Trainium gets Nvidia memory: your silicon hedge narrows

Every AI cost model we've seen from a hopeful operator has the same escape hatch penciled in the margin: if Nvidia stays expensive, we move to the cheaper silicon. AWS just made that hatch smaller. Alongside a very large GPU order, AWS is putting Nvidia interconnect and Nvidia memory technology inside Trainium — the chip that was supposed to be the alternative.

What actually happened

Per Amazon's own announcement and Nvidia's newsroom release on August 26, AWS will deploy 2 million additional Nvidia GPUs across 2027–2028 — Blackwell Ultra, Rubin, and Rubin Ultra. That's on top of the 1 million-plus commitment announced at GTC 2026, which AWS says demand outran.

The part that changes the shape of the market is further down the page. AWS will support Nvidia NVLink Fusion in next-generation Trainium chips, and the two companies will integrate Nvidia's custom high-bandwidth memory (NVHBM) with Trainium for faster, more power-efficient memory. Nvidia Vera CPUs are also coming to AWS, aimed squarely at the CPU-side work behind agentic AI: code execution, tool use, sandboxing, data pipelines, orchestration. The companies also plan AI factories for the US government including 100,000 GPUs on AWS's secure infrastructure.

AWS cites its own figures for GPU-accelerated data processing — up to 3.7x faster with 30% better price-performance, and up to 9x faster vector indexing at a quarter of the cost. Vendor benchmarks; treat them as such until you run your workload.

Why it matters for your business

You are not buying chips. You are buying inference by the token, and the reason the silicon question ever mattered to you is that competition between vendors is what makes tokens get cheaper. When the leading alternative accelerator licenses the incumbent's interconnect and memory, that competition is now partly a supply relationship. Trainium may still be the cheaper line item. It's just no longer independent of Nvidia's roadmap or Nvidia's memory supply.

The practical read: stop treating chip diversity as your hedge. It was never a lever you controlled. Your actual portability lives one layer up, and it's the boring stuff — prompts, eval sets, and routing logic in your repo instead of a vendor console; an OpenAI-compatible interface in front of whatever serves your model; at least one workload you've genuinely run on open weights, so you know the switching cost by measurement rather than by assumption.

Then do the thing almost nobody does: put a date on the calendar to re-benchmark. Price-performance claims move every quarter now, and a hedge you never test is a paragraph in a doc, not an option.

Sources: About Amazon, NVIDIA Newsroom.