The most important number on Advanced Micro Devices ' ( AMD +4.91% ) next artificial intelligence (AI) accelerator may be its memory capacity. The Instinct MI455X, the flagship of the MI400 series the chipmaker introduced this year, is specified at 432 gigabytes (GB) of HBM4, the newest generation of high-bandwidth memory.

Today's top chip, the MI355X, carries 288 GB of the prior generation. The step up is exactly 50%.

The reasoning sits on AMD's own spec sheet, which says the added capacity lets larger AI models and their working data stay in local memory, next to the processor. The timing, I think, is the harder part. AMD is raising its memory content by half at a moment when memory prices are climbing across the chip industry.

Memory is the selling point

AMD isn't shy about the comparison it wants buyers to make. The company's product page stacks the MI455X against Vera Rubin, the new generation of Nvidia 's graphics processing units (GPUs) : 432 GB of memory against 288 GB, slightly higher memory bandwidth, and a claimed 15% edge in peak performance on one AI math format. Those are AMD's own marketing numbers.

But they show where the company believes the fight is -- capacity.

Notably, the whole rack follows the same logic. Helios, AMD's rack-scale system, packs 31 terabytes of HBM4 (about 72 chips' worth). And the company markets that figure right alongside the rack's computing power.

"Helios is now in production with initial shipments on track to begin later this quarter and ramp through the fourth quarter and into 2027 to meet very strong customer demand," AMD CEO Lisa Su said on the company's early August earnings call.

Additionally, AMD has announced a partnership under which Anthropic will deploy up to 2 gigawatts of MI450 series chips in Helios racks.

A seller's market

The memory those racks need is getting more expensive by the quarter. SK Hynix ( SKHY +2.68% ) , the leading maker of high-bandwidth memory, reported second-quarter revenue up 257% year over year, with a 76% operating margin -- a record for the company.

The company said high-performance products for AI servers led price increases, with DRAM and NAND prices rising significantly from the first quarter.

A supplier earning a 76% operating margin is a supplier naming its price. And AMD has chosen to buy more of what these suppliers sell.

Put another way, each MI455X carries 50% more memory than the chip it replaces, of a newer type that likely costs more per gigabyte. That means more units of a costlier input in every chip and every rack AMD ships.

Su's answer on supply came on the same call. "There is very good visibility into HBM allocation for what we expect to deliver in 2027," she said, describing years of work with AMD's memory partners.

However, allocation is volume, not price. A supply agreement means the memory shows up on schedule. What it costs, and how much of that cost AMD can pass along to customers watching their own budgets, is the open variable.

The margin is the test

The business carrying this bet is running hot.

AMD's data center segment produced $6.7 billion of revenue in the second quarter, 107% more than a year earlier and 58% of the company total. Companywide revenue reached a record $11.5 billion, up 50% year over year, and non-GAAP (adjusted) gross margin was 56%. Management guided third-quarter revenue to about $13 billion, up about 41% year over year, with adjusted gross margin holding near 56%.

NASDAQ : AMD

The gross margin is where a memory bill would show up first. Guiding it flat while revenue ramps implies AMD expects to pass its higher input costs through, find offsetting efficiencies, or both.

Sure, demand looks strong enough to carry a lot. Helios shipments are set to start, and the announced Anthropic partnership contemplates up to 2 gigawatts of deployments. But the cost side of this generation is something AMD can manage, not control. The company is more exposed to memory pricing with each new chip, by design.

Still, the stock trades near $479 as of this writing, at about 31 times what analysts expect the company to earn next year -- a price with a smooth ramp already baked in.

I think the 432 gigabytes will sell the racks. Whether the margin survives the price of that memory is a different matter, and the flat guidance says management believes it can pass those costs through or offset them. At about 31 times expected earnings, shareholders are paying as if that turns out to be true.