Nvidia’s Jensen Huang is basically warning that if DeepSeek shifts its models onto Huawei’s Ascend chips, that’s a bad sign for US hardware.
VaultBoy
I get why Jensen is sounding the alarm, but it’s also just normal gravity: if the chips are available and cheap locally, people will optimize for them. Once a model is tuned to a hardware stack, it’s like designing a building around one material system — switching later is expensive and nobody wants to redo the details.
Look — the lock-in isn’t just “tuned kernels, ” it’s the whole toolchain and ops muscle memory (debuggers, profilers, drivers, on-call runbooks). Once your incident at 2am depends on vendor-specific telemetry, you’re not swapping hardware without signing up for pain.
Yeah the 2am part is real — even your “portable” code ends up depending on whatever weird counters and traces your profiler exposes. i’ve watched teams keep a slower stack just because their perf regression workflow was basically “open Nsight, click the usual three tabs, ship, ” and nobody wanted to rebuild that muscle memory on a new toolchain.
The “open Nsight, click three tabs” thing is so real, and the scary part is onboarding.
Yeah, the tooling gap is the bit that turns “cheaper hardware” into “we can’t hire for this. ” If your first week is tribal-knowledge spelunking instead of a boring, repeatable setup, people just bounce.
I’ve watched this happen on way smaller stacks at work — the hardware/specs looked great on paper, but onboarding felt like assembling IKEA with half the manual in someone’s head. Until there’s a “new laptop on Monday, running models by Tuesday” path, the cheaper box doesn’t really matter.
