Topics Pioneers Insight Method Author
Back

Zhou Mo on Big Tech Q3 Earnings: AI Investment Enters Validation

·硅谷坐标 Silicon Valley Vector

Rendering the long image…

Topics Trillion-Dollar AI Capex and Bubble Risks AI Venture Funding and Bubble Risks Autonomous Agents Reshaping Enterprise SaaS Personal AI Assistants Adoption and Monetization Open Source vs Closed Source AI GPU Depreciation and Cloud Market Shakeup AI Data Center Power Bottlenecks
Pioneers 周默

Summary

  • This selloff shifted questions about AI growth from supply to demand and pricing. Zhou Mo said Anthropic’s ARR was about $70B at the end of September, below expectations; the host cited a Wall Street Journal report putting it at about $80B at the end of October. The market is now checking growth month by month. The host said paid inference uses about 30% of compute, with training and research taking the other 70%. Zhou Mo noted that inference compute costs actually fell as a share in Q3. “You can no longer say inference compute is what’s limiting the company’s growth” was the episode’s central valuation challenge.
  • Anthropic is hard to value by simply applying software-company sales multiples. Its roughly $10B in RPO is only a little over 10% of current ARR of about $80B, leaving less revenue visibility than Snowflake or cloud providers, while the company has already committed to about $750B in long-term compute. Zhou Mo’s view: “Grow fast, and margins improve; grow slowly, and margins deteriorate.” Whether 20x sales is appropriate still depends on growth coming through, not just an AI-native-company premium.
  • The three clouds are still accelerating, but that has not reversed the widening absolute gap between AI revenue and capex. Azure, AWS and GCP grew 48%, 39% and 103%, respectively, but higher supply-chain costs and training investment make returns harder to explain; the gap is “still widening.” The market has started discussing flat capex in 2028, or even lower capex in 2H 2027 than in 1H. Zhou Mo stressed, however, that ODMs and OEMs have not yet seen meaningful cuts to complete-server orders; component-level rumors need to be checked against consistent definitions.
  • Microsoft’s enterprise deployment advantage is clearer, while Google’s advertising defenses are starting to come under pressure. Copilot uses existing identity, security and permissions systems to shorten deployment cycles by about three to four weeks, finding “a clever foothold.” Google Search grew only 12%, facing both ad budgets shifting to OpenAI and the challenge of monetizing AI Mode; 103% growth at GCP cannot substitute for assessing the ad business.
  • Meta’s ad growth has come through, but the case for a billion DAUs on Muse remains constrained by cost and retention. Personal Agent inference costs about $30–$60 per active user per month, while Zhou Mo thinks a reasonable current assumption for incremental value may be only $15–$20. He estimates Muse’s 14-day retention at about 15%–16%; it should clear 20% before Meta considers a broad rollout. With capex already at 54% of revenue, incremental purchase intent in advertising is more worth validating than subscription revenue.
  • Open-source technical progress and commercial share are not on the same curve. Zhou Mo thinks the open/closed-source gap may have widened again from three months in midyear to six to nine months, but enterprise expectations for the share of budget routed to open-source and small models have risen from 5%–10% to 20%–25%. Closed-source small and mid-sized models are also cutting prices aggressively, putting pressure on both business models. “Neither side can say its position has clearly improved.”
  • The financing chain bears closer watching at New Cloud and its New Lab customers. Zhou Mo’s rough estimate is that New Labs account for about 20% of incremental New Cloud compute. These customers have little revenue and high valuations, making them more dependent on financing conditions and potentially “the most sensitive link, the leading indicator.” A falling share price may not directly change debt metrics, but GPU prices, customer demand and capital shifting toward the big clouds can materially affect debt-service capacity and access to financing.
  • Compute expansion in 2028 may first run into permitting constraints, increasing the relative advantage of incumbent big clouds. Zhou Mo distinguishes 2027, when approvals are unaffected by this round of state-government changes, from 2028, where the dispute is concentrated. The market worries that planned U.S. capacity could fall from the low-60s GW to the low-40s GW, but he thinks local economic interests mean “it’s not completely off the table for discussion.” Big clouds have existing capacity and overseas footprints; even if space-based compute contributes about 1 GW in 2028, it cannot fill a gap measured in 10 GW.

Deep dive

1. Frontier models are moving from a two-horse race to “4+1”

  • Zhou Mo attributes the shift to RL training increasingly looking like “an assembly-line approach to training”: find data around benchmarks in each domain, optimize the metrics, then combine the results. After its reorganization, Google has focused more on this pipeline and, since Q1 this year, put more emphasis on reinforcement learning; previously, more effort went to scientific research and Nobel-caliber breakthroughs.
  • Google is not catching up by throwing more compute at the problem. Zhou Mo’s comparison: some of the other players may have about 2 GW of training compute, while Google has half that, and much of its TPU capacity also serves external customers, including Apple. But if its in-house training process can already reach a 4.6 or 4.7 level, the path to 5 is relatively clear; it is mainly a question of time and compute.
  • Meta’s advantages are strong compute and well-prepared reinforcement-learning data. SpaceX has even more compute than Google, along with high-quality data sources and assets accumulated through earlier acquisitions. That puts Google, Meta and compute-rich SpaceX in the competitive set alongside OpenAI and Anthropic.

2. The open/closed-source capability gap is widening, even as commercial competition intensifies

  • In July and August, the gap between open- and closed-source models had narrowed from an expected six to nine months to three months. Zhou Mo had said the gap would widen again; he now thinks “that call has basically been borne out over the past quarter”: closed-source iteration has accelerated, while recent open-source releases have been less impressive than those in June and July.
  • In his account, the route by which models learned closed-source capabilities through Signature API and then shifted to smaller models was shut off starting in September. Once closed-source models began using Loop Transformer, imitators also had to infer the recurrent path, potentially taking more steps and using more tokens. Similar capability does not necessarily mean similar inference efficiency.
  • A survey of close to 100 global companies shows Router changing budget allocation: difficult tasks go to closed-source models, while simple ones go to open-source or smaller models. Last quarter, companies typically considered assigning only 5%–10% of their token budget this way; many now expect 20%–25%. Large-scale Router adoption has been underway for only one or two quarters.
  • Closed-source small and mid-sized models are also cutting prices aggressively, bringing their value for money closer to open-source models. Cost-sensitive startup categories such as Personal Agent are beginning to use open-source models; Manus is one example Zhou Mo cited. Some models in areas such as security and biology may be offered only to large industry customers, with KYC reducing the risk of their capabilities being learned. Competition is pressuring both sides, and the gap between public models and enterprise-only models may widen further.

3. Slowing ARR growth undermines the assumption that compute alone drives growth

  • During the summer selloff, model revenue was still opaque, so the market could assume July and August ARR was strong, even discussing a figure near $80B in August. Anthropic’s ARR was about $70B at the end of September; the host cited an end-October figure of about $80B. Zhou Mo thinks Q3 growth was slow and that “more of the concern is coming from the demand side.”
  • Cao Qingyun noted that about 30% of Anthropic’s compute goes to paid inference and 70% to training and next-generation research. Zhou Mo focused on the direction: inference compute costs had risen as a share over the preceding quarters, then fell for the first time in Q3. If inference compute were the main constraint, more compute should have been allocated to inference.
  • His argument is conditional: even assuming ARR doubles to $150B or $160B by the end of next year, a more aggressive compute commitment could still leave inference costs lower as a share of total compute costs. That weakens the logic of forecasting next year’s ARR by adding up new GW and estimating revenue per GW.
  • Zhou Mo also does not simply blame the slowdown on open-source models: the open/closed-source gap has widened over the past three months. More direct short-term variables are OpenAI price cuts and free credits, as well as Opus 5.5 being 20% cheaper than Opus 5. Efficiency gains can explain better value for money, but changes to the pricing structure still hit ARR; the two companies’ growth has begun to show a “seesaw effect.”

4. Anthropic has no public-company valuation template that fits cleanly

  • In discussing valuation, Zhou Mo starts with growth visibility, the right business-model benchmarks and confidence in margin expansion. It was convenient to compare Anthropic with Palantir at about 20x sales or Snowflake at about 100x sales at its IPO, but after the S-1 disclosure, “Anthropic’s financial model doesn’t really look like those SaaS companies.”
  • Anthropic’s roughly $10B in RPO is only a little over 10% of its roughly $80B ARR. By comparison, Zhou Mo cited Snowflake’s next-12-month RPO at about 50% of revenue. Cloud providers also have backlogs from customers signing contracts of three years or longer; Anthropic has less support from that kind of backlog.
  • He acknowledged that this also reflects how Anthropic sells: the company still mainly sells API access and has not pushed multi-year contracts hard. If it changes its sales approach, the RPO ratio could rise. For now, investors can only track revenue month by month, making it hard to judge whether 20x sales is appropriate. The long-term narrative creates a valuation trade-off, but does not resolve the lack of visibility into next year’s revenue.

5. Long-term compute commitments turn margin changes into a growth problem

  • Anthropic has about $750B in long-term compute commitments. Given the company’s belief in AGI and some current research pointing to AGI emerging in 2028, Zhou Mo thinks those commitments may be hard to change. Revenue growth therefore directly affects profit: “Grow fast, and margins improve; grow slowly, and margins deteriorate.”
  • Gross margin was relatively stable in Q2 and Q3, but training costs rose as a share of revenue. Zhou Mo therefore thinks margins next year may be somewhat worse than this year: compute commitments are set first, and if revenue does not accelerate in step, the training burden grows.
  • Cao Qingyun asked whether new models quickly making old models obsolete means every generation must stay ahead. Zhou Mo replied that about 80% of revenue comes from new models, and customers are mainly technology companies and startups with relatively low switching costs, making it hard to stop investing at the margin. Faster training and iteration could help, but he is unsure whether the change will be clear by next year; he also mentioned the METR report from the past week.

6. The two model companies’ ARR captures revenue, but not total compute demand

  • Zhou Mo thinks OpenAI’s and Anthropic’s ARR remains a good measure of overall AI revenue and is not yet materially understating it. Open-source revenue is spread across New Cloud, GPU rentals from CSPs, and model deployment and inference services. Closed- versus open-source revenue may currently be split more sharply than 80/20, closer to 90/10.
  • Revenue share cannot be mapped directly to compute share. Zhou Mo noted Sierra’s gross margin is above 80%, close to 85%, while K models have a much wider gross-margin range of about 20%–70%. Even if revenue eventually splits 80/20, compute costs could be 1:1 or 3:2 because open-source models have somewhat lower gross margins; their contribution to hardware demand would exceed their revenue share.
  • Finance, biology and robotics create another mismatch. Quant funds are starting to lock in data-center compute, memory and supply chains themselves, with ROI easier to calculate. Robotics data-purchasing budgets accelerated sharply in the latest quarter, and companies are more actively locking in New Cloud compute. But these investments may not convert into ARR until next year, and robotics revenue may not show up even then.

7. Personal Agent’s $100B opportunity first has to clear unit economics

  • Asked about $100B-scale opportunities after coding, Zhou Mo said the marginal change that excites him most is Personal Agent. Silicon Valley has not had an exciting consumer application in a long time, and Instinct and Meta Muse have reopened the question: at 1 billion DAU, the user base would be about 30 times the 30–40 million engineers.
  • But inference currently costs about $30–$60 per DAU per month, naturally limiting the scale of rollout. Using Meta’s more than 200 million MAU in North America and about $50 in monthly ad value per user as a rough calculation, he said each source of incremental Personal Agent value—ad leads, subscriptions and revenue—has to be validated.
  • At a $20 subscription price and a 20%–25% take-up rate, subscription revenue works out to only about $5 per user across the full user base. If advertising also contributes, a 30% increase might be worth the teens to $20. Zhou Mo thinks $15–$20 may be a reasonable current assumption for incremental user value; costs still need to fall substantially, and quality needs to improve.
  • Commercial timing, technical progress and cost constrain one another. The five-year opportunity is exciting, but 2027 revenue alone could disappoint. The AI drug-development opportunity Anthropic often discusses still faces clinical-stage bottlenecks; real-world adoption may not come until around Phase II. Robotics also has a long timeline. There is no new use case in sight that can build confidence in $100B of revenue within a year.

8. Faster cloud revenue has not made capex returns easier to underwrite

  • Cao Qingyun listed growth at the three clouds as Azure 48%, AWS 39% and GCP 103%. Zhou Mo cautioned that the absolute gap between AI revenue and capex, which had narrowed for a time early in the year, widened again as prices rose across the supply chain. The rising share of training costs at labs is a clear sign that the gap between ARR and capital spending is widening.
  • His view is that getting closer to AGI may require putting more compute into training, making short-term ROI harder to explain. If AGI cards are actually sold, the assumptions may all have to be revisited, but the path toward AGI is not necessarily getting easier to model. On this view, returns were harder to explain in Q3 than in Q2, and the trend could continue next year.
  • Demand needs to be split into three layers: AI startups are still accelerating commercialization; usage growth at technology companies is slowing, with Meta and Apple starting to limit employee budgets; and enterprises sharply increased AI adoption in the first half of this year and now need time to deploy and validate ROI. Startups are performing best, but they are also the most dependent on financing: “They burn whatever they raise,” which then flows into tokens.
  • The market has started discussing flat capex in 2028, or even lower capex in 2H 2027 than in 1H. But Zhou Mo has not yet seen meaningful cuts to complete-server orders at ODMs or OEMs. Noise around CPUs and optical modules could reflect order timing or technology changes; gaps across sources with different definitions cannot be read directly as order cuts.

9. Model companies are gaining bargaining power, while cloud providers are regaining the enterprise entry point

  • Anthropic’s Q3 margin was better than expected. One reason Zhou Mo cited was that it renegotiated rental pricing with AWS and GCP, particularly for TPU and training compute. As a single model customer has a growing impact on cloud growth, its bargaining power rises: selling 100,000 more or fewer TPUs could change GCP’s growth rate by double-digit percentage points.
  • The customer entry point is shifting. Traditional enterprises care more about security, SLAs and commitments, and want to buy APIs from CSPs. Most of Anthropic’s incremental ARR this quarter came from Amazon, well above the 48% share of company revenue that CSPs represent in its existing base. Technology customers prefer buying models directly; traditional enterprises prefer going through the cloud. The entry point and the profit pool may not concentrate in the same place.

10. Copilot is starting to get enterprise deployments through the trial phase faster

  • The trend of Microsoft’s large customers moving from hundreds or thousands of seats to tens of thousands is continuing. Zhou Mo distinguishes buying 5%–10% of seats, which still looks like a trial, from coverage above 50%, which signals large-scale deployment and treating the product as essential. More customers are crossing that line.
  • Copilot is easier to align with existing identity, security and permissions systems, shortening deployment by three to four weeks compared with enterprise Agent products from model companies. External products also need Microsoft’s help to integrate with Office, with technical work, staffing capacity and speed of cooperation all playing a role. Zhou Mo did not claim to know which factor was decisive, but Copilot deployments are faster. As GPT-5.6 improved overall, the experience gap between Copilot and model-company products also narrowed.
  • No longer reporting seat counts does not by itself mean growth has stalled: moving from “over 20 million” to “over 30 million” does not mean a net addition of 10 million; the actual change could have been from 7 million to 8 million. Another 10 million would then imply sequential growth of about one-third. Some usage was sold in June and July, with more of the revenue impact landing in Q3. Microsoft therefore did not need to repeat the same seat additions this quarter to accelerate revenue growth.

11. Azure’s results were clean, but OpenAI is becoming a bigger single variable

  • Microsoft reported revenue of $92.3B, up 19%, and Azure growth of 48%, with guidance of 49%–50% for next quarter. Free cash flow was $18.6B, the only positive figure among the four companies. Zhou Mo called the results “very clean, very good.” Azure is the main near-term focus, while long-term valuation increasingly depends on Copilot.
  • Commercial RPO was $677B, up 73%, but the portion recognizable within one year grew only 29%. Zhou Mo estimates that OpenAI accounted for about 40% of Azure’s growth this quarter. For next quarter’s roughly 50% growth, he splits it into about 25 percentage points from OpenAI and about 25 from non-OpenAI sources; concentration is set to rise further.
  • OpenAI signed long-term contracts, with stronger commitments than those of many short-term rental customers, and Zhou Mo does not think contract quality is poor. The risk is a change in future bargaining power: will Microsoft also have to renegotiate prices, as AWS and GCP have? Another point of debate is that Microsoft had previously guided to positive cash flow in FY27, yet still needed financing this quarter, exceeding some investors’ expectations.

12. Google’s main issue is back to advertising, not whether cloud growth is high enough

  • Google reported revenue of $127.6B, up 25%, but Search grew only 12%, the slowest ad growth in five quarters. Free cash flow was negative $9.1B; the company stopped buybacks, issued about $41B in equity and debt, and shares fell 12.6% in a day after earnings. Zhou Mo’s pre-quarter report had already expressed concern with the line, “This is the first quarter affected by ChatGPT.”
  • OpenAI’s advertising ARR grew from $100M to $1B, becoming large enough to affect ad budgets. Advertisers and agencies Zhou Mo surveyed were shifting Google budgets to OpenAI, but not moving Meta budgets. As a new channel, OpenAI ads’ ROI is not yet a major focus for advertisers, but long-term budget allocation has started to change.
  • Google itself is shifting from AI Overview to AI Mode. The former may increase ad load; the latter has yet to prove it can monetize as effectively, so it needs better revenue per search and CPMs to make up the difference. The easy-to-monetize piece comes first; the harder monetization shift comes after. External ChatGPT and internal AI Mode are both testing how traditional search adapts to new search behavior.
  • Meta is also gaining share in industries where it was previously weaker, such as travel and finance. Its North American ad growth increased by about 10 percentage points. Even if midterm elections explain four or five points, five or six remain; Zhou Mo therefore does not see this as simply an improvement in the overall ad market. The “gray rhino” of AI’s impact on advertising has re-entered the narrative for Google.

13. GCP’s high growth increasingly depends on TPU sales and revenue recognition

  • Zhou Mo estimates that 100,000 TPUs affect GCP growth by about 12 percentage points: selling 200,000 versus 300,000 matters far more than a two- or three-point swing in traditional cloud business. Rentals spread revenue over about five years, while sales can be recognized in the quarter. That means 103% growth could still meet or slightly miss expectations for investors who had looked for 108%–110%.
  • Google’s compute shortage also has a supply-side cause. Based on the production-plan changes he cited, supplier yield problems could reduce next year’s TPU output from 10 million units to 8.5 million, while demand from external customers is higher than it was six months ago. Gemini’s training resources are therefore being squeezed; Google can no longer rely on “throwing everything at the problem.”
  • Renting compute from SpaceX at a high price adds capacity for Gemini and products such as Gemini API, Gemini Enterprise and Vertex. Zhou Mo thinks Gemini’s Q3 experience did not improve meaningfully relative to OpenAI and Anthropic, nor did it make a clear contribution to GCP. Improvements in the latest models may not show up until Q4.

14. AWS used its medium-term outlook to offset a slight revenue disappointment this quarter

  • Amazon reported revenue of $201.9B, up 12%, and AWS growth of 39%, below market expectations of about 42%–43%. But the company guided to roughly 40% growth over the next two years, improving the outlook for 2028; the stock’s post-earnings reaction was the best of the three clouds. The host also noted that backlog covers about three years, free cash flow was negative over the past year, and debt more than doubled.
  • This quarter included a gap as Trainium 2 transitioned to Trainium 3, as well as Anthropic pushing prices down again. Amazon plans to sell Trainium externally starting in late 2027, but for now it is used mainly by Anthropic and called in the background by products such as Bedrock; it does not yet have many external customers like TPU does.
  • Chip sales may not improve AWS margins. GCP’s operating margin is only in the low 30s, so selling chips at close to 50% margins would help. AWS already has gross margin in the 60s and higher operating margins; it is unclear whether Trainium can achieve margins above 40%. Zhou Mo leans toward a limited margin impact. Not needing to raise financing this quarter is a relative positive, but “not raising now doesn’t mean it won’t raise later.”

15. Meta’s ads had no obvious blemishes; financing changed the investment case

  • Meta reported revenue of $64.6B, up 26%, and capex of $34.9B, exceeding operating cash flow for the first time and reaching 54% of revenue. Shares fell 7% the next day. Zhou Mo’s view of operating performance was direct: Meta’s ad results and guidance were both good, with little to criticize.
  • Channel feedback points to a sharp rise in soft-ad CPMs, as algorithm improvements make more inventory monetizable. Audience Network has also shifted from being relatively overlooked to receiving active investment. Gains in weaker sectors such as travel and finance contributed too, but Zhou Mo cautioned that, at least this quarter, strong ads “don’t necessarily have anything to do with AI.”
  • The real debate is financing above expectations. Zhou Mo expressed concern with the shift from the prior estimate that the company would raise “2 to 3” a year to “4 to 5” now. If Meta has to raise $40B–$50B every year, investors in a large company will worry about its long-term financing needs; the issue is no longer just the absence of buybacks.

16. The easiest Muse value to realize may be purchase intent in advertising

  • Zhou Mo has more confidence in commercialization that adds value within the existing model. Travel accounts for more than 10% of Google’s ad budget, but perhaps only a few percentage points, or less than 5%, at Meta. Applying that difference to Meta’s nearly $10B in ad revenue could mean billions of dollars. “How long would it take to make billions from subscription fees?” That is the arithmetic he uses to illustrate the value of incremental advertising.
  • His example is parents taking their child to Mexico for Thanksgiving: they first search Google for attractions and restaurants, then check flights, Booking, Agoda or hotels. Google knows the destination, as well as the user’s feedback and travel dates. Meta usually has to wait for Booking to get a lead, share it through CRM, match the user and serve an ad; that process is slower than Google’s.
  • If Muse can plan a morning, afternoon and evening itinerary through multiple rounds of conversation, it may understand the user’s travel needs and plans better, narrowing that ad-lead disadvantage. Subscriptions can help too, because costs are hard to measure if the product is free. After two years of user education, people are more willing to pay $30 a month for a product. But Zhou Mo still thinks it would be difficult to make subscriptions alone into tens of billions of dollars in revenue.

17. Muse’s scale inflection depends on retention, not Meta’s existing user base

  • Meta’s advantages are its large user base, big budget, and the feedback, research and operating capabilities that come with its social network. Zhou Mo thinks Muse is somewhat like Manus and less like GPT: these products move quickly in response, operations and iteration. He sees no major technical difference; the competition comes down to who can spot a new need this week and respond with a better product immediately.
  • Cao Qingyun noted that Muse passed 10 million users six weeks after launch. Zhou Mo estimates DAU was about 2.5 million–2.6 million before the earnings call and is now close to 4 million. He stresses that this is not a product where Meta can simply decide to promote it and quickly reach 10 million DAU: countries are opening one at a time, promotion is lighter than it was for Threads, and users still need to learn how to use it.
  • Agents differ from traditional tools: the core feature in the first week may not be the most important feature in the future, so current retention cannot be treated as long-term retention. He cites OpenAI: 14-day retention was about 10% at launch and took nearly two years to approach 20%. Now that it is at 25%, it can be treated as long-term retention.
  • He estimates Muse’s current 14-day retention at about 15%–16% and thinks a broad rollout makes more sense above 20%; otherwise, “you’re wasting potential users.” Moving from 10% to 15%–16% may be easy, but retention could stall at 17%–18%. It is unclear when the threshold will be crossed, and one cannot infer how many users Muse should have from a 2 billion DAU assumption.

18. Financing pressure is concentrated in New Cloud, with New Labs the most sensitive customers

  • Zhou Mo used CoreWeave as an example: it has raised more than $140B and has $130B on its balance sheet; it has already spent about $10B and still has positive cash flow. He described CoreWeave as essentially a zero-capex company for now, though buying AMD in the future could involve capex. Data centers are built mainly by CSPs and New Cloud, so financing and cash-flow pressure flows through to those cloud companies and New Cloud.
  • Against a backdrop in which the 10-year Treasury yield briefly broke 5%, Zhou Mo stressed that lenders look at metrics such as net debt to EBITDA; the share price does not necessarily change those metrics directly. Lower GPU prices or customer demand can change collateral values, EBITDA or financing pressure. Rates, financing costs and maturities also affect the ability to raise money.
  • Share prices still affect relative access to financing: for the same amount of capital, investors may prefer a CSP or SpaceX over a weaker New Cloud company. When Nvidia backs New Cloud, that may be seen as positive when shares rise and interpreted as a problem when they fall. Reflexivity matters, but a falling share price alone does not prove financing will dry up.
  • He is more focused on New Labs. Some companies are valued at tens of billions of dollars despite little revenue, inviting discussion of 200x or 2,000x sales rather than Anthropic’s 20x. Zhou Mo’s rough estimate is that New Labs account for about 20% of incremental New Cloud compute. These companies may not be more efficient than frontier-model companies, yet depend more on financing; they could be the “storm point” for the whole chain.

19. The durability of upstream profits ultimately depends on the capex cycle

  • Cao Qingyun contrasted TSMC raising its full-year growth outlook for the third time this year, to 44%; ASML customers booking deliveries for 2029–2030; memory margins around 80% and TSMC near 60%; and cloud providers’ cash flow turning negative. Zhou Mo’s answer: “Basically, it comes down to whether capex is in an upcycle or a downcycle.”
  • If capex is flat, sectors relying on price increases will find further hikes harder, while those relying on volume or share gains may fare relatively better; bottlenecks will also shift. The direction of the broad cycle is often shared, but each company’s dependence on price, volume or other constraints determines relative performance. Upstream margins cannot all be extrapolated in the same way.

20. Data-center permitting disputes mainly change the 2028 supply outlook

  • In response to Cao Qingyun’s point about the midterms and new governors proposing to delay or pause construction, Zhou Mo stressed that 2027 approvals will still go through; the impact is mainly in 2028. Planned and permitted U.S. capacity is a little over 60 GW, including about 18 GW in Texas. If all Texas projects pass, other states account for about 42 GW. The market worries that the planned total of more than 60 GW could fall to the low-40s, with at least 15 GW affected.
  • That arithmetic strengthens the flat-capex narrative: if the market originally expected about 40 GW next year and another 40 GW the year after, supply growth could be flat even with strong demand. But Zhou Mo thinks it is wrong to apply the most pessimistic estimate across the board. The governor later said the issue was not opposition to all data centers and that specific conditions could be negotiated.
  • The key is county-level economic incentives and negotiations at each level; some Republican counties also want the revenue that construction would bring. Zhou Mo expects the actual impact to be less than 15–18 GW. An 18 GW impact would be very large; the outcome would be different if it ends up at only 5–10 GW. The range remains highly uncertain.

21. Existing capacity and overseas footprints make big clouds better equipped to handle slower construction

  • The first step in the analysis is that existing compute is concentrated at the three clouds, while new Texas projects are more concentrated at New Cloud and SpaceX. If new construction is restricted, the big clouds would be less affected and could gain pricing power. Their overseas presence in Malaysia, Southeast Asia and Northern Europe also gives them more flexibility than New Cloud companies that rely mainly on U.S. sites.
  • Space-based compute can help, but its contribution should not be overstated. Zhou Mo’s supply-chain estimate is about 1 GW of space-based compute in 2028; that is not enough to fill a 10 GW gap and cannot be treated as an equivalent replacement for terrestrial permitting capacity.
  • Behind-the-meter power generation could become a condition for approval: provide subsidies to counties, reduce electricity bills and show that new load will not raise local power prices. Channel feedback is that 100% self-generation is usually a positive signal; communities may not care about the specific generation technology, with many counties focused mainly on the outcome.

22. Enterprise AI scale-up is increasingly constrained by delivery and organizational decisions

  • Zhou Mo had expected security and permissions concerns to make enterprises more comfortable with AI features from traditional software companies, but many software earnings reports this quarter still disappointed. Deployments require substantial FDE capacity, and services take longer than expected, limiting near-term revenue. Software companies with FDE capacity will have an advantage because enterprises often need hands-on help designing and implementing deployments.
  • Enterprises are also less certain about the right direction than expected. Buying a $20 monthly subscription is easy; following peers is enough. A $200 product also requires workflow changes, so CIOs and CTOs have to consider ROI and take on organizational risk. Zhou Mo used McKinsey as an analogy: companies do not expect 10 consultants to automatically raise a 10,000-person company’s gross margin by one point, and they also want consultants to take on some of the risk.
  • FDEs, consulting and services are therefore becoming clear bottlenecks, and deployment timelines cannot be extrapolated from the first half. Enterprises may have spent one or two years just in the Copilot trial phase. It is unrealistic to expect them to deploy larger budgets and restructure workflows within a year; it could take a year or longer.

23. Security and consumer ad software offer more legible sources of incremental AI growth

  • Zhou Mo thinks frequent security incidents are amplifying the problem of hackers and vulnerabilities. Security is one of the few software categories where, as with coding, demand can be explained with a quantitative measure of the opportunity. His example: if the number of hackers is 10 times what it was before, it is easier to explain the market opportunity. Many other software categories lack such a clear metric.
  • The beneficiaries still need to be assessed company by company and in the context of each company’s own cycle. He mentioned Cloudflare and the possibility of a future opportunity similar to CrowdStrike. The trend is clear for good companies, but the entire security sector should not be treated as benefiting in lockstep.
  • Consumer software and consumer internet are also receiving more attention. Zhou Mo mentioned AppLovin: the market had previously worried about Unity taking share, but this quarter he heard bullish views for the first time. Improving competition, a valuation that had been under pressure, the search for net new AI growth and concerns that AI could displace software have together brought ad-related companies back into view.

24. If short-form drama creates a closed ecosystem, it could shift from ad growth to a fight for share

  • Discussing Douyin, content-production costs and film-industry jobs, Zhou Mo cautioned that short-form drama in the U.S. has so far acted more like a tailwind for the internet ecosystem. Dozens of standalone short-drama apps have low retention and need to keep buying users, acquiring customers and monetizing them. They continuously add to ad-platform budgets; it is a business that “needs users to come in, and also needs them to leave.”
  • If a product eventually achieves high retention and time spent, creating an ecosystem where “users come in but don’t leave,” the logic changes: it starts competing directly for users’ time and ad budgets. Money spent on old platforms this year may not be there next year. Zhou Mo thinks this could affect the internet and consumer ecosystem next year; the direction is hard to avoid, while the degree of impact remains to be measured.