ARK: "Anthropic's S-1 Shows That Investment In Training Is Paying Off"
![[2026-11-02-ark-anthropic-revenue-per-compute.jpg]]
Why this is in the vault
It has the cleanest single chart from the Anthropic S-1: revenue per dollar of compute, quarter by quarter, Q1'25 to Q2'26. It is also a useful example of how an advocate turns one ratio into a causal story.
⚠️ Structural Bias
ARK is a public advocate for the AI-infrastructure bull case and has private AI exposure. The data comes from the S-1. The explanation of what caused it ("training pays off") is ARK's own.
Summary (author's terms)
Revenue per dollar of compute, as in-period revenue / compute and infrastructure expenses:
Quarter Ratio Q1'25 0.40x Q2'25 0.57x Q3'25 0.68x Q4'25 0.69x Q1'26 1.40x Q2'26 2.08x (revenue 11.6B, compute 5.6B) That is 5.2x through Q2. Q3'26 fell to 1.72x as short-term SpaceX compute contracts came online. ARK quotes the 0.40 → 1.72 move as "4.3x".
Q2'26 revenue was up 1,368% year over year. Q2 was the first quarter with GAAP and non-GAAP operating profit. Q3 revenue was nearly 17B, more than all of H1.
Anthropic has raised 141B since 2021 and held 132B in cash plus short-term investments at 2026-09-30, so roughly 10B net has been used. ARK says that cushion could fund owned data centers in place of short-term leases.
ARK's thesis: training spend plus "compute multipliers" (algorithmic efficiency gains) raised both demand and serving efficiency. On that view, frontier-lab compute spend is justified.
Mental model
Operating leverage in a hypergrowth quarter, presented as an efficiency curve. The ratio divides revenue in a quarter by compute spent in the same quarter. When revenue goes up 2.5x in one quarter, it runs ahead of compute contracts signed months before. Part of the jump is timing.
Mapping against Ray Data Co
- Cross-check with Stratechery. [[2026-10-27-stratechery-anthropic-s1-cost-of-compute]] puts total compute at 55% of revenue in H1'26. ARK's chart gives (3.4+5.6)/(4.7+11.6) = 55% for the same period. The two sources agree on the totals. Thompson adds the part ARK leaves out: R&D compute is growing faster than inference compute, and it is driven by competition. That cuts against ARK's clean "training pays off" story. Training is the cost line the competitive race controls.
- Investing thesis (capital cycle). The move from leased to owned data centers, funded from free cash flow, is a supply-side capital-cycle signal ([[2026-08-13-ai-capex-lender-exposure-insurance-pension-vehicles]]). Watch Q4: if the ratio climbs back from 1.72x, the SpaceX capacity was spending ahead of demand. If it stays flat or falls, Q2'26 was the peak.
- phData credibility. The demand inflection and the first profitable quarter are public, citable facts. They fit the "agents in production" story for phData conversations, and they carry no disclosure risk.
- phData partnership (public). Founder confirmed 2026-11-02 that the phData and Anthropic partnership and investment are public. The public source is not yet filed in the vault. Ground-game framing: Anthropic builds the model, and partners like phData drive enterprise adoption. ARK's demand inflection only turns into revenue through that adoption work.
- Cert path. This adds to the case that the Claude ecosystem the founder is betting career capital on (Architect Pro 12-22) has durable enterprise pull. It does not remove the R&D-compute exposure Thompson flags.
Open questions
- Does the Q4'26 ratio recover above 2x, or did Q2 mark the peak?
- How much of the 2025 → 2026 ratio jump is price (enterprise mix, Claude Code seats) and how much is serving-cost efficiency? The S-1 line items may separate them. ARK does not.
Related
- [[2026-10-27-stratechery-anthropic-s1-cost-of-compute]]
- [[2026-08-03-dwarkesh-why-smarter-ai-models-could-drive-up-compute-prices-10x]]
- [[2026-08-25-dwarkesh-dylan-patel-two-labs-workforce]]
- [[2026-03-04-stratechery-anthropic-revenue-nvidia-earnings]]