5 min read · 17 sources

The day’s real news is a benchmark. SemiAnalysis’ third-party inference results showing Google’s TPUv7 Ironwood delivering up to 50% better performance per dollar than NVIDIA’s B200/B300 is the first independent confirmation that the custom-silicon discount has become a custom-silicon advantage - and it lands the same day Anthropic disclosed roughly $517bn in compute capacity leases over eleven months, with the majority of the 14.8GW coming from Google TPUs and AWS. Everything else today is either continuation or plumbing.

When your largest customers start signing half-trillion-dollar leases for your competitor’s silicon, the moat conversation stops being about CUDA and starts being about price-performance.

Custom silicon vs. the incumbent

We’ve been positive on NVDA for three straight weeks on the “moat moves up the stack” thesis. We’re stepping down from that today, and the reason is specific: this is not an OpenAI press release about an in-house chip (which we’ve treated as noise), it’s third-party inference measurement, and the customer signing the biggest cheques in history - Anthropic, committed to over one million TPUs - is doing so on economics. When performance-per-dollar gaps reach 50%, CUDA lock-in has to absorb a lot of slack before the next capacity cycle.

The second-order beneficiaries are the ones nobody headlines. MU is the cleanest expression of the Anthropic lease story that doesn’t require you to pick a silicon winner: 14.8GW of new capacity is 14.8GW of memory demand whether it’s served by Ironwood, B300, or Trainium. We’ve been positive there since the $10bn memory-backing story and nothing today weakens it.

GOOGL is the obvious winner and partially priced, but the compounding is underappreciated: the TPU wins on economics, Anthropic’s lease makes Google a compute landlord at scale, and the Accelerator Agents release (MaxCode, MaxKernel - CUDA-to-Pallas conversion tooling) shows Google actively building the migration path off Nvidia’s software stack. That last item is the quiet one. Tooling that converts CUDA to Pallas is aimed at the moat, not the market.

The lease nobody can ignore

Anthropic’s $517bn in leases deserves its own paragraph because the second-order read is who isn’t in it. Microsoft is conspicuously absent from the 11GW that Google and AWS are contributing. We’ve been neutral-to-negative on MSFT all month and today adds a structural datapoint: the largest AI-native customer is building its future on other people’s clouds. Meanwhile OpenAI’s DevDay plan for “Managed Agents” - with ChatGPT ads resolving directly into agent conversations - is an attack on the ad intermediation stack that Microsoft’s enterprise franchise partially sits on.

Robots, folds, and other continuations

XPeng’s automated humanoid line (2,250 TOPS across three in-house Turing chips, mass production targeted end-2026) is a genuine manufacturing first, but we’d note the TAM is still a narrative and Tesla’s Optimus is the same story from a different press office. Tesla’s rare-earth-free Cybercab motor is the more interesting datapoint - for the magnet supply chain, not for Tesla, and it’s week-two coverage of a launch already in revenue service. We hold our neutral.

On AAPL, we stay negative at low conviction: Huawei’s Mate XT 2 shipping two days before Apple’s expected foldable debut, on domestically designed silicon that bypasses export controls, is the China margin-squeeze story getting worse, not better. Consistent with our September 7 call.

Plumbing worth a line

Cursor’s self-hosted agent execution (Lambda MicroVMs inside the customer’s AWS account, with the agent loop staying in Cursor’s cloud) is a smart trust-architecture compromise that should quiet the “my code leaves my VPC” objection - a mild negative for fully-self-hosted alternatives. Amazon Linux 2027 preview with SELinux enforcing by default and AWS-LC crypto acceleration is standard platform hardening. Anthropic’s 80% B2B revenue concentration and the flagged Meta churn risk is worth monitoring but is a risk disclosure, not an event.

A thin day dressed in big numbers. The benchmark and the lease math are the only things that change positions.

Scoreboard

The book

Trades decided 2026-09-08

Filled at the next session’s open, not today’s close.