Quick answer

Nvidia still wins on training. But on inference — which is now 80% of AI compute spend — the field is genuinely competitive in 2026. AMD MI400, Google TPU v7, AWS Trainium 3, and Cerebras WSE-3 each have credible market share. Hyperscaler-managed inference prices have dropped 40% YoY. The duopoly era is ending.

For four years Nvidia GPUs were the only serious option for training frontier AI. In 2026 that's still mostly true for training. But inference — running the model after it's trained, which is where most actual compute spend now sits — has become genuinely contested.

The four serious Nvidia challengers

  • AMD MI400 — finally got the software stack right with ROCm 7. Roughly 90% of H200 inference performance at 70% of the price.
  • Google TPU v7 — Google's flagship internal chip. Powers Gemini 3.5 Pro. Trillium economics let Google undercut OpenAI on price.
  • AWS Trainium 3 — Amazon's in-house chip. Anthropic now serves most Claude traffic on Trainium clusters.
  • Cerebras WSE-3 — wafer-scale specialty chip. Dominates a small but high-value niche: low-latency inference for high-stakes apps.

What it means for prices

AI API prices have dropped 40% on average over the last 12 months. This is mostly because hyperscalers can now serve inference on their own silicon and skip Nvidia's margin. Expect another 20-30% drop in the next 12 months as MI400 capacity scales and Trainium 3 ramps.

What it means for Nvidia

Stock is still up YoY but growth has slowed dramatically. The duopoly with TSMC means revenue is enormous either way — but the gross-margin premium of "the only AI chip" is gone. Margins are compressing. Wall Street is still figuring out what that means.

What it means for builders

  • Cheaper APIs: budget another 25-30% input cost drop in the next 12 months.
  • More backends: model providers will be more willing to do on-prem deals.
  • Lock-in still risky: backend-specific quantisation is more advanced now; switching providers loses some efficiency.
  • Local inference more viable: consumer Nvidia, AMD, and Apple Silicon are all running quantised frontier models in 2026.

If you're negotiating multi-year inference contracts, don't lock in. Prices will drop. Negotiate 90-day renewable terms with re-pricing clauses.

Related reading

The AI Energy Crisis — How Much Power AI Uses AI API Pricing in 2026 AI Pricing Models in 2026

Bottom line

Nvidia still dominates training. Inference is now genuinely contested. Prices are falling, and they'll fall more. The Nvidia monopoly era is ending — and AI builders are the winners.