6 min read · 19 sources

Two stories mattered today, and they are the same story: hyperscalers are converting AI capex from a procurement decision into an infrastructure sovereignty decision. Microsoft announced gigawatt-scale deployments of its custom Maia chips, starting with the Maia 300 for inference, and Meta launched the Muse Wonder family backed by more than 5 GW of infrastructure and over $100 billion of capex. Everything else on the tape - an OpenAI claim of a sub-Cooley-Tukey FFT, a Helix world model running on an H800, an Anthropic IoT worm warning - was research, not revenue.

The interesting trade in a silicon-sovereignty story is never the sovereign; it is the merchant supplier watching its largest customer become a competitor.

The silicon divorce

We flagged Maia 300 when it was introduced last week; today’s increment is the deployment commitment, and it is the part that reprices things. A gigawatt of in-house inference silicon is not a pilot. At Microsoft’s scale, every gigawatt shifted from merchant accelerators to Maia is volume that never reaches NVIDIA’s datacenter line item - and Microsoft is the largest customer in that line item. We are raising our NVIDIA stance from negative at conviction 2 to negative at conviction 3; what changed is that the displacement moved from roadmap to committed capacity.

The second-order read: the memory and design-service suppliers sit on both sides of this divorce. Custom silicon still needs HBM and DRAM, and Micron has already warned supply stays squeezed through 2028 - two hyperscalers committing gigawatt-scale builds in one day makes that squeeze tighter, not looser. Meta’s $100B capex, meanwhile, is largely a transfer to the accelerator design ecosystem, which keeps the Broadcom pipeline at record depth. The displaced incumbent is NVIDIA; the quiet beneficiaries are memory and design services.

For Microsoft itself we stay negative. In-house silicon trims merchant-chip cost over time, but it also deepens the capex commitment we have been fighting since last week, and today’s Exchange Admin Center outage - attributed to a configuration change - is a small but unhelpful data point on execution strain while the company rebuilds its own hardware stack.

Meta's overbuild gets a monetization story

Muse Wonder on its own would be a model launch. What makes it investable is the combination with what we covered earlier this month: Meta preparing an enterprise cloud push to rent out its excess datacenter gigawatts. A >$100B capex commitment only pays if the overbuild has a tenant, and Meta is building the tenant channel in parallel. That makes Meta’s capex a supplier transfer with a monetization path rather than pure burn, which is why we are willing to sit positive while the spend ramps.

Agents reach the storefront

Shopify adding WebMCP to its storefront tools is the first major commerce platform making its catalog and checkout natively callable by browser-resident AI agents. It is early and unmonetized, but the position matters: if agent-mediated commerce becomes the default discovery layer, the platform that is agent-native first captures the flow. Relatedly, the merger of two Base-based tokenized-stock platforms concentrates that activity on Coinbase’s own chain - small numbers today, but it is volume on Base, which is the asset Coinbase actually controls.

The research pile - thin day, honestly

Beyond the two capex stories, today was a preprint day. Amazon’s “Priming” method claims >99% reduced pre-training cost for adapting Transformers to hybrid architectures; NVIDIA’s Semi-OPD improves distillation in 14 of 17 teacher-student pairs; a Helix world model runs real-time audio-visual simulation on a single H800. The OpenAI claim of FFT implementations beating Cooley-Tukey at sub-5 n log₂ n operations would be genuinely significant if it verifies - we would want independent replication before treating it as more than a repository claim. Anthropic’s research on 9B models running on IoT hardware and near-complete self-replication in sub-30B models is an early worm-threat warning; it is a narrative seed for the security vendors rather than a purchase order today. Google’s biomedical showcase (AlphaGenome Atlas, AlphaFold-based Chagas screening) is impressive and continuing coverage of an existing strength, not a new fact.

We are not forcing views on the thin part of the tape. The day belonged to Microsoft and Meta’s gigawatts, and to whoever sells into them.

Scoreboard

The book

Trades decided 2026-10-13

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