AMD chief executive Lisa Su said demand for AI chips will exceed supply for the next several years and that the company is now planning manufacturing capacity three to five years ahead with partners, up from the one to two year horizon that used to be standard. Su made the comments in Taipei on October 6 after visiting Taiwan Semiconductor Manufacturing Company, Hon Hai, Acer and Asus, and said supply will rise substantially in 2027 but still fall short of what customers are asking for.
Planning horizons are the clearest signal of where the semiconductor cycle stands. When Nvidia and AMD were selling gaming chips, one to two years was enough visibility for everyone in the chain. AI datacenter buildouts changed the math entirely. Wafer starts, advanced packaging, high-bandwidth memory and substrates all sit on separate lead times, and a shortage in any one of them caps the whole product line. Su said AMD has to plan across every layer at once, which is why the window has stretched from something a CFO could track quarter by quarter to a multi-year negotiation with foundry partners on three continents.
We have increased our own supply and capacity this year, but demand exceeds supply, Su told reporters in Taipei. We plan to increase supply substantially in 2027, but clearly more will be needed. She also said AMD has already surpassed the $10 billion investment in Taiwan announced in May, because that figure keeps growing as long as demand holds. The company is expanding across the island rather than concentrating in a single campus, and her meetings with Hon Hai point to assembly and rack integration work as well as pure silicon.
Helios and the HBM4 Question
The centerpiece near-term is Helios, an AI server rack AMD starts shipping in volume in the fourth quarter. Helios integrates 72 Instinct MI455X GPUs, which handle AI computation, with 18 Venice CPUs, which run the server, into a single steel rack. Line-ups of that scale put enormous pressure on packaging and memory. Samsung Electronics was selected in March as the priority supplier of HBM4, the sixth-generation high-bandwidth memory, for the Instinct MI455X line, which marks a departure from the SK hynix-heavy memory sourcing Nvidia has leaned on for its own flagship parts.
Customers who have already committed to Helios include OpenAI, Anthropic, Microsoft, Meta and Oracle. OpenAI signed a multi-gigawatt commitment with AMD in October 2025 and earned a warrant for up to 10 percent of AMD shares as the deployment scales. The pattern is straightforward. AMD needs marquee names behind its flagship line, and those names need an Nvidia alternative so negotiating leverage does not collapse to a single supplier. Anthropic’s separate deal with Google for TPU capacity shows the same buyer-side math in motion.
Su also said AMD will tap alternative packaging approaches beside TSMC’s CoWoS, referring to fan-out panel-level packaging, or FOPLP, which raises production efficiency by packaging chips in batches on large rectangular panels instead of round wafers. CoWoS is the process where chip-on-wafer packaging places an AI chip and high-bandwidth memory side by side on an interposer so they operate as a single chip. TSMC’s CoWoS capacity has been the choke point for the entire AI chip market since 2024, and opening a second track is the main way AMD keeps up without waiting for new cleanrooms and equipment to come online.
The CPU Renaissance Nobody Predicted
Su separately highlighted the growing market for AI agents, spurred in part by Meta’s recent launch of its personal AI agent Muse. CPUs are the AI chips that run agents, since inference control and orchestration run better on general-purpose cores than on GPUs optimized for dense math. Citigroup analyst Atif Malik projected an average annual growth rate of 60 percent for the CPU market, from $29 billion in 2025 to $300 billion in 2030, calling AMD the main beneficiary of a CPU renaissance.
AMD has made a point of positioning its Epyc server CPU line as general-purpose rather than siloed into specific workloads. That positioning matters because agents are unpredictable in what they request, and buyers who committed to narrow inference hardware in 2025 have found some of those deployments rigid and hard to repurpose. Su told reporters that AMD and Intel lead the server CPU market and expects that to continue as agent workloads expand into enterprise and cloud environments.
What Comes Next
TSMC’s recently reported investment in Texas is the other side of the same equation. Su declined to answer a question about it directly, saying only that additional advanced wafer capacity is needed to meet strong demand. AMD’s Taiwan spending expansions and TSMC’s US construction both point in the same direction, toward more capacity everywhere as fast as anyone can get permits, labor and materials in place.
Watch fourth-quarter earnings for two numbers. One is Helios rack shipments against targets AMD has already issued. The other is whether AMD raises its MI455X production guidance the way Nvidia did through 2025 and 2026, or whether packaging and memory constraints cap it. Datacenter GPU demand is not going anywhere. The real test is whether AMD turns supply plans made three years out into racks that actually arrive on schedule.
For competitors, the message from Taipei is that Nvidia’s lead in AI accelerators no longer rests on design alone. It rests on locked-up capacity that new entrants cannot easily match, which is exactly where AMD and Intel have started to spend. The next 12 months will show whether that spending buys enough racks to change the standings in the AI datacenter hardware business.