Anthropic’s Compute Costs Surge as IPO Approaches

Anthropic disclosed through an updated securities filing on Monday that the company’s compute economics deteriorated in the third quarter of 2026, even as revenue surged nearly 50 percent. The new filing extended financial disclosures through September 30th, providing the first comprehensive look at how Anthropic’s surging revenue translates into actual profitability—and revealing a widening gap between what the company takes in and what it spends on computing infrastructure.

The numbers paint a picture of rapid growth tempered by mounting costs. Anthropic generated $17 billion in revenue during the third quarter, compared to $11.5 billion in the second quarter. Yet operating losses expanded at the same time, a rare divergence that points to a single culprit: compute expenses.

The Widening Compute Burden

Revenue Being Consumed by Infrastructure

In the most recent quarter, compute costs consumed 58 percent of Anthropic’s revenue, up sharply from 48 percent in the previous quarter. That 10-percentage-point shift in a single quarter underscores how aggressively the company is scaling its infrastructure to meet demand for its Claude models.

The company spent $10 billion on compute and infrastructure in the third quarter alone. Breaking down that figure further reveals the pressure Anthropic faced: $5.0 billion of that total came from compute procured under short-term commitments, which the filing notes currently costs significantly more than comparable compute procured under multi-year commitments. In other words, unable to negotiate longer contracts at favorable rates because demand was growing too quickly, Anthropic was forced to pay premium prices for immediate capacity.

The Declining Return on Compute Spending

A chart included in the filing tracked a metric critical to understanding Anthropic’s path to profitability: revenue for every dollar spent on compute. Since the start of 2025, this figure had been trending upward, peaking slightly above 2x in mid-2026. By the third quarter, it had fallen back to 1.72x, suggesting that each additional compute dollar was generating proportionally less revenue than it had earlier in the year.

For a company preparing for an initial public offering, this deterioration in margins presents a challenge to investors who typically want to see improving operational efficiency, not widening cost structures.

Why Margins Collapsed: Scrambling for Capacity

Read also: Apple’s Audio Intelligence begins Apple Watch beta testing · Squarespace Cuts 300 Jobs as AI Builders Reshape Website Market · A.I. Foundation Models Trigger Second Autonomous Vehicle Boom

The Shortage of Available Compute

Anthropic’s operating environment has been shaped by a scarcity that surprised industry observers. The company’s demand for computing power grew faster than either traditional cloud providers or newer infrastructure vendors could accommodate through conventional long-term contracts. The practical result was reliance on what the industry calls neoclouds—smaller infrastructure providers offering more flexible arrangements than hyperscalers like AWS or Google Cloud. Flexibility comes at a cost. These providers charge significantly higher rates than the volume discounts available to companies committing to multi-year arrangements.

The Partnerships Driving the Costs

Anthropic’s compute partnerships reflect the scramble for capacity. The company has committed to paying SpaceX $1.25 billion per month for compute capacity, part of a broader infrastructure strategy that includes multi-billion-dollar agreements with Nscale, Akamai, and Volta. Each of these deals represents a partial solution to an immediate problem: the company needed more computing power than traditional providers could deliver on terms that Anthropic could negotiate at scale.

These partnerships carry a message about Anthropic’s negotiating position. The company is not choosing these arrangements because they offer superior technology or innovation. It is choosing them because capacity exists and is available immediately, even if the price is substantially higher than long-term commitments would be.

Anthropic’s Path Forward: Direct Infrastructure

Building Proprietary Compute Capacity

The company is pursuing a project called Anthropic Direct designed to build and operate its own compute infrastructure. This effort represents a long-term bet that vertical integration could eventually reduce dependency on external providers and improve margin dynamics. However, the filing makes clear that this strategy faces a timing problem. Building proprietary data center capacity takes months or years, while demand from customers and internal training requirements continues to accelerate quarter to quarter.

The company’s own assessment, contained in the filing, captures this tension: we expect compute and infrastructure expenses to increase in absolute dollars and remain our largest operating expense for the foreseeable future as additional committed compute becomes available to us under multi-year arrangements, and as we continue to invest in additional compute. In other words, compute costs will keep rising in raw dollars for years, even if Anthropic succeeds in negotiating better rates as multi-year commitments mature.

The Timeline Challenge

Building infrastructure that scales to Anthropic’s current demand levels is not a months-long project. The company faces a race between its internal infrastructure buildout and the pace at which existing partners can deliver improved terms. If this race tilts toward the latter, Anthropic will depend on neoclouds longer than management would prefer.

Temporary Crisis or Structural Problem

Several factors suggest the recent margin compression may be temporary rather than indicative of a permanent shift in Anthropic’s economics. Revenue is still growing at roughly 50 percent per quarter. As demand moderates and Anthropic’s negotiating position improves with its larger scale, the company should be able to lock in multi-year commitments at lower rates than the short-term premium prices it is currently paying. The neoclouds that consumed such a large share of Q3 spending may represent a one-time cost of keeping up with demand spikes.

Yet the filing also reveals genuine uncertainty about Anthropic’s ability to control costs going forward. The company’s success depends not just on its own infrastructure strategy but on the broader buildout of global data center capacity. If bond markets tighten and capital becomes scarce, if local opposition to new data centers slows projects, or if infrastructure providers themselves face margin pressure, Anthropic has limited tools to manage the fallout.

The Compute Supply Chain’s Structural Risk

For all of Anthropic’s operational successes—it has built competitive models, secured committed customers, and achieved rapid revenue growth—the company remains dependent on a trillion-dollar infrastructure ecosystem it does not control and can barely influence. Its supply chain is not the traditional kind that companies might diversify or hedge. Compute capacity in the regions where Anthropic operates is concentrated among a handful of providers, and adding new capacity takes years.

A significant slowdown in global data center construction, whether triggered by macroeconomic shifts or regulatory action, would put acute pressure on Anthropic’s already strained compute budgets. The company acknowledges this risk implicitly by emphasizing that its success depends on additional committed compute becoming available under future contracts. That compute must exist before Anthropic can purchase it.

The disclosure of these economics, while uncomfortable in the near term, may ultimately reassure investors about one thing: Anthropic’s revenue growth is real and driven by genuine customer demand, not by unsustainable margin expansion. But it also sends a message that competing at the frontier of AI capabilities is fundamentally a capital-intensive business dependent on infrastructure growth. For Anthropic and other labs, the next five years of financial performance will be written less by business strategy than by the pace of global data center construction and the availability of semiconductors and power.

Source: TechCrunch