Anthropic is still grappling with the economics of compute

Introduction

Anthropic’s latest quarterly filing reveals a growing tension between its impressive revenue growth and the escalating cost of the compute power that fuels its models. While the company posted $17 billion in revenue, compute expenses now consume 58 % of that total, up from 48 % a quarter earlier. This shift underscores the broader challenge AI labs face: turning massive data‑center spend into sustainable margins as they race to meet exploding demand.

Compute Costs Outpacing Revenue Gains

The filing shows that Anthropic’s compute bill jumped to roughly $10 billion for the quarter, driven largely by short‑term contracts that carry higher rates than the multi‑year deals the firm prefers. The company’s reliance on “neocloud” providers such as SpaceX, Nscale, Akamai and Volta—smaller, more flexible but pricier than traditional hyperscalers—has amplified this effect. These contracts allow Anthropic to secure capacity quickly, but they also inflate the cost per unit of compute, pushing the compute‑to‑revenue ratio above the 50 % threshold that analysts consider a warning sign.

Why the Compute Margin Is Worsening

Three factors explain the deteriorating compute margin:

1. Rapid demand growth – Anthropic’s product pipeline and partnership announcements have spurred a near‑50 % quarter‑over‑quarter revenue increase, forcing the firm to acquire additional compute on short notice.
2. Higher short‑term rates – The bulk of the new spend is tied to contracts that lack the volume discounts of longer‑term agreements, inflating per‑GPU costs.
3. Infrastructure lag – Anthropic’s internal build‑out, dubbed “Anthropic Direct,” is still in early stages. Until it can supply a meaningful share of the required compute, the company must rely on external providers.

The Path Forward: Balancing Growth and Efficiency

Anthropic acknowledges that compute will remain its largest operating expense for the foreseeable future. The filing notes a strategic focus on two fronts:

• Securing multi‑year commitments with cloud service providers to lock in lower rates as the market stabilizes.
• Accelerating the Anthropic Direct initiative to bring more of the compute stack in‑house, reducing dependence on external pricing volatility.

Investors appear tolerant of the current cost structure because revenue is still expanding at a robust pace. However, the margin pressure serves as a reminder that AI labs are heavily dependent on the broader data‑center ecosystem. Any slowdown in the trillion‑dollar infrastructure build‑out—whether due to tighter credit markets, regulatory hurdles, or local opposition to new facilities—could tighten supply and push prices higher, squeezing margins further.

Conclusion

Anthropic’s experience illustrates a classic scaling dilemma for AI companies: the need to match soaring demand with equally rapid compute capacity, all while keeping unit costs in check. As the firm pushes toward an IPO, its ability to transition from short‑term, high‑cost contracts to a more balanced mix of long‑term agreements and internal infrastructure will be a key determinant of long‑term profitability. For the AI industry at large, the story reinforces the importance of strategic compute planning—without it, even the most promising revenue trajectories can be undermined by mounting infrastructure expenses.