Skip to main content
Code bridges arcing from an AMD rack across a moat to a fortress of CUDA software On the left of a dark field stands a blue fortress made of circuit blocks, labelled CUDA, released 2007. A vertical channel of shimmering blue water separates it from a tall orange server rack on the right, labelled Helios, 72 GPUs per rack. Five dashed orange lines arc from the rack across the water to the fortress wall, labelled AI-written kernels. Another label notes that bringing up an Anthropic model on AMD MI355X chips took about three days, where Anthropic says it used to take six to twelve months. CUDA · RELEASED 2007 NVIDIA'S PROGRAMMING PLATFORM HELIOS · 72 GPUs PER RACK UP TO 23.3 TB/s MEMORY BANDWIDTH PER GPU AI-WRITTEN KERNELS MI355X BRING-UP · ~3 DAYS WAS 6–12 MONTHS (ANTHROPIC)
AMD · CUDA · The AI chip race

The Moat Is Leaking


For almost two decades, Nvidia's CUDA software kept much of the world's AI tied to its chips. Now AI models are writing the code that crosses over, and AMD has built the racks to catch everyone who wants a second supplier.

The job used to take dozens of engineers and six to twelve months. This time it takes about three days. An AI model at Anthropic writes or ports almost every piece of low-level code needed to get one of the company's models processing prompts on AMD's MI355X chips, with what the company describes as "basically no human intervention."↗ Anthropic put that number in front of investors just before its stock-market debut on 19 November, the largest IPO on record.↗

It sounds like a technical footnote. It's actually a crack in the most valuable wall in the technology industry.

That wall is software. Nvidia's chips are superb, but what has kept so much of the AI world buying them is CUDA, the programming platform Nvidia created in 2004 and released in 2007.↗ Rivals could build fast hardware. What they couldn't easily do was get the world's AI code to run well on it. If AI can now write that code itself, the race for AI chips starts to look very different, and AMD is the company best placed to find out.

I · Why software, not silicon

The moat gets dug


Start with credit where it's due. Graphics chips were built to draw video games: thousands of small calculators working side by side on millions of pixels. CUDA let programmers aim that parallel power at anything else, and when researchers discovered that training neural networks is mostly the same kind of massively parallel arithmetic, CUDA was already there, mature and free to use. The tools, the tutorials, the libraries and a generation of engineers' habits all grew up around it. Tom's Hardware sums up the result: CUDA is "very mature at nearly two decades old."↗

If you were ten An AI model is a recipe. The chip is the kitchen. A kernel is one tiny, very precise instruction card for that kitchen: "multiply these two huge grids of numbers, using these exact shelves and burners." A big AI model needs hundreds of these cards, and every kitchen's layout is different. Nvidia's kitchen came with a complete box of cards already written. Anyone moving to a new kitchen had to rewrite the box first.

AMD has its own equivalent. It's called ROCm, and HIP is the layer programmers actually write in. Code written for CUDA has to be translated before it runs there, and AMD even built a translation tool, HIPify.↗ But a translation that merely runs isn't enough. The kernels that matter most have to be tuned by hand to each chip's memory and layout, and that tuning is slow, expert work. That's the moat: not a legal barrier, just months of skilled labour standing between a lab and a second supplier.

Where the moat sits in the AI software stack Four layers stacked from top to bottom. At the top, the AI model. Below it, frameworks such as PyTorch, which run on any chip. Below that, the kernels, split into two columns: CUDA kernels for Nvidia and ROCm and HIP kernels for AMD. This layer is highlighted as the moat, where code must be rewritten and hand-tuned for each chip. At the bottom, the hardware: Nvidia GPUs on one side and AMD Instinct GPUs on the other. THE AI MODEL a giant file of numbers and the maths that uses them FRAMEWORKS (PyTorch and others) largely chip-neutral: the same model code can target either side THE MOAT: REWRITE AND HAND-TUNE PER CHIP CUDA KERNELS mature since 2007 ROCm / HIP KERNELS translated, then tuned NVIDIA GPUs AMD INSTINCT GPUs
Fig. 1 · The model and the frameworks above it travel easily. The kernels underneath are where switching chips used to cost months.
II · The new variable

The code starts writing itself


AMD has been betting on this for more than a year. In August 2025 it published GEAK, an AI agent that writes GPU kernels. Against 30 kernels written by human experts it got 19 right, running at 0.92 times the experts' speed on average. That fell short of the humans, but not by much. A month later, chief executive Lisa Su said much of the world's AI code needs to be "translated to AMD," and "a ton of that can now be done in software." By July 2026 AMD's senior vice president of AI, Vamsi Boppana, said some AI-written kernels were "better than the most manually tuned versions." Anush Elangovan, AMD's vice president of AI software, had put it more bluntly in April: "Claude is better than HIPify because it has web search built in."↗

Claims from a chipmaker about its own chips deserve a raised eyebrow. What changes things this week is who's talking. Anthropic buys chips rather than selling them, and it told investors in writing: "We can onboard new chip platforms in a fraction of the time it once took." Its slides added the numbers. Its Claude Mythos Preview model "ported or wrote almost 100% of the kernels" to bring up prompt processing for one model on AMD MI355X GPUs "at a useful level of performance" in about three days. The company's own historical estimate for that kind of job, using traditional engineering, was six to twelve months.↗

Time to get a model running on a new chip Horizontal bars on a scale of days. The traditional engineering estimate is shown as a band from about 180 to 365 days, six to twelve months. The AI-written bring-up on AMD MI355X is a sliver at about three days. TRADITIONAL dozens of engineers 6 MONTHS 12 MONTHS AI-WRITTEN Claude Mythos Preview ~3 DAYS 0 DAYS · BARS TO SCALE
Fig. 2 · Anthropic's figures, for prompt processing on one model. It hasn't said whether the answer-generating step was included, or how fast the result ran next to Nvidia hardware.

Read the fine print, because it matters. The three days cover prompt processing, the step where a model reads what it's been given, for a single model. Anthropic hasn't said whether the step where the model writes its reply was included, and it hasn't said how fast the AMD version ran compared with the same model on Nvidia chips.↗ That second number is the one that decides purchases. Still, getting a model running at all used to be the expensive, slow part. It no longer is.

Why would AI be good at this particular job, when it still stumbles on so much else? Because kernel writing is unusually easy to check. Each kernel is short. It has a right answer, since its output can be compared number for number against a version already known to work. And it has a score, its speed, that a machine can measure in seconds. A model can write a version, test it, time it, and try again thousands of times without anyone watching. That's the loop the human experts used to run by hand over months.

The same pattern is showing up elsewhere. At AMD's Advancing AI event in July, OpenAI's Philippe Tillet said that once OpenAI received its Helios racks, a single engineer confirmed everything worked within a few days. AMD's data-centre GPU chief Andrew Dieckmann went further: "I have almost zero conversations with our customers about CUDA at this point in time."↗

DisclosureThis article was researched and drafted with Claude, which is made by Anthropic, the company whose results are described here, and which became a listed public company on 19 November. Every figure in this section comes from Anthropic's investor materials as reported by Tom's Hardware, with that outlet's caveats included.

III · The hardware arrives

AMD ships the whole rack


Software only matters if there's competitive hardware to run it on, and this is the year AMD stopped selling chips and started selling machines. Helios is a complete rack of 72 Instinct GPUs, AMD's direct answer to Nvidia's Vera Rubin NVL72. Dieckmann says AMD has "transformed into a systems company."↗ Anthropic, Meta, Microsoft, OpenAI and Oracle are all now running models and workloads on the first Helios racks, according to AMD.↗

AMD's pitch rests on memory. Big modern models, especially the "mixture of experts" kind that switch between specialist sub-networks, spend much of their time waiting for data to arrive from memory rather than doing maths. So how fast a chip can pull data from its high-bandwidth memory (HBM) often sets the pace. Every chip in AMD's new MI400 family is quoted at up to 23.3 terabytes per second. Nvidia had lifted Vera Rubin to 22 TB/s in January, but its product page now lists 19.2 TB/s.↗

Memory bandwidth per GPU, terabytes per second Three horizontal bars. AMD MI400 family, 23.3 terabytes per second. Nvidia Vera Rubin as announced in January 2026, 22. Nvidia Vera Rubin as now listed, 19.2. AMD MI400 23.3 RUBIN (JAN) 22.0 RUBIN (NOW) 19.2 0 · TB/s PER GPU, BARS TO SCALE
Fig. 3 · Published memory bandwidth per GPU. The AMD figure is 21% above Nvidia's current listing. Source: iTWire, citing both companies' spec pages.

That bandwidth is probably most of the story behind AMD's boldest claim this month: that its flagship MI455X processes 21,333 tokens (roughly, word fragments) per second on the DeepSeek-R1 model, about 30% more than Nvidia's best published Vera Rubin result. iTWire's Alex Zaharov-Reutt checked the maths and it holds: Nvidia's rack works out to about 16,435 per GPU, so AMD's figure is 29.8% higher. But the two numbers aren't equals, and the table below shows why.↗

Fig. 4 · The MI455X vs Vera Rubin inference claim, side by side
AMD MI455XNvidia Vera Rubin NVL72
Tokens per second per GPU (DeepSeek-R1, offline)21,333about 16,435
Who measured itAMD's own labs, October 2026MLPerf Inference v6.1 submission
Independently reviewed?No. Not submitted to MLCommonsYes, peer-reviewed (preview category)
HBM4 memory per GPU432 GB288 GB
Memory bandwidth per GPUup to 23.3 TB/s19.2 TB/s (current listing)

"Offline" measures bulk throughput with every query handed over at once, the friendliest test for a memory-rich chip. The "server" and "interactive" tests, closer to what a chatbot user feels, add time limits. Source: iTWire, 17 Nov 2026.

The buyers are real, and they're big. In October 2025 OpenAI signed up for six gigawatts of AMD GPUs across several generations, starting with a one-gigawatt deployment of MI450 chips in the second half of 2026.↗ As part of that deal AMD gave OpenAI a warrant for up to 160 million AMD shares, which vest as the milestones are met.↗ Meta, which has deployed millions of AMD's EPYC server processors, now has Helios racks installed and under test, and will deploy a custom Instinct GPU designed around its own workloads.↗ Anthropic's agreement with AMD is for up to two gigawatts, with its first MI450 chips in Helios racks expected in the first half of 2027.↗

The layers above the kernels are getting easier too. AMD and Meta have worked on ROCm through PyTorch, the framework most AI researchers write in, so that developers can use AMD GPUs "right out of the box," and AMD supplied same-day support for Meta's newest models.↗ AMD released ROCm 10 in the same quarter.↗ And it says it has lined up the parts to grow. "We have worked across our supply chain to secure the wafers, substrates, advanced packaging, memory, and other key components needed to support much higher growth next year," Lisa Su told analysts, adding that she expects AMD's server revenue to double in 2027.↗

One honest caveat belongs here. AMD is also using its own balance sheet to win customers. It agreed to invest $1 billion in Meta's equity offering.↗ And when Anthropic floated on 19 November, raising a record $81.3 billion at $83 a share,↗ AMD bought about $2.5 billion of its stock at that price: the third-largest order in a $26.0 billion side sale to seven technology companies.↗ Only Google and Nvidia, at $10 billion each, put in more.↗ Separately, AMD has pledged up to $5 billion in Anthropic shares, released as GPU deployments hit their milestones.

The money flows the other way too. If Anthropic takes its full first tranche of AMD capacity in 2027, it expects to spend more than $20 billion on Helios racks over the life of the agreement.↗ Suppliers and customers now hold stakes in each other across this industry, and Nvidia is no exception. It means some of this demand comes with money attached, and readers should weigh it that way.

~3 daysto bring up prompt processing for one Anthropic model on AMD MI355X, down from an estimated 6–12 months
+97%AMD data-centre revenue, July to September 2026, compared with a year earlier
23.3 TB/smemory bandwidth per GPU across AMD's MI400 family, against Nvidia's listed 19.2
IV · Keeping perspective

Nvidia is still the giant


None of this makes AMD the leader. It makes AMD a contender. In the July to September quarter AMD's total revenue was a record $13.4 billion, with its data-centre business at $8.5 billion, up 97% in a year.↗ Nvidia's own results, published on 17 November, show how far that still is from the top. In the three months to 25 October it took in $115.8 billion, with $107.6 billion of that from data centres, up 110% in a year, and it expects about $130 billion next quarter.↗

That's also why the labs matter so much. Nvidia expects frontier AI labs such as OpenAI and Anthropic to account for about 15% of its data-centre sales this year.↗ They are a meaningful slice of the world's biggest chip business, and they are exactly the customers now writing code that runs anywhere.

Quarterly revenue, AMD and Nvidia, billions of US dollars Three horizontal bars to scale. Nvidia, August to October 2026, 115.8 billion dollars. AMD total revenue, July to September 2026, 13.4 billion. AMD data centre revenue in the same quarter, 8.5 billion. NVIDIA Aug–Oct 2026 $115.8B AMD (ALL) Jul–Sep 2026 $13.4B AMD DATA CTR Jul–Sep 2026 $8.5B
Fig. 5 · Bars to scale. The quarters are a month apart because the two companies' financial years differ. Sources: AMD; Nvidia.

Nvidia also has a serious counter-argument, and it deserves a fair hearing. In April, Jensen Huang said Nvidia already uses AI to write its own kernels, and that "nobody knows our architecture better than we do." Asked why labs like Anthropic use other chips, he called Anthropic "a unique instance, not a trend."↗ In other words, the same AI tools that help AMD cross the moat also help Nvidia deepen it. And there's a practical catch the research firm SemiAnalysis has pointed out: an AI agent writing kernels needs real chips to test its code on, and AMD has had a constant shortage of GPUs for exactly that kind of internal work. The bottleneck may simply move from engineers to hardware.↗

Nvidia is also a company in transition. Those were its first results under interim chief executive Colette Kress, after its co-founder stepped down. "Our momentum is real," she said, and she told analysts it was "business as usual at Nvidia."↗↗ But the competitive shift described here was under way long before that, and it doesn't depend on who runs either company. It depends on code.

The biggest buyers have already decided they don't want to depend on one chip. Anthropic runs its models on GPUs from Nvidia and AMD, on Google's TPUs and on Amazon's Trainium chips, and manages the whole fleet as what co-founder Tom Brown calls "one fungible pool of compute." He describes that early decision as "an unconventional bet."↗ The chips in that pool aren't fully interchangeable, and squeezing the best out of each still takes work. But every month that the cost of switching falls, a mixed fleet makes more sense.

So what actually changes? For the labs, a second serious supplier means more chips, more bargaining power and less risk of being stuck if one supply line stalls. For everyone else, competition is how prices fall, and the cost of running AI is ultimately the cost of the hardware underneath it. The moat isn't gone. But for the first time, crossing it is a matter of days and a purchase order, not a year of rewriting.

What to watch

  • Next reportAnthropic's first quarterly results as a public company. Watch for progress on its AMD deployment, and any word on how fast the next chip bring-up goes.
  • Next MLPerfWhether AMD submits its MI455X inference result for independent review, ideally with the latency-limited "server" and "interactive" tests too.
  • 2H 2026OpenAI's first gigawatt of MI450 chips, the first milestone in its six-gigawatt AMD deal.
  • 1H 2027Anthropic's first MI450 chips in Helios racks: a new chip generation, and a chance to see whether the three-day bring-up can be repeated.
Ko-fi Buy me a coffee
Scroll to Top