AMD CEO Lisa Su said on October 6 that the company will substantially increase chip supply in 2027. According to Reuters' interview in Taipei, AMD will keep expanding supply after 2026 and aims to meet AI demand with both CPUs and GPUs. Among large customers, OpenAI's accelerating rollout and Anthropic's start of deployment overlap in 2027. How far the added output can meet demand depends on whether AMD can line up not only compute chips but also memory and entire racks at the same time.
Expanding supply in 2027, with planning that reaches 3 to 5 years out
Su outlined a plan to widen in 2027 the supply that AMD built up in 2026. Reuters reported that one purpose of her Taiwan visit was to prepare the supply chain so AMD can raise CPU and GPU production to match AI demand. She said she had already met Foxconn and would also hold talks with TSMC.
Su also explained that, because AMD sees very strong demand over the next several years, it needs more leading-edge wafer capacity. Its planning horizon extends 3 to 5 years ahead. She also disclosed that AMD is working closely with memory makers to secure supply. Taken together, the places she visited and her remarks suggest AMD is trying to line up not only more chips from its manufacturing partners but also supply on the system side.
However, her remarks did not give a rate of increase or specific shipment numbers. It is also unclear how much CPU versus GPU supply will grow, or how many units can be delivered each quarter. From the phrase "substantially increase," one cannot conclude that supply of any particular product will double, or that shortages across the entire AI chip market will be resolved.
AMD's own plan to increase supply and the question of whether it can satisfy market demand need to be considered separately. The difference becomes clearer when looking at the customer deployment plans that have already been announced.
Major customers' deployment plans overlap in 2027
OpenAI announced a multiyear, multigeneration 6GW agreement with AMD on October 6, 2025. The Meta agreement, announced on February 24, 2026, covers up to 6GW, and the Anthropic partnership, announced on July 22, covers up to 2GW. All are large, but these figures cannot simply be converted into 2027 shipment volumes.
The table below sorts AMD's official announcements into the overall scale of each plan and the timing of initial deployment. For OpenAI, the July 23 update is included to preserve the differences among the stages reached: shipment, operation and start of deployment.
| Customer | Announced overall scale of plan | Timing and stage of initial deployment | Announcement date |
|---|---|---|---|
| OpenAI | 6GW over multiple years and generations | Initially, the first 1GW to be deployed starting in the second half of 2026. The update says Helios is expected to go into operation from Q4 2026, with deployment accelerating in 2027 | October 6, 2025, updated July 23, 2026 |
| Meta | Up to 6GW over multiple years and generations | Shipments of products supporting the first 1GW deployment scheduled to begin in the second half of 2026 | February 24, 2026 |
| Anthropic | Up to 2GW | Deployment of the first 1GW to begin in the first half of 2027 | July 22, 2026 |
Among the announced deployment plans, OpenAI's accelerated rollout in 2027 overlaps with the start of Anthropic's first 1GW deployment. AMD has to expand supply to an existing customer while also ramping up supply to another major one.
The GW figures in the table indicate the electrical power scale of the equipment each company announced it would deploy. It is not a unit for comparing GPU counts or AI processing performance. Moreover, if the total of a multiyear plan and the first deployment stage are mixed in the same column, it can look as if more equipment is already running than actually is. Here, the totals are not added together; the focus is on which deployment stages overlap in 2027.
Meta's initial configuration also has its own conditions. The GPU is a custom MI450-based part tailored to the company's workloads, and the CPU is the 6th-generation EPYC "Venice." Because products and workloads must be matched customer by customer, the number of standard racks cannot be calculated from the contract GW alone.
Adding GPUs alone does not complete a rack
The standard Helios that AMD showed on July 23 combines 72 Instinct MI455X GPUs and 18 Venice CPUs in a single rack. The high-bandwidth memory on the GPU side is HBM4, and total capacity across the rack reaches 31TB. The design houses 18 trays, each carrying 4 GPUs, with networking equipment and ROCm software integrated.
What this configuration shows is that increasing finished GPUs alone does not necessarily mean racks can be shipped at the same ratio. The CPUs must be available, the GPUs must be fitted with the memory they need, and the whole must operate as a system including communications equipment. Su is holding talks with Taiwanese manufacturing and systems companies and is also working with memory makers to secure supply. Her remarks on both CPUs and GPUs can be read as a response to this combination of supply requirements.
On the memory side, concrete collaboration is already under way. The memorandum of understanding Samsung and AMD announced on March 18 includes primary HBM4 supply for the MI455X, as well as collaboration on DRAM for Venice and DDR5 for Helios. It is an effort to prepare the HBM that handles large volumes of data near the GPU, and the memory used on the CPU side, for the next systems.
Still, the announcement of an MOU does not prove that all the memory needed in 2027 has been secured. It does not state supply quantities by product. Also, the foundry collaboration with Samsung mentioned in the same announcement is at the stage of discussing opportunities, and does not mean a manufacturing partner for future AMD products has been decided.
Securing leading-edge wafer capacity and assembling components including memory is only part of the work. Validation still remains before racks can run customers' workloads reliably. This is why the assessment of the added output should extend to operation on the customer side.
The results of added output will be tested by customer-side operation
In its July 23 announcement, AMD described Helios as already in production. In the same announcement, however, OpenAI's start of operation was expected in Q4, and Meta was at the stage of testing and validating workloads on its CPU platform and on Helios. Start of production, shipment to customers and operation of customer facilities are separate milestones.
Software support also bridges that gap. OpenAI and AMD are optimizing GPT-family workloads using Triton and ROCm, and the collaboration with Anthropic includes optimizing workloads on GPUs using Claude and developing ROCm. This is the work of making customers' models run efficiently on racks assembled from the components. Being able to procure chips is not the same as being able to draw the needed processing capability out of them.
AMD has also indicated plans to introduce the MI500 series in 2027. Because large-scale MI455X deployments and the launch of the next generation fall in the same year, not only total supply volume but also which generation goes to which customer will bear on the assessment. The latest remarks on increased output do not newly commit to specific MI500 shipment volumes or mass-production schedules.
Whether prices for consumer GPUs and PC memory will fall also cannot be judged from these remarks alone. No basis was given for directly inferring price or delivery-time effects on other product lines from expanded data center supply.
From late 2026 into the first half of 2027, it will become possible to check OpenAI's expected start of operation and Anthropic's deployment start plan against actual progress. If AMD can deliver racks with CPUs and memory in place and get customers' workloads running, its added output will expand the compute resources and procurement options available to AI companies.