George Hotz's tiny corp Drops a $10M Datacenter-in-a-Box Called exabox

George Hotz's tiny corp just opened preorders for the exabox: a 20-foot shipping container packed with enough GPU compute to run frontier-scale AI training runs, no datacenter lease required. The preorder deposit is $100,000 -- fully refundable and credited toward the ~$10M purchase price -- and tiny corp says it will only take two slots for external customers in 2027.

A datacenter in a box

The pitch is almost aggressively simple. It's a 20ft shipping container that needs a megawatt of power (208V or 415V three-phase) and is self-contained for cooling and weatherproof. As tiny corp put it on X: they considered raising a round to buy a datacenter, but with exaboxes they don't have to and can build out at their own pace -- exaboxes just require a concrete slab and a big plug.

The specs are serious. The exabox delivers ~1 exaflop of compute, powered by 720x RDNA5 GPUs, 25,920 GB of GPU RAM, and 1.2 PB/s of memory bandwidth. The whole box will be connected at at least 400 Gbps and is capable of training as one unit. And unlike a rack of loosely coupled nodes, at 50% MFU (model flop utilization -- the fraction of theoretical peak compute actually used), it can do 3e24 (Kimi-sized) training runs in 10 weeks, and with tinygrad software it will function as one big GPU, though it is made up of normal computers and you can also use PyTorch.

The numbers that matter

• Preorder deposit: $100,000 (fully refundable, ~1% of purchase price)
• Full purchase price: under $10M
• Compute: ~1 exaflop (FP16)
• Memory: 25,920 GB GPU RAM, 1.2 PB/s bandwidth
• Networking: 400 Gbps interconnect
• Power requirement: 1 megawatt
• Form factor: 20ft shipping container, self-cooled
• Target ship date: Q2 or Q3 2027
• External slots available: ~2

At launch, tiny corp claims it should be the absolute best bang for your buck at the price point with respect to FLOPS/$, GB/$ and GB/s/$. That's a bold claim, but not an empty one. The existing tinybox is described as likely the best performance/$ and was benchmarked in MLPerf Training 4.0 against computers that cost 10x as much. MLPerf is the industry-standard benchmark for AI training throughput -- not a synthetic test, but real training on real models.