Part of: this research — view all 4 rows
Anthropic manages its compute as a single fungible pool and publishes no physical allocation Anthropic Margins, Compute Spend, and Losses: S-1 Filing Anthropic, PBC - S-1 (Q2). The best available proxy is its S-1 financial accounting, which separates compute into cost of revenue (substantially all for customer-serving workloads) and R&D (model development, research, and internal use). In the first nine months of 2026, compute and infrastructure cost $18.8B, of which cost of revenue was just $4.8B Anthropic, PBC - S-1/A-1 Anthropic 85% Gross Margin: Serving vs Training Compute. This means paid customer serving accounted for only about 26% of compute spending, leaving the vast majority for R&D and training.
This relatively low spending share for serving might seem at odds with Anthropic's staggering $81B annualized revenue run-rate reported at the end of October. However, 44% of Anthropic’s 2026 revenue is derived from cloud-marketplace sales, which are booked gross Anthropic's journey from DRS to S-1: Zeroing in on accounting .... The token volume—and therefore the physical inference load—per revenue dollar is significantly lower than the headline figure implies. While we must adjust external serving upward to account for free-tier usage (booked under sales and marketing Anthropic 85% Gross Margin: Serving vs Training Compute) and the fact that serving FLOPs often run at higher utilization than R&D dollars, external inference is still unlikely to dominate the total compute pool to the extent typical of mature commercial platforms.
Within the remaining R&D compute, research experiments must be separated from final training runs. Historically, final training runs have represented a minority of frontier R&D compute; Epoch estimated they made up just 9.6% of OpenAI’s R&D compute in 2024 Final training runs account for a minority of R&D compute ... Final training runs account for a minority of R&D compute spending. However, the growing compute intensity of continuous reinforcement learning (RL) post-training for recent models like Opus 5.5, Fable 5.2, and Mythos 5.2 likely pushes the final-training share significantly higher than those 2024 baselines. Offsetting this is the massive scale of Anthropic's internal agentic R&D. Agent hours reached six times human hours by August, and the list-price value of internal Claude usage has exceeded monthly personnel expense since April https://www.businessinsider.com/anthropic-claude-internal-ai-use-exceeds-employee-pay-tokens-2026-11 Anthropic, PBC - S-1/A-2. This internal AI-driven research firmly belongs in the experiment bucket and keeps its share of total compute substantial. There is a high probability that aggressive allocations to internal AI research also constrain final training run sizes: within a shared compute pool, growing post-training demands need not translate into a proportionately larger final-training share when internal research competes for the same capacity.
These dynamics suggest a central estimate of 45% for research experiments, with external inference and final training splitting the remainder. I place the 25th and 75th percentiles at 35% and 55%, and the 10th and 90th percentiles at 25% and 65%. These wide bounds reflect severe measurement ambiguities. Expense data maps poorly to physical operations: Anthropic spent $5.0B in Q3 on short-term compute that "costs significantly more" Anthropic, PBC - S-1/A-2, and if this premium capacity went primarily to R&D, dollars would overstate actual FLOPs. Furthermore, the arrival of massive new capacity in December—such as the $3.1B monthly SpaceXAI lease—could skew the allocation sharply if it is front-loaded for a major year-end training run or dedicated entirely to experimental R&D ahead of future revenue generation.
Set against related questions, the estimate was nudged upward to reflect the high probability of aggressive internal AI research allocations constraining final training run sizes.
Ask a followup
Sign in to run · $20 free credit, no card · every claim cited