AMD sent its top AI accelerator card, the MI300X, to the startup tiny corp, whose founder George Hotz unlocked an iPhone at age 17. Hotz believes NVIDIA’s CUDA moat is not insurmountable; he developed the open-source deep learning stack tinygrad with approximately 20,000 lines of code to challenge the CUDA ecosystem from the software side. As of May 2026, NVIDIA’s market capitalization is about seven times that of AMD, narrowing from 16 times a year earlier. tiny corp also sells Tinybox computers based on AMD hardware, aiming to commoditize computing power.

Author and source: AI New Era

In March 2025, George Hotz published an article on his personal blog titled simply "AMD YOLO," meaning "all in on AMD, no backup plan."

Hotz wrote in the article: AMD is shipping the two MI300X cards (AMD’s top-tier data center AI accelerators) we requested, and they are already on their way.

The MI300X is AMD’s flagship product for AI computing power, typically in high demand with cloud providers and large model companies. Now, it’s being shipped to tiny corp—a small company that didn’t even have an office and operated solely through GitHub and Discord.

The recipient of this package is known in the crypto community as geohot—George Hotz.

Geohot first became known in 2007 through a video in which, at the age of 17, he unlocked an iPhone for the first time.

In 2007, 17-year-old George Hotz demonstrated the world's first unlocked iPhone on camera; the video, which garnered nearly two million views, made the name "geohot" famous throughout the hacking community.

Later, he was sued for cracking Sony's PS3. This time, he aimed to do something even bigger: undermine NVIDIA's CUDA moat from the software side.

CUDA is an ecosystem, not a moat.

"CUDA is an ecosystem, not a moat."

Hotz bluntly stated in his blog that CUDA (NVIDIA's GPU computing platform) is not the moat people imagine it to be—it's merely an early ecosystem.

He posted a screenshot of a Twitter post in his blog. Back in January 2025, he referred to AMD sending the chip as a “cultural test” to see whether AMD was truly willing to invest in software.

By March, in this blog post, he concluded that AMD had passed the test. He believed AMD would not abandon software, and if that were the case, NVIDIA had no reason to be 16 times more expensive than AMD.

This means that at the time, NVIDIA’s market capitalization was about 16 times that of AMD, yet the actual hardware performance between the two companies wasn’t nearly that far apart. AMD even achieved double the throughput of Tensor Cores with its RDNA4 architecture, while NVIDIA artificially halved performance on its own cards.

Where does the 16-fold difference come from? Hotz’s answer is software—more precisely, software complexity and the ecosystem lock-in that emerges from that complexity.

Developers are accustomed to CUDA, with their toolchains built around it, so even AMD’s hardware is left untouched. What has held AMD back isn’t its chips—it’s the lack of a competitive software stack.

Of course, this is Hotz's assessment, not an official conclusion from AMD.

He didn’t just talk—he personally invested $250,000 to buy AMD, publicly expressing a bullish stance with a five-year horizon. He wrote in his blog:

Either NVIDIA is significantly overvalued, or AMD is significantly undervalued.

Hotz wrote this in March 2025, when NVIDIA's market capitalization was approximately 16 times that of AMD.

By May 2026, NVIDIA’s market capitalization was approximately $5.2 trillion, while AMD’s was around $760 billion, narrowing the gap to roughly sevenfold. Over the year, AMD’s stock price surged significantly, driven by demand for AI data centers, outpacing NVIDIA’s gains.

Of course, this does not prove Hotz is right. AMD's reassessment primarily stems from its own GPU shipments and earnings reports, but the direction of market sentiment does confirm Hotz's assessment: the 16-fold gap is not as solid as it seemed.

From Unlocking Your iPhone to Challenging Hash Power Dominance

What gives a programmer who has never designed a chip the audacity to challenge NVIDIA? The story begins when he was 17.

In the summer of 2007, Apple released the first-generation iPhone, exclusively tied to AT&T. At age 17, Hotz, a T-Mobile user who wanted the iPhone without switching carriers, decided to take the device apart.

According to The New Yorker, he used a screwdriver to open the back cover, located the carrier-restricting baseband processor, soldered on a wire, supplied it with voltage, and disrupted its code. The next morning, he announced to the camera: “This is the world’s first unlocked iPhone.”

The video garnered nearly two million views, making him the world's most famous hacker at the time. Two years later, he turned his attention to the Sony PS3, cracking the supposedly unbreakable gaming console, after which he was sued by Sony and eventually reached a settlement.

In 2011, Sony sued George Hotz for cracking the PS3; the case was ultimately settled out of court, with Hotz accepting a permanent injunction and promising not to touch Sony products again.

Hotz offers a simple definition of a hacker: a hacker is to computers what a plumber is to pipes.

His methodology hasn’t changed in over a decade: find the component within a closed system that can “communicate” with you, then figure out how to make it comply.

Rewrote an entire AMD stack with 20,000 lines of code.

The iPhone and PS3 demonstrated Hotz's capabilities.

But to leverage CUDA, one person isn’t enough—you need something real, and that’s tinygrad. It’s an open-source deep learning stack led by Hotz.

In fact, Hotz’s obsession with breaking closed systems had long extended beyond consumer electronics to the automotive industry. In 2015, he founded the autonomous driving company comma.ai and nearly single-handedly modified a system in his garage capable of navigating highways, directly competing with Tesla.

comma.ai is still operational today, and its open-source driving model, openpilot, runs on tinygrad.

The official GitHub describes it as an end-to-end system encompassing everything from tensor libraries, automatic differentiation, IR (intermediate representation), and compilers, to just-in-time compilation, graph execution, and the optimizers and data loading components required for training.

The standout feature of this system is its simplicity.

tinygrad already supports multiple backends, including CUDA, AMD, METAL, QCOM, and WEBGPU. While other frameworks require implementing an entire complex instruction set to support new hardware, tinygrad only needs the new hardware to correctly execute about 25 fundamental operations—such as addition, subtraction, multiplication, and division—to be integrated.

But Hotz's real focus has been on AMD.

In January 2025, Tiny Corp had already developed its own drivers, runtime, libraries, and simulator. At that time, this AMD stack consisted of approximately 12,000 lines of code and was just one RDNA3 assembler away from being "fully autonomous."

By March, Hotz announced on his blog that tiny corp had developed a complete AMD stack from hardware to the PyTorch layer, with the only exception being the LLVM compiler framework.

He also added: Developers don't even need to learn how to write in tinygrad—they can continue using their familiar PyTorch code, and the underlying system will automatically switch to this AMD stack.

12,000 lines only represent the size of the AMD stack in the January 2025 version. The entire tinygrad project has continued to grow, reaching 22,853 lines by v0.13.0; Hotz's own figure at the end of 2025 was 18,935 lines excluding tests.

But whether 12,000 or 18,000, both are astonishing comparisons within the industry context.

A software stack capable of driving a GPU and running training can easily be millions of lines of code. In Hotz’s words, tinygrad is 1,000 times smaller than these.

This means that CUDA's complexity is not grounded in physical laws; it can be rewritten from scratch by a small team with a minimalist approach, which is precisely what Hotz believes.

Tinybox places a mining rig on the shelf.

If the story ended here, tinygrad would still just be a project on GitHub, but Hotz’s ambitions extend beyond code.

Tiny Corp is now selling a computer called the Tinybox. The official website lists several models—Red, Green, Pro, and Exa—with full details on specifications, pricing, and shipping information. Orders are shipped within one week of payment.

Tinybox previously competed against machines costing approximately 10 times more in the MLPerf Training 4.0 benchmark. According to Hotz’s blog post at the end of 2025, this line of computers generates about $2 million in annual revenue, which he uses to fund the entire R&D operation of Tiny Corp.

In the same blog post, Hotz also mentioned that they signed a contract with AMD to use the MI350X for training Llama 405B on MLPerf, and the negotiation process for this contract was largely conducted publicly on Twitter.

Tiny Corp has distilled its mission into one sentence: commoditize the petaflop.

Petaflop is a unit of computing power, representing one quadrillion (10^15) floating-point operations per second, commonly used to measure the performance of supercomputers and AI computing power.

What Hotz wants to do is commoditize petaflops of floating-point operations per second, making AI accessible to everyone.

When computing power can be priced, ordered, and shipped within a week like consumer electronics, the narrative of "computing power scarcity" may begin to loosen.

A new route

The significance of the Hotz story does not lie in the sensational headline of "a single hacker taking on NVIDIA and ending CUDA." CUDA remains the default choice for the vast majority of developers today.

What truly matters is the assumption Hotz bet on: that the complexity of the AI software stack can be compressed to an extremely small size. Once proven correct, the portion of NVIDIA’s valuation supported by CUDA may need to be recalculated.

And this gamble is no longer just Hotz’s alone. In his December 2025 article, “Five Years of tinygrad,” he wrote: “The first line of code was submitted in October 2020; five years later, the company now has six people, and many have invested years of effort.”

In Hotz’s view, AMD’s hardware is not the issue—the problem lies in software. This is the assessment of an entrepreneur, not an official statement from AMD. Of NVIDIA’s 16-fold market valuation premium, how much is attributable to hardware, and how much to the software barrier that is being systematically dismantled line by line?

Hotz set a five-year deadline to revisit his $250,000 bet.

It’s still unknown whether this bet will be seen as a hacker’s fantasy or the starting point of a new path—but some are already using code to remeasure the thickness of the hash rate moat.