AMD CEO Lisa Su: Customer Demand for AI Chips Exceeds Production Capacity
AMD CEO Lisa Su stated in Taipei that customer demand for artificial intelligence chips has outpaced the manufacturing capacity across the industry. The company has increased supply this year, but is still constrained by capacity, with memory being a clear bottleneck.
Su noted that the CPU market has seen significant growth by 2026. AMD has simultaneously increased shipments and capacity, but demand still exceeds supply. She mentioned that supply will significantly increase by 2027, but "there is definitely still room for more." She pointed out that the planning targets include CPUs, GPUs, and other AI computing products.
This is Su's third visit to Taiwan this year. On the same day, she met with Acer, Asus, Foxconn, and Quanta, and in the afternoon, she visited TSMC and other suppliers in Hsinchu. She requested that wafer, backend testing, and substrate capacities be expanded simultaneously to avoid a situation where only chip production is increased while complete systems cannot be delivered.
The over $10 billion investment in the Taiwanese supply chain announced in May is progressing as planned, and due to rising demand, Su stated that the subsequent amount will be adjusted upward, but did not provide a new figure. This investment covers advanced packaging, substrates, and cabinet-level systems, with the capacity lock window extended from 2026 to 2029. The EPYC Venice has entered mass production at TSMC's 2nm process.
Industry reports indicate that agent-based AI has pushed both CPU and GPU demand beyond original forecasts. The Venice series server CPUs are reportedly sold out until 2027, with some orders pushed to 2028. Shipments of MI450 in the second half of the year and MI455 in 2027 depend on whether advanced packaging, testing, and substrate slots can be secured.
Buy orders are coming from cloud vendors and AI server customers, while sellers are constrained by TSMC's wafers, HBM, and packaging, not due to insufficient orders. This is an event-driven supply lock, with funds flowing from chip design to Taiwanese packaging, substrates, and foundries. AMD, TSMC, Foxconn, and Quanta benefit from proprietary rights, while memory manufacturers facing shortages and secondary customers unable to secure quotas are under pressure, with near-term revenues capped by shipment volumes.
Stifel raised AMD's target price from $635 to $700 ahead of the quarterly report, maintaining a buy rating, citing that AI infrastructure is limited by supply chain rather than demand. Su will subsequently visit South Korea, with the market focusing on the progress of collaboration with Samsung on HBM.
Source: Public Information
ABAB AI Insight
When Lisa Su took over in 2014, AMD had sold off its fabs and relied on GlobalFoundries for supply, with server market share being dominated by Intel. She pushed advanced processes to TSMC, transforming data center CPUs from catching up to competing for market share with Zen and EPYC, and then directly challenging Nvidia's H100/H200 with the Instinct MI300 series. The shipment of MI300X in 2024 is hindered by CoWoS packaging, which is a direct historical precedent for this early slot locking.
This investment does not buy a brand but secures packaging, substrates, and cabinet slots from 2026 to 2029. The over $10 billion investment in Taiwan announced in May ties the 2nm EPYC Venice and MI450 on Helios cabinets to the same delivery timeline with TSMC, Foxconn, and Quanta. Her visits to Acer, Asus, Foxconn, and Quanta, followed by a trip to Hsinchu, aim to extend the planning cycle from one to two years for cross-generation locking, avoiding mismatches like in 2024 where chips were available but advanced packaging was not.
This approach is similar to Nvidia's prepayment to TSMC and securing CoWoS for 2023 to 2024, and also resembles Broadcom's use of custom ASICs to lock cloud vendors' capital expenditures into long-term agreements. The difference is that AMD is still competing for market share in both CPU and GPU lines, rather than relying solely on GPUs. The industry is currently in a capacity control phase, where design capability is no longer a bottleneck; only those who can secure 2nm, HBM, and testing quotas can convert orders into revenue.
Essentially, this represents a transfer of pricing power. Once demand exceeds supply, prices and delivery times are determined by the synchronized capacity of wafers, memory, and packaging, rather than solely by chip specifications. Su has clearly stated that supply increments will be significantly increased by 2027, acknowledging that the incremental orders for 2026 will primarily turn into backlogs. Whoever secures Taiwanese capacity first will gain the shipping rights for the next round of AI servers.
ABAB News · Cognitive Laws
Orders are not revenue; the ability to deliver is revenue.
The shortage in chips is not in design but in synchronized capacity.
When demand is overheated, quotas are more expensive than performance.