NVIDIA still owns the AI compute market — Q1 FY2027 revenue of $81.6 billion with $75.2 billion from data center, at a 74.9% GAAP gross margin, and the Vera Rubin platform's seven chips in full production. But 2026 is the first year the alternatives stopped being slideware and started being gigawatts. Here's the honest state of the field.
AMD: From Benchmark Contender to Signed Gigawatts
The evidence that AMD crossed a threshold is contractual, not benchmark-based:
Meta signed a definitive multi-year agreement for up to 6 GW of Instinct GPUs, explicitly framing it as portfolio diversification away from single-vendor dependence.
Microsoft is deploying AMD Helios on Azure for frontier-model inference and Azure AI services.
Helios rackscale is in production: 72 Instinct MI455X GPUs plus 18 6th-gen EPYC "Venice" CPUs per rack — AMD's first credible answer to NVIDIA's rack-as-unit-of-compute model.
The software gap is narrowing at the compiler layer: SPIR-V on ROCm brings compile-once, specialize-on-device portability.
Caveats remain: AMD's Q1 2026 revenue of $10.3 billion (all segments) is a fraction of NVIDIA's data-center number, export controls on MI308 cost it real margin in China, and ROCm's ecosystem depth still trails CUDA's by years of accumulated tooling.
Custom Silicon: The Margin Eater
The hyperscalers' own chips are the structurally biggest threat, because they don't need to win the merchant market — only their owners' workloads:
Amazon's chip business passed a $20 billion revenue run rate, growing triple digits; OpenAI committed to ~2 GW of Trainium capacity, and Anthropic's Amazon agreement includes nearly 1 GW of Trainium2/3 by end-2026, with over a million Trainium chips already in use.
Google TPUs are now a merchant-ish business: Broadcom's 8-K disclosed a long-term agreement to develop and supply future TPUs plus components through 2031, and Anthropic signed for multiple gigawatts of next-gen TPU capacity from 2027. Pichai says TPU systems began shipping to customer data centers.
Meta and Broadcom extended their partnership through 2029, co-developing the industry's first 2nm AI accelerator — multi-gigawatts of custom silicon.
Broadcom is the quiet winner across all of this: $10.8 billion of AI semiconductor revenue in a single quarter, plus networking (Tomahawk 6, the first 102.4 Tbps Ethernet switch, shipping in volume).
What's Still Missing
Three things keep NVIDIA's moat intact for now: (1) training at the frontier — the labs' flagship runs still overwhelmingly land on NVIDIA silicon (xAI trained Grok 4.5 on tens of thousands of GB300s and holds over a million H100-equivalents); (2) the CUDA ecosystem's accumulated debugging, kernels, and talent; (3) rack-level systems integration, where Vera Rubin's full-production status keeps the reference platform NVIDIA-shaped. Alternatives are winning inference and owned-workload gigawatts first — the pattern to extrapolate.
Operator Takeaways
Price your inference on at least one non-NVIDIA path this year; the buyers signing 6 GW deals have concluded the software gap is manageable for inference.
Watch Broadcom's AI revenue as the cleanest proxy for custom-silicon share shift.
Expect training to stay NVIDIA-dominant through 2027; plan portability at the serving layer, not the training layer.
