George Hotz Wants to Sell You a $12,000 AI Supercomputer — and 221 Hacker News Comments Can’t Stop Arguing About It

The idea of running a 120-billion-parameter model in your own office, with zero cloud dependency, sounds like a pipe dream. George Hotz — the teenager who once jailbroke the first iPhone and hacked the PS3 — thinks it should be as normal as buying a workstation. His company, tiny corp, ships the Tinybox: a line of GPU-packed machines designed to bring serious AI compute to anyone willing to plug in two power cables.

On March 21, the Tinybox hit the top of Hacker News with 370 points and over 220 comments. The debate was fierce — some calling it the future of AI independence, others questioning whether a $65,000 box makes sense when cloud APIs keep getting cheaper. Here’s what the fuss is about.

What Exactly Is the Tinybox?

The Tinybox is a 12U rack-mountable (or freestanding) computer stuffed with multiple high-end consumer or professional GPUs, pre-loaded with Ubuntu 22.04, tinygrad (tiny corp’s own neural network framework), and PyTorch. It ships from San Diego, worldwide, typically within a week of payment.

There are currently two models in stock, plus an upcoming beast:

Model	GPUs	VRAM	Compute (FP16)	Power	Price
Red v2	4x AMD Radeon 9070 XT	64 GB	778 TFLOPS	1x 1600W PSU	$12,000
Green v2 Blackwell	4x RTX Pro 6000 Blackwell	384 GB	3,086 TFLOPS	2x 1600W PSU	$65,000
Exabox (2027)	720x RDNA5 units	25,920 GB	~1 EXAFLOP	600 kW	~$10M

Earlier models — the original Red (6x 7900 XTX, $15K), Green (6x 4090, $25K), and Pro (8x RTX 4090, $40K delivering 1.36 PetaFLOPS) — helped tiny corp establish its hardware credibility and appeared in MLPerf Training 4.0 benchmarks, where the team claims performance competitive with machines costing 10x more.

The Exabox, announced for 2027, is the ambitious endgame: an actual exaflop in a shipping container-sized enclosure weighing 20,000 pounds.

The George Hotz Factor

You can’t talk about the Tinybox without talking about the person behind it. George Hotz (geohot) has a track record of doing things people said couldn’t be done — or shouldn’t be done. After the iPhone jailbreak and PlayStation saga, he founded comma.ai in 2015 to build open-source self-driving technology. He left comma.ai’s day-to-day operations in 2022 and departed entirely in November 2025.

Tiny corp was founded in November 2022 with a specific thesis: AI compute is too expensive and too concentrated. The company raised $5.1 million in 2023 to build both the tinygrad framework and the Tinybox hardware. Tinygrad itself now powers the driving model in openpilot (comma.ai’s software), running on Snapdragon 845 GPUs — replacing Qualcomm’s own SNPE with better performance.

The pitch is what Hotz calls “commoditizing the petaflop.” Instead of renting GPU time from hyperscalers at unpredictable prices, you buy a box, own it outright, and run whatever models you want with zero API costs and complete data privacy.

How the Tinybox Stacks Up Against Alternatives

The local AI hardware market has gotten crowded. Here’s how the Tinybox competes at different price points:

Under $5,000 — NVIDIA DGX Spark & Mac Studio

The DGX Spark ($3,999) packs 128 GB of unified memory and one petaflop of FP4 performance on a Grace Blackwell chip. It’s great for prototyping and inference on models up to ~70B parameters. The Mac Studio with M3 Ultra offers up to 512 GB unified memory with 819 GB/s bandwidth — ideal for large-context inference. AMD’s Strix Halo-based systems (like the Framework Desktop at $2,348) offer 128 GB unified memory at competitive bandwidth.

These are inference-focused machines. For training or running the largest open-weight models at production speeds, they hit a wall.

$12,000–$65,000 — Tinybox territory

This is where tiny corp plays. The Red v2 at $12,000 targets cost-conscious buyers who want raw FP16 throughput on AMD hardware. The Green v2 Blackwell at $65,000 is for organizations that need 384 GB of VRAM — enough to run very large models without aggressive quantization.

DIY builds

The Hacker News crowd pointed out — loudly — that you can theoretically build a comparable rig yourself for less. One commenter estimated building a Red v2 equivalent for under $9,000. This is true in the same way you can build your own rack server: possible, but you’re handling thermal design, power distribution, driver compatibility, and ongoing maintenance yourself. The Tinybox ships as a tested, integrated unit with tinygrad pre-configured.

Cloud GPUs

For intermittent workloads, renting H100s or A100s from cloud providers may still be cheaper. But for teams running inference 24/7, the math shifts. A Tinybox Red v2 at $12,000 pays for itself in a few months compared to continuous cloud GPU rental — assuming you can keep it utilized.

What the Hacker News Crowd Actually Said

The March 21 thread surfaced several recurring themes:

The power problem. The Green v2 Blackwell requires two 1600W power supplies, exceeding what a standard US 120V/15A circuit can deliver. You need two dedicated circuits or a 240V outlet. Commenters noted this is a minor expense for a $65,000 machine, but it’s a real consideration for smaller offices and home setups. Tinybox does include a power-limit script — running sudo power-limit 150 lets you operate the Red v2 on a single 120V circuit at reduced performance.

The 120B parameter claim. The Hacker News title mentioned running 120B-parameter models, but this drew skepticism. With 64 GB VRAM on the Red v2, running a 120B model requires heavy quantization and limits context windows to around 4K tokens. The Green v2 Blackwell with 384 GB VRAM handles this much more comfortably.

The website’s tone. Several commenters found tinygrad.org’s writing abrasive — particularly the hiring page that requires existing contributions to the project. Others saw it as refreshingly authentic in an AI industry full of corporate speak. One comment praised the site for being “extremely NOT AI-generated.”

Who is this for? The thread identified the core audience: organizations that cannot or will not send data to third-party cloud providers. Healthcare, finance, defense, and companies building proprietary models topped the list. For hobbyists and researchers, the Red v2 at $12,000 sits in an interesting sweet spot — serious enough for real work, cheap enough to justify as a capital expense.

The Bigger Picture: Why Local AI Hardware Matters Now

The timing of the Tinybox buzz isn’t accidental. Several trends converge:

Open-weight models like DeepSeek, Llama, and Qwen have reached quality levels where running them locally is genuinely practical — not just a privacy flex. Inference API costs from providers like OpenAI and Anthropic keep dropping, but they still add up at scale, and you’re still sending your data to someone else’s servers. Regulatory pressure around data residency (GDPR, HIPAA, emerging AI regulations) is pushing more organizations toward on-premise solutions.

The Tinybox isn’t the only player in this space, but it represents a specific bet: that there’s a meaningful market for turnkey, high-performance AI hardware that sits between a gaming PC with one GPU and a million-dollar data center installation. Whether that market is large enough to sustain a hardware company remains an open question — but the Hacker News reaction suggests the appetite is real.

FAQ

How much does a Tinybox cost?
The current lineup starts at $12,000 for the Red v2 (4x AMD Radeon 9070 XT, 64 GB VRAM) and goes up to $65,000 for the Green v2 Blackwell (4x RTX Pro 6000, 384 GB VRAM). Earlier models like the Pro (8x RTX 4090) were priced at $40,000. Payment is by bank transfer only, with a 20% restocking fee on returns within 30 days.

What models can the Tinybox run?
The Red v2 with 64 GB VRAM can comfortably run models up to ~70B parameters at reasonable quantization levels. For 120B+ parameter models with full context, you’ll want the Green v2 Blackwell and its 384 GB of VRAM. Both machines come pre-loaded with tinygrad and PyTorch, so you can run any model compatible with those frameworks.

How does the Tinybox compare to NVIDIA’s DGX Spark?
Different price brackets, different strengths. The DGX Spark ($3,999) is a compact inference-focused machine with 128 GB unified memory. The Tinybox Red v2 ($12,000) has less total memory but significantly more raw FP16 compute — 778 TFLOPS vs. roughly 200 TFLOPS on the Spark. The Tinybox is better suited for training workloads and high-throughput inference; the Spark is more accessible for prototyping and smaller models.

Does the Tinybox require special electrical setup?
The Red v2 runs on a single 1600W PSU and can operate on a standard 120V circuit using the built-in power-limit utility. The Green v2 Blackwell needs two 1600W PSUs and realistically requires either two dedicated 120V circuits or a 240V outlet. For a $65,000 machine, the electrical upgrade is a minor additional cost.

Who is the Tinybox designed for?
Organizations and individuals who need always-on, private AI compute: companies with data sensitivity requirements (healthcare, finance, legal), AI startups building proprietary models, research labs, and power users who want to avoid recurring cloud costs. It’s not aimed at casual hobbyists — the DGX Spark or a single consumer GPU is a better entry point for experimentation.

Top AI Product

Leave a comment Cancel reply