Anthropic filing reveals R&D consumes majority of compute spend

Anthropic’s S-1 filing indicates that research and development consumed the largest share of its $8.9 billion compute expenditure in the first half of the year.reported Customer inference accounted for less than a third of that total, according to the document.reportedThe document shows most money spent on computing power went to research, not running the chatbot for users.The S-1 reveals that inference costs are now profitable, but operating losses persist due to heavy R&D and upcoming hardware capex.

The evidence

The filing details three primary cost categories: revenue costs for inference, R&D for training and research, and sales and marketing. Inference costs for serving customers reached $2.4 billion against $16.3 billion in revenue for the period.reported This structure allows the company to retain an estimated 85 cents for every dollar spent after covering serving costs.reported By comparison, the previous year saw only 12 cents retained per dollar.

R&D expenses increased fivefold year-over-year, with 66 percent or more of compute directed toward training and internal use.reported This heavy investment is cited as the main driver of the operating loss. Sales and marketing costs rose sixteenfold, primarily due to partner fees. These deductions reduce the retained amount to 66 cents per dollar.

The decision

Revenue generation per compute dollar improved significantly, jumping from $0.40 in the first quarter of 2025 to $2.08 in the second quarter of 2026.reported The company attributes this efficiency to capacity purchased 12 to 18 months in advance. However, future hardware commitments present a financial challenge. Five compute deals with the largest stated amounts would cost an estimated $103 billion a year once delivered.reported

This projected annual cost is more than double the revenue pace of roughly $46 billion recorded in the second quarter.reported The filing warns that margins may compress as new hardware lands before the revenue it supports arrives. Developers must consider this gap between current inference profitability and future capital intensity.

The takeaway

Serving Claude currently covers its direct costs, but the broader business model relies on scaling revenue to match massive hardware orders. The shift from a 12-cent margin to 85 cents on serving costs shows improved operational leverage.reported Yet, the five major contracts create a high fixed-cost baseline that could strain margins if revenue growth stalls.

Building on this

Developers can leverage Claude’s inference capabilities knowing the underlying serving costs are currently covered by revenue. The documented 85-cent retention per dollar suggests stable pricing for API usage in the short term. However, the warning about margin compression implies potential cost adjustments as new hardware comes online. Builders should design applications that are resilient to future price fluctuations in compute resources.