The Acceleration

Three triples in one quarter is not three triples in three quarters. Anthropic's $9B → $19B → $30B trajectory could reflect pulled-forward Q2 enterprise budgets closing into Q1 rather than a sustainable new baseline; the Q2 ARR growth rate is the single number that matters. The $40 billion Google commitment is structured as $10 billion cash plus up to $30 billion contingent — meaning $30 billion is a target, not a transfer, and 5 gigawatts of TPU capacity delivered "over five years" is back-end weighted. And Anthropic still spends roughly $7 billion annually on compute against undisclosed gross margins; "fastest ramp on record" can coexist with "deepest losses on record." The tell will be Q2 — if the rate drops below 2x quarter over quarter, the trajectory flattens fast.

One caveat: the 87.6% number runs on Anthropic's partner BenchLM.ai harness, not an independent SWE-bench Pro evaluation — classic Benchmark Theater territory where the harness operator and the model vendor share incentives. SWE-bench Verified is the easier sibling of SWE-bench Pro; the gap between them on Opus 4.5 was 36 percentage points (80% vs 44%), and 4.7's Pro number has not landed. METR's task-horizon measurement — the closest thing to a vendor-independent capability gauge — has not yet re-tested 4.7 either. A 10-14% release-to-release gain is within the normal Claude variance (4.4 to 4.5 was about 12%); the structural question is whether the 89-day doubling cadence has held. If the Pro and METR numbers, when they arrive, show smaller deltas than the Verified jump suggests, the headline overshoots the reality.

Self-reported benchmarks from Chinese model labs have historically replicated at about 85% of the claimed numbers when independent third parties retest — Benchmark Theater territory until DeepSeek's V4 numbers land on a third-party harness. Assume the headline scores will sit around 75-80% Verified rather than 80.6%. Benchmark parity is not deployment parity: US hyperscalers still hold roughly 5-10x the inference infrastructure China does, and 5 gigawatts of new TPU capacity routed exclusively to Anthropic in a single deal exceeds the total capacity of Alibaba's flagship Zhenwu cluster. The Huawei Ascend performance per watt remains 30-40% below Hopper-class silicon, which means V4 economics improve only when measured per token; per training run, the gap remains large. And the export-control bypass story is undermined by the DOJ's $2.5 billion Wally Liaw indictment April 16 alleging Supermicro hardware routed to China through Malaysia — domestic alternatives are growing because the smuggling channel is closing.