Amazon edits Nvidia's lock-in out of PyTorch and points its $25 billion chip business at CoreWeave and Nebius
Amazon's TorchNeuron runs PyTorch apps on its Trainium chips after one configuration change, and Amazon is donating code upstream to strip PyTorch of its chip-specific ties. The same day, Citi's Tyler Radke told CNBC that bond yields, not portability, are what price neoclouds like CoreWeave and Nebius.
Vincent Jiang · 3 min read
Amazon+5.38% — Amazon, up 5.38 percent today announced TorchNeuron on Monday October 19 and opens the pip-installable public beta on Tuesday October 20: applications written in PyTorch, the open-source framework most AI code is built with, then run on Amazon's Trainium AI chips after one configuration change, model.to('neuron'), no rewrite required 15. Hours later on CNBC, Citi's−1.94% — Citi, down 1.94 percent today co-head of US software research Tyler Radke argued the reverse: bond yields, not portability, set prices for the neoclouds, the specialized AI-computing renters CoreWeave−0.11% — CoreWeave, down 0.11 percent today and Nebius+2.78% — Nebius, up 2.78 percent today 34.
The port is routine; the upstream edits are the attack
Config-level ports to AMD+2.03% — AMD, up 2.03 percent today or Apple−0.13% — Apple, down 0.13 percent today chips already existed, so the easy port itself is routine 5. The new move is inside the framework. AWS has filed pull requests to PyTorch itself that strip out hardcoded assumptions favoring CUDA, the Nvidia+1.52% — Nvidia, up 1.52 percent today software layer that made its chips the default, and it registered Trainium through PrivateUse1, the extension point the PyTorch community built for custom hardware 1.
Colin Brace, VP of technology software at Annapurna Labs, Amazon's chip-design unit, named the market he is hunting: founders betting on "single-purpose compute providers" like CoreWeave and Nebius 5. Amazon counts 85% of AI researchers and developers writing their code in PyTorch 2.
A $25 billion run rate, and OpenAI's 2 gigawatts
The prize is already sized. Amazon's custom-silicon business clears a $25 billion annual revenue run rate, and OpenAI$1.18T — OpenAI, private, latest valuation $1.18T signed for 2 gigawatts of Trainium in February inside a deal in which Amazon is investing $50 billion in the lab 26. Deepgram, a speech-AI startup, measured about 60% more audio seconds per dollar on Trainium than its GPU baseline, per its CTO; Brace cites 70% 15. "The real story is removing the lock-in that kept developers from having a real choice at all," Brace said 5.
Nvidia's quarterly revenue has risen more than fivefold, the default now under attack
Data
| Nvidia quarterly revenue | |
|---|---|
| Q3 '23 | $18.12B |
| Q4 '23 | $22.1B |
| Q1 '24 | $26.04B |
| Q2 '24 | $30.04B |
| Q3 '24 | $35.08B |
| Q4 '24 | $39.33B |
| Q1 '25 | $44.06B |
| Q2 '25 | $46.74B |
| Q3 '25 | $57.01B |
| Q4 '25 | $68.13B |
| Q1 '26 | $81.62B |
| Q2 '26 | $96.22B |
Citi's desk: the scarcity is real, the debt is the problem
Radke's arithmetic runs through rates. The 10-year Treasury yield printed 5.09% on October 16, which he puts 30% above its March lows 37. That repricing, not weak demand, cut CoreWeave more than 25% last week to $61.39; its 2032 bond yields 15.77%, and five-year credit default swaps, debt insurance, trade near 1,140 basis points 89.
CoreWeave debt nearly quintupled in seven quarters while revenue more than tripled
- CoreWeave debt
- CoreWeave revenue
Data
| CoreWeave debt | CoreWeave revenue | |
|---|---|---|
| Q4 '24 | $10.62B | $0.75B |
| Q1 '25 | $11.9B | $0.98B |
| Q2 '25 | $14.56B | $1.21B |
| Q3 '25 | $18.81B | $1.37B |
| Q4 '25 | $29.82B | $1.57B |
| Q1 '26 | $35.15B | $2.08B |
| Q2 '26 | $51.61B | $2.58B |
Demand is strong, and posted prices keep rising
Demand on his tape is strong: H100 spot rentals at record highs, labs locking in contracts at the highest prices they have seen 3. Anthropic$350B — Anthropic, private, latest valuation $350B, the AI lab behind the Claude models in which Amazon has invested billions, posted $16.8 billion of quarterly revenue, up from $11.6 billion 8.
Posted price lists tell the same story. Nebius lifted its posted on-demand H100 rate 16.9% to $4.50 an hour effective October 1, and CoreWeave is signing three-to-six-month contracts near $40 million per megawatt of annualized revenue 10. Wells Fargo−2.25% — Wells Fargo, down 2.25 percent today models CoreWeave net debt above $200 billion by 2030 and still rates the stock Overweight 9.
Nebius raised posted Nvidia rental rates up to 21% on October 1
Data
| Value | |
|---|---|
| H100 | 16.9% |
| H200 | 20% |
| B200 | 18.9% |
| B300 | 21% |
No document prices the switching-cost premium
No document prices how much of a neocloud's market value, or Nvidia's, is a switching-cost premium, and no source yet shows a customer moving chips because of TorchNeuron; the record shows code changes, not migrations 15. Gartner+1.15% — Gartner, up 1.15 percent today sets the bear ceiling, predicting more than 80% of neocloud providers cease independent operations by 2030 8. Radke keeps his own conviction in software, Microsoft+0.72% — Microsoft, up 0.72 percent today and Palantir+4.45% — Palantir, up 4.45 percent today, the data-analytics maker, because their value does not require a view on GPU depreciation or financing costs 3.
A free port names the winners and losers
A free port lowers the cost of leaving, and that names the winners and losers: Amazon's Trainium and cloud business gain the exit, while neocloud equity and Nvidia's software lock-in carry the premium at risk. Three readings settle it: whether the PyTorch maintainers accept Amazon's CUDA-stripping changes after Tuesday's open-source launch 1; whether Nebius's posted $4.50 hour actually bills 10; and CoreWeave's November report, where a $200 billion debt path meets a $40 million megawatt 9.
Deepdive
AI-generated from this story and its cited sources. Not investment advice.
