AI Startup TinyCorp Demands AMD Build a 96GB RDNA 5 GPU — But There’s a Catch

It’s bold. It’s audacious. It might also be completely unrealistic. AI startup TinyCorp — the company behind the open-source Tinygrad framework — has gone public with a business plan that hinges on one very big ask: AMD releasing a 96GB VRAM RDNA 5 consumer GPU at around $2,500 a unit.

TinyCorp’s Pitch: The Numbers Behind the Idea

Why This Plan Is Fascinating — and Flawed

The core idea is genuinely clever. TinyCorp’s business model revolves around selling compute power on platforms like OpenRouter using AMD’s consumer GPUs — essentially becoming a low-cost AI compute provider by buying consumer cards instead of enterprise datacenter silicon. With NVIDIA’s only 96GB GPU being the RTX PRO 6000 Blackwell, currently retailing between $8,000 and $10,000 per unit, a $2,500 AMD alternative would be a genuine market disruption.

The problem? That GPU doesn’t exist yet — and may never exist in the form TinyCorp needs. TinyCorp’s entire plan depends on AMD releasing 96GB RDNA 5 SKUs, which analysts describe as a ‘flawed’ assumption given the current global GDDR7 memory shortage and supply constraints. Undeterred, TinyCorp has even suggested it will build its own board with RDNA 5 silicon if AMD won’t — a statement that’s equal parts ambitious and unhinged.

To AMD’s credit, the ask isn’t entirely without logic. TinyCorp argues that an RDNA 5 GPU with a 512-bit bus using 3GB GDDR7 modules double-sided like Blackwell could theoretically hit 96GB — and Lisa Su would be “dumb not to make one,” in TinyCorp’s own words. For anyone tracking AI GPU availability and the growing demand from smaller AI compute startups, this story highlights a real gap in the market — even if TinyCorp’s solution is, for now, more wishful thinking than business plan.

FAQs