AMD launched its Helios rack in San Francisco on July 23: seventy-two GPUs, thirty-one terabytes of shared memory, roughly 5,000 pounds of hardware. Anthropic committed to as much as two gigawatts of them, and AMD committed up to $5 billion of equity in return. The strangest clause in the deal points Claude at the software that has kept AMD out of serious AI work for a decade.

Mark Chubb used to build servers for other companies' chips. AMD acquired his employer, ZT Systems, in 2025 for $4.9 billion, kept the platform architects, and later sold the manufacturing arm. On Thursday in San Francisco, Chubb finally showed what the purchase was for.

"Traditional AI racks are fundamentally discrete GPU servers that are connected through scale-out networks where each server owns its own memory and communicates with other servers over a scale-out network," he told reporters at a press briefing. "Helios changes that model."

Behind him stood a double-width frame roughly 1,200 millimeters across, weighing an estimated 5,000 pounds, holding 72 Instinct MI455X GPUs and 18 sixth-generation EPYC "Venice" processors. AMD says the whole thing behaves as one machine rather than nine servers wearing a trench coat: 31 terabytes of shared HBM4 memory, reachable in a single network hop.

The hardware is not why this week mattered.

AMD Has Never Really Lost on Silicon

Nvidia holds upward of 95% of the data center GPU market. AMD holds roughly 4.5%, according to Futurum Group estimates cited by CNBC. That gap has almost never been about transistors. It has been about the fact that when a machine learning engineer types pip install and expects things to work, the ecosystem underneath assumes CUDA.

AMD's answer to CUDA is ROCm, an open software stack that has spent years playing catch-up on kernels, libraries, and the thousand small compatibility details that decide whether a training run starts or stalls. Practitioners have noticed the gap closing: Zyphra trained an 8B reasoning model entirely on AMD hardware and beat a frontier Anthropic model on math. That was a proof of concept. Gigawatt deployments are not.

So AMD did something no chip company has done before. It hired a language model.

The Software Clause Is the Real Story

Buried in the July 22 partnership announcement with Anthropic, past the gigawatts and the equity, sits a multi-year engineering collaboration with a very specific job description. The two companies will use Claude to optimize workloads for AMD Instinct GPUs and to accelerate ROCm development. AMD will also deploy Claude across its own engineering and product teams.

Read that as a circle. Anthropic will train and serve Claude on AMD silicon. Claude will help write the software that makes AMD silicon worth training on.

AMD extended the same idea to developers on Thursday with ROCm.ai, a development platform whose stated purpose is to let coding agents including Claude, Codex, and Cursor understand AMD platforms and ROCm natively. The company also confirmed that PyTorch, Hugging Face, vLLM, and SGLang are already enabled on MI455X.

For anyone who has burned a weekend porting kernels, that list is the headline. The gigawatt numbers are for investors.

Five Days That Filled AMD's Order Book

MONDAY, JULY 20

Microsoft commits to Helios on Azure

Microsoft says it will deploy the rack for frontier model inference, giving AMD its first hyperscale reference customer before the conference even opens.

WEDNESDAY, JULY 22, 9:00 AM ET

Anthropic signs for up to two gigawatts

AMD commits to a strategic equity investment of up to 5 billion dollars in Anthropic. The first gigawatt of deployment begins in the first half of 2027.

WEDNESDAY, JULY 22

Advancing AI 2026 opens in San Francisco

Sixth-generation EPYC "Venice" makes its commercial debut alongside the Instinct MI400 series.

THURSDAY, JULY 23, 2:30 PM ET

Helios launches, in full production

AMD names OpenAI, Anthropic, Meta, Microsoft, Oracle, HUMAIN, Tensorwave, Vultr and Cirrascale as customers, and announces a joint inference stack with Cerebras.

FRIDAY, JULY 24

Nvidia shares edge lower

Jefferies analyst Blayne Curtis reaffirms a Buy on AMD and lifts his price target from 515 dollars to 640 dollars.

Deployment timing varies by customer, and none of it is immediate. Microsoft gets shipments in the second half of 2026. OpenAI expects Helios online beginning in the fourth quarter of 2026, with deployments accelerating through 2027. Meta is validating EPYC platforms in its labs and has started testing workloads on Helios racks. Anthropic's first gigawatt does not begin until the first half of 2027.

AMD says Helios entered full production this week, with shipments expected by the end of the third quarter.

The Benchmark Numbers Come With Fine Print

AMD compared Helios against published specifications for Nvidia's Vera Rubin NVL72, the rack Nvidia introduced in January and recently moved into full production for late-year customer deployment.

Metric	AMD's claim for Helios vs Vera Rubin NVL72
Peak FP4 performance	15% higher
HBM capacity	50% higher
HBM bandwidth	6% higher
Scale-out bandwidth	50% higher
Tokens per dollar	Up to 30% more

A single Helios rack delivers up to 2.9 exaflops of peak FP4, 1.4 exaflops of peak FP8, and 1.7 petabytes per second of memory bandwidth. AMD also measured the MI455X at 34 times the token throughput of the previous-generation MI355X, running DeepSeek V4 Flash with FP4 serving.

Every one of those figures came from AMD Performance Labs. The tokens-per-dollar claim rests on a Kimi K2 Thinking workload at 32K input and 8K output, priced against "hourly pricing projection of system GPUs based on market conditions," according to AMD's own footnote. That is a vendor benchmark against a competitor's spec sheet, not an independent bake-off between two racks in the same room.

The Other Side

The skeptics have a straightforward case, and some of them have been right before.

DA Davidson analyst Gil Luria described AMD earlier this year as a marginal player in the AI accelerator market that is "playing catch-up," a characterization the market share numbers still support, though his firm has since upgraded the stock to Buy. Nvidia's consensus rating on Wall Street remains a Strong Buy, built on 36 Buys and one Hold over the past three months, with an average price target implying roughly 48% upside.

Timing is the second problem. AMD spent this week collecting commitments that mostly convert to revenue in 2027. Nvidia is shipping Vera Rubin into late-2026 deployments now. A gigawatt promised for the first half of 2027 is a purchase order, not a moat.

Then there is the software argument, which cuts both ways. Using Claude to accelerate ROCm development is an admission that ROCm needed accelerating. AMD is betting that an AI coding agent can compress a decade of ecosystem work into a couple of years. Nobody has demonstrated that yet at this scale.

Investors were not universally convinced either. AMD shares slipped on Thursday even as the announcements landed, with reporting attributing the muted reaction partly to the fact that analysts had already predicted the Anthropic deal ahead of the event. The pattern rhymes with the rest of Big Tech's AI infrastructure spending, where enormous commitments and delayed returns keep colliding.

Jefferies took the other view. In a note released Thursday, Blayne Curtis pointed to "a roadmap that potentially puts AMD ahead of Nvidia," citing the Venice CPU's lead on the processor side.

The Bottom Line

Strip out the gigawatts and the equity and one fact remains: five of the largest buyers of AI compute on earth, including two frontier labs that compete directly with each other, put their names on the same non-Nvidia rack in the same week. That has not happened before, and it happened because the buyers want a second supplier badly enough to fund one, the same impulse that has pushed money toward challenger silicon startups like Etched over the past year.

Whether they get one depends on something no press release can measure. AMD can build silicon that matches Nvidia's on paper, as this week's spec comparison shows. What it has never built is the software gravity that makes engineers reach for a platform without thinking. The company's answer to that problem is now to hand the job to Claude, which is either the most efficient use of a language model anyone has proposed or a very expensive way to discover the limits of one.

Tom Brown, Anthropic's co-founder and chief compute officer, framed his side plainly. "Running across a diversified range of hardware lets us map the right workloads to the right hardware," he said.

Diversification is a strategy, not a verdict. The verdict arrives in the first half of 2027, when the first gigawatt turns on.