AI hardware is starting to look less like a one-cycle boom and more like a full-stack race for compute, pricing power and capacity.
@NVIDIA’s H100 rental price has climbed 22% in a month to $3.28/hour, suggesting older GPUs are still generating meaningful revenue.
But there’s a catch.
The same H100 once rented for $7–$8/hour, while hyperscalers depreciate GPUs over roughly 5–6 years.
So rising rental prices don’t automatically prove older chips are becoming more valuable.
They may simply show that demand is still outrunning available newer capacity.
And the capital behind that demand is getting massive.
CoreWeave reportedly carried $35B in debt, while NVIDIA has agreed to rent back unused capacity from partners.
Then there’s @Broadcom.
OpenAI’s new inference chip, Jalapeño, could give Broadcom another route into the AI compute stack as inference demand grows.
Meanwhile, @Intel is trying to regain pricing power, with reports of a ~10% PC CPU price increase, while Washington’s 9.9% stake has risen to roughly $36B in paper gains.
Even @Apple is entering another hardware cycle, launching its first foldable iPhone alongside the A20 Pro as competition from Huawei and Xiaomi intensifies.
The common thread?
AI is no longer just about who makes the fastest chip.
It’s becoming a battle over compute economics:
→ who owns the hardware
→ who finances the capacity
→ who controls pricing
→ who captures inference demand
→ and who can keep that infrastructure earning
@NVIDIA may still be at the center, but @Broadcom and @Intel are fighting for different pieces of the same expansion.
The real question:
Will AI hardware keep generating enough demand to make aging silicon profitable, or are today’s rental prices simply delaying the depreciation bill?
