Advanced Micro Devices, Inc. AMD

Key data

The business

AMD designs chips and builds none of them. TSMC in Taiwan does the manufacturing, which makes AMD's fortunes a bet on getting enough of TSMC's most advanced capacity. Four lines of business. Data Center is the one that matters now: server processors (EPYC) and AI accelerators (the Instinct MI series), sold to the hyperscalers building AI clusters. The other three are the Ryzen chips inside laptops and desktops (Client), the semi-custom parts inside PlayStations and Xboxes (Gaming), and the Xilinx programmable chips embedded in industrial and networking gear (Embedded).

The moat is two-sided and uneven. On CPUs, AMD and Intel are a durable duopoly around the x86 instruction set, and AMD has spent a decade taking server share on merit. On AI accelerators there is no moat yet. Nvidia owns the customer through its CUDA software, and AMD is the challenger trying to become the credible alternative. That challenge is the whole story. The thing a buyer actually holds is a rack: AMD's Helios rack-scale system, built to plug in against Nvidia's, one row at a time.

The numbers

The last five quarters show a business inflecting hard, and the newest print is the cleanest in years.

The step from Q1 to Q2 2026 is the tell: revenue up 13% in a single quarter and net income up 66%, because gross margin snapped back to 54% from the 40% a year earlier that carried an inventory charge on MI308 chips stranded by the China export rule. Adjusted earnings have beaten consensus four quarters running, the latest $1.66 against $1.62. Against the year-ago quarter, revenue grew about 50% and profit about 163%.

The full-year picture is a company that stalled in 2023, when the PC market collapsed and profit fell to $854 million, then found the AI cycle: revenue from $22.7B in 2023 to $34.6B in 2025, net income up more than fivefold. The first half of 2026 already booked $3.7B of net income, more than any full year before 2025.

Now the multiple. On trailing earnings of $3.89 a share the stock changes hands near 123 times, and on Q2's run rate, $1.38 annualized to $5.52, about 87 times, right where it sat a quarter ago. Both sit near the top of anything AMD has fetched in twelve years, against a band that has usually run 37 to 82 times. The market is not paying for this year's earnings. It is paying for the estimate that has EPS reaching roughly $22.62 by 2028, which would drop the multiple to about 21 times that far-out number. What this memo believes the market underweights is supply: the two-customer warrant structure only makes sense if AMD and those buyers both expect Instinct volume far above what today's models carry. The print that settles it is Data Center revenue and gross margin over the next two quarters.

The balance sheet funds the fight: $13.1B in cash and short-term investments with minimal long-term debt, and 23% of every sales dollar going back into R&D.

Management

Lisa Su has run AMD since 2014 and turned a near-bankrupt also-ran into a server-share taker. The record earns the benefit of the doubt on execution, not on valuation. Insiders have sold, not bought: 44 sales worth about $45 million over the past year against zero purchases, the largest a $7.7 million sale by Paul Grasby in mid-August and two sales totaling $7.7 million by CTO Mark Papermaster days later. Plan status is not disclosed on these, so read them as routine but unconfirmed. Against a $779 billion company the dollars are noise. Capital return is modest by megacap standards, about $1.3 billion of buybacks last year, because the cash is going into R&D and long-term supply commitments instead. Pay tracks the earnings beats.

How it fails or surprises you

Two customers, one story (downside). The warrants dangle up to 320 million shares, about 16% dilution, in front of two buyers tied to Instinct milestones. That concentration cuts both ways: if AI capex cools or those two slow their orders, revenue and the entire second-source narrative crack together. Watch Data Center revenue sequentially over the next two prints. A flat quarter there is the first crack.

The China license (downside). MI308 shipments to China run on licenses the U.S. granted and can pull, and the D5 country rules can widen. Last year's export rule already triggered an inventory charge. A tightening removes a revenue slice at the worst multiple to absorb it. The tell is any 8-K on license status or a guidance footnote naming China.

MI450 runs ahead (right tail). The market pays AMD to be a respectable number two, not a co-leader. If the MI450 and Helios racks ship on time and hyperscalers over-order to secure a second supply, Data Center revenue and gross margin run well past model, and the warrants convert a demand signal into reported volume. The number that reveals it first: gross margin holding above 55% while Data Center revenue climbs, two quarters in a row.

Closing thoughts

Strip it down and AMD is one wager on the AI build-out continuing, with a specific quarterly print riding on top. A stumble in that build-out is the left tail, and at 123 times trailing earnings the price leaves no room for one. The permanent-loss case is not that AMD is a bad company, it is that AI capex normalizes while the multiple still assumes it compounds. The right tail is unusually fat here, because the people closest to the demand, the warrant customers, are signaling volume the estimates have not caught. On balance the odds favor the upside on fundamentals and the downside on price, a real tension, not a resolved call: the business is winning and the stock already knows.

The bet is still that AMD becomes the clear second source for the chips that train and run AI, the Instinct accelerators that go up against Nvidia. What breaks it is Data Center revenue stopping its climb quarter over quarter, or gross margin slipping back under 50%. Either one, two quarters running, means the second-source story is stalling and the 123 multiple has nothing under it. Until then, the numbers are going the right way.