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.

Who is tiny corp, and why should you care?

George Hotz founded tiny corp in November 2022, with the aim of porting machine learning instruction sets to hardware accelerators. In May 2023, tiny corp announced it had raised $5.1M to build computers for machine learning and develop tinygrad. Hotz believes NVIDIA's CUDA moat is not insurmountable and developed tinygrad -- an open-source deep learning stack with approximately 20,000 lines of code -- to challenge the CUDA ecosystem from the software side.

The company's existing hardware already ships. Tiny corp currently sells the red v2, powered by four AMD 9070XTs for $12,000, and the green v2 Blackwell, which costs $65,000 and has four RTX Pro 6000 Blackwell GPUs. The exabox is the logical next step up the product ladder -- from a workstation to a full compute cluster in a box.

For extreme scale, the exabox is detailed with 720x RDNA5 AT0 XL GPUs, aiming for ~1 exaflop of performance and over 25,000 GB of GPU RAM. The GPU choice is notable: tiny corp has historically leaned AMD, and the company was famously known for butting heads with AMD due to driver issues, with AMD CEO Dr. Lisa Su stepping in just to get things right. That relationship has since evolved into something more collaborative.

The real story: skipping the datacenter arms race

The broader context here is a datacenter market that has become nearly impossible to enter at small scale. Standard shell-and-core data center facilities often land around $10 to $12 million per MW, with a 2026 forecast commonly modeled near $11.3 million per MW depending on region and specification. AI-focused facilities can exceed $20 million per MW once liquid cooling, higher-voltage distribution, and dense rack layouts enter the design. The exabox, by contrast, delivers a full megawatt of AI compute for under $10M -- and you can drop it on a parking lot.

This is the crux of tiny corp's bet: if tinybox commoditizes the petaflop, exabox commoditizes the exaflop. The goal is to make serious AI compute available outside of hyperscaler lock-in. Over two-thirds of organizations are actively repatriating AI workloads from cloud to on-premise environments to regain control over spending -- and the exabox is a direct play into that trend, just at a scale that was previously only available to well-funded labs.

The software angle is just as important as the hardware

The exabox isn't just a pile of GPUs. Like the tinybox, it's 100% ready to run with tinygrad, PyTorch, and others. The tinygrad framework -- which compiles a custom kernel for every operation and aggressively fuses operations to minimize memory overhead -- is what lets the box behave as a single unified compute surface rather than a cluster you have to manually orchestrate.

The story is no longer just "buy a GPU"; it is about owning the stack: hardware tuned for a framework that compiles custom kernels, fuses ops aggressively, and aims to beat PyTorch on selected research workloads. That said, tinygrad remains alpha -- production users exist, but the framework's contract is closer to "move fast, expect sharp edges" than "enterprise SLA."

Who wins, who waits, and what comes next

The exabox is not for everyone. At $10M, the buyer profile is narrow: well-funded AI labs, sovereign compute initiatives, or enterprises that have done the math on cloud costs and decided to own the iron. On-premises infrastructure can achieve a breakeven point in under four months for high-utilization workloads, and owning the infrastructure can yield up to an 18x cost advantage per million tokens compared to Model-as-a-Service APIs over a five-year lifecycle.

The losers in this scenario are cloud GPU providers and traditional datacenter builders who rely on the complexity and capital intensity of the current model as a moat. This commoditizes petaflop-scale compute, challenging established hardware vendors and lowering AI development entry barriers. If tiny corp can actually deliver on the FLOPS/$ promise at exaflop scale, the pressure on hyperscaler pricing becomes real.

The risks are real too. Until tiny corp ships verifiable systems, the exabox program is roadmap-stage: treat FLOPS tables as narrative targets, not procurement specs. The GPU supply chain for 720 RDNA5 units is non-trivial, and the company remains small. But with only two external slots available for 2027, tiny corp isn't overcommitting -- they're stress-testing the manufacturing process at controlled scale before opening the floodgates.

The preorder page is live now. Whether or not you're in the market for a $10M shipping container, the exabox signals something important: the infrastructure layer of AI is being disaggregated, and the next wave of compute sovereignty may arrive on a flatbed truck.