AMD is being challenged to build a GPU that doesn’t exist yet

A small startup is betting its entire business model on hardware nobody has made, and the question it raises is bigger than the GPU itself.

There is a gap in the AI hardware market that most people aren’t talking about. NVIDIA’s 96GB GPU starts at $8,000. For small AI labs, independent developers, and startups trying to run large language models without a hyperscaler’s budget, that price is effectively a wall. TinyCorp, a small AI startup founded by George Hotz, just decided to run straight at that wall.

On March 7, 2026, TinyCorp published a fundraising proposal built entirely around a GPU that AMD hasn’t built yet. The plan: raise $20 million, buy a building in Oregon, pre-order 3,000 units of an RDNA 5 GPU with 96GB of VRAM at $2,500 each, and sell AI inference computing power to the market. Clean, simple, and dependent on AMD making a product it has given no indication it plans to make.

What makes this more than just a bold pitch is who is behind it. Hotz was the first person to unlock an iPhone at 17, cracked the PlayStation 3’s security system, and built Comma.ai’s self-driving stack largely on his own. When he says “we’ll build the board ourselves if AMD won’t,” it doesn’t read like a bluff. It reads like a man who has done harder things.

The bigger obstacle isn’t AMD’s willingness , it’s the global memory crisis. In 2026, AI data centers are consuming memory at a pace the industry wasn’t designed to handle. GDDR7 is in short supply worldwide, NVIDIA has been cutting RTX 50-series production by up to 40% because of it, and VRAM now accounts for more than 80% of the bill of materials on some high-end GPUs. Getting a 96GB card to $2,500 in this environment requires memory prices to drop significantly. Possible? Yes. Likely by mid-2027? That’s the real gamble.

What TinyCorp is really doing , beyond the fundraising and the hardware specs — is forcing a question AMD hasn’t had to answer publicly: is the company willing to compete at the high end, or has it permanently ceded that territory to NVIDIA? The RDNA 5 generation could be AMD’s clearest opportunity in years to challenge that assumption. TinyCorp is simply making sure the pressure is visible.

Whether this specific plan succeeds or not almost doesn’t matter. The fact that a startup is designing a business model around a GPU that doesn’t exist yet — and that the math could work if the right decisions get made, says something important about where AI computing is heading and who gets left out if prices don’t come down.

If you want the full breakdown , the technical details behind the 96GB spec, the history between TinyCorp and AMD, and why the GDDR7 shortage makes this harder than it looks, the complete article is here.