George Hotz doesn't do things quietly. The hacker who once jailbroke iPhones from his parents' New Jersey home and later built a self-driving car startup has spent the past few years trying to make a different kind of statement: that serious AI compute doesn't have to cost a small fortune.
His latest venture, tiny corp, now sells what amounts to a provocative counter-argument to NVIDIA's enterprise pricing. While a DGX H100 system can run upward of $320,000, Hotz's TinyBox—a multi-GPU deep learning machine built from consumer graphics cards—starts at $12,000. It's an audacious pitch, perhaps more audacious than even Hotz anticipated when he launched the company in November 2022 with the goal to "commoditize the petaflop."
Whether it's working is harder to say.
The Hardware Proposition
TinyBox isn't trying to be subtle about what it is. The current lineup includes four configurations, though availability has been spotty. The entry model, dubbed TinyBox red v2, packs six AMD Radeon RX 7900 XTX gaming GPUs into a 12U chassis for $12,000. Move up the stack and you'll find the green v2 at $30,000, equipped with four NVIDIA GeForce RTX 5090s. Need more video memory? A $65,000 variant swaps in four RTX Pro 6000 Blackwell cards, delivering 384GB of VRAM. There's also a TinyBox pro v2 listed at $60,000, though the exact GPU configuration remains somewhat opaque in current documentation.
These aren't the sleek rack-mounted servers you'd find in a data center. TinyBox ships in a relatively compact form—about 16 inches deep, 90 pounds, designed to fit under a desk or squeeze into a small server closet. Dual 1600-watt power supplies keep the GPUs fed. A PCIe 4.0 x16 fabric connects everything, enabling GPU-to-GPU communication without forcing data through CPU bottlenecks. An AMD EPYC 7532 processor and 128GB of system RAM handle orchestration duties. The box arrives running Ubuntu 22.04, pre-loaded with both tiny corp's homegrown tinygrad framework and PyTorch for those who prefer familiar territory.
Early models shipped with six GPUs; newer green variants dropped to four. Blame GPU availability, evolving thermal considerations, or simple design iteration—tiny corp hasn't offered exhaustive explanations. What the company does emphasize: privacy. No mandatory cloud connection. No telemetry phone home. Just local compute you control, with BMC and IPMI access configured for remote management.
After raising $5.1 million in May 2023, tiny corp began shipping systems in mid-2024. Production has been made-to-order, with lead times quoted at anywhere from two to eight weeks. This isn't Dell churning out thousands of units a month. It's a startup run by a hacker who built something he wanted to exist.
The Math, Sort of
Positioning TinyBox against enterprise AI accelerators requires some mental gymnastics. When the red variant first shipped in August 2024 at $15,000 (later dropped to $12,000), tiny corp touted aggregate FP16 performance around 738 teraflops. The TinyBox Pro, announced in November 2024, claimed 1.36 petaflops for $40,000. Respectable numbers on paper, though direct comparisons get murky fast. Consumer GPUs can deliver impressive peak throughput, but they lack the memory bandwidth and capacity of purpose-built accelerators like NVIDIA's H100 or AMD's MI300X.
For small teams running experiments or fine-tuning models that actually fit in VRAM, though, the economics start making sense. Cloud GPU instances accumulate costs quickly. Enterprise systems often sit idle, depreciating while waiting for workloads. A $30,000 box you own outright begins looking reasonable when measured against ongoing rental fees—particularly for teams that value data sovereignty or simply want hardware they can physically touch.
The competitive landscape is getting crowded. Lambda Labs sells single-GPU desktops under $5,500. Exxact and BIZON offer four-GPU workstations ranging from $5,000 to north of $30,000 depending on configurations. Tenstorrent announced the TT-QuietBox 2 in March at a $9,999 starting price, using its own Blackhole RISC-V ASICs. The "AI in a box" market has clearly caught venture capital's attention, which means more players are coming.
Whether TinyBox's pricing holds up depends heavily on what you're running. Training large language models from scratch? Probably not. Fine-tuning open-source models, running inference at scale, or prototyping new architectures? Maybe. The use case matters more than the spec sheet suggests.
Turbulence and Pivots
Getting to market wasn't smooth. In March 2024, tiny corp publicly halted development of the AMD-based red variant, citing GPU instability under sustained AI workloads. Hotz aired his grievances on social media, complaining about driver bugs and firmware issues. It wasn't the kind of thing most hardware vendors admit publicly, but Hotz has never been most vendors.
The company pivoted to NVIDIA RTX 4090s for the green variant, announcing the shift "reluctantly" according to coverage in The Register. Intel Arc GPUs also got floated as a potential option. By August 2024, enough issues had been resolved to ship both AMD and NVIDIA configurations. Hotz posted that 13 units were in stock—a small batch, but tangible evidence of actual product.
Then, in January 2025, Phoronix reported that tiny corp was nearing a "completely sovereign software stack" for AMD GPUs through tinygrad. The instability narrative, it seemed, had given way to iterative software refinement. Whether that means the early AMD problems are truly solved or just better managed remains an open question. Formal third-party reviews of TinyBox performance remain scarce; most information comes from the company itself or scattered community discussions on Reddit and Hacker News.
Tiny corp did submit TinyBox systems to MLPerf Training v4.0 in June 2024, appearing among benchmark participants dominated by enterprise vendors. The results didn't break records, but they demonstrated the hardware could complete standardized training tasks without catching fire. For a newcomer, that counts as validation.
Software as Differentiator
Tinygrad—tiny corp's in-house deep learning framework—is arguably more central to the value proposition than the hardware itself. Unlike PyTorch or TensorFlow, which rely on vendor-specific backends like CUDA or ROCm, tinygrad compiles directly to native GPU code through its own runtime layer. The company claims this approach often outperforms CUDA Graph execution by using prebuilt hardware command queues that bypass heavier driver overhead.
The framework supports NVIDIA PTX, AMD HIP, Apple Metal, and Qualcomm runtimes. Documentation describes "HCQ" (hardware command queues) as a method to strip away abstraction layers and schedule compute more efficiently. Whether these claims hold up under rigorous testing is harder to verify. The default OS image ships with both tinygrad and PyTorch, offering an escape hatch for users who'd rather stick with familiar tooling.
Tiny corp positions TinyBox as the "best tested platform" for tinygrad, which makes sense when you control the full stack. Setup documentation covers power limit scripts, BMC configuration, and basic provisioning. But the software bet is risky. Asking developers to adopt a new framework—even one that promises performance gains—faces an uphill battle against ecosystem inertia. PyTorch isn't just code; it's tutorials, community support, and years of accumulated Stack Overflow answers.
The Target Customer
TinyBox isn't chasing enterprises shopping for GPU clusters to power trillion-parameter models. The intended audience is researchers, startup founders, and small AI teams who need capable hardware but can't justify six-figure purchases or prefer to avoid cloud lock-in. University labs running student ML courses. Indie developers fine-tuning open-source models. Companies handling sensitive data that won't touch AWS.
At $12,000, the red v2 undercuts many alternatives for multi-GPU setups. At $65,000, the Pro 6000 Blackwell configuration starts overlapping with high-end workstations from established vendors, raising questions about whether tiny corp's brand and unproven software stack justify the premium.
Availability, however, has been intermittent. All models currently show as out of stock on the company's website. Made-to-order production with multi-week lead times suggests small-batch manufacturing rather than inventory at scale. That's understandable for a small startup, but it also limits reach. Customers comfortable building their own rigs from off-the-shelf parts might just do that instead.
The Bigger Bet
What Hotz is really selling isn't hardware. It's a worldview—one where serious AI development doesn't require enterprise budgets or dependence on hyperscale cloud providers. Whether that vision scales beyond early adopters willing to tolerate rough edges and limited support remains uncertain.
For now, TinyBox occupies an odd middle ground: serious compute for teams who want to own their hardware, packaged with software designed to squeeze performance from consumer GPUs that were never meant for this kind of work. It's scrappy, occasionally unstable, and unmistakably a George Hotz project. Whether the market rewards that combination or demands something more polished will tell us a lot about where AI infrastructure is headed—and who gets to participate.
