Anthropic is still grappling with the economics of compute

Anthropic’s Q3 revenue rose to $17 billion from $11.5 billion in the prior quarter, but compute costs reached $10 billion and absorbed 58% of revenue, up from 48%. The filing attributed some of the higher expense to $5.0 billion of compute under short-term commitments that cost significantly more than comparable multi-year arrangements; revenue per dollar of compute spend fell to 1.72x after peaking above 2x. Anthropic expects compute and infrastructure expenses to remain its largest operating expense, leaving margins exposed to compute prices and infrastructure buildout risks as it approaches an IPO.

Analysis

The key signal is not simply rising compute expense; it is that Anthropic is using higher-priced, short-term capacity to protect growth. That makes revenue quality and delivery capacity more exposed to infrastructure bottlenecks than headline growth alone suggests. If capacity becomes available under cheaper, longer-term contracts, compute intensity could improve—but the filing’s expectation of continued absolute cost growth means investors should not assume near-term margin recovery.

For ANT, the IPO risk is a valuation mismatch: investors may underwrite software-like operating leverage while the marginal unit of growth is still being served by expensive infrastructure. The decisive metrics are compute cost per unit of revenue, utilization, and the share of capacity on short-term commitments—not revenue growth in isolation. A demand slowdown could help margins only if committed capacity can be reduced or productively redeployed; otherwise, fixed commitments could turn today’s capacity scramble into underutilization risk.

SPCX and Akamai may benefit from scarce-capacity pricing, but the article does not disclose contract-level economics, duration, or the providers’ incremental costs. Do not translate Anthropic’s spending into provider profit or valuation upside without those details. The broader second-order risk is financing: tighter data-center funding or delayed power/interconnection approvals could constrain available capacity and raise prices, impairing labs’ margins while benefiting already-operational capacity providers. Conversely, faster hyperscaler capacity delivery or weaker AI demand could unwind scarcity pricing.

Near term, this is a caution on IPO underwriting rather than proof of a broken model. Over 1–3 months, watch updated filings and contract disclosures; over 6–18 months, power availability, financing conditions, and utilization determine whether compute intensity normalizes. The contrarian angle: strong demand can mask poor incremental economics, but a temporary margin dip is also not evidence that long-term unit economics are structurally deteriorating.

Key Decisions for Investors

• ANT: avoid underwriting the IPO on revenue growth alone; require disclosure on compute cost per revenue dollar, short-term versus multi-year capacity mix, and utilization before treating margin recovery as credible. Reassess if those measures improve while growth remains strong.
• Do not initiate a directional SPCX or AKAM trade from this disclosure alone. The named commercial relationships are not enough to establish material earnings contribution; verify contract size, term, capacity utilization, and incremental infrastructure costs first.
• Use data-center financing conditions, power/interconnection delays, and capacity availability as a 6–18 month risk dashboard. A material slowdown in new capacity delivery would strengthen the margin-risk thesis for ANT; evidence of ample lower-cost capacity and improving compute intensity would falsify it.
• For an ANT position around the IPO, size for execution risk and consider waiting for post-listing disclosures rather than chasing initial demand. The thesis weakens if compute intensity falls without sacrificing growth; it strengthens if costs rise faster than revenue or management signals persistent reliance on costly short-term capacity.