Lisa Su delivered a message this week that carried two truths. AMD has boosted production through 2026. It will substantially expand output again next year. Yet demand still runs hotter. “We’ve been able to increase our supply as we’ve gone through 2026, and we’re going to substantially increase our supply in 2027, but we can definitely use more,” the CEO told reporters in Taipei on Oct. 6, per Reuters.

She spoke while visiting key partners. Foxconn. TSMC. Meetings in Taiwan, then South Korea with Samsung and SK Hynix. The focus? Scaling capacity for CPUs and GPUs that power AI data centers. Not laptops. Not gaming consoles. That distinction matters. It explains why PC buyers and console makers may see little relief.

AI Demand Outruns Every Expansion

Su’s comments came amid reports of sold-out server lines. AMD’s EPYC Venice CPUs are reportedly booked through 2027, with some orders slipping into 2028. Customers like OpenAI, Anthropic, Meta, Microsoft and Oracle have locked in massive commitments for the Helios AI server rack, which begins volume shipments this quarter. “There’s very, very high demand for the next several years,” Su added. “We can see very high demand, and so we need more advanced wafer capacity.” The company now plans capacity three to five years out. Previously it looked one to two years ahead.

This shift reflects reality. AI training and inference workloads keep surprising upward. Agentic AI has accelerated requirements for both accelerators and general-purpose server CPUs. Memory stands out as the stubborn bottleneck. High-bandwidth memory for AI servers remains tight. Advanced packaging like CoWoS is “absolutely critical.” AMD has committed more than $10 billion to Taiwan’s supply chain and plans to increase that figure further, according to multiple reports including The Verge.

But. The ramp targets data-center silicon first. Consumer products sit lower in the priority queue. And that creates ripple effects across the market.

Industry analysts don’t expect quick normalization. “The AI compute supply crunch is expected to remain acute till the first half of 2027,” experts told Network World. Shortages extend beyond chips to power, data-center space and packaging. Lead times for premium GPUs still stretch 36 to 52 weeks. Prices sit 25-35% above MSRP in many cases. Cloud providers have already hiked instance rates by 15-25%.

Pareekh Jain, CEO of EIIRTrend & Pareekh Consulting, spelled out the dependencies. “AMD designs the chips, but relies on TSMC for advanced chip manufacturing and packaging, and SK Hynix and Samsung for HBM memory. Its 2027 supply increase therefore depends heavily on these partners adding capacity on time. Any delays in new wafer, packaging or HBM capacity could constrain AMD’s GPU ramp.”

Consumer fallout looks predictable. Radeon GPUs and Ryzen processors for PCs and laptops share manufacturing resources with AI parts. When data centers claim the newest wafers and the best packaging lines, older or less profitable lines get squeezed. Prices stay elevated. Availability stays spotty. Console makers, already dealing with memory constraints and higher component costs, face the same headwinds.

The original TechRadar coverage captured the frustration well. PC owners worry they’ll receive “whatever the data centers consider obsolete.” No direct promise of cheaper laptops or refreshed consoles emerged from Su’s remarks. Gaming segment revenue has already slipped in recent quarters, partly due to softer semi-custom chip sales for consoles.

Recent leaks add pressure. Next-generation gaming GPUs from both AMD and Nvidia face delays, with most RDNA 5 and Rubin-class parts now eyed for 2028 rather than 2027, according to reports citing memory constraints and priority given to HBM for AI. That leaves the current generation stretched longer than planned. Potential TSMC price hikes, rumored around 10% for certain lines, could push board partner costs higher still.

And yet AMD’s position strengthens. Its AI business has driven the stock up nearly 190% year to date. Big hyperscalers sign multi-year deals. Helios racks ship on schedule. The company even explores expanded foundry ties with Samsung. Demand signals remain “very strong” across the portfolio.

So the 2027 increase will arrive. It just won’t flood the consumer channel. Data centers will absorb the bulk of new capacity. Enterprises may see moderate easing by late 2027. Premium pricing, however, looks sticky. Shortages in supporting infrastructure — memory, power, real estate — won’t vanish overnight.

PC and console buyers sit at the back of the line. Again. Su’s Taipei tour locked in wafers and packaging for AI first. The rest follows. That’s the semiconductor order of operations in the age of massive model training. Supply grows. Priorities don’t shift. Consumers pay the price — literally — for longer.