Inference

The GPU Revenue Ladder

The GPU Revenue Ladder

AI infrastructure

Yesterday brought two data points about the health of inference businesses : SpaceX’s third-quarter report1 & DeepSeek’s inference post.2

Overall, they demonstrate that inference businesses can be quite healthy even at current prices & efficiencies. Plus there is a tremendous amount of upside assuming demand persists at higher prices per card hour.

Horizontal bar chart of revenue per GPU card-hour: DeepSeek maximum $38.00, B200 spot $8.77, SpaceX negotiated $5.27, SpaceX realized $2.44
Rung $/card-hour Change from above Provenance
Theoretical maximum revenue, DeepSeek $38.00 — company claim, theoretical
Spot price, B200 $8.77 −77% ORNN, settled 2026-10-20
Negotiated price, SpaceX $5.27 −40% S-1, implied
Realized revenue, SpaceX $2.44 −54% 8-K, derived

DeepSeek calculates a $38 maximum per GPU card-hour.3 It describes a GPU running under optimal conditions : maximum utilization for external customers rather than internal work, & 100% of its output usable & billable, sold at full list price.

Read more
Price War in the Chip Aisle

Price War in the Chip Aisle

AI economics

Five frontier models now sit within five points of each other in terms of intelligence, but their cost per unit differs by 3.8x.

The AI grocery store aisle: a grouped bar chart of weighted average cost per Artificial Analysis Intelligence Index task for five frontier models, grouped by index score; the three models tied at 53 points span $1.99 to $7.63

Google’s Argon prices at the bottom of the cluster at $1.99 a task1, 74% below the most expensive model, Fable 5.1. It is also 67% below Opus 5.5, 56% below Muse Wonder & 39% below Astra.

Read more