NextFin News - Anthropic is about to tell the market a story in three numbers: $16.3 billion of revenue in the first half of 2026, an $80.7 billion net loss over the same period, and roughly $750 billion of computing capacity it plans to buy in the years ahead. Read separately, each figure is staggering. Read together, they frame the central question of the artificial-intelligence boom: is this a company scaling into a durable franchise, or a company spending its way toward a bet that demand never stops compounding?
The answer matters because Anthropic's planned initial public offering is on track to become one of the largest listings in history, with a valuation target approaching $2 trillion. The company's draft IPO prospectus, filed with U.S. regulators and disclosed in October 2026, lays bare the economics of building a frontier AI lab at the edge of what capital markets have ever been asked to fund. This is not a story about whether Claude is a good product. It is a story about whether the math of AI infrastructure can ever produce a return on capital that justifies the price investors are being asked to pay.
Number One: Revenue Growing Faster Than Almost Anything at This Scale
The first number is the one Anthropic wants investors to lead with. Revenue in the first half of 2026 reached $16.3 billion, up from $1.1 billion in the same period a year earlier — a fourteen-fold increase that already exceeds the $4.6 billion the company generated in all of 2025. By the end of July, the company's annualized revenue run rate had crossed $65 billion, more than seven times its roughly $9 billion pace at the end of 2025 and well ahead of the $47 billion it was running in May. In the third quarter alone, revenue reached $16.8 billion, up 46% from the prior quarter, and the company told investors it had reached profitability on an adjusted basis during that period.
Few companies in history have grown this fast at this scale. The trajectory implies that Anthropic moved from a niche research lab into a genuinely large enterprise-software business inside roughly eighteen months. The driver is straightforward: every major cloud provider and large technology company needs frontier-model capacity, and Anthropic, alongside OpenAI, controls one of the two model families that enterprise customers trust enough to build mission-critical systems on.
But the composition of that revenue matters as much as the total. Nearly half of first-half revenue — 48% — flowed through cloud marketplace sales from Google, Amazon, and Microsoft, up from 11% in 2023, according to the filing. Roughly one-fifth of first-half revenue came from just two customers. That concentration is the flip side of hypergrowth: when your buyers are the hyperscalers themselves — the cloud platforms that both distribute your model and compete with you — your growth is real, but your bargaining power is not what a diversified software vendor's would be. A single renegotiation with a cloud partner can move the revenue line by billions.
The growth also rests on a metric that rewards optimism. Run rate annualizes current sales; it assumes the present moment repeats for twelve months. In a market where AI spending is front-loaded and contract terms are still being renegotiated at pace, run rate is a direction indicator, not a guarantee. The $65 billion figure tells you where demand is pointing, not where GAAP revenue will land.
Number Two: An $80.7 Billion Loss That Is Mostly an Accounting Mirror
The second number is the one that will dominate headlines. Anthropic reported a net loss of $80.7 billion in the first half of 2026, up from $13 billion a year earlier. On its face, losing five dollars for every dollar of revenue looks like a business model that cannot exist outside of cheap capital.
It is not that simple. Roughly $72 billion of the loss — nearly nine-tenths — is a non-cash accounting charge reflecting the rising fair value of bonds convertible into Anthropic shares. In other words, the "loss" is largely the market's own verdict that the company is worth more today than it was when it issued the debt. The liability on the balance sheet grows as the valuation rises, and accounting rules require that increase to flow through the income statement. The higher the market thinks Anthropic is worth, the bigger the reported loss.
This is the paradox at the heart of the IPO: the very metric that makes the company look unprofitable is, in large part, a mirror of the optimism that makes the offering possible. Strip out the charge, and the operating picture looks materially different — though still loss-making, because the third number exists.
The real cash cost is compute. Anthropic spent nearly $9 billion on computing capacity and infrastructure in the first half of 2026, up from $2.3 billion a year earlier — almost four times the year-ago pace and already ahead of the $7.3 billion it spent in all of 2025. That is a real outflow, and it is the engine of everything else: without it, there is no model improvement, no capacity to serve customers, and no revenue. The question is whether each incremental dollar of compute buys a proportional dollar of durable revenue, or whether the industry is trapped in a race where the finish line moves every time someone builds a faster car.
Number Three: $750 Billion of Compute — The Bet That Demand Never Stops
The third number is the one that defines the risk. Anthropic has signaled it could spend up to $750 billion in the coming years securing computing capacity, including deals still under negotiation. Of that total, $450 billion is already locked in as non-cancelable cloud commitments. Against those obligations, the company held $131.7 billion in cash and short-term investments at the end of June — more than Nvidia and Apple combined, and behind only Berkshire Hathaway and Alphabet among non-bank U.S. companies.
Do the arithmetic and the asymmetry is stark. Cash covers only a fraction of the pledged and planned compute obligations. The company is committing to spend multiples of its cash pile on the assumption that revenue keeps compounding fast enough to fund the difference. This is not a balance sheet designed for a downturn; it is a balance sheet designed for a world in which AI demand grows without interruption for years.
The specific commitments show how deeply Anthropic is bound to the three cloud giants. The filing shows at least $114.5 billion committed to Google through 2033, $110 billion to Amazon through 2036, and $44.9 billion to Microsoft for computing power — while those same companies have taken equity stakes of roughly $23 billion and $18 billion respectively, with billions more conditional on performance targets. The filing itself describes the cloud providers in a single phrase that captures the entire tension of the business:
investors, customers, distribution partners and competitors
That is why the three numbers cannot be read in isolation. The $16.3 billion of revenue justifies the spending only if growth persists. The $80.7 billion loss is tolerable only if it is mostly accounting. And the $750 billion compute plan is rational only if the revenue curve keeps outrunning the cost curve. Break any one of those links and the story changes.
The Second-Order Question: Who Actually Captures the Value?
The conventional reading of Anthropic's numbers is that AI demand is insatiable and the company is simply building the capacity to meet it. The second-order question is different: in a stack where the model layer is a duopoly but the infrastructure layer is a triopoly, who captures the surplus?
The cloud providers — Amazon, Microsoft, and Google — finance Anthropic's compute, distribute its models, and sell the inference capacity back to customers. They are simultaneously Anthropic's customers, its landlords, and its competitors. If AI economics compress — if pricing per token falls faster than compute costs — the model layer absorbs the margin pressure first, while the infrastructure owners still collect rent on the chips. Anthropic's $750 billion commitment is, in effect, a leveraged bet that the model layer retains enough pricing power to stay ahead of that squeeze.
There is evidence the squeeze is already beginning. Anthropic recently introduced a lower-cost model variant aimed at price-sensitive enterprise workloads, a move that signals competitive pressure on pricing as well as performance. Meanwhile, estimates of its main rival's annualized revenue have been revised down, suggesting the growth curve may be less uniform across the sector than the headline numbers imply.
The Counter-Thesis: This Is Not a Bubble, It Is a Land Grab
The strongest case against the skeptical read is that every transformative technology looks absurd on early financials. Amazon lost money for years while building the infrastructure that would dominate e-commerce and cloud computing. The question is not whether Anthropic is profitable today; it is whether the AI buildout is a one-time capital expenditure or a permanent toll road.
Proponents argue that once the capacity is built, the marginal cost of inference falls, operating leverage kicks in, and the $750 billion becomes a defensible moat rather than an anchor. They point to the adjusted profitability the company reported in recent quarters as evidence that the underlying business, excluding accounting charges and front-loaded capacity investment, is already working. Wall Street is already valuing the company on that premise: with a $2 trillion target resting on a 2028 revenue projection of $190 billion to $200 billion, investors are pricing Anthropic roughly two years into a future that has not happened.
This argument is coherent, but it rests on a single assumption: that demand growth does not decelerate before the capacity comes online. Compute orders are placed twelve to eighteen months before delivery. If revenue growth slows — as the sequential deceleration in the third quarter hints it might — Anthropic could find itself paying for capacity it does not need, at the exact moment pricing power is eroding. That is the classic capital-cycle trap: the industry builds for the demand that existed when the order was placed, not the demand that exists when the chips arrive.
What to Watch: The Signal That Would Break the Story
The three-number story holds only if revenue growth stays ahead of compute cost growth. The falsifying signal is specific: if Anthropic's revenue growth decelerates below the pace implied by its $190 billion to $200 billion 2028 projection while non-cancelable compute obligations remain near $450 billion, the capital-cycle thesis breaks. A more immediate tell is customer concentration — if the share of revenue from the top two customers rises rather than falls as the base grows, the diversification story is failing even as the top line expands. A third tell is the cloud partners themselves: if Google, Amazon, or Microsoft begin pulling back on conditional equity tranches or marketplace commitments, the funding model that underwrites the $750 billion plan comes under pressure.
Scenarios split cleanly by horizon. In the short term, sentiment will track the IPO pricing and the initial lock-up dynamics. In the medium term, the base case is continued rapid growth with losses that narrow as a share of revenue but remain large in absolute terms. The upside case is that inference costs fall faster than pricing, turning the model layer into a high-margin utility. The downside case is a growth deceleration that leaves the company committed to hundreds of billions of capacity it cannot profitably fill.
This is a structural shift, not a cyclical wave — and that is precisely what makes it dangerous. The AI buildout is a regime change in how technology consumes capital, and it will not revert on its own. But structural shifts do not guarantee structural returns. The capital cycle has a long memory for industries that overbuilt ahead of demand, and the distance between "the future is real" and "this stock was a good investment" is measured in the return on every dollar of compute deployed.
Data as of October 30, 2026, based on Anthropic's draft IPO prospectus and company disclosures.
Anthropic's three numbers tell a single story: this is not a company being valued on what it earns, but on what it convinces the market it will never stop becoming. The bet is not that AI is real. The bet is that the curve stays exponential long enough for the balance sheet to catch up.
Explore more exclusive insights at nextfin.ai.