last decade the value in mobile accrued first in semis, then in infra layer, and finally at the software layer: 

Similarly in the cloud, we first witnessed a datacenter buildout followed by the rise of the cloud service providers. AWS started in 2004 and got its first customers in 2010-12 (Amazon itself switched to AWS in 2010 and in 2012 Netflix joined them). 

In similar fashion, I expect Gen AI to follow suit. We are in Inning #1 (Semis) and I expect we’ll get to Inning 3 (Applications) by the end of the decade. As a corollary, given we are stating a low base, I see the biggest open opportunity to be in the application stack! 

There are a few critical questions along the way on this transition that I think about: 

A) Will NVIDIA continue to extract 85%+ gross margins? 

I believe not. It appears that NVIDIA’s margins have peaked and are on a downward trajectory. From SemiAnalysis: “We believe NVIDIA’s margins have peaked. We expect B100 and future families to have slightly lower margins, but furthermore, over the next few quarters H100 margins will also come down due to the H200 and H20.”

The key question I track as leading indicators of NVIDIA supremacy are: 

1. Lead times for GPU supply? Currently ~6 weeks

2. Trends in GPU rental prices? (Pricing sheet here)

B) Cloud apps earn 75-80% GMs but AI apps earn 0-50% margins, how will that evolve?

I believe AI apps will have improve profitability over time. There are a few levers that that will help drive better profitability in the future 

• Better pricing / value alignment: it’s widely known that In some cases, they are not profitable at all - especially for power users given the COGS are linked to the amount of usage.

• Custom silicon chips → lower TCO: all the hyperscalers and Meta are working on their own versions of semiconductors (Google, Microsoft, Amazon, and Meta). In due time, this should reduce total cost of ownership (TCO) not only because it will remove the margin stacking but also because allow them to specialize on workloads 

• Improvements in model architecture: I am seeing various non-transformer approaches such as state-space models (good for long context window use-cases such as coding), JEPA (good for video models), etc. 

• Reduction in model costs: driven by various techniques such as batching, distillation, quantization, mixture of experts, etc. models are getting cheaper at a rapid pace. As Bill Gurley points out: 

C) What about Gen AI in consumer? 

I expect a similar transition in consumer starting from the hardware layer. Much like the data center, consumer devices will be upgraded to AI PC / Smartphone / novel form factors (e.g. Meta glasses, Humane Pin, Rabbit R1). Consumer applications have three segments: information (search), entertainment (gaming, media), and transaction (travel, e-com etc.). Search queries are increasingly moving from informational to LLM-based searches, as my colleague Vivek calculates: 

Similarly in entertainment: between gaming and media both we expect value to shift from creators/producers to tech enablers. Again from my colleague Vivek: 

Thanks to Brad Gerstner, Sud Bhatija, Sanjiv Kalevar, Cobi B-Gantz, Omar Shaya, Jamin Ball, Vivek Goyal and Shreya Bhargava for their input on this article. 

The information presented in this newsletter is the opinion of the author and does not necessarily reflect the view of any other person or entity, including Altimeter Capital Management, LP ("Altimeter"). The information provided is believed to be from reliable sources but no liability is accepted for any inaccuracies. This is for information purposes and should not be construed as an investment recommendation. Past performance is no guarantee of future performance. Altimeter is an investment adviser registered with the U.S. Securities and Exchange Commission. Registration does not imply a certain level of skill or training.

This post and the information presented are intended for informational purposes only. The views expressed herein are the author’s alone and do not constitute an offer to sell, or a recommendation to purchase, or a solicitation of an offer to buy, any security, nor a recommendation for any investment product or service. While certain information contained herein has been obtained from sources believed to be reliable, neither the author nor any of his employers or their affiliates have independently verified this information, and its accuracy and completeness cannot be guaranteed. Accordingly, no representation or warranty, express or implied, is made as to, and no reliance should be placed on, the fairness, accuracy, timeliness or completeness of this information. The author and all employers and their affiliated persons assume no liability for this information and no obligation to update the information or analysis contained herein in the future.

1. Meet MANG (Microsoft, Amazon, Nvidia, Google)

2. The Demand Side: Investing in Customers

3. The Supply Side: Digital Tollbooths Printing Cash

4. What This Means for Entrepreneurs

#1 Meet MANG (Microsoft Amazon Nvidia Google)

Say hello to the newest group of prolific venture capitalists in data and AI: Microsoft, Amazon, Alphabet, and Nvidia. Total capital raised in their 2023 deals was $25B+, which is ~8% of the total venture capital raised in North America1

The MANG investments are particularly concentrated in Data and AI: their deals raised $23B, a staggering ~30% of all deals in Data and AI2. Notable 2023 investments: 

#2 The Demand Side: Investing in Customers 

MANG investments have evolved post the ChatGPT launch: 

• Before ChatGPT: Mobility startups: Waymo, Cruise, Rivian, Uber

• After ChatGPT:  Large Language Model (LLM) startups: OpenAI, Anthropic, Inflection, Adept, Coreweave

LLM startups are some of the largest customers of compute. It turns out that cloud providers funding their customers is a great business model. The investment capital makes a round trip back to the P&L in the form of revenue and helps them break even on the investment faster. Here’s how: 

From Semafor: 

“Only a fraction of Microsoft’s $10 billion investment in OpenAI has been wired to the startup, while a significant portion of the funding, divided into tranches, is in the form of cloud compute purchases instead of cash, according to people familiar with their agreement.”

Let’s assume that fraction is 50%. And let’s assume this investment will last them 2 years. That would generate $2.5b of revenue for Azure in year 1 (50% of $10b spread across 2 years). The enterprise value that creates at the median software revenue multiple of ~6x is $15b. 

…

$10b EV invested for $15b EV gain in year 1 and equity in a high-potential startup. 

…

$10b for $15b + equity. 

…

A traditional VC would have a different equation. Simply: $10b for equity.

Setting aside Bill Gurley’s concerns around revenue recognition, these are clearly NPV-positive investments for MANG!

While this business model of investing in your customer is not new (Illumina has done this with their spinoffs and Palantir with their investments into customer SPACs), it does raise a set of interesting ramifications for VCs.

• As MANG sets prices on these rounds (e.g. OpenAI by MSFT), what does the valuation represent?

• Does it make sense for a financial investor to participate in these rounds?

• For early-stage VCs, do I want my companies to raise $ from MANG?

#3 The Supply Side: Digital Tollbooths Printing Cash

The MANG businesses have evolved into modern-day digital tollbooths: earning fees from a vast number of customers. Whenever a new data or AI project launches, the cash register first rings at the hyperscaler. In fact, for every $1 of inferencing that customers pay their LLM provider (e.g. Anthropic hosted on AWS Bedrock), we estimate that >50% of that revenue goes to the hyperscaler to cover infrastructure costs + gross profit share. 

The combined operating profits for MANG in 2023 were $276B3, representing  ~5x growth in the last 10 years. What do you do with all this cash? 

One use of this cash is capital expenditures. Their business is also capital-intensive, necessitating continuous investment in infrastructure. The combined Capex for MANG in 2023 was $108B, which grew ~12x in the last 10 years4. 

By investing in infra-consuming startups, MANG can ensure the optimal utilization of their cloud services, reducing the risk around their voracious capital deployment. 

MANG now represents $7.6 trillion of combined market capitalization (18% of US market cap) today versus up 16x in the last 13 years from $480 billion (3% of US market cap) in 20105.

#4 What does this mean for entrepreneurs?

If you are an entrepreneur considering raising from MANG, there are a few tradeoffs you should consider.

Pros

• Access to lower cost of capital: if you are strategically aligned with the MANG, you might be able to access large pools of lower-cost capital (vs traditional VCs). This will manifest in your fundraise as a larger total round and perhaps one with lower dilution. However, raising at a higher valuation is not always a good thing. As our friends from HBO’s Silicon Valley put it:

• Shared GTM: being able to leverage MANG’s sales force would be a dream come true for most entrepreneurs. But has this worked out in practice for anyone successfully? What are the economics you need to share with MANG? What would a high concentration of revenue coming from one cloud mean for your business? 

• Access to models/team: getting access to state-of-the-art models, infrastructure or research could help improve your performance vs competitors, especially if the model layer is not your core competency.

Cons

• Alienating customers: if you raise significant capital from any one of the MANG, you will most likely get aligned on infrastructure. In doing so, you may forego a certain set of customers who are predominantly built on the other clouds. Relatedly, having a major strategic MANG investor can be a turnoff to potential customers or business partners who view themselves as competitors to the MANG or one of their offerings. For instance, enterprises need to go via Azure OpenAI service to access OpenAI and can’t access it via GCP/AWS as of this writing

• Multi-round dynamics: Do you need to raise more capital after this round? Will one of the MANGs be around for that next round? If not, what does this mean for your future rounds? What if the price set by a strategic partner cannot be matched by a financial partner? 

• Cap table construction: If you do aim to go public, having a large strategic investor on your cap table can create negative signaling for your stock. Ideally, your largest shareholders are investors who are known to provide steady long-duration capital, who are typically financial investors. 

• Misaligned incentives: MANG aren’t purely financial investors and will have strategic directives that may be misaligned with other purely financial investors. They may limit/block acquisition offers and financing rounds that would have been advantageous from a financial standpoint but didn’t align with their strategic interests.

Thanks to Brad Gerstner, Cobi B-Gantz, Sud Bhatija, Angad Singh, Govind Chandrashekar, Eric Trac, and Shreya Bhargava for their input on this article. 

The Problem

BLUF (Bottom Line Upfront): The goal/problem/mission is to recreate the human intelligence at a machine level (Artificial Intelligence). Breaks down to broadly three atomic steps Receive input → Understand → Act.

• Step 1: Receive inputs (video, images, sound, text and soon hopefully olfactory data too)

• Step 2: Understand (humans use different parts of the brain to process different sensory signal, machines use various custom architected GPUs/supercomputers optimized to solve specific problems (vision or speech or text))

• Step 3: Act (drive a car based on road conditions).

Why is that hard? Consider a simple problem — someone across the room throws a ball at you and you catch it. Sounds simple, but actually, this is one of the most complex processes we’ve ever attempted to comprehend — let alone recreate at a machine level. What actually happens in the human circuitry is: the image of the ball passes through your eye and strikes your retina, which does some elementary analysis and sends it along to the brain, where the visual cortex more thoroughly analyzes the image. It then sends it out to the rest of the cortex, which compares it to everything it already knows, classifies the objects and dimensions, and finally decides on something to do: raise your hand and catch the ball (having predicted its path). This takes place in a tiny fraction of a second, with almost no conscious effort, and almost never fails. So recreating human-esque vision in machines, as an example, is a set of interconnected problems each of which relies on the other.

The Solution — Deep Learning

BLUF: Deep Learning refers to a specific kind of Machine Learning technique — multi-layer Artificial Neural Networks (ANN) (other synonyms: multilayer perceptrons, deep neural nets, etc.) — which is very good at learning patterns and using that intel to classify or predict outputs given a set of inputs.

Demo:

1. Tinker with a Deep Learning framework (ANN) in your browser here.

2. Checkout Clarifai — a demo in object recognition.

ANN is actually a pretty old machine learning model that was inspired by the human brain (invented in 1957 by Frank Rosenblatt). A neural network is an interconnected set of nodes (each with an activation function) that are connected by edges, with differing weights, to solve problems like classification, regression, etc. For classification problems, for instance, each node will have a similar classifier but each node’s input is modified by different weights (edges) and biases. If you stack a lot of layers in a neural network (more than 3), that is suddenly considered deep learning. For simple pattern recognition problems, you could use simpler frameworks (Logistic regression or Support Vector Machines), but for anything with medium to high complexity deep learning will outperform other frameworks. Patterns are presented to the network via the ‘input layer’, which communicates to one or more ‘hidden layers’ where the actual processing is done via a system of weighted ‘connections’ as shown in this image to the right. ANNs are able to break down the bigger pattern into smaller patterns. Image recognition, for instance, is performed in small chunks to see if there’s a human face (starting from looking for simple things like vertical and horizontal edges for instance).

The Building Blocks (BB) of Deep Learning

There are three basic characteristics of any ANN : data, computational hardware and an algorithm.

BB#1 Data Scale (needed for accurate training)

The more data you let the ANN train on, the better the outcome is. To get a sense of scale, Google’s face recognition algorithm trained on about 260M images in 2015.

However for most real world problems, you probably need somewhere around 10k to 100k data points. If you don’t have a lot of data you could

• Fine Tune: Take a strong pre-trained network that has trained on a large dataset (easy to find) and use it as an initial network for your problem. This was not true until a couple years ago, but now is possible given a much better ecosystem.

• Augment Data: You can add random transforms (such as crops, mirrors or rotations) to artificially increase the amount of training data.

No matter what learning algorithm you use, performance gets better with more data.

BB#2 Computer Hardware (needed for quick iterations)

Value in AI is driven by empirical testing. Quick iteration speed on various algorithms helps researchers try out more ideas. E.g. How did scientists discover that Recurrent Neural Network (RNNs) is the best architecture for speech recognition? 7 years ago, Andrew Ng and team started out using an audio clip, cut it into steps, feed into a vanilla NN (simplified diagram below) to get the output transcript. This didn’t work and after some experiments they realized they needed to move to a Recurrent NN that led to the creation of Deep Speech

CPU → GPU → Cloud → HPC.

• At the core of ANN algorithm is linear algebra. Linear algebra is inherently a parallelizable problem and hence parallelizable GPUs (such as those from NVIDIA) have become common hardware for training large scale NNs. Here’s a good primer on the linear algebra behind cost functions

• With the rise of High Performance Computing (HPC) experiments that took weeks and months are now being done in days and hours. Bleeding edge of AI is moving to HPC, bringing down the iteration speed to days. Instead of testing dozens of ideas a year, we can now test hundreds of ideas a year.

BB#3 Algorithms for Feature Representation: Each node in each layer is a computational units. This means each node performs some computation (normally nonlinear like a sigmoidal function, hyperbolic tangent nonlinearity, or rectifier linear unit) given its inputs from the previous layer. The feature representation / loss-function that works best for a specific problem is a result of teams of 1000s of researchers spending dozens of years. Example — image from SIFT paper for computer vision that took 10 years to come up. Thankfully, this hard-work has now been abstracted out (for vision, audio and text). The growth here is very empirical, hence the need for better computing infra to move the research faster.

Unsupervised learning using unlabeled data.

• Feature extraction

• Unsupervised learning

• Pattern recognition

• Best algo — RBM (Restricted Boltzman Machine), auto-encoders

Text processing (RNTN Recurrent Neural Tensor Net (character level parsing))

• Image (DBN Deep Belief Net, Convolutional Net)

• Object (RNTN, Convolutional Net

• Speech (Recurrent Net). Examples of NLP features: Stemming, Anaphora, Named entity recognition, Ontologies (WordNet)

General

• General Classification. → RELU

• Time series analysis → recurrent net.

So what’s the big hype?

BLUF: Since 2012, there has been a huge jump in accuracy of ML techniques (~74% → 84%), because they started using ANNs and realizing the giant leap provided by ANNs, the industry is investing resources (hardware, software, data, money) at an unprecedented rate.

DL became feasible only in recent years because of four broad advancements:

• Large data sets image-sets given trillions of sensors everywhere. To build face recognition models that are as good, if not better then humans: FB used 4.4M to train in 2014 and Google used 260M images to train in 2015.

• GPUs and HPC — at the heart of deep learning is Convolutional NN which does repeatedly similar computation, so you can push much of the logic to the hardware layer (in Silicon). So more and more specialized hardware is being built, though still in very exploratory state

• Algorithmic improvements (Lots of potential detail here, skipping for the summary) are abundant given compounding from higher processing power, larger datasets and more investment into the area.

• Capital investments, MnA activity is fueling the resources pouring in

Applications

Today, Deep Learning is the state if the art approach in almost every computer vision (ANNs are better than a human for face recognition, for instance), voice recognition (Siri, Alexa) and Natural Language Processing tasks. Application of Deep learning today is easy.

Step 1 Choose your weapon (package) -

• Caffe/Digits (Matlab + Python)

• TensorFlow (Python)

• Torch 7

• Theano/Keras/Lasagne (Python)

• MatConvNet (Matlab)

Step 2: Get a GPU or use cloud resources (AWS, IBM Softlayer)

Today it’s used for (non-exhaustive list below, a much more comprehensive one here):

• Machine vision — Image classification, Object recognition

• Video recognition — Theft prevention and Self-driving cars

• Speech recognition — Siri, Alexa

• Text processing — fact extraction, machine translation, sentiment analysis (movie ratings), caption creation.

• Healthcare — Earlier Cancer detection, Drug discovery based on chemical structure

• Radiology — classifying malignant tumor versus benign tumor. Using MRI, FMRI, CT scans data

• Finance — short and long term triggers to create alpha, portfolio allocation, etc.

• Digital advertising — Segment users to make more relevant ad, Based on historical prices, optimal bidding

• Fraud detection — credit card charges

• Customer intel in marketing and sales — Upsell and cross-sell

• Agriculture — predict and insure against problematic environment issues.

• Play Games optimally (especially ones with large number of permutations) — AlphaGo

My take on the next frontiers / next big waves:

1. Generative Learning: Learning about learning, generative learning is the next step. Task: Not just answer the question, but make machines ask the right questions. That way you’re able to creating AI with some understanding of context (e.g. the Boston driverless Tesla versus the SF one). A lot of current AI is Task-driven work but the sense of General AI (understanding intention / purpose / communication / context) is not getting anywhere. We are very far from generative learning: we are able to do captioning and even QnA style work. But we need to get the algorithm to learn to explore to ask the right question.

2. Chip Design, ASIC and FPGAs : Generally speaking, the more function specific the hardware is (custom designed ASICs, FPGAs) the more efficient it is at that task. At the same time, however, it loses plasticity and cannot adapt to changing software demands. So the right hardware for Deep Learning, and other AI algorithms, has to have the ability to both specialize and morph. This would not be a general purpose computer, or an ASIC device, this would be a particular new type of neuromorphic hardware. Hardware, however, is only half of the problem, and the real game-changer I think will be a platform that is the right combination of software and hardware. This platform would allow you to code and run applications on neuromorphic hardware as easily as you would run a Python or C++ application.

Companies worth looking at:

Healthcare

1. Sano: Continuously monitor biomarkers in blood using sensors and software.

2. Enlitic / MetaMind / ZebraMedical: Vision systems for decision support (MRI/CT). Enlitic has developed deep learning networks that analyze medical imaging data such as x-rays and MRIs. Their networks increase diagnostic accuracy in less time and at reduced cost when compared to traditional diagnostic methods. Enlitic’s software also allows comparison of an individual patient’s radiological data with millions of other patients who received the same diagnosis in order to identify and track treatment outcomes for the most similar cases.

3. Deep Genomics / Atomwise: Learn, model and predict how genetic variation influence health/disease and how drugs can be repurposed for new conditions. Deep Genomics uses deep learning networks to predict how both natural and therapeutic genetic variation changes cellular process such as DNA-to-RNA transcription, gene splicing, and RNA polyadenylation. Applications include better understanding of diseases, disease mutations and genetic therapies. Spidex, their first product, provides “a comprehensive set of genetic variants and their predicted effects on human splicing across the entire genome. “

4. Flatiron Health: Common technology infrastructure for clinics and hospitals to process oncology data generated from research.

5. Google filed a patent covering an invention for drawing blood without a needle. This is a small step toward wearable sampling devices.

Enterprise Automation

1. Sigopt — SigOpt is the optimization platform that amplifies your research. SigOpt takes any research pipeline and tunes it, right in place. Our cloud-based ensemble of optimization algorithms is proven and seamless to deploy, and is used by globally recognized leaders within the insurance, credit card, algorithmic trading and consumer packaged goods industries.

2. Arimo — empowers companies with a new way to grow revenues and better understand their business. With innovative Distributed Deep Learning architectures, Arimo’s Behavioral AI solution allows companies to incorporate customer behaviors and digital assets into their business, marketing, and strategic decisions. Team includes executives who held held senior positions at Google, Yahoo, Amazon, SAP, Microsoft, and VMware. Arimo received $13M in a Series A raise from Andreessen Horowitz, Lightspeed Ventures, and Bloomberg Beta.

3. Ditto Labs has built a detection system that uses deep learning networks to identify company brands and logos in photos posted to social media. Their software also identifies the environment in which brands appear and whether a brand is accompanied by a smiling face in the photo. Clients can use Ditto to track brand presence at events or locations, compare brand performance with competitors, and target advertising campaigns.

4. Nervana provides a deep learning framework called neon that allows users to build their own use-case deep learning networks. They also provide a deep-learning cloud service where users can import and analyze their data using one of Nervana’s deep learning models or a model of their own design that uses neon. They suggest solutions in the areas of finance, energy and online services among others.

5. Databricks — Databricks is a company founded by the creators of Apache Spark, that aims to help clients with cloud-based big data processing using Spark. It brings DL techniques into enterprise customer’s application to make it faster, predictive and user-friendly.

Machine Vision, Speech recognition

1. Comma.AI (George) — Self-driving car add-ons.

2. Affectiva — MIT Media Lab spinoff that uses deep learning networks to identify emotions from video or still images. Their software runs on Microsoft Windows, Android and iOS and works with both desktop and mobile platforms. Applications include real-time analysis of emotional engagement with broadcast or digital content, video games that respond to the player’s emotions, and offline integration with a range of biosensors for use in research or commercial environments.

3. Gridspace uses deep learning networks to drive sophisticated speech recognition systems. Their networks begin with raw audio and build up to topic, keyword and speaker identification. Their Gridspace Memo software is designed to identify speakers, keywords, critical moments, and time-spent-talking, along with providing group take-aways from conference calls. Gridspace Sift provides similar information about customer service and contact calls.

Security

1. Deep Instinct is a cybersecurity company that uses deep learning networks to detect, predict and prevent advanced persistent threats in real time. Their software operates on all servers, operating systems and mobile devices. A program tailored to the customer’s security needs serves as the middleman between Deep Instinct’s deep learning network and the customer’s desktop and mobile platforms. The customer’s endpoints are protected whether or not they are logged into the company network.

Appendix

Some Glossary notes for common terms that one may come across while reading about Deep Learning

• Gradient Descent — Gradient descent is a first-order iterative optimization algorithm. To find a local minimum of a function using gradient descent, one takes steps proportional to the negative of the gradient (or of the approximate gradient) of the function at the current point.

• Back Propagation — common method of training artificial neural networks used in conjunction with an optimization method such as gradient descent.

• Convolutional Neural Network — Great summary here

• Deep Belief Network — generative graphical model, or alternatively a type of deep neural network, composed of multiple layers of latent variables (“hidden units”), with connections between the layers but not between units within each layer.

• Restricted Boltzmann Machine — a generative stochastic artificial neural network that can learn a probability distribution over its set of inputs. Stack together to create a DBN

• Deep Reinforcement Learning — Generative Intelligence, Great summary here

The below write-up are my jottings in appreciation of Amy. The team has built an incredible product, I liked it enough to throw Amy some curve-balls (also called edge-cases). In addition to those, I’m writing aboutsome enhancement requests below that I’ve passed on to the engineering team. The goal is to make Amy a better product, so please feel free to suggest features that come to your mind.

I also have some questions for you embedded in the writeup, where I’ll appreciate your take on whether you think the response was AI-generated or human-generated. In answering these, you’ll hopefully challenge your understanding of the capabilities of man vs. machine.

Installation and Setup

Installation: Unlike traditional software, there is no installation involved for Amy (no sign-in, passwords, or downloads”. I signed up for a beta preview on Jan 29, 2016 and was invited to try it out on Mar 23, 2016 (welcome email attached). Together with the welcome email, you get a chance to configure all scheduling preferences, which I’ll cover next.

Setup:

With the link above, you’re ready to go with the following settings within minutes of hanging out on their website -

• Calendars (I added my google calendar. Might add work calendar if Amy’s utility remains +ve)

• Scheduling preferences (in person meetings need 1hr + 15min buffers, calls need 30mins, etc.)

• Email addresses

• Contact details (phone line, Skype handle, etc.)

• Meeting Location preferences (office address for breakfast, lunches and coffees; home address for dinners, etc.)

As of today, Amy or Andrew are available for beta testing via their website, go sign up!

Features

The Basics: You copy Amy on an email thread with the person(s) you’re scheduling with. Amy takes care of the rest (you get dropped, Amy continues conversation with recipient(s)). Amy does so in an extremely personable way and I’ve had receivers on the other side actually mistake her for a human (until I pointed out her signature to them, where she clarifies her bit-flipping origins). This is a WIN! Once both Amy and your recipient(s) come to an agreement on a time, she sends out a calendar invite to all parties. You can track everything Amy is working on at any given time by visiting their website. It looks like this (names hidden)

Amy’s response to your recipients looks something like (names hidden):

…quite personable right?

Some features I found missing in the current release:

1. Amy doesn’t send out timezone-sensitive meeting times. This has been the biggest problem for my scheduling requests, since most of them tend to be international calls. This is especially hard when the timezone conversions are over 5 hours (California to New Zealand, Singapore, India, Europe, New York, etc.). Plausible limitation: There’s no good way for Amy to know which timezone the other person is in. Also, timezones are the bane of any software engineer who has to deal with them. 2nd best solution: Perhaps add functionality for the user to let Amy know recipient’s time-zone?

2. Amy doesn’t accept timezone-sensitive responses. If she proposes 9am PT, if you in New York interpret that as 12pm ET and respond back “Yes 12pm ET works”, Amy gets confused. However, if your recipient asks Amy to provide times in another timezone (say GMT), then she’ll understand and make the conversions accurately.

3. Updating the meeting is a hassle, especially if managing a tight calendar. A call with a friend in London needed to be pushed by 30 minutes, 10 minutes before it started. Because the calendar invite was not owned by me (it’s owned by amy@x.ai), I could not move it. My best bet to write to Amy and wait for her to adjust it. This is not as instantaneous as you would expect software to be.

4. Amy is unable to setup recurring calls so far.

5. There’s no way for you to see the conversation Amy is having with your recipient or to intervene it. For instance, if you look at meeting #5 in the screenshot above, Amy is trying to request a venue a very busy friend and I know it’s going to be a phone call. It’d be cool to have a way to override that request and tell Amy it’s just a phone call.

6. (H/T Chirag Ahuja for this suggestion) Amy sends out a couple of time-slot suggestions instead of open blocks of time for the recipient to pick from. More discussion on this with Amy below, but it would be cool if it could send a soft-link to the latest calendar (since the calendar may have moved from the time it suggested time-slots and the recipient acted) for the recipient(s) to pick a time slot. I understand this comes with engineering trade-offs.

7. Amy uses emailIDs as the unique identifier. If the same person has multiple email addresses, Amy will be busy trying to setup an n-way conference call where n is the number of email addresses. Resolution to this is actually probably non-trivial given the quest for a true unique identifier can be endless.

What other features would you like your PA to have that are achievable by software?

Support and Privacy (has an embedded question that’ll provoke thought on boundary between man and machine)

Support here is of two types given the nature of product, first is directed to Amy (via e-mail) and the second type is via the support page on x.ai’s website. I’ll talk about my experience with the former since the latter is not particularly out of the blue.

This is where the true boundary between what a human can handle and machine can handle comes to play. Because at the back of my mind, I think of Amy as a real person, I wrote to Amy asking for a very simple support question such as the following —

On Sun, March 27th, 2016 at 4:50 pm (EDT), Apoorv Agrawal wrote:

The response (Amy’s?) to this was:

On Mon, Mar 28, 2016 at 9:56 AM, Amy Ingram <amy@x.ai> wrote:

Hi Apoorv,

I avoid putting tentative holds on your calendar and wait until all your guests have confirmed a time before sending out an invite. This way, you can always scan your calendar and know that everything on it is a firm commitment, and your calendar won’t be cluttered with holds that don’t end up being the final meeting time.

That said, if a time I’ve proposed to your guests becomes unavailable, I’ll be sure to find a new time that works to ensure you’re never double booked.

Amy

Question for you: given all information in the above exchange (timestamp, rhetoric, etc.), do you think the response was human-generated or computer/AI-generated?

I think it was human generated, but I’m not a 100% sure. But if my hunch is right, this is a huge problem! Here’s why: if like me, you accidentally forwarded a thread with an issue to Amy trusting no human would read the email content, your confidence has probably been breached.

This is a tricky support zone because when is the user expected to know whether Amy’s got this or this needs human intervention. Perhaps it’s best if Amy be transparent and reply back with a couple of clickable options like

“I’m unable to answer your question and need my creators to take a look. To protect your privacy, I need your permission. Is it ok if I forward this thread (with the email content) to my creators (engineers at x.ai)? If not, please feel free to write to them at support@x.ai .”

Performance

Here is some data on the outcomes

• Of the 11 requests I made to Amy, it has completed acting on 9 of them.

• Of the 9 on which it acted, 3 were scheduled perfectly (win!). 3 were scheduled but needed some intervention from me (sort of a win!). 3 were canceled due to a non response (it waits 2 days).

Happy to provide a more detailed breakdown, but skipping here for brevity.

Conclusion

Amy is effective, incredibly personable and on track to create a shake up. My recipients and I both forgot that Amy was software, at least once in the process and engaged with it like we would with humans. That’s a huge win on sentimental grounds for the team. Congrats x.ai team!

Watch-out world! Amy Ingram is ready to take on the scheduling world (or at least a part of it) with the scale, malleability and effectiveness that software can provide in the long run. Just like the physical bookstore industry went through a shuffle due to its software rival Amazon, online-scheduling seems to be another one in the queue of a re-shuffle!

If you have any feedback on the write-up, I gladly welcome all of it. Fire away!