AMD promete aumento substancial de chips de IA em 2027, diz CEO Lisa Su

AMD CEO Lisa Su said on Tuesday (Oct. 6), in Taipei, that the company will "substantially increase" its supply of artificial intelligence chips in 2027. The statement was made to reporters during a trip to Taiwan whose goal, according to her, was to ensure that AMD's supply chain can scale production of CPUs and GPUs to meet growing AI demand, according to a Reuters report published by ET Tech.

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.

Lisa Su, CEO of AMD

AMD is today Nvidia's main rival in the AI GPU market, and its market value recently surpassed US$1 trillion amid a broad rally in companies tied to artificial intelligence. Su's trip included a meeting with Foxconn and a scheduled meeting with TSMC, as well as a stop in South Korea, home to memory makers Samsung and SK Hynix, with whom AMD is working to secure sufficient supply of memory chips.

The bottleneck isn't just GPUs, it's advanced wafers

Asked whether TSMC had discussed a possible investment in Texas with AMD, Su did not answer directly. But she was direct about the reason for any expansion: the company needs more advanced wafer capacity to sustain projected demand.

There's very, very high demand for the next several years. We can see very high demand, and so we need more advanced wafer capacity.

Lisa Su, CEO of AMD

Su said AMD is now planning three to five years ahead, a horizon typically used when the real constraint isn't chip design but the physical capacity to manufacture it. Reuters had previously reported, the week before, that TSMC is weighing an investment in Texas, which would add more advanced manufacturing capacity on American soil, something both AMD and Nvidia depend on to scale production.

What this means for those deciding on architecture today

For those who architect AI systems, train, or serve models, the accelerator supply bottleneck is nothing new: waiting lists for cutting-edge GPUs and high prices have been part of the capacity planning for any team handling training or inference workloads at scale since the start of the generative AI boom. Su's announcement is relevant because it signals when (and whether) this bottleneck starts to ease specifically on AMD's side.

In practice, this changes the calculation for those deciding today between staying locked into Nvidia's CUDA ecosystem or diversifying with AMD accelerators via ROCm. If 2027 supply really grows as promised, cloud providers and companies that today struggle to secure GPU allocation gain a more available competing supplier, which historically pushes prices and lead times down across the whole chain, including on Nvidia's side.

This doesn't solve the software maturity problem, though. Anyone who has tried migrating training workloads from CUDA to ROCm knows hardware availability is only half the equation: libraries, compatibility with frameworks like PyTorch, and response time to stack-specific bugs still weigh on the decision. Su's announcement talks about silicon; it doesn't, by itself, resolve the ecosystem gap.

Memory is also a bottleneck, and Brazil feels it indirectly

Su's mention of working with Samsung and SK Hynix to secure memory supply is a detail that tends to go unnoticed in coverage focused solely on GPUs. High-bandwidth memory (HBM) chips are, along with TSMC's advanced wafer capacity, one of the main limiters of how many AI accelerators can roll off the production line per quarter.

For infrastructure teams in Brazil, who generally don't buy hardware directly but instead contract cloud instances with AMD or Nvidia GPUs, this kind of supply constraint shows up indirectly: in instance pricing, in the availability of specific machine types in regions like São Paulo, and in wait times to provision additional capacity during demand spikes. A real supply relief in 2027, if it comes, tends to reach end users with a delay, through contract renewals and data center expansion by cloud providers.

What remains unclear

Su didn't detail numbers: there is no production increase percentage, additional unit count, or quarterly timeline in the reported remarks. "Substantial" is a CEO's word at a press event, not a planned capacity figure, and the semiconductor industry's recent track record shows that advanced wafer capacity expansion tends to lag behind the original announcement, since it depends on third parties like TSMC.

It's also unclear whether TSMC will actually invest in Texas, and what share of that additional capacity (if it materializes) would be reserved for AMD versus other customers like Nvidia itself and Apple. These are two variables that determine whether the 2027 promise translates into more real supply in the market or stays concentrated in specific contracts closed before competitors show up.

AI safety also entered the conversation

Outside the supply chain discussion, Su said AI can be "incredibly good for the world," but argued that companies need to stay alert to risks and work together on AI safety. She cited the summit between US President Donald Trump and tech executives the week before as an important step, mentioning that participating companies agreed to work with "independent auditors" to assess whether AI systems operate as planned by their designers, according to an agreement announced by Trump himself on his social media platform.

For those building AI products, this is a sign that external system audits may stop being the exception and become a contractual or regulatory requirement in key markets, something engineering teams that currently treat model governance as a secondary roadmap item would do well to watch closely.