For years, the AI chip conversation has had one dominant name and a long list of challengers trying to close the gap. Most of them haven’t.
One of them is starting to look different, and Bank of America just made that case in writing.
Bank of America analyst Vivek Arya, who covers semiconductors for the firm, spent July 23 in San Francisco watching AMD make its most ambitious AI pitch to date. By the following morning, he had a new price objective and a significantly expanded view of how large AMD’s AI market opportunity could become.
The bank raised its price objective on AMD to $620 from $560, reiterating a buy rating, with AMD shares trading at $554.23 on July 24. But the more significant part of the note isn’t the number. It’s how the bank is now describing AMD’s place in the AI market.
How Bank of America is reframing AMD’s AI market position
The language Bank of America used in the note is worth paying attention to. The bank said AMD has “successfully transitioned from a merchant GPU vendor to a full-stack, rack-scale accelerator systems supplier.”
That’s a meaningful distinction. AMD is no longer being compared only to Nvidia on chip-by-chip specifications. It’s being evaluated as a systems-level competitor with an integrated hardware and software offering.
More Bank of America:
- Bank of America warns America now has 2 economies
- Bank of America answers a tough stock market question
- Bank of America gives stock market investors a summer reality check
Analyst Vivek Arya set the $620 price objective at 47 times the bank’s 2027 estimated non-GAAP earnings per share, toward the middle to upper range of AMD’s historical 13x-to-58x multiple. He said that multiple is supported by AMD’s potential for 50% or higher annual EPS growth and its share gains in AI CPUs and GPUs.
AMD also updated its total addressable market forecasts at the event. The company now targets a total compute market of more than $2 trillion by 2030, up from a prior estimate of $1 trillion.
Its AI accelerator TAM estimate rose to $1.4 trillion from $500 billion, and its server CPU TAM target increased to $220 billion from $120 billion earlier this year, Investing.com reported.
Why AMD’s Helios rack is central to the Bank of America thesis
The product at the center of the upgraded thesis is AMD’s Helios rack, which the bank said is now in full production with deployments beginning in the third quarter and ramping into the fourth. Each rack integrates 72 MI455X GPUs with 18 EPYC Venice CPUs, Pensando networking chips, and ROCm software support. AMD’s shift into full rack-scale AI systems has been building all year, as TheStreet reported.
The specs are notable. Each Helios rack delivers roughly 2.9 exaflops of FP4 compute, 1.4 exaflops of FP8, and 31 terabytes of HBM4 memory.
Against Nvidia’s Vera Rubin NVL72, AMD says Helios offers about 15% more FP4 compute for training, 50% more HBM capacity, and up to 30% more tokens per dollar on memory bandwidth, all at a rack price of around $5 million, compared to $6 to $7 million for Nvidia’s system.
David/Getty Images
Why AMD’s customer pipeline is what investors should track most closely
The customer announcements are what give the Bank of America note its momentum. Anthropic will deploy up to 2 gigawatts of Helios, with the first gigawatt beginning in the first half of 2027. AMD is also investing up to $5 billion in Anthropic as part of the deal, CNBC reported.
Because Anthropic already runs AMD’s MI355X chips, the bank views this as the start of a multi-generational relationship extending through the MI500 and MI600 product cycles rather than a one-time win. The Meta relationship follows a similar pattern, with a six-gigawatt, four-year AI infrastructure partnership already in place, as TheStreet reported.
OpenAI and Meta each appeared at the AMD event with initial contractual deployments of 1 gigawatt set to begin in the second half of 2026, out of 6-gigawatt total agreements, respectively. Bank of America’s view is that AMD does not expect any customer to go through the design-in effort for a single generation of systems. If that holds, these wins could become recurring infrastructure relationships across multiple product cycles.
The roadmap underpinning those relationships now extends through 2028. MI450 begins shipping in the second half of 2026. MI500 arrives in 2027, paired with the Verano CPU. MI600 follows in 2028 with the Zen 7 Ferrara CPU and the next generation of Helios.
What the software picture means for AMD’s long-term competitiveness
Hardware specs alone won’t close the gap with Nvidia. Bank of America acknowledged that and pointed to AMD’s ROCm software ecosystem as an increasingly important part of the investment case.
AMD introduced ROCm.ai at the event, an AI-native developer experience built on ROCm 7, which delivers roughly 3.5 times the inference performance and 3 times the training performance compared to ROCm 6.
Third-party validation from SemiAnalysis and InferenceX showed AMD’s MI355X achieving up to 40% lower cost-per-token than Nvidia’s B200 on the SGLang FP8 framework.
That kind of independent confirmation matters. It shifts the software conversation from a narrative about catching up to a data point about competing on economics.
Bank of America’s $620 target reflects a view that AMD’s combination of hardware, software, and customer relationships is becoming more durable than the market has historically assumed.
The bank still lists execution risk on its first rack-scale product among the key downside risks, along with reliance on a single manufacturing partner. But the overall tone of the note is that AMD’s AI window is opening wider, and the bank thinks the market is underestimating how quickly that’s happening.
Related: Wells Fargo doubles down on AMD stock after Anthropic deal
