NextFin News - An AI expert has called Anthropic's growth projections "fantasy," and the charge lands on the most important day of the year for the artificial-intelligence trade. Nvidia reports fiscal second-quarter earnings after the close, and the valuation structure of the entire AI buildout — from the chipmaker's roughly $5.2 trillion market capitalization to Anthropic's $965 billion private valuation — rests on a single question: can the revenue generated at the model layer ever justify the hundreds of billions being spent on the infrastructure layer?
The tension is unusually sharp because both sides of the argument are supported by hard numbers. Anthropic, the maker of Claude, has posted the fastest revenue ramp in modern technology history. Its second-quarter 2026 revenue exceeded $11.5 billion, a fourteenfold increase from the year-earlier period, according to documents the company showed to prospective investors. Its annualized revenue run rate has climbed from about $9 billion at the end of 2025 to more than $65 billion by mid-August, a sevenfold increase in roughly eight months. Yet the same company's chief executive, Dario Amodei, has warned that if AI growth forecasts are off by just one year, "then you go bankrupt." Both statements are true. The market is now being asked to decide which one matters more.
If AI growth forecasts are off by just a year, then you go bankrupt.
That warning, from Anthropic's own chief executive, is the clearest admission that the company's trajectory is a high-wire act rather than a straight line. It is also the starting point for the skepticism now being directed at the entire AI capital cycle.
The Arithmetic Behind the Skepticism
The skepticism is not aimed at whether Anthropic is growing. It is growing faster than almost any company ever has. It is aimed at the arithmetic that connects that growth to the capital being deployed in its name.
Combined capital expenditure across the major cloud and AI infrastructure providers is projected to exceed $690 billion in 2026, nearly double the prior year. Anthropic's own projections, first reported early this year, pointed to revenue of as much as $18 billion for 2026 and $55 billion for 2027. Even accepting the company's more recent run-rate figure above $65 billion, the model layer's revenue remains a fraction of the infrastructure being financed on its behalf. The gap is the story.
Anthropic's valuation math depends on that gap closing quickly. The company closed a $65 billion funding round in late May at a $965 billion post-money valuation, surpassing OpenAI as the most valuable artificial-intelligence startup. At that price, investors are underwriting a trajectory in which Anthropic's revenue continues to multiply for years without a meaningful slowdown, gross margins expand toward traditional software-company levels, and the company reaches positive free cash flow — a milestone management itself does not project until 2028 at the earliest. Internal projections still anticipate a loss of roughly $14 billion for 2026.
There is a further wrinkle that the skepticism seizes on. Technology press reporting has documented that both Anthropic and OpenAI have missed their gross-margin targets as inference costs climb faster than expected. Anthropic's 2025 gross margin came in near 40 percent, below the 50 percent the company had guided for a year earlier, after actual inference costs ran roughly 23 percent above forecast. Current guidance points to margins near 63 percent for this year and 70 percent next year — well below the 75 to 85 percent typical of traditional software companies, and under constant pressure from the cost of serving every additional token.
Why Nvidia Is the Real Barometer
Anthropic's projections matter less as a standalone investment case than as evidence for the broader AI thesis. If a model lab can grow revenue sevenfold in eight months, the argument goes, then the hyperscalers spending $690 billion on compute are not building cathedrals in the desert. They are building to meet demand that is already visible.
That is why all eyes are on Nvidia. Consensus calls for fiscal second-quarter revenue of about $92.2 billion and adjusted earnings of roughly $2.09 a share, which would represent revenue growth of about 97 percent year over year. For the full fiscal 2027 year, analysts expect revenue near $370 billion, up from about $216 billion in fiscal 2026 — growth of roughly 70 percent. Nvidia has already guided investors to expect $500 billion in combined revenue from its Blackwell and Rubin architectures through the end of 2026, and its finance chief, Colette Kress, has said there is "definitely an opportunity" to exceed that figure.
The mechanism that ties the two companies together is straightforward but unforgiving. Anthropic does not own the data centers. It rents compute, buys chips, and signs multi-year capacity agreements, including a reported $1.25 billion-per-month compute commitment running through 2029. Every dollar of Anthropic's revenue growth implies several more dollars flowing upstream to Nvidia, TSMC, and the cloud providers that own the physical infrastructure. The model layer captures the attention; the infrastructure layer captures the cash flow. That asymmetry is precisely what the skepticism targets: the revenue may be real, but the economics may accrue to someone else.
Cyclical Cadence, Structural Direction
The central analytical question is whether this is a cyclical surge that will revert or a structural regime change that will not. The answer is both, and they operate on different time horizons. Getting this wrong flips the conclusion.
The cyclical leg is the deployment wave. Every major technology platform cycle — the PC, the mobile internet, cloud computing — begins with a capital-intensity spike as the new infrastructure is built out ahead of demand. Capacity is ordered years in advance; utilization catches up later. If AI adoption slows even modestly, the industry finds itself with billions in underutilized graphics processors and stranded power commitments. That is the classic cyclical pattern, and it is mean-reverting by construction. The hyperscalers' capital expenditure is a leading indicator that will eventually roll over.
The structural leg is different. Agentic AI — systems that can execute multi-step tasks rather than merely produce text — appears to be expanding the total addressable market rather than redistributing existing software spending. Anthropic's own estimate of its addressable market exceeds $30 trillion, above the $28.5 trillion figure SpaceX presented before its public listing. If autonomous agents become the interface through which enterprises run workflows, the software layer's economics could eventually rival the infrastructure layer's. That would be a genuine regime change: value migrating from the picks and shovels to the intelligence running on top of them.
The evidence floor matters here. A cyclical claim requires demonstrated mean reversion, and there is none yet — utilization remains tight, lead times remain long, and pricing power remains with the suppliers. A structural claim requires proof that demand is durable rather than speculative, and that proof is still being written quarter by quarter. The honest call is that the infrastructure buildout is cyclical in its cadence but structural in its direction. The capex wave will crest; the installed base of AI systems will not shrink back to zero.
The Second-Order Question Nobody Is Asking
The first-order question is whether Anthropic hits its numbers. The second-order question is what happens if it does. Suppose Anthropic reaches $100 billion in revenue by 2028, exactly as its most optimistic backers project. At a $965 billion valuation today, that is still a price-to-sales multiple near ten on forward revenue — expensive, but not absurd for a company growing this fast. The problem is that the denominator keeps moving.
Every percentage point of Anthropic's growth requires additional capital expenditure from its suppliers. The more successful the model layer becomes, the more bargaining power accrues to the scarce inputs: advanced packaging capacity at TSMC, high-bandwidth memory from SK Hynix and Micron, power and land for data centers. Revenue grows, but the margin may not grow with it. This is the second-order trap: success at the model layer intensifies competition for the infrastructure layer, and competition is how excess returns get arbitraged away.
There is a further transmission channel the market has not fully priced. If Anthropic and its peers prove that agentic AI substitutes for human labor at scale, the ultimate customer for AI is not the enterprise software budget — it is the global wage bill. That expands the addressable market by an order of magnitude, but it also invites a regulatory and political response that could cap pricing, mandate data retention, or restrict deployment in sensitive sectors. Anthropic has already clashed with the U.S. government over model guardrails and has moved to let enterprise customers retain their data on their own infrastructure. The commercial progress is real; the friction is real too.
The Strongest Counter-Thesis
The bull case deserves a serious hearing, because it is not built on faith alone. Investors such as Gavin Baker of Atreides Management have argued that Anthropic is the most credible long-term candidate to eventually surpass Nvidia as the world's most valuable company, on the view that AI economics will ultimately concentrate in the model intelligence rather than the compute layer. Venture investor Tomasz Tunguz has placed the crossover somewhere between seven months and three years. The evidence they point to is the revenue curve itself: no company has ever scaled from $9 billion to more than $65 billion in annualized revenue in eight months.
The counter-thesis has one fatal vulnerability: it assumes the margin structure of software without the cost structure of software. Traditional SaaS companies achieve gross margins of 75 to 85 percent because distributing another copy of software costs almost nothing. Frontier AI is the opposite — every additional token consumed costs electricity, memory, and silicon. Anthropic's guidance points to gross margins near 63 percent for this year and 70 percent next year, below traditional software and under constant pressure from inference costs. The company has already missed an earlier margin forecast when actual inference costs ran roughly 23 percent above guidance.
The falsifying signal is specific and observable. If Anthropic's gross margin does not reach at least 70 percent within four quarters while revenue continues to grow above 50 percent annually, the software-multiple valuation framework breaks down. The company would be growing like software but earning like a capital-intensive utility, and the $965 billion valuation would need to be repriced against infrastructure-like multiples rather than software-like ones. Watch the margin, not the revenue. Revenue can grow forever and still destroy value if the cost of producing it grows faster.
What To Watch Next
The immediate catalyst is Nvidia's earnings report after the close. A beat-and-raise quarter would reinforce the demand-side evidence and temporarily quiet the skepticism. A guide that disappoints, or any hint that hyperscaler customers are moderating their capital expenditure, would do the opposite — because Anthropic's projections are only as credible as the infrastructure spending that underwrites them.
Over the medium term, three data points will decide the debate. First, Anthropic's free-cash-flow trajectory toward its 2028 positivity target. Second, the gross-margin path against the 70 percent threshold named above. Third, the hyperscalers' capital-expenditure guidance for 2027 — any deceleration would signal that the cyclical leg of the buildout is cresting even if the structural leg continues.
Scenarios, not a single line. In the base case, Anthropic prices its public listing between $1.5 trillion and $2 trillion, Nvidia continues to beat conservative guidance, and the capex wave extends through 2027 with a gradual deceleration. In the upside case, agentic AI drives enterprise adoption faster than expected, margins expand toward 75 percent, and the model layer captures a durable share of the software budget. In the downside case, inference costs stay stubbornly high, gross margins stall in the low 60s, and the market re-rates frontier AI companies from software multiples toward infrastructure multiples — a compression that could erase hundreds of billions of paper valuation even while revenue keeps climbing.
The closing judgment is this: Anthropic's growth is not a fantasy, but the valuation built on top of it assumes a margin structure the company has not yet earned. The revenue is real; the question is who keeps the profit. Until the model layer proves it can capture margins like software rather than incur costs like infrastructure, the safest position in the AI trade remains the one that owns the scarcity — the chips, the packaging, and the power — not the one that rents it.
Explore more exclusive insights at nextfin.ai.
