The Nvidia Chips Everyone Wanted Are Sold Out Again

Warren Pies / 3Fourteen Research (@3F_Research), citing Silicon Data: Nvidia B200 rental rates rose from ~$5.30 to ~$5.77 over the past week, and B200 availability collapsed back to roughly 0%. Falling availability signals strong compute demand, and historically availability has led rental pricing. Key nuance from the report: the entire rise in the broader availability index is from older-vintage GPUs (H100, H200, A100), not B200s — current-gen supply is fully absorbed. 3Fourteen's bottom line: they do not see frontier token pricing weakening (based on OpenRouter data); as open-source improves it has captured more share, which they read as logical and "not a threat to the compute story." GPU availability is low (high demand) and led by B200 tightness. Source: Warren Pies / 3Fourteen Research (@3F_Research), via Silicon Data

This contests Layer 3, and the honest move is to flag it. The bear case assumes AI compute demand and token pricing are softening. Two independent readings this week say otherwise: B200s are sold out with rising rents, and frontier token prices are holding rather than falling. Both cut at the monetization-gap leg. The fair rebuttal is that current-gen tightness confirms the boom is live — the same way a supplier's record quarter does — without proving the capex is financed sustainably (Layer 4) or earns its return (Layer 2). One detail cuts slightly the other way: availability is loosening first in older GPUs, which is at least consistent with an eventual demand plateau starting at the trailing edge. On balance, though, this is evidence against the near-term "demand is cracking" read, not for it.