One 24GB Card Ate a $400K Rack

A lab at Tsinghua quietly pushed a project to GitHub, and it does something faintly absurd: it replaces a $400,000 H100 rack with a single 24GB graphics card. The trick is quantization, squeezing a large model down to lower precision so it fits in a fraction of the memory. The old rule was simple. Want to run a big model, rent a big rack. That rule just met a card sitting in somebody’s drawer, while Paxis and Metis argue it out at a cafe under the Eiffel Tower.

Source: RT @dunik_7: a Tsinghua University lab just put a project on GitHub that replaces a $400,000 H100 rack with a single 24GB graphics card · twitter

What this means for ThakiCloud

Here is why this cheers us up. Every time the bar for running a big model drops, running it yourself on hardware you own stops being a fantasy and starts being a plan. That is the whole point of ThakiCloud’s bet on on-prem and sovereignty, which just means keeping your models, data, and infrastructure under your own control. Metis serves these squeezed-down models efficiently inside your own facility, and Paxis orchestrates the agents on top. No rack to rent, no invoice unrolling across the floor. You can start with a single card, and the sovereignty was in your drawer the whole time.

---

An auto-generated comic riffing on this week’s industry news.