rationale:

This assessment estimates the true operational compute allocation at Anthropic as of December 31, 2026, specifically the share dedicated to "research experiments." Under this three-way taxonomy, this bucket captures all internal compute that is neither external inference serving nor the final training runs of released models. This includes ablations, scaling and derisking runs, synthetic data generation, reinforcement learning environment rollouts, and internal agentic inference workflows.

The strongest downward pressure on this share comes from Anthropic's massive commercial deployment scale. The company's annualized revenue run rate reached roughly $65B by mid-2026, driven by enterprise traction and Claude Code adoption 2 sources, with significant capacity agreements tied to serving demand anthropic.com. However, high inference revenue does not mechanically dictate compute-share dominance. Inference gross margins have reportedly improved to ~70% klover.ai, meaning direct serving compute consumes a much smaller fraction of total costs than revenue figures imply. While external serving likely constitutes the largest single compute bucket, massive ongoing capital investments in training fleets and reported R&D-to-inference compute ratios ensure R&D retains a substantial share of total capacity 2 sources.

Within that R&D allocation, final training runs are historically a small minority of total compute. Benchmarks from other leading labs suggest final runs consume roughly 10% to 23% of R&D compute (e.g., Epoch estimates 9.6% for OpenAI in 2024) epoch.ai. The remainder lands in the research experiments bucket. Furthermore, Anthropic explicitly notes that its internal R&D is highly automated, with Claude authoring a large majority of merged code and persistent agent deployments directly burning substantial experimental compute 4anthropic.com. This heavy internal reliance on AI loops and automated research agents structurally inflates the non-training R&D compute load.

Synthesizing these dynamics, and weighing expected enterprise API market shares against automated AI research trends, balances massive external serving loads against the growing resource demands of internal agentic workflows to yield a central estimate of 37%. Assuming R&D accounts for roughly 40% to 45% of total compute and final training runs consume a small-to-moderate slice of that, research experiments naturally settle around this mark. The broad uncertainty interval—spanning from 18% at the 10th percentile to 56% at the 90th percentile, with an interquartile range of 27% to 47%—reflects several latent variables. The share could drop significantly if a massive final training run happens to be active around the target date, if end-2026 partner-channel serving load absorbs nearly all new chips, or if RL-heavy production runs are strictly categorized as "training." Conversely, if internal AI-agent R&D loops ramp even more sharply, or if massive compute blocks are reserved for pre-training derisking ahead of a next-generation model, the experimental share could push well into the 50s.