Agentic AI shifts the GPU-CPU ratio to 1:1: why Wall Street is already choosing between AMD and ARM

By Zendoric · September 27, 2026.

Analyst Geoffrey Seiler, of The Motley Fool, has published a clear recommendation: buy AMD shares and sell those of Arm Holdings, the two big stock market bets tied to central processing units (CPUs) for agentic AI. The underlying argument is not financial but technical: according to figures cited in the article itself, the ratio of graphics processing units (GPUs) to CPUs in AI data centers is changing radically. It goes from 8 GPUs per CPU in servers dedicated to training large language models, to 4:1 in inference servers, down to 1:1 in infrastructure designed specifically for AI agents.

The technical reason is easy to understand. GPUs are extraordinary at the massive parallel computation required to train a model or generate a response, but they are clumsy at sequential logic: deciding which step to take first, calling an external tool, connecting to an API, chaining reasoning step by step. That is CPU territory, the processor that governed computing for decades and that the generative AI wave had relegated to a supporting role next to the GPU. With agents —AI programs capable of carrying out multi-step tasks autonomously, not just answering questions— that role is regained. As sector context, Grand View Research projects that the enterprise agentic AI market will grow at an annual rate above 48% to reach $83.4 billion in 2033, and the article recalls that Meta's personal agent app Muse became the most downloaded on Apple's App Store this month, a sign of real traction beyond the lab.

On that basis, Seiler builds his thesis. AMD, with a market capitalization of $1 trillion and a share that over the past 52 weeks has moved between $159 and $639, already sells CPUs designed specifically for agentic AI, with more cores —each a kind of virtual workstation that executes tasks— and faster memory access, which allows it to handle more agents in parallel. The company calculates that the data center CPU market will reach $220 billion and aims to take half of it. Added to that is its inference GPU business, backed by separate $100 billion deals with OpenAI and Meta, and a third significant agreement with Anthropic, according to the article.

Arm Holdings, by contrast, fares worse in the analysis, not for lack of opportunity but because of price. The company, owner of the low-power architecture that dominates the mobile market and that already underpins the custom data center chips of Nvidia, Alphabet and Amazon, announced in March its direct entry into the server CPU business, with the goal of capturing 15% of a market it estimates at $100 billion in 2031 and $25 billion in revenue of its own. But it trades at a multiple of nearly 140 times forward earnings —a price/earnings ratio that measures how much the market pays for each dollar of expected profit— and it also faces an expected decline in mobile volumes, its traditional business. Seiler argues that such a multiple made sense when Arm was a pure intellectual property company, with gross margins close to 94%, but that manufacturing its own chips compresses margins and does not justify the current premium.

It is worth placing this piece exactly where it belongs: it is the opinion of an analyst at a research firm aimed at retail investors, not a market verdict or a done deal. The article itself discloses, as required, that Seiler holds personal positions in AMD, Alphabet, Amazon and Meta —though not in Arm— and that The Motley Fool as a firm recommends both companies at once. It is an investment thesis with its biases declared, not a technical ruling on which architecture will win the agent race.

That said, the figure underpinning the article —that GPU:CPU ratio falling from 8:1 to 1:1— does deserve to stay on the radar beyond the specific stock market bet. It is a sign that the industry is starting to design infrastructure genuinely intended for agents that execute tasks, not just for models that generate text, and that is an indicator that agentic AI is moving from demo to real deployment, with physical capital —chips, plants, multi-year contracts— committed behind it. In general, this kind of large-scale infrastructure commitment tends to precede, not follow, mass adoption: first the plumbing gets built, then the traffic arrives. It fits with what we have already noted about tech employment: the winners are the profiles that design and integrate this architecture, the loser is the routine work that agents end up absorbing.

At the same time, the tangle of cross deals worth hundreds of billions of dollars between chipmakers and AI labs —AMD with OpenAI, Meta and Anthropic, in this case— is part of a pattern the sector is already scrutinizing: the more compute customers and suppliers intertwine in gigantic commitments, the harder it becomes to tell organic demand from self-reinforcing contractual commitments. And the valuation contrast between AMD and Arm —a 140-times multiple versus a company already billing billions in real chips— is a useful reminder that at this stage of the cycle the market pays generously for the promise of agents, not just for their execution. That, ultimately, is the short-term risk worth keeping in sight while the abundance agentic AI promises is built, literally, chip by chip.