4 hours ago
15
AMD has a problem most companies would love to have. Its customers want more AI chips than it can physically make.
On October 6, 2026, AMD CEO Lisa Su said from Taipei that demand for artificial intelligence chips is running ahead of the semiconductor industry’s ability to produce them.
What Lisa Su said in Taipei
Su’s message was blunt. AMD has worked to boost its chip output, but the company is still supply constrained.
Memory production stood out as a particular pressure point. Su also pointed to bottlenecks in overall manufacturing capacity across the supply chain.
AMD’s response, according to Su, is coordination. The company is working with both suppliers and customers to align strategic plans aimed at significantly increasing supply in 2027.
Su’s comments came alongside meetings with Foxconn, the contract manufacturing giant, with discussions planned with TSMC as well.
Sold out through 2027, with orders reportedly sliding to 2028
Su’s remarks landed at the same time as reports of a crunch in server CPU and AI accelerator supplies. Those are the processors and specialized chips that power data centers running AI models.
The driver, per those reports, is agentic AI. Workloads tied to agentic AI have reportedly surged beyond prior estimates, pushing requirements for both GPUs and CPUs to unprecedented levels.
AMD’s EPYC Venice series, its server CPU line, is reportedly sold out through 2027. Orders for those chips are reportedly being pushed into 2028.
AMD has ramped up production throughout 2026 and has secured significant commitments from major clients. It has also committed over $10 billion in Taiwan to support supply chain expansion.
AMD has seen significant growth in data-center revenue as demand for AI applications climbs.
Why manufacturing, not demand, is now the limit
The growth of AI infrastructure now depends heavily on manufacturing capacity and memory availability. Foxconn handles assembly at massive scale. TSMC manufactures chips for a long list of designers, AMD included.
When Su singles out memory as a constraint, it means even a perfectly designed processor can sit idle without the memory components needed to ship it as a finished product.
What this means for AMD, customers and the AI build-out
Supply constraints cap how much of that demand AMD can convert into revenue in the near term.
The 2027 supply push is the date to circle. If AMD and its partners deliver the significant increase Su described, it could ease current tensions.
If those agentic AI workloads keep exceeding forecasts, as reports suggest they already have, even the expanded 2027 supply could end up stretched.
Su’s Taipei visit, the Foxconn meetings, the planned TSMC talks and the more than $10 billion committed in Taiwan all point to the same island as the center of gravity for AI hardware production.
Su’s own framing suggests constraints are industry-wide. She described the gap as one between AI demand and the industry’s manufacturing capability, not just AMD’s.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.
English (US) ·