Nvidia confirmed on Tuesday that its next-generation Rubin GPU has entered volume production at TSMC's Arizona and Hsinchu facilities, with first shipments to hyperscaler customers scheduled for late May.
Spot pricing for the older H100, the workhorse of the 2023-25 training boom, has fallen 12% in the past fortnight to roughly $24,800 per unit on the secondary market — the steepest decline since the chip launched.
Microsoft, Meta, and Oracle have each disclosed Rubin allocations in the tens of thousands of units. Smaller buyers, including most of the AI-native startups that filed S-1s in the past year, will not receive units before Q4.
Pricing dynamics
The transition is unlikely to be smooth. Industry analysts expect a brief glut as cloud providers retire H100 capacity, followed by renewed scarcity as Rubin demand outstrips Nvidia's stated 2026 ship target of 1.4 million units.
"You're going to see a six-week window where renting an H100 is genuinely cheap," said one capacity broker. "Then it's gone."