Author: Shenchao TechFlow
GitHub code "leaked," Anthropic becomes an AMD customer? NVIDIA's GPU monopoly faces its biggest challenge
DeepInsight Summary: An AMD executive uploaded a YAML code file to GitHub listing Anthropic as a "customer" with the highest priority level, equal to Meta. Jefferies previously noted that Anthropic is hiring ROCm engineers. If AMD officially announces a partnership at its Advancing AI conference on July 22, this would mark AMD’s third top-tier AI lab client after OpenAI and Meta—and potentially without having to give up 10% of the company’s equity. NVIDIA still holds 80% to 90% of the AI accelerator market, but supply chain diversification has shifted from a “strategic vision” to “actual procurement.”
A YAML code file on GitHub brought to light the previously undisclosed customer relationship between AMD and Anthropic.
According to Stocktwits on July 19, the semiconductor research firm SemiAnalysis found that in a code file submitted by Anush Elangovan, AMD’s Vice President of AI Software, on GitHub, Anthropic was listed as a "customer" and assigned the highest priority boost points of 30, matching the treatment given to publicly known hyperscale customers like Meta.
AMD's stock rose 1.3% in after-hours trading.
SemiAnalysis also noted that Anthropic is still in the evaluation phase. “If AMD fails to announce Anthropic at the upcoming Advancing AI conference, it means Elangovan’s on-site engineering team has not yet addressed all of Anthropic’s concerns regarding software quality.”
The AMD Advancing AI 2026 conference is scheduled for July 22–23 at the Moscone Center in San Francisco, just two days after the code leak.
Multiple clues converge in the same direction
GitHub code is not an isolated signal.
Back in April this year, SDxCentral reported that Anthropic posted a job opening for an engineer on its reinforcement learning team, explicitly requiring candidates to have experience with ROCm (AMD’s AI software stack) and the ability to migrate workloads across different types of accelerators. The starting salary is $350,000.
Jefferies analyst Blayne Curtis further noted in a research report on July 17 that Anthropic has been hiring engineers with ROCm experience, indicating the company is preparing to further diversify its computing infrastructure. Curtis views this as a key signal that customer announcements may emerge at the AMD Advancing AI conference.
Anthropic’s current public compute architecture consists of three pillars: NVIDIA GPUs, Amazon’s custom Trainium chips, and Google TPUs. In October 2025, Anthropic reached an agreement with Google to gain access to up to one million seventh-generation TPUs. In November 2025, NVIDIA and Microsoft invested $15 billion in Anthropic, in return for which Anthropic committed to purchasing $30 billion worth of NVIDIA compute power through Microsoft Azure.
AMD's "stock-for-stock" model has reached a crossroads
Previously, AMD won over two massive clients, OpenAI and Meta, with an aggressive equity incentive program.
In October 2025, AMD signed a 6-gigawatt GPU deployment agreement with OpenAI, issuing up to 160 million warrants with an exercise price of just $0.01, subject to vesting only after AMD’s stock price reached $600 (then approximately $210) and delivery milestones were met. In February 2026, AMD signed an almost identical agreement with Meta: the same 6 gigawatts, the same 160 million warrants, and the same $600 stock price threshold.
The two transactions combined imply that AMD may be surrendering approximately 20% of its company equity in exchange for locked-in GPU orders totaling 12 gigawatts. Estimating the value based on NVIDIA’s GPU pricing, each gigawatt is worth roughly $35 billion, meaning the potential revenue from 12 gigawatts far exceeds AMD’s full-year 2025 revenue of $34.6 billion.
Jefferies, in its research report on July 17, explicitly stated that since AMD has already committed 20% of its equity to OpenAI and Meta, future deals will require smaller incentive packages. A more traditional Anthropic agreement—without substantial equity stakes—would bolster market confidence and demonstrate AMD’s ability to secure top-tier clients without diluting shareholder equity.
As of the close on July 18, AMD's stock price was $495.76, approximately 21% below the $600 warrant strike price.
NVIDIA remains the absolute leader, but the #2 player is closing in.
For AMD, securing Anthropic holds strategic significance beyond revenue alone.
NVIDIA's data center revenue for the first quarter of fiscal year 2027 reached $75.2 billion, a 92% year-over-year increase, equivalent to approximately 13 times AMD's data center revenue of $5.8 billion during the same period. NVIDIA still holds an 80% to 90% market share in the AI accelerator market.
The gap remains substantial. However, AMD’s growth trajectory is steepening. In Q1 2026, AMD’s data center revenue increased by 57% year-over-year, with full-year revenue consensus estimates at approximately $49.6 billion, representing about a 43% increase from 2025. GPU orders of 6 gigawatts each from OpenAI and Meta have secured AMD with a multi-year pipeline of visible revenue. Microsoft has also become a customer of AMD’s MI400 series.
SemiAnalysis's previous analysis indicated that the AMD MI355X has become competitive in cost-effectiveness for inference with small and medium-sized models. The total cost of ownership for the MI355X is approximately 33% lower than NVIDIA’s HGX B200 and offers greater HBM memory capacity. However, in frontier model training and rack-scale inference scenarios, the MI355X still cannot compete with NVIDIA’s GB200 NVL72.
AMD plans to begin mass production of the MI450 series GPUs based on the CDNA 5 architecture in the second half of 2026. The next-generation MI500 series, built on a 2nm process and HBM4E memory, is expected to launch in 2027.
The real battlefield is in software, not just in chips.
The analysis released by SemiAnalysis on July 13 refocused the competitive spotlight on software. In performance tests of the open-source inference engine vLLM, NVIDIA's GB200 NVL72 significantly outperformed AMD's MI355X, with the gap being especially pronounced in large model scenarios.
This conclusion contrasts with SemiAnalysis’s analysis two weeks ago, which noted that Anthropic’s Claude model training ran heavily on Google TPUs, while Claude Code inference was increasingly deployed on Amazon Trainium chips, with NVIDIA GPUs gradually losing market share within Anthropic to proprietary chips.
Elangovan, Vice President of AMD AI Software, previously told SDxCentral that, following a series of updates, ROCm has "nearly caught up" with NVIDIA, and even leads in FP4 (4-bit floating point) scenarios. ROCm 7 delivers 3.5 times the performance of ROCm 6 and now supports all major AI frameworks.
However, there is still a gap between "nearly catching up" and "production-ready." The core bottleneck in Anthropic’s evaluation phase, as mentioned by SemiAnalysis, is likely whether the ROCm software stack can meet the stability and performance requirements for its production workloads.
For investors, the AMD Advancing AI conference on July 22 was the latest validation milestone. If Anthropic is announced as an official customer, AMD will demonstrate, without diluting equity, that it has become a "trusted second choice." If not, it means the software gap still exists, and the parameter advantages on AMD's GPU hardware are insufficient to translate into actual orders.
Twitter: https://twitter.com/BitpushNewsCN
BitPush Telegram community: https://t.me/BitPushCommunity
BitPush TG subscription: https://t.me/bitpush