Amazon's Trainium can now train AI models, not just run them, testing Nvidia before Amazon's October 29 report

Trainium, Amazon's▼−0.73% — Amazon, down 0.73 percent today in-house AI chip, has crossed the line that defined it: it can now train models, not just run them. SemiAnalysis, an independent chip research firm just off weeks on Trainium3 racks, reports that PyTorch, the framework most researchers already use, now runs natively↗: code executes in eager mode, line by line the way researchers expect, the torch.compile speed-up tool works, and about the only change is swapping `.to("cuda")` for `.to("neuron")` ¹.

The limits are stated plainly too: the stack is "nowhere close to the maturity of the CUDA ecosystem," Nvidia's▼−0.54% — Nvidia, down 0.54 percent today software layer, and the largest model tested was 32 billion parameters, research scale rather than a frontier run ¹.

Amazon chief executive Andy Jassy spent the same day on demand, telling the venture firm a16z that inference, the computing paid for each time a model answers, is still an order of magnitude too expensive, and that Trainium delivers 30 to 40 percent better price-performance than comparable processors ². Nvidia's training prices have rested on the software friction SemiAnalysis just measured. That friction is eroding.

The switching cost was the moat

AnthropicPVT$350B — Anthropic, private, latest valuation $350B, the AI lab behind Claude and Trainium's anchor customer, already runs production inference and reinforcement-learning rollouts on Trainium2 and Trainium3 ¹. Research training was the holdout, because researchers reshape model architectures daily, and the old Trainium path ran through an XLA-style graph compiler that made anyone coming from CUDA relearn their training loop ¹.

SemiAnalysis trained the open-weight Qwen3-32B, edited its architecture code, and ran the gpt-oss recipe, OpenAI'sPVT$1.18T — OpenAI, private, latest valuation $1.18T open-weight training setup, out of the box ¹. The result: 14,156 tokens per second at 30.47 percent MFU in compiled mode, up from 11,099 at 23.89 percent in eager mode, a 1.28x gain ¹. MFU measures how much of a chip's peak compute a run actually uses.

Behind that sit up to 80,000 physical cores of automated quality control, checking numerics against Nvidia baselines across the top 100,000 models on Hugging Face, the repository where researchers share code ¹. AWS also hired the engineer who led Google's effort to run PyTorch natively on its TPU chips to run the same play here ¹. Every test that passes is one more model a researcher can move without rewriting code.

The margin evidence arrived first

The money moved before the test did. AWS posted a 39.4 percent operating margin in the second quarter of 2026 even as capital spending climbed ³. Its contracted backlog stands at $496 billion↗, with capacity effectively sold out for 2026 and largely booked for 2027 ³.

Disclosed future Trainium commitments from Anthropic and OpenAI exceed $225 billion ². Earlier this month we described Amazon selling the risk on $8 billion of Nvidia chips; the missing half of that trade was proof Trainium could win work on its own software, and that proof went public this week ².

The bottleneck moved to the compiler

What limits Trainium now is software polish, and Amazon is paying to close it. Compiling the gpt-oss recipe took about 27 minutes, dead time for researchers who iterate hourly; AWS says a rebuilt compiler targets up to a hundredfold cut in that wait ¹.

It is also nearing a $7 billion acquisition of Decart↗, an Israeli startup whose optimization stack compresses the months-long port from CUDA into weeks ⁵. Anthropic is hedging too, with its own chip team, though production silicon would not arrive before 2028 ⁶. For engineers, the premium skill is drifting from writing CUDA to making models run well on whatever silicon is cheapest.

October 29 is where the trade reprices

The facts run as far as the training test and AWS's margin. What follows is projection: cheaper training stacked on cheaper inference opening what Jassy calls the middle of the barbell, the everyday business software he expects to become AI's biggest source of demand ².

Amazon reports third-quarter results on October 29, and traders on the prediction market Kalshi put 96 percent odds on Trainium dominating the call ². The question for Amazon is whether investors stop treating its chips as a side business. The question for Nvidia is whether training, its richest business, starts getting priced against a chip researchers can move to without a rewrite.