the tiny corp

The tiny corp is George Hotz’s computer company behind the deep learning framework tinygrad and the tinybox computer for local AI training and inference. Its mission is “to commoditize the petaflop.” Hotz’s strategy begins with software for existing GPUs, with custom chips as an eventual ambition; computer sales fund the work.[2][1][5]

The framework predates the company’s funding announcement. Its earlier development explains the software foundation that tiny corp brought to its computer business.[2]

Hotz started tinygrad in October 2020 as a toy project to learn about neural networks; he later dated its first commit to October 17. By May 2023, it was running the model in openpilot. In his June 2023 interview, Hotz recalled imposing a thousand-line limit to discourage wasted code, then removing it because the restriction had become “a little code golfy at the end.”[2][5][1]

Tinygrad’s PyTorch-like frontend builds on a small set of operations. Hotz’s May 2023 description gave its intermediate representation 12 operations, expressing matrix multiplication and convolution through multiplication, sums and movement operations such as reshape and permute. Lazy execution enables fusion: in his June explanation, A × B + C becomes one kernel that computes the intermediate in local registers and writes the result to memory, avoiding a separate intermediate write and read.[2][1]

Hotz also emphasized visibility and correctness testing. In June 2023, he described graph output showing the kernels actually dispatched to the GPU and command-line debugging that reported each kernel’s FLOPS and memory accesses. His testops.py suite compared roughly a hundred small tinygrad and PyTorch examples in both forward and backward passes.[1]

On May 24, 2023, Hotz announced that tiny corp had raised $5.1 million, stating that the money was in the bank. He described the business as selling computers for more than they cost to make. The initial technical plan was to build a framework, runtime and driver for AMD chips, with the short-term goal of getting AMD onto MLPerf using tinygrad. His argument was that an AI chip company should establish usable, performant software before developing its own silicon.[2]

The June 20, 2023 Latent Space interview introduced tinybox as a $15,000 local AI computer with 738 FP16 TFLOPS and 144 GB of GPU memory. Its proposed out-of-the-box support for 65B FP16 LLaMA was explicitly subject to software development risks. Hotz described the practical challenges of fitting six GPUs into an under-desk machine: maintaining PCIe 4.0 connectivity, controlling fan noise and using one power outlet. Limiting GPU power was part of that design.[1]

AMD software stability became an immediate obstacle. In June 2023, Hotz said AMD demonstration applications had caused kernel panics on his machine. After he emailed AMD CEO Lisa Su, she connected him with engineers; according to Hotz, they reproduced the problem with two GPUs and supplied a prerelease ROCm 5.6 driver that fixed it.[1]

Recruitment centered on demonstrated work in tinygrad. The May announcement sought engineers in person in San Diego and required high-quality pull requests. By June, Hotz said the company should operate remotely. He outlined a progression from completing bounties, to proposing useful bounties, to full-time employment with equity.[2][1]

Driver and firmware difficulties later changed the product lineup. Matthew Connatser reported on March 25, 2024 that tiny corp would add an NVIDIA RTX 4090 tinybox while continuing the AMD RX 7900XTX version after finding a useful umr repository. The announced prices were $25,000 for NVIDIA and $15,000 for AMD, with June launches planned. The company presented NVIDIA as the choice for buyers who wanted something that “just works,” while acknowledging continuing crashes and hangs in the AMD version.[8]

In his December 29, 2025 retrospective, Hotz reported a six-person company and a computer-sales division making about $2 million in revenue a year. He described a stack spanning frontend, graph compiler, runtimes and drivers, and said work had begun to remove LLVM, aiming to drive AMD GPUs with only pure-Python dependencies. He also reported an AMD contract to get MI350X onto MLPerf for Llama 405B training, negotiated mostly in public on Twitter.[5]

On February 15, 2026, Hotz identified large-LLM training and generic GPU infrastructure as the company’s current focus. His longer-term tinybox vision was a local system that would update model weights through interactions with its user. He explicitly placed that adaptive product “many years away.”[6]

In its MLPerf Training v6.0 supplemental contribution, tiny corp reported submitting LLaMa 8B pretraining results on single-node AMD MI350 machines. That submission concerned a different model from the Llama 405B training goal in its reported AMD contract.[7][5]