Sign in Subscribe
AI Intelligence

Anthropic's IPO Bets on Revenue That Doesn't Exist Yet

Vijay Bhagwati

• 9 min read
Share

Bankers are pricing a $965B valuation on a 2028 forecast four times today's run rate, while Nvidia quietly halves its OpenAI backstop.

August 17, 2026 · 11 minutes · Issue #239

Anthropic's IPO is being priced against a number that has not happened yet. According to Reuters, bankers and investors are building valuation models around a 2028 revenue forecast of $190–200 billion — a figure that would require the company to roughly quadruple its current $47 billion annualized run rate (as of May 2026) within two years. The resulting valuation range under discussion, per multiple derivative reports, sits near $965 billion, which would make Anthropic's public debut the largest of 2026 by a wide margin, ahead of OpenAI's own listing plans.

The mechanics of that forecast matter more than the headline number. Anthropic's revenue has grown largely on enterprise API consumption and Claude Code adoption, and the company has been explicit that coding and agentic workloads are the primary growth lever. A forecast four years out that assumes continued exponential growth is not unusual for a pre-IPO growth story — Snowflake and Palantir both leaned on similar multi-year curves — but the gap between a $47 billion run rate and a $200 billion target is unusually wide even by AI-industry standards, and it arrives the same week Nvidia's own financial exposure to the sector came under scrutiny.

A 13F filing disclosed that Nvidia halved its $250 billion compute-financing backstop to OpenAI, a commitment first reported in mid-August alongside similar arrangements involving SpaceX and Intel. Nvidia has not explained the reduction publicly, but the filing places it beside stakes in two other capital-intensive, revenue-forward companies — a pattern that reads less like retreat and more like risk-spreading across a compute-financing web that increasingly resembles vendor financing more than pure product sales. Anthropic's prospectus will need to explain its own capacity commitments in similarly forensic detail once the roadshow starts, and every number in that forecast will now be read against Nvidia's own hedging.

Nvidia's Compute Guarantees Get a Second Look

Nvidia's halved OpenAI backstop is the more concrete of the two data points this week, precisely because it came from a regulatory filing rather than a forecast. The original $250 billion compute-financing arrangement, structured to backstop OpenAI's data-center buildout, was already unusual for how directly it tied Nvidia's balance sheet to a single customer's expansion plans. Cutting that exposure in half — while simultaneously disclosing new stakes tied to SpaceX and Intel — suggests Nvidia is diversifying counterparty risk across its biggest AI customers rather than deepening any single bet. Separately, Nvidia is reportedly in talks to invest up to $3 billion in SB Energy, a SoftBank-backed data-center developer central to OpenAI's own campus plans. The company that sells the chips is now also financing the buildings that house them, and hedging both sides of that ledger at once.

Stripe Buys Its Way Into Model Routing

Stripe has finalized an acquisition of OpenRouter, the AI model marketplace, for more than $7 billion — nearly six times the $1.3 billion valuation OpenRouter carried as recently as May. OpenRouter's core product, a unified API layer that lets developers route requests across dozens of model providers, has become critical infrastructure as the number of viable frontier and open-weight models has multiplied. Stripe's payments infrastructure and OpenRouter's routing layer solve adjacent problems for the same customer: developers who need to meter, bill, and route AI usage across providers they don't want to integrate individually. The six-fold valuation jump in three months says more about how fast infrastructure plays are re-rating this cycle than about OpenRouter's own disclosed growth.

Higgsfield Triples Video-Gen Valuation in Months

AI video generation startup Higgsfield raised $400 million from DST Global, Goldman Sachs, Liberty Global, and Intel at a $5.4 billion valuation, up from $1.3 billion earlier this year. The round continues a pattern across the video-generation category, where enterprise demand for branded, controllable video output is outpacing the underlying models' actual differentiation. Goldman Sachs' participation as a named investor, rather than through a fund vehicle, is one of the bank's more direct bets on generative media infrastructure this year.

Open-Source Pulse: Qwen Passes Google and Meta on Downloads

Hugging Face's "State of Open Models Summer 2026" report puts Alibaba's Qwen family past 3 billion cumulative global downloads over the past six months, ahead of Google's 418 million and Meta's 227 million for the same window. Developers have built more than 151,000 derivative models on top of Qwen checkpoints, more than any other open foundation, according to the same report. The download gap is not simply about model quality — Qwen ships at more parameter sizes, with more permissive licensing terms, and with faster cadence than Llama or Gemma, and Alibaba has treated open weights as a distribution strategy for its cloud business rather than a research artifact. Separately, developer Simon Willison highlighted Qwen 3.8's 27B variant, a 17GB open-weight model he found capable of long-context reasoning, tool calling, and competent code generation — evidence that the mid-size open-weight tier, not just flagship releases, is where competitive pressure on closed labs is sharpest. Mistral is now betting the same dynamic plays out in Europe: the company announced it will host third-party open models, starting with Z.ai's GLM-5.2, on its own regional infrastructure alongside its proprietary models, explicitly framing the move as "AI sovereignty" — enterprises and governments should be able to choose where inference runs without switching infrastructure providers. Mistral also said it is forming a coalition to secure up to 1 gigawatt of dedicated European AI compute capacity by 2030, a scale commitment that puts a number on Europe's chronic complaint about compute dependency on US hyperscalers.

Policy & Power: Washington Tells Allies to Pick a Side

The US State Department is preparing to tell 35 partner countries they will be excluded from the Pax Silica critical-minerals and AI initiative if they also join a rival Chinese-led program, according to Reuters, which reviewed a letter and cited a US official confirming the ultimatum. Pax Silica, launched to reduce allied dependence on Chinese critical minerals and to coordinate AI model and chip export controls, has so far recruited partners including Kazakhstan. The letter's language — "you can't have it both ways" — reflects Washington's judgment that Chinese open-weight models have closed enough of the capability gap with US labs that non-aligned countries can no longer treat AI infrastructure choices as commercially neutral. The timing is pointed: it lands the same week Z.ai's GLM-5.3 posted a cybersecurity benchmark score edging out Anthropic's Mythos 5, and the same week Alibaba's open models passed 3 billion downloads globally — evidence Washington can point to that the capability gap it is trying to legislate around is narrowing in public, not just in classified assessments. Anthropic co-founder Dario Amodei used a lengthy public post to defend his own policy positions on the same day, arguing that open-weight models will not meaningfully decentralize power and endorsing pre-launch vetting requirements for frontier systems — a stance that puts him closer to the export-control camp than to the "let capability diffuse" argument favored by much of the open-source community he sits adjacent to.

Eastern Front: Beijing's Benchmarks Get Harder to Wave Off

Z.ai's GLM-5.3 scored 84.5% on the CyberGym cybersecurity benchmark against Mythos 5's 83.8%, according to the company, though its most sensitive offensive-security capabilities remain restricted to verified users — a hedge that lets Z.ai claim parity on paper while avoiding the liability of publishing a fully unrestricted cyber-capable model. The claim arrives days after Taiwan's government disclosed it was targeted by what officials called an "abnormal," AI-assisted hacking campaign last month, with the Guardian and Financial Times both reporting China-linked involvement in what several outlets are calling an unprecedented use of autonomous AI agents in offensive cyber operations. Separately, chipmaker SMIC raised prices across its AI-chip lines after Q2 earnings beat estimates, with wafer shipments up 14% on what the company's co-CEO described as sustained demand outstripping supply — a rare case of a Chinese semiconductor firm exercising pricing power rather than discounting to win volume, and a signal that domestic chip capacity, not just model capability, is now a genuine constraint rather than a talking point. Apple, meanwhile, is reportedly training a China-specific large language model with Alibaba's technical support, a move that would make it the first major foreign company to offer a proprietary AI model inside China's regulatory perimeter rather than relying entirely on licensed domestic partners.

India Brief: TCS Tells Its Workforce What's Coming

Tata Consultancy Services launched ADD AgentHub on Monday, an enterprise agentic AI platform aimed at pharmaceutical clients for clinical-trial management and drug-safety workflows, built on what TCS calls a "human plus AI" model where agents execute defined tasks under human oversight and auditability requirements. The launch is not an isolated product move: TCS chair N. Chandrasekaran said separately this month that AI agents could match the company's human headcount within three years, and that TCS will hire fewer people going forward as agentic deployment scales. That is a striking admission from India's largest IT services employer, whose business model has depended for three decades on scaling headcount to match client demand. The tension shows up elsewhere in the country's AI economy: Goldman Sachs and Nomura both published research this month arguing India is less exposed to AI-driven job losses than other major economies, citing AI-related hiring that is currently outpacing displacement, while LinkedIn's CEO cited 51% annual growth in AI engineering roles in India. Whether that hiring growth holds once agentic platforms like AgentHub move from pilot to production at scale is the open question TCS's own chair is now raising in public. Separately, Sarvam AI closed a $74–75 million Series B extension with Nvidia as a strategic investor — the first Indian AI firm to bring Nvidia on as a named backer — while Reliance chairman Mukesh Ambani used his company's annual shareholder meeting to commit roughly $120 billion (₹10 lakh crore) toward AI infrastructure, insisting publicly that the investment "won't kill jobs."

The View

Three numbers from this week don't reconcile cleanly, and that's the story. Anthropic's IPO math assumes 2028 revenue four times its current run rate. Nvidia just cut its own compute-financing exposure to OpenAI in half. And TCS's chair is telling investors that agents could replace the equivalent of his entire workforce within three years — a workforce that exists specifically to sell the labor-augmentation services his own clients are buying instead of hiring. Read individually, each is a rational actor managing its own exposure: Anthropic's bankers need a growth story to justify a valuation, Nvidia needs to diversify counterparty risk after concentrating too much of it in one customer, and TCS needs to signal cost discipline to investors nervous about margin compression. Read together, they describe an industry where the companies closest to the compute layer are quietly hedging their bets on the very growth curve that the companies selling intelligence-as-a-service are asking public markets to underwrite at a trillion-dollar multiple. Somebody's forecast is wrong, and given who is hedging and who is still projecting straight-line growth, it's worth asking which side of that trade you'd rather be on before the roadshow starts.

The Miss

Two Financial Times items from this week's river landed with almost no follow-up: OpenAI's repeated executive reshuffles — including disbanding its "preparedness" safety function — are reportedly frustrating staff as the company prepares its own IPO, and a Washington Post investigation found nearly 40% of 2026 midterm races now feature explicit AI or data-center policy positions on candidate websites, the first cycle where that's been true. Both point to the same shift — AI governance moving from lab concern to electoral politics — but neither got the scrutiny a genuinely novel structural change deserves.

Pull Quotes

"It's difficult to see how a country can credibly position themselves as trusted partners in one technology ecosystem while simultaneously signing up for an initiative designed by China to advance a competing vision for AI." — unnamed US official, via Reuters

"You can't have it both ways." — State Department letter to 35 partner countries, on dual alignment with Pax Silica and China's rival AI initiative

Dario Amodei, defending his policy proposals in a public post: open weights "won't decentralize power," and pre-launch vetting for frontier systems is warranted regardless.

Anthropic's roadshow will spend a lot of time explaining a number nobody at Nvidia seems willing to underwrite at full size anymore.