Anthropic’s compute costs rose to 58% of revenue last quarter, filing shows

Anthropic’s compute costs rose to 58% of revenue in the third quarter, up from 48%, indicating that compute spending grew faster than revenue. Revenue increased to $17 billion from $11.6 billion in the prior quarter, according to figures identified in the company’s amended IPO filing.

Analysis

The key signal is operating leverage moving the wrong way: compute intensity rose even as revenue scaled. If the filing’s definitions and periods are comparable, compute expense grew faster than revenue sequentially, challenging a simple “revenue growth will fix margins” IPO narrative. This ratio is not necessarily gross margin: verify what the filing includes and separate training from inference, fixed capacity commitments, and other cost of revenue.

Near term, expect investors to scrutinize revenue quality and unit economics more than headline growth; a revenue multiple without a credible path to lower compute per dollar of sales is vulnerable to compression. Over 1–3 months, the catalyst is further filing detail or guidance on compute mix, utilization, and cost per token. Over 6–18 months, the thesis turns on whether model efficiency and pricing power improve faster than usage and capacity costs. NVIDIA and cloud-compute providers may see demand support, but this disclosure alone does not establish Anthropic’s supplier allocation or incremental economics.

Contrarian angle: rising compute intensity can reflect deliberate capacity investment or a costly training quarter rather than structurally poor inference economics. A reversal in the ratio would matter more than one quarter’s level. No valuation, IPO pricing, or cost breakdown is supplied, so avoid a directional position based solely on this excerpt.

Key Decisions for Investors

• Treat ANT as a watch, not an automatic short: do not underwrite a revenue-growth multiple until the filing clarifies compute-cost definitions and shows a path to improved compute intensity.
• Before taking IPO exposure, verify quarterly compute spend, training versus inference split, committed capacity obligations, and whether revenue is recognized on a comparable basis across periods.
• Use a 1–3 month catalyst checklist: a sustained further rise in compute as a share of revenue, or guidance that fails to show improving cost per unit of usage, would strengthen the margin-risk thesis; a clear decline would weaken it.
• Avoid using this item alone to buy NVIDIA or cloud providers: confirm supplier exposure and capacity economics first. Reassess if disclosures link Anthropic’s spending to identifiable contracts or materially change supplier outlooks.