Tiny Corp makes tinybox configurable, starting at $7,000 without GPUs

Tiny Corp, George Hotz’s hardware company and the team behind the tinygrad framework, has opened its deep-learning workstation to customer-selected memory, storage and GPUs. The configurable tinybox starts at $7,000 with no accelerators, while one published build with Blackwell GPUs costs about $64,000. The change replaces Tiny Corp’s fixed configurations and gives teams a way to reuse compatible cards or add compute over time.

Four slots, zero GPUs at entry

• Platform: G4 server platform with PCIe 4.0 connectivity.
• Processor: Water-cooled, 32-core AMD Epyc CPU.
• Motherboard: Server board with a baseboard management controller for remote monitoring, power control and console access.
• Memory: 32 GB of DDR4 in the base system, configurable up to 192 GB.
• Storage: 1 TB boot SSD in the base system, with larger options available.
• GPU capacity: No GPUs at the $7,000 price; up to four supported cards, depending on their power requirements.
• Power supply: 1,700 W.

Each GPU slot receives a dedicated PCIe 4.0 x16 connection to the CPU instead of sharing an upstream link through a PCIe switch. That topology preserves host-to-device bandwidth. Multi-GPU scaling still depends on peer-to-peer support, drivers, the machine-learning framework and the workload.

The configurator exposes memory, storage, CPU and GPU choices. Published examples include a build with 192 GB of system memory, two NVIDIA RTX PRO 6000 Blackwell GPUs, 16 TB of storage and an Epyc Genoa processor for about $64,000. Actual GPU count depends on the cards’ power draw and the combinations Tiny Corp supports.

Fixed configurations loosen up

Tiny Corp previously sold tinybox systems as sealed configurations. The original documentation described red, green and green v2 models with six Radeon RX 7900 XTX, six GeForce RTX 4090 or four GeForce RTX 5090 GPUs. The company’s FAQ also rejected customization as a way to limit costs and simplify quality control.

Customer-selected accelerators reduce Tiny Corp’s dependence on the availability of any single GPU model. That flexibility is relevant after the company publicly complained about inconsistent RTX 5090 supply, which complicated production of its fixed green systems.

The chassis also targets office use with low-speed Noctua fans, a water-cooled CPU loop and GPU installation that requires few tools. Tiny Corp claims an acoustic target below 50 dB. The published figure lacks a stated measurement distance and workload, which limits comparisons with other workstations.

A step below the eight-GPU rack

The configurable model sits below the tinybox pro v2, a 5U rack-mountable system with eight RTX 5090 or RTX PRO 6000 GPUs. The pro model provides PCIe 5.0 x16 connectivity, two 128-core AMD Epyc processors, 384 GB of system memory and four 2,000 W power supplies. It requires 208 to 240 V electrical service.

The FP16 column reports theoretical half-precision throughput in trillions of floating-point operations per second. These are manufacturer figures; observed performance will vary with model architecture, numerical precision, software support and communication overhead between GPUs.

Ordering brings wire transfers and export checks

• Payment: Tiny Corp accepts bank wire transfers and requires payment within five days of order confirmation.
• Lead time: Systems are built to order and are expected to ship within one to eight weeks.
• NVIDIA compliance: NVIDIA orders may require a Know Your Customer form covering the intended use and installation address. Availability is subject to U.S. export controls.
• Returns: The return window is 30 days, with a 20% restocking fee and return shipping paid by the buyer.
• Power planning: The base power supply is rated at 1,700 W, so circuit capacity and sustained load requirements depend on the selected GPUs.

A lower entry point for staged builds

The GPU-free configuration changes how teams can procure a multi-GPU workstation. Organizations with compatible accelerators can reuse them, while others can buy the server platform first and add supported cards as budgets and supply allow. The $7,000 starting price excludes the system’s largest compute expense, so total cost depends heavily on GPU choice, power consumption and required memory capacity.

Teams evaluating the system will need to match aggregate GPU memory to model size, system RAM to data-loading requirements and PCIe behavior to the intended training or inference workload. Framework compatibility, driver support and GPU-to-GPU transfer performance may have more effect on throughput than the chassis’s four-slot capacity.

Hotz has separately previewed an “exabox” concept with a proposed exaflop of performance, 720 RDNA5 AT0 XL GPUs, 25,920 GB of GPU memory and an estimated $10 million price. Tiny Corp has provided no launch date or ordering details for that system.