Anthropic's Compute Margins Narrow as Short-Term Deals Eat Revenue

Revenue Grew, but Compute Ate More of It

Anthropic disclosed on 3 November that it generated $17 billion in revenue for the quarter ending 30 September 2026, a 48 per cent jump from $11.5 billion the previous quarter. Yet the company's operating loss widened in the same period. The culprit was compute: infrastructure expenses climbed to $10 billion, claiming 58 per cent of revenue compared with 48 per cent the quarter before.

The updated S-1 filing extended financial disclosures beyond the 30 June cut-off in the initial document filed last week. For a company preparing to go public, the trajectory is uncomfortable. Where investors typically want to see margin expansion as scale arrives, Anthropic is moving in the opposite direction. A chart in the filing tracks revenue per dollar of compute spend; it peaked above 2x earlier in 2025, then slid to 1.72x in the September quarter.

At Opentechwire, we have tracked similar disclosure patterns from AI labs as they approach public markets, and the compute margin question consistently surfaces as the single largest variable in valuation models. Anthropic's numbers make plain why: infrastructure is not just the largest line item but one that resists the economies of scale investors expect from software businesses.

The Neocloud Premium

The filing attributes much of the margin compression to short-term capacity commitments. Anthropic spent $5 billion on compute procured under such arrangements during the quarter, "which currently costs significantly more than comparable compute procured under our multi-year commitments," the company noted.

Demand for Claude models outpaced the capacity Anthropic had locked in under longer-term contracts with hyperscale cloud providers. To keep inference latency acceptable and avoid turning away customers, the lab turned to what the industry calls neoclouds: smaller infrastructure vendors such as Nscale, Akamai, and Volta that offer flexible terms at higher rates. Anthropic has also been paying SpaceX $1.25 billion monthly for compute, a deal that became public earlier this year.

These vendors occupy a distinct niche. Hyperscalers such as Amazon Web Services, Microsoft Azure, and Google Cloud require long lead times and multi-year commitments in exchange for lower per-unit pricing. Neoclouds can spin up capacity in weeks rather than quarters, but charge a premium that can exceed 50 per cent. For a lab racing to capture market share whilst models remain differentiated, that premium is the cost of staying in the game.

The result is a two-tier compute architecture: cheaper baseline capacity from hyperscalers, topped up with expensive spot capacity from neoclouds when usage spikes. Anthropic's margin deterioration in the September quarter reflects the latter category growing faster than the former.

Anthropic Direct and the Scale Problem

Anthropic has begun building its own data centre capacity under an initiative it calls Anthropic Direct. The goal is to reduce reliance on both hyperscalers and neoclouds by owning more of the stack, from server procurement to facility management. The company discusses the effort at length in the filing, framing it as a path to better unit economics over time.

Yet the same filing makes clear that demand is growing faster than Anthropic can build. "We expect compute and infrastructure expenses to increase in absolute dollars and remain our largest operating expense for the foreseeable future," the company stated, citing both new multi-year arrangements with cloud service providers and continued investment in additional capacity.

Building data centres is a multi-year process involving site acquisition, power contracts, cooling infrastructure, and hardware procurement, all of which require lead times measured in quarters or years. Fine-tuning a large language model or scaling inference to meet a traffic surge happens on a weekly cycle. The mismatch means that even as Anthropic Direct comes online, the lab will continue buying expensive short-term capacity whenever usage outpaces owned infrastructure.

This is not unique to Anthropic. Across the frontier labs we have followed in Asia and North America, the pattern repeats: in-house infrastructure projects are strategic necessities but operational luxuries, useful for baseline load but insufficient for peak demand. The neocloud layer persists because it solves a timing problem that capex alone cannot.

What Slowing Buildout Would Mean

The filing's tone on compute is cautiously optimistic, but the risks are structural. Anthropic depends on a trillion-dollar global infrastructure expansion continuing at pace. If bond markets tighten and make data centre financing more expensive, or if local opposition delays permitting in key markets, the neocloud premium could widen further or capacity could simply become unavailable at any price.

Neither risk is far-fetched. Data centre projects in Northern Virginia, Singapore, and parts of Europe have already faced permitting delays due to grid capacity concerns and community opposition. Higher interest rates have made sale-leaseback financing for new facilities more expensive, slowing the pace at which neoclouds and hyperscalers alike can add capacity.

Anthropic has limited leverage over these variables. The lab can optimise inference efficiency, negotiate better terms with existing providers, and accelerate its own buildout, but it cannot will new power substations into existence or override local zoning boards. The compute layer sits beneath the business model, and turbulence there propagates upward.

Champagne Problems, For Now

The company is growing revenue at nearly 50 per cent quarter-on-quarter, a pace that overshadows most margin concerns in the eyes of growth-focused investors. If demand were to slow, the compute picture would brighten immediately; Anthropic could shed expensive short-term contracts, let Anthropic Direct absorb a larger share of load, and watch margins expand. The problem is only a problem because the business is succeeding.

Yet the filing offers a glimpse of how precarious that success is. AI labs do not control their primary input cost, cannot easily substitute suppliers when prices rise, and face demand volatility that makes long-term capacity planning nearly impossible. The result is a business that scales revenue quickly but struggles to scale margins at the same rate.

For Anthropic, the next few quarters will test whether the neocloud reliance was a temporary spike driven by unexpectedly strong Q3 demand, or a structural feature of operating at the frontier. The answer will shape not only the company's IPO valuation but the broader investment case for AI labs as a category. At current growth rates, compute costs remain manageable. If growth stalls whilst infrastructure expenses stay elevated, the economics become far less forgiving.

The filing does not resolve that uncertainty. It simply makes visible what was previously opaque: the degree to which Anthropic's fate is tied to variables it does not control, in markets where supply cannot easily flex to meet demand. For now, the lab is paying the premium and growing into it. How long that remains viable depends on forces well outside its San Francisco headquarters.