Topics Pioneers Insight Method Author
Back

Deep Dive: Anthropic S-1 Myths and Mythos — Jamin Ball, Altimeter

·Latent Space

Rendering the long image…

Topics Battle for Frontier AI Moats AI Inference Cost Optimization AI Venture Funding and Bubble Risks Enterprise AI ROI and Cost Optimization Trillion-Dollar AI Capex and Bubble Risks Compute Cloud Debt and Macro Liquidity Specialized SLMs vs General LLMs
Pioneers Jamin Ball swyx

Summary

  • Ball reads Anthropic’s S-1 as evidence that frontier AI can generate substantial cash while still growing rapidly. He cites 85% gross margins, 21% free cash flow margins, and 8% GAAP operating margins. Against $141 billion raised, Anthropic holds $134 billion in cash and equivalents; Ball’s subtraction implies $7 billion of lifetime burn, alongside separately reported Q3 revenue of $16.8 billion.
  • Inference efficiency explains the profitability surprise, but its future pace remains an open question. Serving-compute spending rose 2.3x while revenue rose 14x, which Ball describes as roughly a 7x efficiency gain and “almost accidentally profitable.” Swyx presses on how much optimization remains physically possible; Ball expects margins could turn negative again as Anthropic invests more aggressively in compute.
  • The headline convertible liability and accounting loss are not cash burn in Ball’s explanation. Roughly $6 billion of outstanding convertible funding has appreciated to a $108 billion liability; that liability is not an additional $108 billion raised. Ball attributes the entire $81 billion first-half 2026 net loss to revaluation of converts, SAFEs, and notes: appreciation in promised shares creates an accounting expense.
  • Enterprise penetration leaves substantial room for spending to deepen, even after extraordinary revenue growth. Half the Fortune 500 uses Anthropic, but the cohort spends only 0.1% of its expenses with the company. Ball contrasts that with John Collison’s expectation that Stripe’s token spending could rise from 1% to 15–20% of expenses over several years; swyx’s preferred denominator is “what are you spending on humans?”
  • The durable business depends on a factory that repeatedly produces frontier models. Models six months old or younger generated 80% of revenue in the three months ending September. Ball expects open weights to win token volume while frontier models capture the most dollars, but swyx preserves the failure risk: an unsuccessful successor run can undermine the economics of rapidly depreciating models. A strong business also still needs a sensible purchase price.
  • The $856 billion commitment headline needs a spending schedule and a revenue trajectory. Ball’s illustrative annualization asks whether Anthropic can fund roughly $100 billion a year, with some commitments back weighted or less firm, including a SpaceX arrangement terminable in 90 days. His case depends on continued demand, efficiency, and capacity expansion; swyx’s framing is that “the game is not to reduce the commitments, it is actually to get more commitments.”
  • Credit appetite is becoming a separate constraint on the compute build-out. Ball describes multibillion-dollar GPU debt markets as “No bid. No bid.” Lenders see through different NeoCloud borrowers to the same end customers, OpenAI and Anthropic, while construction delays and expensive hedges compound exposure. He disputes swyx’s suggestion that debt markets are running out of money: the immediate issue is appetite, concentration, and terms.
  • Compute rationing helps explain Anthropic’s coding focus and leaves openings for specialists. Allocating capacity to Sora means capacity unavailable elsewhere; Ball argues the labs are pursuing “the next trillion,” leaving smaller markets attractive to startups. He expects applications to need aligned product and research teams, while swyx challenges the evidence for Anthropic’s vertical ambitions: a design tab or legal Markdown repository is still a limited product commitment.

Deep dive

1. The IPO’s financing value extends beyond another equity raise

  • Ball discloses that Altimeter has a large Anthropic position and that he is expressing personal views. The IPO’s size remains undecided: he cites a $75–100 billion fundraising rumor. Historically, companies floated 10–15% of their equity; larger companies can sell a smaller percentage and still raise enormous sums.
  • At the October 29 recording, Ball sketches roughly 15 days between filing and roadshow, followed by about a week and a half of investor meetings, suggesting trading around November 10–20. Filing does not guarantee listing; swyx cites Cerebras filing, stopping, and returning. A 50% first-day pop also raises Bill Gurley’s objection: the issuer could have priced higher.
  • Employee liquidity and acquisition currency remain reasons to list, although regular tenders already provide restricted liquidity. Ball thinks leading private companies’ shares increasingly resemble cash to counterparties. Private fundraising capacity is also considerable: he cites OpenAI raising roughly $120–130 billion, weakening the argument that only public markets can supply large equity checks.
  • Ball’s more consequential explanation is credit access. Lenders underwriting Anthropic directly, or a NeoCloud building for it, need detailed financial disclosure throughout syndication. Public reporting makes that openness easier, supports credit ratings, and can improve borrowing terms. His explanation of “why now” remains speculation, with several motives potentially contributing.

2. The cash balance makes the profitability surprise concrete

  • Ball’s headline financial profile is 85% gross margins, 21% free cash flow margins in the latest quarter, and 8% GAAP operating margins. Earlier leaks mostly exposed revenue or ARR; the filing adds cash generation and operating economics. That changes the discussion from whether demand exists to whether supplying it produces an attractive business.
  • His favorite comparison is “$141 billion of capital raised to date,” including equity, converts, and SAFEs, against a $134 billion cash and cash equivalents balance. Ball infers $7 billion of lifetime burn, then contrasts it with $16.8 billion of quarterly revenue in separately reported Q3 figures that are not in the S-1.
  • Free cash flow adds back stock compensation, which Ball acknowledges is a real economic expense: employees might not join without equity. Treating it as an expense still leaves approximately an 8% margin in his calculation. Ball calls this “very rare”; swyx’s stronger contention is that this profitability does not happen at a software IPO.
  • Swyx flags distribution economics beneath the gross-margin headline: he puts cloud-partner revenue at 48%. Ball believes partner revenue sharing sits in sales and marketing, which swyx puts at $3 billion. Gross margin would look lower with those payments included, although Ball emphasizes that operating and cash-flow margins already capture broader expenses.

3. Inference efficiency has changed the cash-flow curve

  • Only 27% of Anthropic’s compute currently serves revenue-generating models, according to Ball’s chart. The remainder supports internal work, including training, research, safety, and alignment. The company therefore has substantial compute expenditure that does not directly correspond to tokens being sold today.
  • Serving-compute spending rose 2.3x over a year while revenue rose 14x. Ball calls the embedded improvement roughly 7x and describes the business as “almost accidentally profitable.” His monetization illustration moves from $10–15 million per megawatt to $70–100 million: substantially more revenue from a unit of capacity, rather than proportional spending growth.
  • Traditional SaaS promised an “annuity of cash flows”: acquire customers while losing money, then enter harvest mode when acquisition spending slows. Anthropic has reached positive cash flow during exceptional growth. Ball’s interpretation is that efficiency improvements have accelerated the point at which the business can support itself.
  • That does not make 21% free cash flow margins a permanent floor. Ball guesses Anthropic will invest more aggressively, with upfront compute payments potentially pushing margins negative again. Swyx stresses separating core operating economics from investment choices; Ball’s takeaway is that Anthropic has demonstrated the capability to operate profitably.

4. Enterprise adoption is broad while spending remains shallow

  • Half the Fortune 500 uses Anthropic, yet that cohort spends only 0.1% of its expenses with the company. Ball treats this as the distinction between breadth and depth. Moving from 0.1% to 1% would be a substantial expansion even without assuming adoption or spending patterns comparable to the most aggressive technology companies.
  • Ball cites John Collison saying Stripe spends 1% of expenses on tokens and expects 15–20% over the next few years. Stripe is unusually forward leaning, he cautions. The relevant possibility is a similar magnitude of growth elsewhere, not the assumption that an ordinary Fortune 500 business will reach Stripe’s proposed spending mix.
  • Swyx’s denominator is knowledge-worker compensation, including coding, administration, operations, and legal work. He cites median AI-company token spending at 50% of salaries and a high-end Anthropic employee example around 10x salary. Physical businesses differ until robotics improves, but the ratio of token spending to human knowledge-work spending has moved from zero and can keep rising.

5. Frontier models can capture dollars while open weights capture volume

  • Ball reduces the investment debate to three questions: can Anthropic remain at the frontier, is there a durable business there, and what does maintaining that position cost? The last two are linked: the revenue opportunity must cover the expense of staying competitive. His central conviction concerns Anthropic’s ability to keep producing leading models.
  • In the three months ending September, models six months old or younger generated 80% of revenue—more than Ball expected. He anticipates two simultaneous outcomes: open-weight and inexpensive models gain the largest share of tokens, while frontier models take the largest share of spending. Token-volume leadership therefore need not imply dollar-revenue leadership.

6. Compute commitments need a schedule before they can be judged

  • The approximately $856 billion commitment headline is a multiyear total. Ball constructs an illustrative annual schedule, spreading disclosed deals across their stated durations; this is his interpretation, not a company-provided year-by-year forecast. The funding question becomes whether Anthropic can support roughly $100 billion annually, rather than treating the entire commitment as immediately payable.
  • Timing and firmness also differ. Some obligations are back weighted, and Ball separates harder commitments from less binding intentions. Swyx highlights the SpaceX arrangement that Anthropic can walk away from in 90 days. Adding everything into a single headline loses distinctions that matter when assessing how much capacity and financial exposure the company has actually secured.
  • Ball’s operating thought experiment starts with only roughly a quarter of compute serving customers. In theory, doubling that allocation without adding capacity could double revenue. It illustrates flexibility, rather than a forecast: Anthropic must still decide how much compute belongs in inference versus the research that produces its next competitive models.
  • Even after subtracting all compute expenditure, including internal work and training, Ball’s chart shows compute margins above 50%. Continued revenue growth and efficiency underpin his confidence in covering commitments. Swyx goes further on management intent: “the game is not to reduce the commitments, it is actually to get more commitments.”

7. Compute scarcity can restrain overbuilding and increase defensive demand

  • Ball says nobody he speaks with at the major labs feels adequately supplied. OpenAI is buying TPUs and developing Jalapeño; Anthropic has staffed an internal chip effort, while hyperscalers pursue their own silicon. Swyx calls the resulting scramble favorable for TSMC, NVIDIA, and ASML. More architectural options do not eliminate the immediate capacity shortage.
  • Local opposition adds another constraint: noise, pollution, electricity bills, and fear of AI make data-center construction political. People also raise water-use concerns, which Ball says are misplaced; he identifies electricity use as the real issue. Ball sees permitting friction as a governor against the fiber-style overbuilding of the internet bubble. “We have no dark GPUs today.” That protection against a supply glut also slows capacity the industry wants.
  • Safety, alignment, and verification could absorb an increasing share of compute, Ball argues, both in research and during inference. More powerful attackers also require more capable defenders. His personal belief is that compute constraints last longer than the world expects; swyx says establishing the difference between market expectations and actual duration could be materially valuable.
  • Cybersecurity supplies a concrete example. Swyx corrects Ball that the PHP vulnerability work used GPT-5.6, rather than a frontier model. Ball says discovery-to-exploitation windows have collapsed toward zero. Swyx questions spending billions of tokens finding and fixing problems that otherwise remained hidden; Ball’s response is that “you can’t put the toothpaste back in the tube.”

8. Falling task prices do not settle the remaining efficiency question

  • Swyx repeatedly distinguishes current progress from future headroom. A poorly optimized inference stack can improve dramatically, but physical movement of energy and data imposes limits. Was the reported improvement partly a one-time transition to competent optimization? His question is how much remains to squeeze out, not whether models have stopped improving today.
  • Swyx uses Mark Chen’s chart accompanying a Navier–Stokes result to argue that an internal OpenAI model was well ahead of Astra. Public frontier-versus-open-weight comparisons can therefore understate the actual gap. Ball calls the point fair and says each model generation still shows “quantum leaps”; he also notes that no one is saying released models have plateaued today.
  • Swyx returns to the financial question: Anthropic’s revenue efficiency improved without a corresponding price increase. His chart shows a 25–40% reduction in cost to complete the same task across successive Fable and Opus model families. Lower task costs coexist with rising revenue. The speakers discuss Jevons’ paradox: falling prices can unlock enough additional usage to expand spending substantially. That demand argument does not quantify the remaining engineering headroom.
  • Ball’s cloud analogy is S3 and EC2: prices fell about 99% from launch while the businesses expanded enormously. Swyx supplies the investing mistake from his own hedge-fund experience: avoiding Amazon because of AWS margin concerns and startup-customer exposure. “You can do all this work and make worse decisions,” especially by fixating on the wrong variable.

9. Frontier leadership still has a price and a product cycle

  • Swyx pushes back on translating a successful business into a stock purchase: “there’s a right price for that.” Ball agrees; the IPO price is not yet known. Company execution also cannot insulate shares from correlated market declines. He gives little weight to private-stock proxies, although they might reveal sentiment.
  • Ramp spending data makes competitive turnover visible. Ball links the summer slowdown to Astra and stronger Kimi K3, GLM, and DeepSeek models. After Opus 5.5 and Sonnet 5.5, spending outside those models fell by half or more over two or three months. He would not be surprised by October reacceleration, but treats third-party data as an indication.
  • Swyx feels labs’ reserves of unreleased models are building again, although he does not know why; a backlog also gives executives commercial launch options, including around an IPO. Ball emphasizes safety and alignment reasons for delay. Their disagreement concerns interpreting release timing, not whether internal models can be more powerful than public products.

10. The model factory matters more than a single successful run

  • Swyx frames the depreciation risk with a hypothetical $1 billion model lasting six months. His answer is a qualified no: rapid replacement works while the successor beats the current state of the art. A disappointing run can disrupt ROI; he cites GPT-4.5’s less successful large run and OpenAI’s subsequent recovery.
  • Ball’s semiconductor analogy distinguishes taping out one competitive chip from building an organization that does it every year. NVIDIA’s advantage is “the best factory,” not simply the best current chip. Swyx calls this “the machine that builds the machine.” Applied to labs, the asset is the repeatable process for producing successive competitive models.
  • That is also Ball’s bear case for new labs. One industry heuristic he hears is that challengers cannot fall more than one order of magnitude behind leading labs in compute and remain competitive. The question is whether they can repeatedly attract the talent and capacity required, not merely finance one impressive training run.
  • Ball thinks unknown entrants face a difficult generalized-frontier catch-up problem, while domain-specific models retain opportunities. Anthropic must keep running because OpenAI and Gemini will continue if it stops. He expects longer, more valuable tasks to sustain opportunity, while acknowledging his disposition: “maybe I’m drinking too much of the Kool-Aid.” OpenAI is also trying to compete at the low end with smaller, cheaper Luna-class models.

11. Coding focus reflects compute rationing and a route into knowledge work

  • Ball calls this a “compute innovator’s dilemma”: if Sora demand explodes, does OpenAI divert Codex capacity to serve it? Swyx says that tradeoff explains Sora’s shutdown. Coding has delivered unusually high economic value, making competing allocations difficult. The labs are pursuing “the next trillion”; a $10 billion outcome can simultaneously be enormous for a startup and too small for them.
  • Anthropic’s strength is focus: coding models plus Claude Code, rather than simultaneous expansion into images, robotics, and every other modality. Ball argues coding is also a framework for generalizing knowledge work. He uses Claude Code without coding; the same harness can perform broader tasks. Swyx contrasts Anthropic’s narrower enterprise-and-coding focus with OpenAI’s broader exploration and Gemini’s willingness to try many areas. Both speakers see life sciences as another area of expansion, with swyx saying Anthropic seems to be investing more there and Ball agreeing.
  • Specialists can exploit capabilities the generalized labs underprioritize, including visual reasoning and decision models. Ball’s analogy: “you don’t need to drive the Ferrari engine to the grocery store.” A cheaper model suited to the task can suffice. His enthusiasm for frontier economics therefore coexists with an expectation of better, less expensive domain-specific alternatives.

12. Labs and application companies are integrating from opposite directions

  • Swyx asks what Anthropic’s product segmentation actually amounts to beyond Claude Code, Cowork, and the platform. Where do consumer Claude, legal, and design fit? For an IPO-stage company—“much less a $2 trillion company,” as swyx puts it—three SKUs is strikingly sparse. He wants evidence of distinct businesses, not merely a list of domains the underlying model can address.
  • Ball describes an earlier hypothesis: labs might withhold their best models and build superior applications around them. A coding agent using an unavailable model could threaten Cursor; swyx supplies Harvey as the legal equivalent. That possibility helps explain application companies bringing models in-house to gain more control over their future.
  • Ball’s application-investing requirement is a strong product organization paired with strong research, with improvements feeding directly into the product. Research for its own sake is insufficient. Swyx distinguishes model labs, where “research is king at Anthropic,” from agent labs that prioritize customer needs and integrate backward into models when useful. He thinks both can coexist.
  • Ball expects a larger vertical-application push; swyx challenges the evidence so far. Claude Design is a tab, not an Adobe replacement, and Claude Legal is a Markdown repository. Ball concedes the point while noting organizational inertia: salespeople naturally favor the established product they know how to sell over unfamiliar offerings.

13. Convertible revaluation explains the headline debt and loss claims

  • Ball puts convertible notes payable near $108 billion. That is not $108 billion of additional funding. Approximately $11 billion was originally raised through the relevant converts; $5 billion converted into equity, leaving $6 billion outstanding. The promised shares’ appreciation raised that remaining funding’s accounting value to approximately $108 billion.
  • Why use converts? Ball offers hypotheses, not established motives: strategic investors Google and Amazon might avoid ownership thresholds triggering additional reporting, while Anthropic might prefer particular voting arrangements. The financing occurred years earlier at then-current valuations. It was not a new grant of $108 billion in discounted shares, and the original funding is included in the $141 billion total.
  • The income-statement counterpart is first-half 2026’s approximately $81 billion net loss. Ball attributes $72 billion to revaluation of the AWS and Google converts and $9 billion to another group of SAFEs and notes. His categorical explanation is that 100% of that loss reflects appreciation of those instruments, rather than operating cash consumption.
  • A larger liability requires an offset through the linked statements: the revaluation expense reduces net income and retained earnings. Straight equity financing would not create that recurring effect; conversion removes it. Ball accepts the GAAP loss while rejecting it as a representation of operating performance. His balance-sheet conclusion is equally explicit: “there is no debt on the balance sheet.”

14. Credit appetite has become a separate bottleneck

  • Ball describes current multibillion-dollar GPU debt issuance as “No bid. No bid.” Credit providers examine the ultimate offtaker, not just the NeoCloud borrower. Loans to CoreWeave, Nebius, or another operator can all lead back to OpenAI and Anthropic. After substantial lending against those customers, a lender can reasonably say, “I’m feeling kind of full.”
  • Swyx interprets this as debt markets running out of money. Ball disagrees: “Debt markets are super deep”; appetite is the issue. Expensive credit-default swaps and other hedges make additional exposure harder to manage. There is also a price for risk—his illustration is $5 billion borrowed at 20%—but availability on such terms differs from attractive financing. Ball also says data-center debt deals are now often getting done above 10%.
  • Project risk adds to concentration. A NeoCloud can contract with OpenAI, order NVIDIA hardware packaged by Dell, and raise financing, only for community opposition to delay construction. Interest still needs paying. Delayed-draw facilities help, but have limits; contracts must allocate responsibility among suppliers, developers, operators, and other causes of delay.
  • Ball will not conclude Anthropic cannot borrow: obtaining a rating and choosing to assume debt are separate questions. Labs may prefer NeoClouds carrying construction and balance-sheet risk. His analogy is fintech businesses that prospered at low rates and suffered when rates rose. Anthropic has undrawn facilities, and future borrowing would not surprise him.

15. Gigawatt forecasts need physical and financing constraints

  • Ball declines to project FY ’27: the future is too “unknown and unknowable.” Swyx proposes a more useful analyst exercise: estimate remaining inference efficiency, plausible competitive pricing, secured GPU supply, and the tokens that supply can produce. Extending historical growth by dragging a spreadsheet cell misses the physical ceiling on deliverable revenue.
  • Ball puts one gigawatt’s build cost around $50–70 billion and says both labs have discussed reaching roughly five gigawatts by year-end, while acknowledging uncertainty. Swyx notes OpenAI’s Stargate ambition of 30 gigawatts by 2030 and says he has not heard an equivalent Anthropic plan. Ball does not know whether their capacity is equal.
  • Ball’s deliberately rough scenario assumes 100 gigawatts of industry additions over five years at $50 billion per gigawatt: $5 trillion, or $1 trillion annually. He compares that scale with roughly 5% of GDP and the railroad build-out. It is illustrative arithmetic, not a forecast, and requires exceptionally receptive credit markets and global construction.
  • Constructed capacity also differs from live capacity generating tokens: next year could mean 50 gigawatts, or 20–30. Ball hears orders exceeding supply-chain capacity. Kimi K3 illustrates memory demands; swyx says he had previously expected model sizes to cap and optimization to take over. They agree open-weight models seem to cap around 100 billion active parameters, with inference speed, memory availability, and GPU generations shaping that ceiling.

16. Infrastructure leads this cycle while applications may lead later

  • Ball’s six core beliefs concern more consumption of models, power, compute, networking, and storage, plus more data centers to supply intelligence. He looks for businesses directly downstream of those convictions. His examples include a stealth chip company and American Terawatt, which aims to build an industrial grid serving large industrial electricity customers.
  • Swyx sees compute trading as another route to capacity: improve utilization of hardware already installed. His illustration moves utilization from roughly 30–40% toward 60%, potentially creating a large increase in useful compute without a corresponding increase in GPUs. Ball sees opportunities for innovation at silicon, capital-market, and product layers simultaneously.
  • Ball’s triangle analogy puts hardware first in early market cycles, followed by software infrastructure and applications. In mature markets, applications become the largest opportunity; he expects agents and apps eventually to do the same here. Swyx contrasts Cursor’s roughly $60 billion outcome with the $2 trillion lab discussion. Ball also acknowledges his personal infrastructure bias.

17. Broader equity ownership is part of the political argument

  • Public listings also broaden participation beyond accredited investors and institutions. Swyx links exclusion from AI wealth to opposition and regulation. Ball’s closing example is Brad’s work on Invest America: government-funded investment accounts at birth, supplemented by families, employers, and philanthropy, intended to give more people firsthand participation in equity appreciation.
  • Ball says the initiative has passed into law, newborns receive $1,000, and funds are inaccessible until 18. He cites Michael Dell’s $6 billion donation and Musk’s pledged shares, while hedging the precise targeting of Dell’s contribution. Swyx calls it a “baby 401(k),” then raises the objection: someone in immediate poverty may need the money today.
  • Ball acknowledges the difficulty and emphasizes the current execution problem: parents must claim accounts already created for children. Older children do not automatically receive the newborn grant, although donations may fund their accounts. The work has shifted from legislation and infrastructure to converting created accounts into claimed accounts through awareness and outreach.