Deedy Das claims $10-38 per GPU-hour margins serving open models

Builders are arguing about how much money is actually in serving open weights. The spark is a high-velocity X post from Menlo Ventures partner Deedy Das claiming $10-38+ per GPU-hour in provider margins; the one verifiable floor underneath it is that OpenRouter publishes token-share rankings and per-model prices anyone can read.

Key facts

• @deedydas claims providers earn $10-38+ per GPU-hour serving open models — unverified.
• The post drew 1,495+ likes, 267k+ views and 180+ replies on X.
• OpenRouter's public rankings and per-model pricing pages are the checkable cross-reference for share and rates.

What's driving the conversation

The Deedy Das post lays out a land-grab story: providers reportedly earning $10-38+ per GPU-hour serving open models such as Kimi K3, DeepSeek, and GLM against GPU rental costs, Chinese labs reportedly demanding revenue share from the servers, and low barriers pulling new entrants like inference.net and DeepInfra into a market where commodity pressure keeps pushing prices down. The engagement numbers — 1,495+ likes and 267k+ views within the cycle — mark it as the week's dominant inference-economics thread.

Replies have split the usual way: some operators sharing numbers that roughly match, others arguing the spread collapses once you price idle capacity, failed requests, and quantization engineering. Nobody in the thread has published a verifiable ledger.

The substance

What is verifiable today: OpenRouter's public rankings page lists token share by model, and its per-model pages list published per-token prices. Those are the demand-side and price-side anchors. What remains claim-only: the $10-38+ per GPU-hour margin band, the revenue-share demands, and the cost side, which would need spot GPU rental rates and provider cost disclosures to pin down. Until a provider discloses unit economics or a credible third party reproduces the math, every figure in the thread stays a claim.

Why builders are watching

If the claimed spread is even roughly right, serving open weights is currently the most profitable seat in AI infrastructure, which explains the flood of new providers and implies today's aggressive per-token pricing is subsidized margin that commodity pressure will compress. For anyone buying capacity, that argues for treating current list prices as negotiable and declining — DeAI's tracking of the inference cost curve shows the trend, and provider coverage like our DeepInfra pricing page shows how fast the floor moves. It also rhymes with router competition like You.com's inference router launch: when routing is a commodity, margin lives at the GPU, briefly.