AMD's 14-Gigawatt AI Bet: The Demand Is Real, and So Is the Leverage

At a Citi investor conference this week, AMD's chief financial officer, Jean Hu, laid out the specific shape of what the company's AI business is about to look like. The MI450, its newest accelerator, enters volume production at the end of the third quarter. The fourth quarter is where she said to expect "a very significant step up," with the ramp continuing into 2027 and a data-center business she expects to double next year.

It is a confident, specific roadmap. The problem — and the question for anyone weighing the stock — is that AMD's shares have already moved to meet it. Up 136% on the year, the price is set as though that ramp is already happening. Today, most of it is still ahead of the curve.

From roadmap to rack

To see why this matters, you have to know where AMD sits in the current AI cycle. AI compute has two jobs. Training is the occasional, brute-force act of teaching a model — it rewards the fastest, most tightly interconnected machines, and it is Nvidia's home turf, protected by a software moat that took a decade to build. Inference is the constant, high-volume act of running a trained model to answer millions of users — and it rewards cost per answer, memory, and speed. It is also where the action is shifting. At the Citi presentation, management noted that inference demand has already outpaced training demand globally.

That shift is what makes AMD a credible second supplier rather than a permanent also-ran. On the pure training game, Nvidia still leads. On inference, the contest is closer, and AMD's new Helios rack — a double-wide cabinet holding 72 of its MI455X GPUs — competes on the dimensions that matter for serving answers: more memory per chip, more bandwidth, and a pitch of roughly 30% more tokens per dollar than the competition.

This is no longer theory. In the second quarter AMD posted $6.7 billion in data-center revenue — more than double the year-ago figure and now 58% of the whole company's sales. And the commitments behind the ramp are named, not hypothetical: OpenAI has signed for up to 6 gigawatts of AMD accelerators, Meta for up to 6 more, and Anthropic for up to 2 — roughly 14 gigawatts of committed compute in total. (A gigawatt is a unit of electricity; in AI it has become shorthand for "a very, very large data center.")

How AMD is paying for that demand

This is the part the headline glosses over, and the one worth understanding. Gigawatt-scale AI deals are not financed by customers writing checks that arrive on schedule. A good deal of them is financed by the supplier handing over its own equity and cash to lock the customer in.

To OpenAI, AMD issued a warrant — a right to buy shares later at a fixed, low price — for up to 160 million shares. To Meta, a similar warrant for up to 160 million. At AMD's roughly 1.6 billion shares outstanding, each warrant is worth close to a tenth of the company; together they are close to a fifth. For Anthropic, AMD committed to invest up to $5 billion in the company it is trying to win as a customer.

None of that is a trick. It is a two-way bet that lines up incentives: the warrants vest only in tranches, as the customer actually deploys the hardware and AMD's stock clears agreed share-price thresholds, so both sides are motivated to make the gigawatts real. But it is also leverage. AMD is, in effect, financing its own demand. And that demand is running into a wall of its own making. At Citi, management was candid that supply is constrained across wafers, advanced packaging, and high-bandwidth memory, and that the MI450 ramp will dilute gross margin — meaning each new dollar of revenue keeps a little less profit than the last.

Real demand and added leverage are the same set of commitments, seen from two directions. That is the tension to hold.

What the multiple is telling you

Here is the judgment the evidence supports. The demand is real. A data-center business that more than doubled in a year, backed by roughly 14 gigawatts of named commitments from three of the largest AI labs on earth, is not a story about a company still waiting for the AI tailwind. This is a company already riding it.

But "real demand" and "a good entry price now" are different questions, and this stock has already answered the second one in the direction that matters to you. It is up 136% on the year, and the earnings that justify that price arrive on a back-half-weighted curve — the "very significant step up" management described is the fourth quarter and 2027, not today. A trade that has run this far, and that sold off after a strong earnings beat in August, tends to be one where the easy money has already been made and the bar for the next leg is high.

The long-term case — AMD establishing itself as the second source in AI and capturing the shift to inference — is intact and, in my view, a good one on a three-year horizon. The near-term question is a different one, and it is the one that decides whether to deploy capital here rather than elsewhere in the AI trade: is the return you can still get from a stock priced for 2027 better than what the rest of the market is offering you right now? I would not rush to fill that gap on enthusiasm. The commitments AMD has locked in are proof the ramp is real — and, at the same time, the leverage and the margin cost of getting there.