Summary

The Big Four will spend more than $630B on capex in 2026, with capex as a share of operating cash flow rising from 40% in 2023 to nearly 90% projected, but Zhou Mo’s core call is that the anxiety is easing: the larger the model company, the faster it grows (“体量越大，增速越快”). Anthropic’s ARR was approaching $45B at the end of April, with monthly growth of 50%+ for several consecutive months and $100B in sight by year-end; a rumored $900B valuation equates to roughly 9x year-end ARR, or “roughly 30x P” assuming eventual margins of 20%-30%—versus Microsoft’s long-standing 35x, which means it does not look particularly expensive on a longer-dated view.
Big tech has underperformed semiconductors since last July because, once the market ran the 2026 capex math, it found that “these companies would have almost no cash flow left,” raising the bar for revenue acceleration. The subsequent rotation from mega caps into the semiconductor supply chain created a double hit from fundamentals and positioning; Amazon was the most extreme case, with TTM free cash flow falling from $38B to $1.2B.
There is hard evidence of supply tightness: after GPU long-term contract prices fell by $0.05-$0.1 per month for a year, they reversed in December, while a rising share of long-term contracts squeezed the short-term market—“short-term contracts are a bit like hotel prices.” Storage long-term agreements were signed in bulk from mid-March to late April, locking in floor prices for the next several quarters and even the next 3 years, forcing a one-time capex increase this quarter; “the overwhelming majority of the increase is inflation, while demand will matter more the further out you go.”
The power structure is reversing: model companies are replacing public clouds as the distribution gateway, and customers “do not much care which cloud the API runs on.” In Q1, 80%-90% of Anthropic’s cloud API sales ran through AWS Bedrock, with distribution expanding to Vertex and Azure AI Foundry; “the more fragmented it is, the stronger its bargaining power.” That makes custom silicon a necessity for cloud providers: Trainium shifted from being externally discounted to being supply-constrained this quarter, while Andy Jassy says it saves tens of billions of dollars in annual capex and delivers a margin advantage of several hundred basis points.
Zhou Mo sees Google as the hyperscaler with the strongest moat: at the same FLOPS, TPU training costs only 20%-30% as much as GPU training, can train models 2-3x larger within the same generation, and carries higher gross margins than CPU sales, without the internal conflict of “an ICE-car maker building an EV.” More than 70% of Gemini revenue is booked to GCP and contributed 4-5 points of acceleration this quarter; the bottleneck is CoWoS, which Nvidia has locked up—TPU volume will be under 4M units this year and near 10M next year, creating an internal-versus-external demand conflict that “will not be easy to solve this year.”
Microsoft has the least favorable setup: Maia 200 can run only GPT-4o-class small models, not coding models such as 5.2, 5.3 and 5.4, while Copilot’s native integration, permissions and compliance advantages are being dismantled by Computer Use, harnesses and stronger multimodality, and the Responsible AI layer can even “dumb down” the product. The path forward is Copilot Studio’s consumption model: revenue is already one-fifth to one-quarter that of Office Copilot, with some customers spending $10K-$20K per month rather than $30 per seat.
Meta’s post-earnings collapse was a capex mismatch, not a revenue problem: true product acceleration was only 3 points and barely reached the high end of guidance, while $10B of incremental capex pushed Q4 10%-20% above the original assumption, prompting investors to linearly cut 2027 earnings. Meta is nevertheless “the most aggressive company at using AI to improve efficiency”—PIP evaluations include code volume, while sponsored-content inventory and the LATTE model are producing real incremental gains; its in-house Avocado reached Gemini 3 levels within months, “already very good,” with a self-use mandate rather than a challenge to OpenAI and Google.
The endgame math is improving: of roughly $700B in cloud capex this year, the GPU- and AI-related portion can receive a 60%-70% discount under Zhou Mo’s framing; model companies could reach roughly $200B of combined ARR by year-end, covering about two-thirds of the relevant capex, and $400B next year would lift that to three-quarters. Coding has already achieved “90% of AGI’s capabilities,” while other domains may follow in 2028-29; substitution of OPEX—which is about 30% of enterprise revenue and 7x IT spending—should accelerate, making token taxes and transfer payments unavoidable issues.

Deep dive

1. Last July Was the Inflection Point: Big Tech Underperformed Semiconductors Once 2026 Capex Was Counted

Cao Qingyun opened with the headline numbers: the Big Four will spend more than $630B on 2026 capex, a level close to the combined scale of major infrastructure projects of the 20th century relative to GDP. Capex as a share of operating cash flow rose from 40% in 2023 to 65% in 2025 and is expected to approach 90% in 2026; Amazon’s TTM free cash flow fell from $38B to $1.2B.
Zhou Mo’s review: before last July, mega-cap tech and semiconductors moved together. Afterward, the market began running the 2026 capex math: “once you include 2026 capex, you find that by 2026 these companies will have almost no cash flow left.” That raised the requirement for visible revenue acceleration, but Microsoft and AWS did not accelerate meaningfully in 2H last year; AWS only showed clear acceleration in Q1 this year.
The second force was positioning. The alpha in this cycle has been concentrated across the semiconductor value chain, while investors had relatively small semiconductor allocations; adding semiconductors necessarily meant cutting mega caps. The selloff was driven by “both fundamentals and positioning.”

2. Chips Are the Scarce Input, but Model Companies Command the Highest Valuations: The Gateway Is Changing Hands

Cao Qingyun framed the industry-economics puzzle: Nvidia has 70%+ gross margins, clouds have 35%-45% operating margins, and model companies have gross margins in the 30%-40% range. Scarcity sits at the chip layer, yet valuation sits at the model layer. Zhou Mo’s answer: everything looks expensive on P/E, but not necessarily on a forward view. Anthropic’s ARR was near $45B at the end of April and could reach roughly $100B by year-end; a $900B valuation would be about 9x year-end ARR, or roughly 30x P at a 20%-30% profit margin, versus Microsoft’s long-standing 35x.
The more structural answer is bargaining power. In the SaaS era, public clouds distributed all software traffic and turned CPU rental with 40%-plus gross margins into a 60%-plus-margin business. Now customers buy APIs through model companies and “do not much care which cloud the API runs on.” The cloud is being pushed downstream, while model companies are assuming something like the role public clouds once held.

3. The Capex Narrative Has Moved Through Several Rounds: The Larger the Company, the Faster the Growth

Zhou Mo rejected the “prisoner’s dilemma” framing and instead described a sequence of narrative shifts. At the end of 2024, the concern was insufficient data and poor model generalization—“that was genuinely true at the time,” which led to the DeepSeek moment in January. Then RL, mid-/post-training and the labeling ecosystem began to work; by the end of last year, model companies could allocate more than half of their compute to reinforcement learning.
The key change over the past 3 months is that Frontier Labs are scaling without slowing down—and are instead accelerating. Anthropic has posted monthly ARR growth of 50%, while OpenAI and Gemini are approaching 50% quarter-on-quarter growth from much larger bases than last year. The pattern has become: the larger the company, the faster it grows.

4. Hard Evidence of a Shortage: GPU Prices Reverse and Storage Contracts Lock In

The leading indicators all point to tight supply: GPU rental lead times are lengthening, while long-term contract prices fell by $0.05-$0.1 per month for the past year before reversing in December; both spot and contract prices are now rising. The mechanism is the structure of the capacity pool: Frontier Labs are locking in more long-term supply because of RL and coding, with the long-term share rebounding to around 80% and squeezing the most elastic short-term contracts. “Short-term contracts are a bit like hotel prices.”
Storage was the direct catalyst for the capex jump this quarter. Public clouds negotiated long-term agreements in bulk from mid-March through late April, locking in floor prices for the next several quarters and even the next 3 years. This round of price increases “is not a one-off,” requiring a one-time capex increase for subsequent quarters; and because only the floor was locked, not the upside, further increases may still be needed. The increase this quarter is “overwhelmingly inflation,” with demand becoming more important over time.

5. Custom Silicon Rewrites the Profit Pool: Trainium’s Inflection Point and the Graviton Template

Trainium’s inflection point is here. Its main customers are Amazon itself and Anthropic, and Amazon heavily discounted the product externally last year. Q1 was the strongest quarter for hyperscaler bargaining power in a year, and Trainium 2 was also supply-constrained. The business moved “from something that did not generate much profit to something that generated a lot of incremental profit this quarter.” Andy Jassy has said it saves tens of billions of dollars in capex annually and provides a several-hundred-basis-point operating-margin advantage.
Graviton 2 offers the historical template: a 30% improvement in price-performance, with 20% passed through to customers and 10% retained by AWS as spread. “The public cloud business model is essentially deflationary”: custom silicon lets a provider cut prices while maintaining higher margins than competitors, which helped AWS defend share at the time.
ASICs are difficult to deploy. Software, libraries and compatibility create a long list of problems, and TPU has been the smoothest implementation. Collaboration with Gemini has already resolved many adaptation issues, while Anthropic and Meta have encountered relatively few; “the communication between Trainium and Anthropic is certainly not as seamless as that between TPU and the Gemini team.” Within the TPU supply chain, OCS optical suppliers are gaining incremental share.

6. “AI Revenue” Means Four Different Things

The definitions vary by company. Microsoft’s disclosure of how much AI contributes to revenue growth includes GPU rental, helping OpenAI sell APIs, and AI-related growth from third-party products such as Fabric and Databricks. Amazon mainly means Bedrock, which grew 170% quarter on quarter this quarter, largely from its share of Anthropic API revenue. Google does not report a standalone AI number but offers the clearest picture: GPU/TPU rental, direct TPU sales and Gemini revenue, roughly 70% of which is booked to GCP. Its AI mix is the highest.
Meta never discloses an AI revenue figure because the contribution of recommendation algorithms cannot be isolated: “it is difficult to quantify how much of its revenue is actually generated by AI.” The common fact is that, starting in Q1 this year, AI revenue began accelerating as a share of revenue across the public clouds.

7. Three Monetization Paths: Cloud Has the Highest Ceiling, Agents Are Hardest to Prove

The starting point is that model companies have largely abandoned 1P data centers. OpenAI found building its own to be “very challenging” and shifted to 3P rental. Anthropic’s nominal 1P setup is actually GCP placing TPU, racks and OCS as a bundle inside Anthropic’s data center. Only xAI may still be building a genuinely 1P data center. Compute is therefore settling into public clouds and neoclouds, giving cloud the highest revenue ceiling.
The concern for cloud is channel fragmentation. In Q1, 80%-90% of Anthropic’s API sales through the cloud ran via Bedrock; Vertex came online in Q4 last year, followed by Azure AI Foundry in Q1. Foundry’s share of related APIs could jump from a few percentage points to the teens or even 20% in Q2. “The more fragmented it is, the stronger its bargaining power,” while the terms for connecting directly to data centers are becoming more demanding.
The one-line problem with selling Copilots and agents is: “it is difficult to prove that your Agent product can do a better job than a Frontier Lab.” Office has 400M users and is the world’s largest software product, but Claude Cowork and Gemini Enterprise can call Office through Computer Use, with multimodal PowerPoint output that may be even better. The iteration cycles are fundamentally asymmetric: a startup can ship a major iteration in a week, while at a large company “the process from the PowerPoint to the presentation may not even be finished after 3 weeks.”
Recommendation algorithms face diminishing returns. Once the low-hanging fruit is gone, the business must turn a 7-8-layer pyramid into a cylinder with 10-20 layers, making each additional 1-2 points of ROI contribution less visible. The market’s focus has already shifted from “AI’s contribution to advertising” to “AI’s contribution to cloud revenue growth.”

8. The Debate Over $37B of AI Revenue: Watch Growth; RPO Quality and OpenAI Dependence Matter

Asked how much of Microsoft’s $37B annualized AI revenue is merely relabeling, Zhou Mo did not dwell on definitions: “the most important thing is growth.” The figures across companies are all highly credible; the difference is the cadence of acceleration. GCP has the cleanest setup because more than 70% of Gemini revenue is booked to GCP and it has the largest TPU volume. Microsoft is at the greatest disadvantage: Maia 200 can run only GPT-4o-class small models, while “you cannot run the models related to Codex or 5.2, 5.3 and 5.4 on Maia,” and APIs are not OpenAI’s primary growth engine.
Cloud positioning should be evaluated through growth percentages and each quarter’s share of incremental revenue. RPO quality is high across the board, but the accounting differs: part of GCP’s backlog doubling reflects the start of direct TPU sales, which are recognized upfront net of Broadcom’s cost rather than amortized as one-fifth of rental revenue over 5 years.
Microsoft’s 45% OpenAI share of RPO is a real concern. Model companies are diversifying cloud providers, and OpenAI updated its agreement just 2 days before Microsoft’s and Google’s earnings reports. Some AWS capacity in Q1 may have been held for OpenAI, and the capacity reserved for OpenAI may not be recognized as revenue. Still, “the market is actually quite forgiving: if your acceleration is strong enough, the concern is much smaller.” The worst combination is flat growth, a rapidly expanding base and a high share of revenue from one major customer.

9. Google’s Moat: TPU as a Cost Weapon and Incentives That Do Not Cannibalize

TPU’s cost advantage is formidable. At the same FLOPS, TPU training costs only 20%-30% as much as the GPU infrastructure used by OpenAI, while training models 2-3x larger. Parameters increase roughly 3x with each generation, effectively comparing the next generation’s parameter scale with the current generation’s. That makes TPU exceptionally strong in benchmarks and pretraining. Coding requires RL and labeling, however, and Google underinvested last year; “coding may still need another 3-6 months to catch up.” That is why Gemini’s GCP acceleration has been less explosive than Anthropic’s.
The financial translation is favorable: almost all Gemini revenue accrues to Google, while GCP has a smaller base. Gemini contributed 4-5 points of acceleration this quarter alongside TPU rental and direct sales, and cloud operating margins rose from just above 17% to above 30%.
The incentive structure is the key distinction. When an ICE-car maker builds an EV, “it ends up fighting with its original ICE-car business,” weakening commitment. AWS wavered once at the end of 2024 and consequently did not secure enough data-center capacity in 2025. Google’s CPU gross margins were only in the low 40s, while TPU margins are higher, so the ROI is better than what the company earned in the CPU era.

10. Severe TPU Shortage: CoWoS Is Locked Up, and Internal Demand Will Clash with External Demand This Year

Asked whether Google prioritizes TPU for Search or external cloud customers, 劈柴 offered an ROIC framework without giving a split. Zhou Mo’s explanation is that TPU is currently the most supply-constrained chip, with the bottleneck being CoWoS capacity locked up by Nvidia. Google will receive fewer than 4M units this year, while new capacity allocated next year will approach 10M units.
Demand is already visible: 2 of the 3 SOTA models are trained on TPU—the previous-generation Gemini 3 and the Mistral pretraining model due out next. Only OpenAI’s 5.5 uses GPU. Google plans to allocate more than 900K cards to Anthropic this year, mostly TPU v7, while its own Gemini 3.5 and Gemini 4 training also require v7. “Internal demand and external demand will clash more and more intensely”; the problem will not be easy to solve this year and should improve somewhat next year.
Departmental incentives are not aligned. The TPU team wants to enter Frontier Lab adaptation workflows as early as possible, because doing so increases its future share: “every company has inertia; the earlier you get in, the more that inertia works in your favor.”

11. Search Advertising’s Second Curve: The 60%-70% of Queries That Are Not Monetized Is a Gold Mine

The counterintuitive starting point is that 60%-70% of Google’s search queries generate no revenue because they are too long or too short to match advertisers’ keywords. AI Overview and AI Max change that by inferring purchase intent over multiple turns. When the user is ready to choose “the exact brand they want most,” Google can serve a large banner at the final stage of the decision, where CPMs can be extremely high.
The opportunity is incremental rather than cannibalistic. On each refresh, Google can show 3 traditional ads or layer an AI ad onto different combinations. During the Los Angeles wildfires last year, the value of fire-insurance queries spiked; Google could keep 3 traditional ads and add an AI Overview ad, increasing AI penetration without damaging the traditional business.
The timing still has runway. Last year was the first AI year for advertisers in North America and Asia-Pacific, while Europe lagged by 2-3 quarters because of heavier compliance reviews. This year is Europe’s first year, so acceleration should continue.

12. Microsoft and OpenAI Unbundling Is Win-Win; Copilot’s Three Defenses Are Being Dismantled

Unbundling benefits both sides. Selling Codex only through Azure meant “losing the largest developer market,” because developers are on AWS; coding products must reach the developer market rather than only enterprise buyers. Microsoft also wants to sell the strongest model APIs. As model companies gain bargaining power, decoupling is an inevitable trend.
Copilot’s original three selling points—native Office integration, reuse of the identity and permissions stack so users avoid “10 authorization procedures,” and deeper contextual understanding—are being weakened one by one. Gemini and Claude Cowork can call Office through Computer Use; they have not integrated permissions, but their output quality may already be higher.
The technology gap is concrete. Copilot’s multimodal capability uses the prior-generation Co-Interpreter, producing unattractive PowerPoint output. When harnesses emerged in February and March, “there was simply no time to adapt before your legacy product was already expected to become fully harness-based.” Even after catching up, “if a few months later it is no longer the harness but the next-generation technology, will you have to adapt all over again?”
Microsoft is constrained by its own enterprise stack. Compliance adds an orchestration layer and a Responsible AI layer that process value-sensitive queries for compliance. “Some questions could actually receive a more creative answer, but because of compliance handling, the product ends up being dumbed down.”

13. Copilot Studio: From Selling Seats to Selling Tokens

Office has 400M users, but Copilot has only about 20M; seat-based penetration is roughly 5% and may not rise quickly next year. The biggest growth opportunity is shifting Copilot to consumption-based token pricing through Copilot Studio, a low-code internal enterprise Agent Builder. Revenue is doubling each quarter, already reaching one-fifth to one-quarter of Office Copilot revenue; some customers spend $10K-$20K per month rather than $30 per seat.
The product’s initial traction comes from Microsoft’s installed base: permissions, compliance, identity and the sales network are already connected, so it can move faster than other Agent Builders. Zhou Mo compares the path to Palantir’s AIP, which moved from customized cases toward standardization. OpenAI, Anthropic and Gemini Enterprise are now converging on the same use case: Microsoft has a first-mover advantage, but it may not be durable.

14. Amazon: From “Unable to Release Revenue” to a Narrative Reversal—and Two Layers of Lock-In

Last year’s recurring theme was that Amazon had bought power capacity but could not turn it into revenue. Amazon is a hardware-like cloud business: its data centers must deploy both GPUs and Trainium, while cooling, power and resource-allocation requirements caused repeated delays. The issue has been resolved. Amazon bought 4GW of power last year, more than Microsoft, and “there is no power shortage this year.”
The validation chain appeared this quarter. Bedrock grew 170% quarter on quarter, broadly in line with Anthropic’s 160%-170% sequential revenue growth. More importantly, Trainium utilization rose without the need for discounting, protecting margins. Revenue the market had viewed as low quality was shown to be high quality—accelerating without hurting margins—and operating margin reached a record 13.1%.
The blemish was AWS growth: the market expected 30%, while the actual figure was 28%. Zhou Mo believes Q1 was affected by the Blackwell ramp, potentially through capacity constraints and possibly through the OpenAI relationship—OpenAI had wanted to move onto AWS for some time, but its constraints with Microsoft were not resolved until April. “This problem will most likely be resolved next quarter.”
Investing in model companies creates 2 layers of lock-in. Amazon has invested more than $33B in Anthropic in exchange for a 5GW Trainium commitment, then invested $50B in OpenAI. The first layer is: “I give you money so you use my chips and help me solve these problems.” The second locks API and inference share onto Amazon’s cloud. Google’s relationship with Anthropic works the same way: TPU has already completed part of the job, with the Mistral model able to pretrain on TPU, while Trainium can still run only experiments.

15. Meta Advertising Deconstructed: 19% More Impressions and 12% Higher Pricing Are Not Contradictory

Zhou Mo decomposed Meta’s 9-point year-over-year acceleration into 4 points from foreign exchange, 2 points from the election comparison base and 3 points of real growth. Half came from recommendation algorithms and half from products. Inventory growth was driven by format: since Q3 last year, Meta has aggressively pushed Overlay Ads, renamed Partnership Ads this year, a form of sponsored native content resembling the path Douyin took in China in 2020-22—“a streamer talks for 5 minutes, then suddenly promotes Zhuanzhuan or Guazi used cars.” U.S. creators and agencies have not yet learned to integrate sponsored content, but once the ecosystem is established, Instagram inventory can scale rapidly and still command decent prices.
LATTE is a rare example of a paper tied to a real product. Most AI-company papers have little to do with actual algorithmic iteration; “people tend to publish papers that are already outdated or make no significant contribution.” Meta launched LATTE at the end of December and then published the paper. It contributed visibly in Q1, although the impact was still in the low single digits. Before foundation models, “a single major algorithm iteration producing a 2-point improvement in conversion rate would already have been an industry-shocking event.”

16. Meta’s Monetization Depth: The Economics of Replacing Labor and Expanding Categories

Advantage+ and Advantage+ Creative work together to rotate AIGC assets and personalize creative at scale; third-party AIGC materials perform poorly by comparison. The more interesting math is in Business Messenger. In a lead-generation ad’s CPA, half is advertising cost and the other half—or more—is human customer-service labor. The model previously worked mainly in Southeast Asia, where labor is cheap. If an agent cuts labor by 80%, then under the assumption that labor represents 50% of the original cost, total cost falls 40%. Most of the savings can be recycled into ad spend: instead of selling a $1,000 B2B lead, advertisers can sell a $100 product, expanding the addressable customer base to smaller merchants.
The long-term opportunity is category expansion. Travel accounts for nearly 10% of Google’s ad budget and auto for another several percentage points, but Meta has lacked the search leads to compete. Once a search chatbot is connected to Meta’s own models, users can start searching within Meta for which car to buy or where to travel. “This will not happen quickly, but it is important.” After Muse Spark and Avocado launch, Meta can also use its own models and capture more lead data.

17. Meta’s Selloff Was a Capex Mismatch; Avocado’s Self-Use Positioning Is Already Better Than Expected

The math was straightforward: revenue “just reached the high end of guidance,” whereas it typically beats by more than 1 point. Capex increased by $10B, or 7%, and Q1 capex came in below expectations, pushing the entire shortfall into the next 3 quarters. Q4 could therefore come in 10%-20% above the original assumption, leading a linear extrapolation to lower 2027 depreciation and earnings. Layoffs offset depreciation this year, but the market asked: “Can you make layoffs of the same magnitude next year?” That is why the reaction was so severe.
The internal model needs to be placed on the right coordinate system. For commercial viability, it must break into the top 3—perhaps only 2 models remain in the first tier, with Google 0.25-0.5 of a step behind but likely to catch up quickly as compute arrives. Otherwise, “you have neither the price of an open-source model nor the capability of a first-line model company.” Meta assembled its team from multiple sources, and the next generation may move beyond the Llama framework. Reaching Gemini 3 levels within months is “already very good”; the relevant test is whether it fits recommendation algorithms and accelerates Meta’s own products and advertising, which would reduce concern over capex.

18. Coding Productivity Is Improving Monthly; Layoffs Are Only Beginning

Meta is the most aggressive example: PIP evaluations include code volume, and employees are incentivized to use coding agents. The bar rises every month. At the end of last year, having AI write 50% of incremental code was considered excellent; by the end of January, all code was expected to be AI-written; after Claude 4.6 became widespread in February, engineers were expected to orchestrate 5-10 coding agents; by March, remote coding and issuing commands from a phone had arrived. “You have to leave a task running before bed, otherwise those 8 hours of sleep are wasted.” Aggressive users moved from 5 agents to 100 and then 2,000; “one product manager at our company keeps telling me he has already created 14,000, so I still have not figured out how many are actually usable.”
The macro data are already showing the effects. Since last July, U.S. employment data have missed expectations in most months—“because it has been very visible since last July.” Engineers’ effective working time has fallen from 5-6 hours a day to 1-2 hours. Companies can either accelerate product iteration, as Meta has with sponsored content, DSP and Audience Network development, or continue cutting headcount. Engineer performance is increasingly judged by the multiple on coding API consumption. “It is actually a new kind of occupation,” and the emergence of a new occupation always brings a transition between the old and new.

19. The Endgame Math: 7x OPEX Replacement Potential and Convergence Between Capex and Revenue

Cao Qingyun challenged the AI monetization story as mostly efficiency gains without entirely new markets. Zhou Mo acknowledged that “we are indeed not seeing that many use-case demands right now,” but pointed to OPEX: enterprise IT spending is only 3%-4% of revenue, while OPEX is 30%, “a 7x concept.” Coding has achieved “90% of AGI’s capabilities,” and other domains may arrive in 2028-29, so substitution should accelerate from here. Token taxes, AI layoff taxes and transfer payments will have to be addressed; only afterward can remaining time and disposable income generate a larger consumer market. “It may have to take 2 turns.”
Cao Qingyun’s framing is $100B-$150B of AI revenue against $630B of capex, or roughly 4-6x. Zhou Mo’s math is more favorable: this year’s cloud capex is about $700B, but the GPU- and AI-related portion can receive a 60%-70% discount under his methodology. By year-end, Anthropic could reach $100B ARR, OpenAI $60B-$70B and Gemini $30B-$40B, for roughly $200B of combined model-company ARR—about two-thirds of the relevant capex. Capex is an investment pulled forward by 4 years; if model-company ARR reaches $400B next year, the ratio rises to three-quarters. “The numbers look a little more comfortable,” as the market begins to price convergence between incremental costs and incremental revenue.

Market Overview | April 21, 2026 Waymo's Dmitri Dolgov: 20 Million Rides and the Road to Full Autonomy