Executive Summary
Today's sources describe a market where model capability is getting cheap and turning over fast, while the physical capacity to run it is scarce and expensive to finance. Anthropic's S-1, read by an investor who holds the stock, shows frontier AI is now a profitable business: 85% gross margin, about 7x better serving efficiency, and price per task down 25–40% per generation. METR's assessment, written with Anthropic controlling access, says AI is now the main driver of Anthropic's own capability gains. SemiAnalysis shows China delivering datacenter shells at one-fifth of US cost, with chips as its only binding constraint. The US has the reverse problem: power, permitting and credit. For enterprises, the risk is not that models stall. It is getting locked into one model version, one capacity provider, or a facility design that will be obsolete by 2028.
What Changed
- Model cadence has a structural driver. METR puts Anthropic's capability progress at about 2x the 2024–25 rate (range 1.3–3x) and attributes 80% of that to AI assistance. Ball's revenue data matches the timing: 80% of last quarter's revenue came from models less than six months old.
- Vendor solvency is less of a concern. Anthropic has $134B in cash, has burned only about $7B in its lifetime, and runs a 21% FCF margin. The alarming headlines ($108B "debt", $81B loss) are non-cash markups on convertible notes.
- Financing is now the brake. Large GPU debt deals are "no bid" because lenders see too much concentration in two offtakers. Datacenter debt is pricing above 10%.
- 800V DC at the rack is running in production. Galaxy runs a 5MW pilot of solid-state transformers going straight from 10kV AC to 800V DC, plus chiller-free shared water loops. About 97% of US facilities still use AC UPS.
Cross-Expert Synthesis
The sources agree that compute is the constraint, even where they claim otherwise. Berman reads METR as saying compute is not a bottleneck. Anthropic's own researchers say they are bottlenecked on ideas and compute. Ball names compute, power and credit as the binding constraints. SemiAnalysis shows that even a country that builds fast and cheap is limited by chips. The accurate reading: AI has removed the code bottleneck inside the labs and pushed the pressure down to the physical layer, where everyone is constrained.
The second link is unreleased models. METR sees internal models on a steeper curve than public ones. Ball argues the real gap to open weights is wider than benchmarks show and expects labs to move into vertical applications using models they never release. Together this is a moat that compounds with every generation: labs use better private models to build the next model faster. Application vendors without their own research teams are the exposed layer.
Miessler's Core demo is unverified as a product, but it describes the right defensive architecture for this environment. Agents compose only inside a human-built foundation that owns auth, roles and billing. Every app exposes an MCP server matching its UI. A thin abstraction falls back across Bedrock, Anthropic direct and OpenAI. That design treats the model as a part you expect to swap, and fast turnover makes that the correct assumption.
Enterprise Implications
Token budgets will grow, not shrink. Cheaper capability drives more usage: Fortune 500 token spend is about 0.1% of expenses, while Stripe is at 1% and expects 15–20%. Spend on tokens relative to headcount is becoming a board-level metric.
Private AI facility plans being specced today for 2027–28 GPUs need an 800V DC path and warm-water cooling, or they will need a costly retrofit. China's deployment is evidence the design works at scale, not a guess about the future.
Capacity scarcity and firm pricing will last. If credit stays tight, neocloud capacity is at risk of delay. A contract with a GPU provider is only as good as that provider's financing.
What To Do About It
- Re-baseline roadmaps to 6–12 month model cycles. Shorten model-pinning commitments, write model-substitution rights into contracts, and require an abstraction layer with provider fallback.
- Build on an opinionated internal platform layer. Standard auth, scoped per-agent tool access, and MCP parity with the UI by default. Agent-generated code without that foundation drifts, and you pay the integration cost again for every app.
- Treat power and cooling architecture as an AI decision. For any facility or colo refresh, require a 45°C liquid-cooling and 800V DC readiness assessment before signing.
- Check the financing behind GPU capacity providers. Ask about lender exposure and offtake concentration, not just delivery dates.
- Customer question: "If your AI application vendor's model supplier ships a competing vertical app on a model you can't buy, what is your exit path?" It exposes vendor dependency in one sentence.
Risks and Blind Spots
The two most important claims come from interested parties. METR's report ran on 2.5 FTEs, Anthropic held a veto over publication, Anthropic redacted the compute evidence, and it cancelled the compute-allocation verification partway through, ahead of an IPO. Ball holds Anthropic stock. Both may be right, but the acceleration story and the profitability story are exactly what an IPO roadshow would want told. Plan for faster cadence. Do not plan on 3x.
Ball expects Anthropic's free cash flow to turn negative again as it reinvests in compute. Low lifetime burn does not mean low future burn, and the build-out math (about $5T for 100GW) depends on credit that is currently refusing GPU deals.
The China cost advantage may not carry over to the US. It rests on vertical integration, $0.055/kWh power and no local opposition. SemiAnalysis itself says it is too early to call that a durable edge. Use the engineering, not the cost per MW, as the benchmark.
Core is invite-only and published no outcome metrics. Take the architecture pattern from it, not the product claims.