# HBM demand — X 热门讨论 (2026-09-30 12:03 UTC)
## @jeffreyajewett (JJ) · 09-30 11:53 · ♥85 ↻7 💬3 In 1817, David Ricardo noticed something strange about farmland.
Rent on the best fields kept rising. Not because the soil got better. Because nobody could make more of it, and more mouths kept showing up.
That's the B200 in 2026.
In early March it settled at $2.31 per GPU-hour. The market had written it off. The B300 was shipping with 50% more memory and roughly 1.5x the FP4 compute. Why buy last year's model when the new one is on the lot?
Economists call that the Osborne effect. A better version on the horizon drains demand from the current one.
Then the story turned.
New allocation shifted to the B300, so the B200 base barely grew. HBM3e contract prices rose about 20%. Much of the existing fleet was already locked into long-term contracts. And inference demand arrived, the kind that doesn't hand capacity back when a job ends.
The B200 became farmland.
On September 28 it settled at $8.11. 3.5x in about seven months.
Not smoothly. $6.11 at the end of May, $4.22 three weeks later, then a new high. When supply can't move, every shift in demand lands entirely on price.
Nothing about the chip changed.
Its scarcity did.
Now the second act.
The B300's median on-demand listing has sat flat at $7.89 for 90 days. The B200's median listing is $6.49. That's about a 22% premium for 50% more HBM, which makes the B300 roughly 19% cheaper per GB of memory.
The market is pricing the B300 the way it priced the B200 in March: as the chip that's about to be replaced.
The skeptic has two good replies.
"Rubin changes everything." It will change the frontier. It will also absorb the allocation, which is exactly what froze B200 supply.
"The B200 fell 30% in June." It did, and then made a new high. Volatility is what a thin float looks like, not a reversal.
Ricardo's lesson was never really about land. It was about what happens to the price of anything you can't make more of. https://x.com/jeffreyajewett/status/2105264828264710158