Anthropic’s IPO Filing and Big AI Budgets with Fusion Fund’s Zhang Lu
Topics
Battle for Frontier AI Moats
AI Venture Funding and Bubble Risks
Enterprise AI ROI and Cost Optimization
Enterprise Multi-Agent Orchestration Architecture
Open Source vs Closed Source AI
GPU Depreciation and Cloud Market Shakeup
Startup Playbook for Vertical AI Applications
Pioneers
张璐
Summary
- Anthropic’s enterprise spending profile suggests its products are becoming foundational infrastructure in corporate workflows. Zhang Lu recalled that roughly 11,000–12,000 customers spend more than $100,000 a year, including about 2,000 spending more than $1 million; Cheng Manqi added that the S-1 discloses net revenue retention of 900%. Zhang Lu is focused on whether these customers keep expanding their spend: “Model performance” gets them to try a product, but embedding it in existing workflows determines whether they stay.
- An IPO could reset valuations for AI assets, but a valuation pullback and growth in underlying demand are separate questions. Zhang Lu leans toward a discussion range around $1.5T, though the final price will depend on the roadshow and subscriptions; Cheng Manqi is watching whether October’s run rate supports the roughly $100B year-end expectation. More than the IPO price, Zhang Lu is cautious about the relationship between revenue growth and capital investment, particularly operating margins, cash flow, training spend and long-term compute contracts. She does not think an IPO itself will reverse the AI trend: “Overvalued prices and real AI demand are two different things.”
- Large companies are setting aside enormous AI budgets, but budget size is not proof that returns have arrived. Zhang Lu heard from the chief AI officer of a well-known US bank that its budget for the year exceeded $10B; in her survey of the world’s 20 largest pharmaceutical companies, 60%-70% ranked AI as their top strategic priority. Asked by Cheng Manqi about ROI, she answered plainly: “We’re still in the trial phase.” First embedding AI in existing workflows, then optimizing the business, could take 3–5 years just for the first step.
- Post-coding growth may be spread across enterprise agent platforms and a wide range of vertical use cases. Zhang Lu is unsure there will be another standalone product with growth as fast and powerful as coding, and sees more potential in Cowork connecting software through MCP to become an enterprise workflow entry point. Meanwhile, requirements around local data, regulation and customization in sectors such as healthcare and finance leave room for startups; it remains an open question whether model giants can—or should—go deep in every niche.
- Open source is reshaping enterprise buying mixes, forcing closed-source vendors to address cost pressure. Zhang Lu cited AT&T’s public statement that 60%-70% of its internal AI budget would go to open source; she also recounted an anonymous case of a company’s Claude bill approaching $500M in a single month, underscoring that such spending is unsustainable. But she does not see open source’s rise as necessarily hurting closed-source providers: AI is moving beyond tech into more industries, and “the overall pie is still getting bigger.”
- Compute demand can rise even as rents for older GPUs and their asset values fall. Zhang Lu described a shift from roughly 80% of compute going to training in the past to perhaps 80% going to inference in the future, redirecting infrastructure priorities toward cost per task, local deployment and distributed computing. H100 availability does not mean customers who need B300s, H200s or a specific cluster architecture can use them; idle capacity and shortages can coexist, and used-GPU resellers may not benefit from demand growth.
- The startup opportunity is shifting toward sector-specific barriers, data trust and governance, while consumer AI remains dominated by distribution advantages. Zhang Lu thinks model companies or tech giants could ultimately capture 80%-90% of the US consumer AI market, but stresses that this does not mean startups cannot build strong products. Highly regulated companies may share data with startups they believe they can keep accountable, while worrying that a model provider could become a competitor. Physical AI also makes governance a question that must be answered before large-scale deployment: industry-leading CTOs are urgently asking whether technical AI governance solutions need to come first.
Deep dive
1. Enterprise customer depth says more about Anthropic’s position than revenue growth alone
- The customer spending mix is the metric Zhang Lu finds most compelling. By her recollection, roughly 11,000–12,000 customers spend more than $100,000 a year, including about 2,000 spending more than $1 million. That is a change from the earlier, more concentrated customer base and suggests the product is embedded inside companies, “a bit like infrastructure within the enterprise.”
- Her logic: model performance determines whether a company tries a product; long-term retention depends on whether it fits into existing workflows. The deeper the integration, the greater the potential for upselling and broader adoption. Cheng Manqi added that the S-1 reports net revenue retention of 900%, reflecting ongoing customer expansion.
- What makes Zhang Lu more cautious is the relationship between revenue growth and capital investment: abundant cash and financing capacity also come with a strong capacity to burn. She will keep watching operating margins, cash flow, training spend and future compute contracts. Cheng Manqi also noted that after the recording, the company updated its official third-quarter revenue and total long-term compute contract figures; the discussion should be understood on the basis of what was known at the time of recording.
2. The IPO will test the price, but need not reverse the AI trend
- Asked whether an IPO could mark a turn from boom to bust for AI, Zhang Lu said she does not think the listing itself will cause a reversal. She acknowledged that there is “obviously some bubble” in the market, but distinguishes valuations inflated by hot money from industry demand: “Overvalued prices and real AI demand are two different things”(价格虚高和 AI 的真实需求增长是两回事).
- She sees the market at the boundary between 2 phases. For the past 5 years, the priority was deploying resources and compute to pursue the strongest models; next, the focus shifts to embedding AI in industry at low cost, broad scale and high efficiency. “Model capability is no longer the most important factor”; cost and integration are moving up the priority list.
- As the market discusses valuations of $2T and $1.5T, Zhang Lu leans toward a discussion range around $1.5T, which she thinks could lead to healthier market feedback after the listing, though the result still depends on the roadshow and subscriptions. If the stock performs poorly, other model companies, infrastructure businesses and highly valued unicorns could also be reassessed; she thinks a rational pullback would actually benefit the startup ecosystem.
- Cheng Manqi offered another read-through from public markets: some investors are waiting for October’s undisclosed run rate to judge whether the company can reach the roughly $100B year-end expectation set at midyear. If it falls short, the IPO price could also come in below $1.5T, rather than necessarily landing in the current discussion range.
3. The race to list first puts a claim on limited capital, while employee unlocks test internal conviction
- Zhang Lu thinks both Anthropic and OpenAI want to be the first to list. On OpenAI’s delay to 2027, she is unsure its internal readiness and financial position support a listing that year, so the delay cannot be read entirely as a voluntary choice; Anthropic may want to go public while growth and market enthusiasm are still high.
- Being first also matters for access to capital. A company valued in the trillions will require a huge pool of buyers; the first to list may absorb capital looking for AI exposure, leaving the next company with a less ample pool. “There isn’t an infinite supply of money in the market,” which is one reason she thinks OpenAI faces more uncertainty.
- The S-1 allows non-core employees to unlock 10% of their options on the IPO date. Zhang Lu acknowledged potential selling pressure, but said the company is relatively young and allowing employees some early liquidity could also support long-term holding. Cheng Manqi added that the more than 5,300 employees share fairly similar values and that relatively few may actually sell; both see employees’ willingness to hold as a possible signal of confidence.
- The IPO’s impact could spill over to other companies. Zhang Lu relayed what the head of Nasdaq’s West Coast operations said: Anthropic’s listing has led some companies to delay plans, while Oura Ring withdrew about 24 hours ago. According to that Nasdaq executive, a last-minute cancellation so close to listing has happened only about twice in Nasdaq’s history.
4. Anthropic’s market consensus followed an early contrarian bet and validation from coding
- Cheng Manqi noted that Spark Capital invested in Anthropic in May 2023 at a valuation of about $4.5B, just as ChatGPT was setting off a global frenzy. Zhang Lu recalled that the market had not yet reached a consensus that there was room for a second frontier-model company; AI safety was not a top priority, either, and Anthropic’s early fundraising was difficult.
- Zhang Lu sees backing from strategic investors such as Google and Amazon in 2023 as both a model bet and a way to avoid a market with only OpenAI. By 2024 and 2025, Claude had earned recognition for coding, complex reasoning and context capabilities, while its safety-first positioning helped it win enterprise deployment trust and clarified its B2B strategy.
- For her, the more important inflection point was coding commercialization in January this year: frontier models began participating directly in high-value production, rather than just finding information and answering questions. Cheng Manqi added that Dario had said Spark’s Yasmin pushed the team to make coding a priority; that kind of strategic support matters more than simply backing a fast-growing company.
- Zhang Lu cautioned against using Spark’s frequent mentions in the S-1 to rank investors’ contributions. Yasmin is a board member, which itself brings disclosure requirements. She credited Spark for early conviction, long-term ownership and continued investment, but stressed that the number of mentions in the filing does not measure the scale of an investor’s contribution.
5. B2B’s advantages lie in feedback, willingness to pay and the data flywheel
- Zhang Lu first disclosed her own bias: she has invested only in B2B for the past 11 years, so she has always leaned toward Anthropic. She sees a focus on high-value enterprise needs as a rational choice for the second model company with limited resources. Consumer needs are fragmented and feedback inconsistent; enterprise workflows are more structured and yield higher-quality information that can be used to improve the product.
- Enterprises also offer 3 data advantages: high-quality training data, feedback generated through use, and new data created as the product is used. Their willingness to pay is clearer, too. The last time she checked, the share of ChatGPT users in North America who paid was about 5%, and likely no more than 10%; usage intensity is higher in companies. Coding also makes ROI easier to measure, giving vendors more clarity on pricing.
- She does not think Anthropic has to enter every consumer market, though she acknowledges that personal agents could make the consumer entry point more important, with OpenAI, Google and Apple all strengthening their positions. Anthropic can still reach users through enterprise APIs and third parties, just with a weaker advantage at the point of entry. Her view is that the market is large enough to support multiple companies with different approaches.
6. There may not be another coding breakout; enterprise platforms are worth building out first
- Asked by Cheng Manqi what the next coding product would be, Zhang Lu did not name a new product with comparable breakout potential. She thinks there may not be another single product that grows as quickly, and that Anthropic should strengthen its enterprise platform and become a key infrastructure provider for agent workflows.
- Cowork is one example she is exploring; Fusion Fund uses it internally as well. If it connects different software through MCP and becomes a middleware layer through which workflows run, Cowork could strengthen Anthropic’s position as an enterprise platform, rather than simply adding another way to sell model capabilities.
- Cheng Manqi broke down the S-1 figures: in the first half of 2026, usage-based revenue grew 1,200% year over year and subscription revenue 1,690%, which could mean more non-programmer knowledge workers are using Claude. But she offered another explanation: subscriptions may previously have been the better deal, which could have boosted growth; the new policy also requires companies with more than 150 employees to pay based on usage. There is no official breakdown of coding versus non-coding users, so revenue categories alone do not settle the question.
7. A $10B budget signals preparation, not proof of ROI
- Zhang Lu’s observations come from the Fusion CXO network she started building in 2018, which includes more than 40 and nearly 50 technology leaders at Fortune 500 companies. Over the past 6 months, more companies have appointed chief AI officers; in 2026, members have been reaching out to her noticeably more often, moving from quarterly gatherings to board and management discussions of AI and data strategy.
- In a recent survey of the world’s 20 largest pharmaceutical companies, she found that 60%-70% ranked AI as their top internal strategic priority. She cited Lilly’s multibillion-dollar partnership with Nvidia and Merck’s partnership with Google Gemini as examples of how quickly competitors follow: “Once one company in an industry starts moving, the others catch up quickly.”
- The chief AI officer at a well-known, unnamed US bank told her that its AI budget for the year exceeded $10B. Cheng Manqi asked whether that reflected cost savings or revenue growth already achieved. Zhang Lu’s answer was direct: “We’re still in the trial phase.” A budget and strategic consensus do not mean financial returns have already materialized.
- Companies are often more worried about competitors than about startups. Zhang Lu imagines a market where 30 companies once held 80% of the market, but 5 or 7 could hold that share in the future as AI amplifies product and service capabilities. Every company wants to be among the few that remain, driving follow-on deployment across industries.
8. Companies adopt automation first, then move toward business optimization
- Zhang Lu distinguishes business automation from business optimization. The former boosts efficiency, reduces costs and improves workflows, but may not show up on financial statements immediately; the latter goes further, increasing revenue and restructuring the business. Different departments are at different stages, so the ROI they see will naturally differ.
- She uses an insurer’s CRM to explain the resistance. Tech companies are accustomed to “breaking the old process first, then rebuilding it”; an insurer does not want to rebuild its CRM. It “doesn’t want to; it just wants to keep using what it has.” Companies may accept agents, but want them to keep working with existing software, so large-scale deployment starts with fitting into current workflows.
- A common path is to start with a department, team or specific business unit, expand company-wide, and then move from individual tools to an enterprise agent platform. Zhang Lu estimates the first step alone could take 3–5 years, with optimization taking longer; it is unrealistic to expect the whole economy to become AI-enabled in a single year, next year.
9. Coding’s breakout pace cannot be projected directly onto other knowledge work
- Cheng Manqi’s question was about pace: coding spread quickly from 2025 to 2026; can finance, law and healthcare move just as fast? If not, could the next few years be uneven? Zhang Lu agreed that this is a risk and stressed the need for reasonable, rational expectations, rather than simply assuming the end market will be large.
- Coding combines a tech sector that adopts new tools quickly with high task frequency, high value per task, direct product output and fast validation. Zhang Lu cited one definition from Kai-Fu Lee: AI can do 90% of the work and do it better than 90% of people. She narrowed the definition to a specific application: an agent approaches AGI in a given field when it can complete more than 90% of the work and do it better than more than 90% of people. For now, only AI coding meets that bar. Many legal, medical and financial tasks can only produce advice, with no immediate way to verify the result.
- Rather than looking for another coding breakout, Zhang Lu thinks enterprise back-office agents may reach adoption sooner, because needs are more consistent across companies and more standardized. Finance, from gathering and consolidating data to producing reports, is a complete use case for AI automation; compliance, procurement and inventory management offer similar opportunities.
- Data strategy is another central theme. Zhang Lu recalled saying at FII in 2019 that data would become companies’ most important asset. Large companies are now cleaning and integrating data that was previously unusable to build high-quality databases for AI. Model companies pay for data, while much of the highest-value enterprise data remains in traditional industries.
10. Open source is taking share in procurement, but closed-source demand can still grow with the market
- Zhang Lu cited a report she saw about 2 weeks ago showing the share of large companies using closed-source or frontier models falling from 57% to 14%, as more companies look to deploy open-source models. She also understands why public-market investors want to see Anthropic’s October data rather than extrapolate from its past growth.
- AT&T is the public example she cited: 60%-70% of its internal AI budget will go to open source. She has also seen Nvidia actively help enterprise customers deploy open-source models, and thinks the US open-source ecosystem has developed quickly since July. China already has a strong open-source ecosystem as well, giving companies more options when they buy.
- Cost is a major driver. She cited an anonymous client of a consulting firm: a large company’s Claude bill reached hundreds of millions of dollars in a single month, close to $500M, shocking the CFO and clearly proving unsustainable. Tech companies are changing their stance, too: from encouraging heavier token use to limiting excessive use and asking how to control costs.
- She still does not think the rise of open source necessarily hurts closed-source companies. Tech accounts for less than 10% of US GDP, while AI is moving into more industries representing 30% or 50% of economic activity. “The overall pie is still getting bigger”(整体的蛋糕还是越来越大); the market just will not be dominated by 1 or 2 companies, but shared by open and closed models, large companies and startups.
11. Healthcare’s opportunity comes from standardized workflows, local deployment and data trust
- Zhang Lu is especially bullish on healthcare: it is close to 20% of US GDP, has large volumes of high-quality data, and many of its workflows are relatively standardized. She estimates AI penetration in the industry may still be below 5%-6%, and stresses that it is certainly below 10%. If companies can address data sensitivity, regulation and specific functionality, adoption could move faster than people expect.
- She cited CommonSpirit Health Foundation, where she serves as a director. In May this year, its CTO told her that thousands of hospitals and clinics could begin deploying agents. Once institutions of that scale enter a real usage and feedback loop, she believes there is still room for rapid growth. She added that many healthcare organizations already use startup solutions, while some large institutions may deploy localized open-source models internally.
- Some data in highly regulated industries has to stay local and cannot be uploaded directly to the cloud, creating real demand for small local models, open-source deployment, privacy and regulatory features. Zhang Lu stressed that this is not just about whether a company wants to use a general-purpose model; legal risk is also involved. Several healthcare AI companies she has invested in build these capabilities directly into their products.
- Startups may also have an advantage on trust: companies may be willing to share data with a startup because they believe they can control the relationship and build AI applications suited to their workflows and use cases. By contrast, Anthropic’s move to build a wet lab has made some pharmaceutical companies uneasy: “You used all my data, and now you’re competing with me.” The same concern could spread to other industries.
12. New infrastructure must be rebuilt around inference costs and physical AI
- Zhang Lu thinks infrastructure was previously optimized to “train the strongest model”; the priority now should be lowering inference costs, especially “the cost of inference per task.” The shift she describes is from roughly 80% of compute going to training in the past to roughly 80% going to inference in the future. Inference is closely tied to local enterprise deployment, creating new startup opportunities.
- The mode of use is shifting from passive to active: users used to ask questions and products would answer; agents proactively call on resources and complete work. Cheng Manqi added that Fusion Fund has invested in a heterogeneous inference-chip company, showing how changes in applications and models could also change the compute architectures needed underneath.
- Physical AI requires discussion of specialized models from the outset, because devices have limited compute and power, and may not be able to run large models. Zhang Lu agreed that world models are powerful, but emphasized that running them consumes enormous compute and electricity. In a phase where practical costs matter, how to deploy them at low cost and large scale is itself an infrastructure innovation opportunity.
13. AI for science can train reasoning, but its near-term revenue contribution is uncertain
- Asked about frontier model companies investing in mathematics and theoretical science, Zhang Lu first explained the technical rationale: mathematical proofs can be formally verified, making them easier to judge as right or wrong than open-ended natural-language tasks. That matters for reinforcement learning, reasoning training and automated evaluation. She recently spoke with a group of mathematicians generating their own data, which model companies including OpenAI are buying at high prices.
- The longer-term aim is for AI to move from summarizing existing knowledge to producing new knowledge: a model that can pose questions and produce new, verifiable scientific results is beginning to create knowledge. Asked by Cheng Manqi about commercial returns, Zhang Lu reserved judgment: this research may not contribute revenue in the short term, but could improve reasoning and scientific capabilities over time and open up new technical paths.
14. Drug development’s bottlenecks extend beyond target discovery to clinical costs and success rates
- Zhang Lu pushed back on the emphasis on drug discovery: going from finding 2 or 3 targets to finding hundreds does not mean you have a drug. Clinical trials still come next, and that is where drug development costs the most. “Are we going to take all these hundreds of candidates into clinical trials now?” More candidates could mean higher spending, without necessarily improving success rates.
- She sees more potential in AI-assisted clinical trials, and Fusion Fund has invested in companies in the space. Her vision is to select the 2 or 3 candidates with the highest clinical success rates from 300–500 targets, then move them into experiments and try to raise the success rate to above 50%-60%. This is the value proposition she described, not an outcome already achieved.
- Another portfolio company has a vertical model for cell therapy that uses digital twins of human cells to study different diseases. Zhang Lu said it recently reached a substantial partnership with Novo Nordisk in Copenhagen for a Parkinson’s indication. She sees healthcare as a major AI-for-good opportunity, with both economic value and the potential to improve human welfare.
15. Total compute demand can rise as older GPUs lose value
- Cheng Manqi raised the possibility that mid-sized unicorns could sign 3–5-year compute contracts, then sublease capacity at a discount if their business falls short of expectations. Zhang Lu confirmed that isolated cases exist, but does not think they amount to an overall surplus. She said companies such as Together AI and Cohere had essentially booked all their compute for the year, with customers needing to wait until next year; training, pretraining and post-training still face shortages.
- She recalled that more compute suddenly became available late last year, only for a shortage to return quickly at the beginning of this year, suggesting a cyclical market. Pricing can also push companies toward long-term deals: she heard that SpaceX’s quote for a one-year contract “seemed” to be double, making companies feel they might as well sign for longer. That incentive can amplify the pressure of future commitments.
- GPUs are not interchangeable. H100s may be available to order, but many customers need B300s or H200s; training requires large-scale, high-bandwidth, low-latency clusters, while inference is more distributed. Different GPUs, network architectures and data center locations serve different workloads, so capacity released by one company cannot automatically fill another’s shortage.
- Zhang Lu’s distinction is that “compute demand growth and falling prices for compute assets can happen at the same time.” The volume of new compute can rise while rental rates for a generation of GPUs keep falling, forcing resellers of older cards to cut prices. Cheng Manqi closed by relaying another guest Rob’s bottom line: if demand does not keep pace, debt on the supply side is when the problem emerges.
16. Model-as-product is squeezing wrapper apps, while consumer AI puts a premium on distribution
- Zhang Lu thinks it will be very difficult for startups to compete directly in areas where Anthropic is already strong, such as coding. Tech applications that simply “wrap” an existing model are also becoming less resilient against model companies. She is more interested in horizontal infrastructure and vertical applications in highly regulated industries, where regulation itself creates a barrier to entry.
- Zhang Lu publicly argued several years ago that model companies or tech giants could ultimately capture 80%-90% of the US consumer AI market. Google’s advantage comes not just from high-quality user data, but also distribution channels such as Chrome; Gemini Flash can use them to gather extensive real-world feedback and improve the product, at a scale and speed most startups cannot match.
- Cheng Manqi asked about personal-agent startups that have recently caught on. Zhang Lu said the user experience “is still pretty good for now” and they still have plenty of users, but whether the momentum lasts remains to be seen. Giants such as Meta bring more than product competition: they can reach large numbers of users directly. She left room for exceptions—consumer AI will still produce good products, and the US and China markets are different.
17. Governance must be addressed before physical AI can scale
- The issue Zhang Lu is most focused on, but still sees no clear answer to, is AI governance. If a language model gets something wrong, a user may laugh it off; an error in physical AI could hurt someone or cause a major accident. CTOs at leading industrial companies are urgently asking her whether technical AI governance solutions must be in place before physical AI can be deployed at scale; healthcare faces the same question.
- Governance may have to start with visibility. She asks company CEOs whether they know how many AI tools their employees use and what core data those tools can access. If they lack even that basic understanding, she says, there is no point talking about AI safety. Transparency into tools, data and permissions is therefore another layer companies need for deployment.
- The AI for Good Commission she is part of recently spent more than 2 hours discussing the issue. The challenges include accounting for future technology, cross-regional and cross-industry coordination, and deciding between sector-specific regulation and cross-industry standards. Zhang Lu sees both an opportunity and a challenge, and urged startups, enterprises and investors to pay attention; she did not offer a settled answer.