Amazon is spending $8 billion a month on AI chips it has nowhere to plug in, while simultaneously shipping tools to make sure nobody needs Nvidia chips in the first place.

The Summary

The Signal

Amazon is running the most expensive hedging strategy in tech history. The company created an Early Rack Receipt process to secure billions in Nvidia GB300 systems before they can even be deployed. That's not normal procurement. That's panic buying dressed up as supply chain optimization.

The numbers tell the story. AWS expected roughly $5 billion in initial GB300 deliveries to US facilities in August, with projections of $3 billion more arriving monthly "for the foreseeable future." Some of those racks never made it to data centers. They went to intermediary warehouses instead, finished equipment sitting dark because AWS sites had physical space but not electrical capacity to power them on.

"At least one AWS team did not have a clear forecasted path for getting the incoming racks into data centers."

This is the infrastructure bottleneck everyone knew was coming but nobody wanted to price in. You can manufacture AI chips faster than you can build power substations. The result is what industry insiders call "dark GPUs," purchased hardware that sits idle not because of technical problems but because the electrical grid can't keep up with compute ambition.

Meanwhile, Amazon is building an exit. The same week internal emails showed warehouses full of unused Nvidia systems, AWS launched TorchNeuron, a developer tool designed to make migrating PyTorch applications off Nvidia GPUs trivial. PyTorch is the framework behind most AI applications. Whoever controls the easiest path for PyTorch developers controls where AI workloads run.

Amazon isn't alone in this strategy. Google released its own PyTorch migration tool the next day, a coordinated attack on Nvidia's developer moat. Both companies want the same thing: to redirect AI workloads from Nvidia's H100s and GB300s to their own silicon, Amazon's Trainium and Inferentia chips, Google's TPUs.

The timing is not coincidental. Hyperscalers are:

  • Spending billions securing Nvidia inventory they can't immediately use
  • Shipping tools to help developers avoid Nvidia entirely
  • Racing to bring online enough power capacity to activate dark GPUs

This is what desperation looks like at scale. Amazon CEO Andy Jassy has called AI infrastructure "a major opportunity," but the constraint isn't capital or even silicon supply anymore. It's watts. You can't cloud your way out of physics.

The Implication

Watch the power situation, not the chip counts. If Amazon is storing billions in finished AI systems because it can't power them on, every other cloud provider is facing the same math. The companies that solve electrical capacity first will capture the next wave of AI workloads, regardless of whose chips they're running.

For developers, this creates real optionality. If PyTorch migration tools from AWS and Google actually work as advertised, you're no longer locked into Nvidia's ecosystem by default. That's leverage you didn't have six months ago. Test the alternatives now while Nvidia's supply constraints and power bottlenecks are giving competitors room to prove themselves.

Sources

TechCrunch AI | Business Insider Tech