Capacity Reached: The New 404
It's happening again. The 'Capacity Reached' messages are everywhere. AWS p5 instances? Gone. Lambda Labs? Waitlisted. CoreWeave? Don't even ask. The spot price for H100s has tripled in 30 days. Small labs are pausing training runs, and startups are burning runway waiting for compute.
The 'Project Stargate' Rumor
Insiders suggest that Microsoft and OpenAI have activated 'Project Stargate'—a massive training run for GPT-6 (or whatever comes after Orion). They aren't just using their own clusters; they're buying up every spare GPU hour on the spot market. It's a 'Compute squeeze' designed to starve competitors.
The Black Market for Compute
Desperate times call for desperate measures. Startups are turning to shady brokers on Telegram who sell access to 'gray market' clusters in non-sanctioned regions (or sanctioned ones, if you don't ask too many questions). Prices have spiked to $8/hr for an H100, up from $2.50/hr last month. It's the Wild West.
- Spot Price (Jan '26): $2.50 / hr
- Spot Price (Feb '26): $8.00 / hr
- Availability: < 1% globally
- Impact: The 'Indie AI' movement is effectively paused.
The Rise of 'Inference-Only' Clouds
This shortage is forcing a shift. Providers like Groq (LPU) and Cerebras are booming because they are optimized for inference, not training. If you can't train, you might as well serve. The training game is becoming a billionaire-only club. If you don't have your own hardware by now, you are tenant farming on Microsoft's land.
