AMD Stock After Q3 2026 Earnings: Its AI Chips Earn Less Than Its Average

I think AMD is winning AI as the cheaper second source, the supplier big buyers keep so they aren't stuck paying Nvidia's price, and that isn't a moat yet. Data Center sales nearly doubled in Q3 and Lisa Su now expects total revenue to grow more than 80% in 2027. But asked about margins as Helios, AMD's full-rack AI system, ramps up, CFO Jean Hu said the AI business carries "a margin a little bit dilutive to corporate average, but the gross margin dollars are increasing very significantly."

Gross margin is what's left of each sales dollar after the cost of making the chip. AMD's company average is 56%. Nvidia sells almost nothing but data center chips, so its 75% is close to its AI margin. AMD's AI chips sit somewhere below 56%. That's a wide gap on chips that do the same job.

Two things pull against each other inside that flat line. EPYC server CPUs are AMD's richest products, and Su expects server revenue to double in 2027. The Instinct GPUs are, in Hu's words, below average. Data Center's operating margin, which also takes out research and sales costs, rose from about 25% to about 34% over the year. I think most of that is costs growing slower than sales, since Hu says 2027 operating costs will grow at less than half the rate of revenue. That's a bigger business, and it doesn't show the GPUs fetching a better price.

AMD isn't alone. Broadcom, whose custom chips for Google and others are the other way off Nvidia, told investors the same thing in September.

Every one of these chips carries stacks of expensive high-bandwidth memory. Broadcom and AMD earn a margin on top of that memory that shrinks the more of it they ship. Nvidia ships the same memory and its margin went up. The difference is how much the seller can charge above its costs, and that's the moat. Nvidia gets it from CUDA, the software AI developers have spent fifteen years learning and building on, which makes leaving expensive even when a rival chip is cheaper.

Broadcom has a different answer, and Hock Tan gave it on the call: "stop focusing on gross margin." Its custom chips are designed with one buyer over several generations, so the revenue is locked in and its operating costs barely rise with it. Its CFO said operating margin rose 2.4 points over the year to 67.9%, even as gross margin fell. Broadcom accepts a thinner margin on each chip in exchange for being built into the buyer's systems for years, and Marvell sells custom chips the same way on a smaller scale.

Asked how AI labs choose between AMD and their own custom chips, Su didn't talk about software. She said "the key is for us to continue to show that TCO advantage across workloads," meaning total cost of ownership. That's a price-for-performance pitch. It's working: Anthropic, Meta, Microsoft, OpenAI and Oracle all have brought their first Helios racks online, Su said, and quickly. But a buyer who chose you on cost runs the comparison again at the next order.

AMD's closest thing to Broadcom's lock-in is OpenAI's multi-year deal for 6 gigawatts of GPUs, gigawatts being how AI buyers now size data centers by the power they draw. AMD paid for that with equity: a warrant for up to 160 million AMD shares that vests as OpenAI's purchases scale and as AMD's share price hits set targets. A supplier with pricing power doesn't hand its biggest customer stock to buy from it.

The grid below is InvestMoat's live rating of where each company's moat comes from. Learned interfaces and network effects are the software habit and the developer community. Transaction embedding is how deeply the product is wired into a customer's operations. Bundling is selling the pieces as one package.

Nvidia is rated strongest where the software lives. Broadcom and Marvell are rated strongest on embedding, which is where their co-designed chips come from. AMD is rated below Nvidia on the software rows and below the custom-chip makers on embedding. It has a real rack-level bundle, but nothing yet that makes a buyer pay more to stay.

The strongest objection is timing. MI450 only started shipping in volume this quarter, and first-year racks carry yield, assembly and memory costs that fall as production matures. If AMD's AI margin climbs through 2027, I'm reading a launch cost as a pricing problem. The bull would add that memory is a cost every seller passes through. Su said AMD is working with customers "to ensure that those costs are shared," and Broadcom's own margin is falling for the same reason, so a lower percentage may be mostly memory.

Second, Su said "demand is higher than supply." A seller that can't make enough isn't obviously discounting. And being the second source has real value. Every large buyer wants one, AMD is the only merchant GPU maker at gigawatt scale, and Su argues GPUs have "more fungibility" than custom chips, which matters to a lab that doesn't know what next year's model needs. Hu's point stands too: gross margin dollars are rising fast, and a company can compound for years as the indispensable alternative without Nvidia's margin.

Here's why I still hold the view. Nvidia's own ramp dip was small. Its margin was 72.5% a year ago and is 75.0% now, so a full Nvidia-style recovery still leaves AMD's AI chips far behind. If memory explained the whole gap, Nvidia's margin would be sliding too, and instead it rose over the year and is guided down only about a point. Selling out in a shortage proves demand for AI chips, not that buyers would pay AMD more once supply catches up. And buyers keep a second source because it holds Nvidia's price down, which also limits what that second source can charge.

This isn't a price call. The scorecard carries the live valuation and the AMD page the scenarios. Su expects data center to be more than 80% of AMD's revenue in 2027, so the company margin will increasingly be the data center margin. If it stays in the mid 50s while AI keeps outgrowing servers, AMD is selling volume at a second source's price. If AMD starts saying its AI margin is above average, a moat is forming.