George Hotz and tinygrad: The Hacker Who Thinks NVIDIA Is a Scam

George Hotz is the kind of person the technology industry can’t quite figure out what to do with. He’s too talented to ignore, too chaotic to employ, and too controversial to celebrate without caveats. He’s also, depending on who you ask, either the most important AI infrastructure developer working today or a brilliant programmer who mistakes audacity for strategy.

His current project — tinygrad, a minimalist deep learning framework, and tiny corp, the company building AI hardware around it — is his most ambitious bet yet. It’s a direct challenge to NVIDIA’s monopoly on AI compute, built on the principle that the entire AI software stack is needlessly bloated.

The Resume of Chaos

Before tinygrad, Hotz compiled a resume that reads like a hacker’s hall of fame:

2007 (age 17): First person to carrier-unlock the iPhone. Apple’s lawyers sent threatening letters. Hotz posted the technique online anyway.

2010 (age 20): Jailbroke the PlayStation 3. Sony sued him. The case settled, but not before it became a landmark in digital rights law.

2015 (age 25): Founded comma.ai, a self-driving car company. When he announced it on Twitter, regulators questioned whether a startup could safely build autonomous vehicles. Five years later, comma.ai’s OpenPilot has been installed on over 500,000 vehicles and is arguably the best open-source self-driving system available.

2022 (age 32): Started tinygrad and tiny corp. The pitch: build a deep learning framework so simple it fits in ~10,000 lines of code, then build custom hardware to run it on.

Every project shares a common thread: Hotz identifies something he considers unnecessarily complex, controlled by a powerful gatekeeper, and attacks it with an absurdly small team using first-principles thinking.

What Is tinygrad?

tinygrad is a deep learning framework — the same category as PyTorch and TensorFlow. It lets you define neural networks, train them, and run inference. The difference is philosophical: while PyTorch has grown to millions of lines of code, tinygrad is intentionally kept under 10,000 lines.

# tinygrad: Train a simple neural network
from tinygrad import Tensor, nn

class SimpleNet:
    def __init__(self):
        self.l1 = nn.Linear(784, 128)
        self.l2 = nn.Linear(128, 10)

    def __call__(self, x):
        x = self.l1(x).relu()
        return self.l2(x)

model = SimpleNet()
optimizer = nn.optim.Adam(nn.state.get_parameters(model))

# Training loop
for x, y in dataloader:
    out = model(Tensor(x))
    loss = out.sparse_categorical_crossentropy(Tensor(y))
    optimizer.zero_grad()
    loss.backward()
    optimizer.step()

The code looks similar to PyTorch. But underneath, tinygrad’s architecture is radically different.

The Key Innovation: Lazy Evaluation and Hardware Abstraction

tinygrad doesn’t execute operations immediately. It builds a computation graph lazily, then optimizes and compiles the entire graph before execution. This approach enables:

1. Hardware portability. tinygrad generates kernels for NVIDIA GPUs (CUDA), AMD GPUs (ROCm/HIP), Apple Silicon (Metal), Intel GPUs (OpenCL), and even specialized accelerators — all from the same model code. You don’t rewrite your model for different hardware; tinygrad handles the translation.

2. Automatic kernel optimization. The graph compiler searches for optimal kernel configurations automatically. Traditional frameworks rely on hand-tuned kernels (cuDNN for NVIDIA, MIOpen for AMD). tinygrad generates and benchmarks its own.

3. Simplicity as performance. With a smaller codebase, there’s less overhead. tinygrad’s scheduler can find optimizations that PyTorch’s sprawling codebase misses because there are fewer layers of abstraction to work through.

Current Performance

In benchmark tests (as of early 2026), tinygrad’s performance varies by hardware:

Hardware	vs. PyTorch Performance	Notes
NVIDIA A100	75-90%	Close but CUDA/cuDNN is optimized over years
AMD MI250	95-110%	tinygrad’s hardware abstraction shines here
Apple M3 Max	100-120%	tinygrad’s Metal backend is excellent
Intel Arc	80-95%	Limited optimization so far

The AMD and Apple numbers are the story. On non-NVIDIA hardware, tinygrad is competitive or superior because it doesn’t depend on NVIDIA’s proprietary optimization stack. This is the entire strategic thesis: if you break NVIDIA’s CUDA lock-in, you can make non-NVIDIA hardware competitive.

tiny corp: The Hardware Play

In 2023, Hotz announced tiny corp — a company that would sell AI accelerator hardware running tinygrad’s software stack. The pitch was audacious: build a box that competes with NVIDIA’s AI hardware at a fraction of the cost.

The tinybox

The first product was the tinybox — a desktop AI workstation:

tinybox specs:
- 6x AMD Radeon RX 7900 XTX GPUs
- 144 GB total VRAM
- ~738 TFLOPS FP16
- Price: ~$15,000

Comparable NVIDIA setup:
- 6x RTX 4090 (if you could buy them)
- 144 GB total VRAM  
- ~600 TFLOPS FP16
- Street price: ~$12,000-15,000

NVIDIA A100 server (8x):
- 640 GB HBM2e
- ~2,496 TFLOPS FP16
- Price: ~$150,000-200,000

The tinybox isn’t competing with enterprise NVIDIA hardware on raw performance. It’s competing on the value proposition: “You can train and run serious AI models on commodity AMD GPUs using tinygrad, and you don’t need CUDA.”

The tinybox Pro

Announced in 2024, the tinybox Pro uses AMD’s Instinct MI300X accelerators — the same chips used by enterprise data centers. This is tiny corp’s bid for the serious AI training market.

The Strategic Vision

Hotz’s theory is this:

1. NVIDIA’s dominance is based on CUDA software lock-in, not hardware superiority alone
2. AMD and Intel hardware has competitive silicon, but their software stacks (ROCm, oneAPI) are immature
3. tinygrad provides a mature software stack that runs on ALL hardware
4. Therefore: tinygrad + commodity hardware = viable alternative to NVIDIA

If this theory is correct, it’s worth tens of billions. NVIDIA’s market cap is based on their moat being permanent. If tinygrad or something like it makes switching costs near-zero, NVIDIA becomes a commodity hardware company with commodity margins.

The Hotz Method

What makes Hotz unique isn’t just his technical ability — it’s his approach to building companies and technology.

Radical Transparency

Hotz streams his coding sessions live on Twitch. He discusses company strategy, financial decisions, and technical challenges in real-time with his audience. He publicly tracks tinygrad’s performance benchmarks and invites the community to contribute improvements.

This isn’t just marketing. It’s a development methodology. By streaming code review and development, Hotz gets immediate feedback from experienced programmers worldwide. The tinygrad community has contributed significant optimizations that no small team could have developed alone.

Small Teams

comma.ai has ~40 employees and produces software used in 500,000+ vehicles. tiny corp has ~15 employees and is building a competitive deep learning framework and hardware products. Hotz’s thesis: small teams with talented people outperform large teams with average people. Bureaucracy is the enemy of innovation.

Confrontational Communication

Hotz regularly calls out NVIDIA, AMD, Intel, and other companies on Twitter and in streams. He’s called CUDA “a prison,” described NVIDIA’s pricing as “a scam,” and publicly challenged Jensen Huang to benchmark competitions.

This communication style is polarizing. Critics call it unprofessional. Supporters call it honest. From a business perspective, it generates attention for tiny corp that billions in marketing spend couldn’t buy.

The Critics

Not everyone is convinced.

”tinygrad Can’t Match cuDNN”

NVIDIA has invested billions in CUDA and cuDNN over 15+ years. The hand-tuned kernels represent thousands of person-years of optimization. tinygrad’s auto-generated kernels, while impressive, haven’t matched NVIDIA’s peak performance on NVIDIA hardware. Critics argue this gap will never fully close because NVIDIA’s kernel engineers have access to hardware documentation and optimization opportunities that tinygrad can’t replicate.

”The Hardware Business Is a Trap”

Building hardware is a capital-intensive business with thin margins. Companies with far more resources — AMD, Intel, Qualcomm — have struggled to compete with NVIDIA in AI hardware. A 15-person startup taking on this challenge raises legitimate questions about sustainability.

”Hotz Gets Bored”

The most persistent criticism: Hotz starts ambitious projects, pushes them to a certain point, then moves on. comma.ai has persisted, but Hotz stepped back from day-to-day operations. Will the same happen with tiny corp?

Hotz’s response: “I don’t get bored. I get productive. When something works, I hire people to run it and move to the next problem.”

Why It Matters

Regardless of whether tinygrad “wins,” the project matters because it proves a point: NVIDIA’s moat is software, not hardware. If a 15-person team can build a competitive deep learning framework that runs on commodity hardware, the strategic implications are enormous.

Even if tiny corp never becomes a major hardware company, tinygrad’s existence pressures NVIDIA to keep CUDA competitive and — crucially — prevents them from raising prices unchecked. Competition, even from a scrappy underdog, benefits the entire AI ecosystem.

George Hotz is 36 years old. He’s already been a hacker legend, a self-driving car pioneer, and now a deep learning infrastructure builder. The common thread across all his work: find the biggest locked door in technology, and pick the lock. NVIDIA’s CUDA fortress is the biggest lock he’s tried yet.

Whether he picks it or not, watching him try is one of the most entertaining shows in technology.