Everyone built their own chip. Nvidia still made $75 billion.

Last week three of the world’s biggest cloud buyers shipped their own AI chips to escape Nvidia’s pricing — and Nvidia still booked its biggest quarter ever.

01 — Everyone is building now. Microsoft shipped Maia 200 (TSMC 3nm, 216GB HBM3e) at a claimed 30% better performance per dollar than its current fleet; AWS made Trainium3 generally available at 4.4x its predecessor and up to 50% lower cost; Google’s Ironwood TPU anchors an Anthropic deal for up to 1 million chips. The hyperscalers no longer rent all their compute.

02 — But it’s margin defense, not a moonshot. Custom silicon targets inference — the high-volume, cost-sensitive workload — not frontier training. The point isn’t to beat Nvidia; it’s to stop paying its ~75% gross margin on the cheapest half of the stack.

03 — Nvidia’s quarter says the panic is early. $75.2B in data-center revenue (Q1 FY2027, +85% YoY); Huang calls demand “parabolic,” with ~$1 trillion in Blackwell and Rubin orders through 2027. Building your own chip and buying more Nvidia aren’t contradictions this cycle.

04 — The “$5B” everyone misread. That figure is Microsoft’s equity stake in Anthropic (alongside Nvidia’s own $10B and a $30B Azure commitment) — an investment that runs on Nvidia, not a budget to escape it.

05 — The board call. Custom ASICs hit 27.8% of AI server shipments in 2026, a record. The question isn’t “Nvidia or not” — it’s which workloads you own the silicon for, and which you keep renting at frontier prices.

The cost center is quietly shifting from GPUs to memory — SemiAnalysis puts memory at ~30% of 2026 hyperscaler capex. The HBM makers (SK Hynix, Micron, Samsung), not just Nvidia, increasingly set the price of AI.

The “independence” trade is circular: Microsoft puts $5B into Anthropic, Nvidia $10B, Anthropic commits $30B back to Azure — much of it running on Nvidia. The same dollars loop through the same balance sheets.

Custom silicon ships late: Microsoft’s Maia slipped roughly six months. Owning your chip buys margin — but also roadmap risk Nvidia doesn’t carry.

For which workloads do we own the silicon — and where are we still paying a vendor’s margin we could be compounding ourselves? If we’re renting frontier compute, name the supplier setting that price, and the date we stop.

Christophe Fouquet — CEO, ASML. While hyperscalers fight over whose chip wins, ASML sells the lithography machines that make all of them — Maia, Trainium, TPU and Nvidia alike. This week Fouquet pressed Europe to build its own chip supply and demand, from datacenters to cloud. The quiet tell of the custom-silicon era: the real leverage sits one layer below the chip — with the company no one can design around.

Fouquet on LinkedIn: https://www.linkedin.com/in/christophe-fouquet-asml/