George Hotz's Tiny Corp Slashes Exabox Price in Half to $5M

George Hotz’s Tiny Corp has revised exabox, its containerized AI cluster. The advertised price has fallen from about $10 million to $5 million, while 256 AMD MI300X accelerators replace the planned 720 unreleased RDNA5 GPUs. Delivery remains scheduled for Q3 2027. According to the current listing, two customer slots remained at publication, each requiring a refundable $100,000 deposit.

Replacing a future GPU with hardware that has shipped since late 2023 reduces procurement and scheduling uncertainty. The revision also lowers the advertised compute figure, nearly doubles aggregate GPU memory, halves the stated power requirement, and exposes several specifications that buyers will need clarified before purchase.

MI300X replaces the RDNA5 bet

Tiny Corp’s original plan called for 720 RDNA5 GPUs and roughly one exaflop of compute. Because AMD had yet to release that architecture, the design depended on future pricing, performance, power consumption, and availability.

AMD’s MI300X is a released CDNA 3 data-center accelerator. Published MI300X specifications list 192 GB of HBM3 and 5.3 TB/s of local memory bandwidth per accelerator. Across 256 units, the theoretical totals are:

• 49,152 GB, or about 49.2 TB, of HBM3
• 1,356.8 TB/s, or about 1.36 PB/s, of aggregate local memory bandwidth
• About 335 PFLOPS of dense FP16 or BF16 matrix compute
• About 669 PFLOPS with supported structured sparsity, or at dense FP8

The compute math has a gap

AMD publishes a peak of about 1.307 PFLOPS per MI300X for dense FP16 or BF16 matrix operations. The 2.615-PFLOPS figure requires supported structured sparsity, which skips operations on zeros arranged in a hardware-compatible pattern. The same 2.615-PFLOPS figure also corresponds to dense FP8, a lower-precision format. Tiny Corp’s 669-PFLOPS total therefore needs a precision and sparsity label before it can support direct comparisons.

The preorder page also estimates that exabox can complete a Kimi-sized, 3 × 10²⁴-FLOP training run in roughly 10 weeks at 50% model FLOPs utilization. MFU measures the share of theoretical accelerator throughput sustained after communication and other overhead. The published figures produce different estimates:

• At 335 PFLOPS dense and 50% MFU, the run takes about 29.6 weeks.
• At 669 PFLOPS and 50% MFU, the run takes about 14.8 weeks.
• A 10-week run requires about 496 PFLOPS of sustained compute, equivalent to a 992-PFLOPS peak at 50% MFU.

The 10-week estimate aligns more closely with the former one-exaflop design. A purchase agreement should define the precision, sparsity assumptions, FLOP-counting convention, and benchmark used for the updated system.

A container still needs a site

Tiny Corp’s infrastructure rationale is to package compute, networking, power distribution, and cooling inside a shipping container. The updated listing says the chiller is included and describes the system as ready to run. Installation still requires a suitable pad, a 500 kW 480 V electrical service, network connectivity, permits, and an operating plan for maintenance and failures.

A continuous 500 kW draw would consume 4.38 GWh per year. Electricity alone would cost about $438,000 annually at $0.10 per kWh or $657,000 at $0.15 per kWh, before demand charges, staffing, financing, maintenance, and site work. Actual consumption will depend on utilization and whether the quoted 500 kW represents typical or maximum draw.

One box, 256 memory domains

The listing says the system has at least 400 Gbps of connectivity and can train as one unit. It does not specify whether that figure applies per accelerator, host, switch, or fabric uplink. A 400 Gbps link carries 50 GB/s before protocol overhead, more than 100 times less than one MI300X’s local HBM bandwidth. That ratio is common in clusters, making topology, oversubscription, collective-communication performance, and model sharding central to real throughput.

The 49.2 TB memory figure and 1.36 PB/s bandwidth figure are sums across all accelerators. Each MI300X accesses its own 192 GB at local HBM speed, while cross-GPU data travels through the interconnect. Large models can span the cluster, but the aggregate totals do not behave like one physically shared memory pool.

The software promise covers tinygrad and PyTorch out of the box. Tiny Corp has also invested in open driver work for AMD hardware. Standardizing on MI300X gives that effort a fixed target and offers an established PyTorch path through ROCm. Production readiness will still depend on exact framework versions, RCCL support, orchestration, checkpointing, monitoring, fault recovery, and upgrade policies.

Ownership math depends on utilization

At $5 million, the advertised system works out to about $19,500 per accelerator and $102 per gigabyte of HBM. Those allocation ratios cover the container, hosts, fabric, power distribution, and cooling alongside the GPUs.

Public rental listings have placed MI300X capacity around $2.73 per GPU-hour, with some reserved offers in the low $2 range. Renting 256 GPUs continuously at $2.73 would cost about $6.1 million per year, while a $2 rate would cost about $4.5 million. Headline rental capacity may be unavailable at that scale, and the ownership calculation must include electricity, staffing, maintenance, financing, downtime, and the cost of idle hardware. The purchase case is strongest for organizations with sustained utilization, suitable facilities, and firm data-residency requirements.

Terms buyers need in writing

A $100,000 refundable deposit secures a place in a delivery schedule more than a year away. Before converting that reservation into a purchase, buyers will need contractual answers covering:

• Final price, payment milestones, shipping, taxes, installation, and refund timing
• GPU substitution rights, procurement guarantees, and remedies for delivery delays
• Host CPU count, system RAM, local storage, and checkpoint capacity
• Fabric topology, bandwidth per endpoint, oversubscription, and switch redundancy
• Measured all-reduce performance and end-to-end training benchmarks
• Supported precisions, sparsity assumptions, and MFU methodology
• PyTorch, ROCm, tinygrad, driver, and collective-library versions
• Warranty coverage, spare parts, remote support, and hardware replacement times
• Site requirements, acceptance tests, and performance guarantees

On the published specifications, the revision lowers the entry price, reduces silicon schedule risk, increases aggregate HBM, and cuts the stated power requirement. It also lowers the advertised compute figure and leaves the training-time estimate unreconciled. Exabox remains a Q3 2027 preorder whose distributed performance, support model, and total operating cost require contractual detail and independent measurement.