All news

AMD's Helios ramp gives AI compute buyers a second price

Two steel taps on a graphite wall pour water into one white stone basin, a cobalt glass on its edge

AMD's data-centre sales rose 97% as its first racks went live at five AI leaders, and a slipping margin shows the buyers, not the chip maker, keep the gain.

AMD's data-centre revenue nearly doubled in a year, up 97% to a record $8.5 billion in the third quarter, and the company guided to about $17 billion of total revenue for the fourth, against $13.4 billion just reported. The results, out on 3 November, also place the first Helios racks, AMD's complete rack-scale AI systems, inside Anthropic, Meta, Microsoft, OpenAI and Oracle's cloud.

That changes who sets the price of compute. Rack-scale AI systems have had one dominant seller, Nvidia, and its margin has sat at the bottom of the AI-native cost stack. A second supplier now ships racks to the labs that set demand, and it is paying for the position with its own margin. AMD expects a non-GAAP gross margin of about 55% in the fourth quarter, down from 56%, in a quarter when revenue is set to rise about 27%. The dip comes from mix: AI accelerators carry a below-average margin, so every extra rack pulls the average down. When the fastest-growing product earns less than the rest of the company, the buyer keeps the difference.

The buyers have reason to care. Anthropic's amended prospectus, filed a day earlier, showed what the premium for short-term compute costs: capacity bought at short-term rates turned an operating profit into a loss in one quarter. Every rack a second supplier delivers narrows that premium, and the orders behind it are already placed. AMD expects total revenue to grow more than 80% in 2027, with data-centre chips at more than 80% of sales. A year ago the market's answer to scarce compute was a chip supplier financing its own demand. Now the labs can play two sellers against each other on terms as well as price.

AMD's own mix is moving the same way. Its gaming business, console chips and graphics cards, fell 40% to $773 million, and the company plans for PC market shipments to fall by a double-digit percentage next year. Data-centre chips already bring in close to two-thirds of revenue, against under half a year earlier.

For founders, the gain shows up only in the unit economics of teams that can use it. Inference is a core variable cost of AI-native products, and a two-supplier market lowers it for software that runs on both. AMD has just released ROCm 10, its developer stack; Oracle's cloud already runs workloads on Helios, and Vultr will deploy it next. Build portable, and set prices against the cost curve of 2027 rather than this year's invoice. Fanaura looks for AI-native companies whose gross margin widens as compute gets cheaper, not ones whose edge is a negotiated GPU allocation, and that is what Fanaura invests in.

The strongest objection is that the thinner margin on AI racks reflects the cost of ramping a new product, not price concessions, and that five customers running workloads on first racks is a pilot, not a market. The quarters after the ramp settle it. If AI racks still dilute AMD's margin once Helios ships at volume through 2027, AMD is buying share and the buyers pocket it; if the margin climbs back above 56% at that scale, the labs face two suppliers and one price.

The compute bill for AI-native companies now has two possible senders, and that, more than any benchmark, is what brings it down.

For founders

Building AI-native products that run on any accelerator? Send us the project.

Fanaura invests its own capital and the capital of its limited partners (LPs), usually through a separate SPV for each deal. If we see a fit, we get in touch.

Apply in five minutes
Subscribe

Get the news by email

One short note each weekday. No marketing, and you can unsubscribe at any time.