Is Nvidia’s dominance starting to wobble? A once-in-a-generation alliance emerges—and China is on edge

AMD–Meta sign AI infrastructure supply deal

Supply-chain diversification; reduced reliance on Nvidia

AI competitive landscape shifts from “training” to “inference”

Focus on stronger bargaining power versus Nvidia

Chinese media “closely watching” the deal

AMD and Meta have signed an artificial intelligence (AI) infrastructure supply agreement of up to 6 gigawatts (GW). Under the deal, AMD will begin shipping an initial 1GW tranche in the second half of this year, centered on custom Instinct graphics processing units (GPUs) based on the MI450 architecture, 6th-generation EPYC central processing units (CPUs), and the Helios rack-scale architecture. The platform will run AMD’s ROCm software stack.

Industry watchers are also focusing on AMD’s decision to grant Meta warrants for up to 160 million newly issued shares—equivalent to roughly 10% of AMD’s outstanding stock. By awarding warrants in stages that allow Meta to purchase AMD shares at $0.01 per share depending on conditions such as Meta’s actual purchase volume and AMD’s share price, the companies are seen as elevating their partnership to the level of an “alliance.” As cooperation among U.S. big tech firms strengthens, there are signs that rival countries including China are growing concerned their domestic companies could face a disadvantage in the contest for AI infrastructure dominance.

AMD–Meta team up on AI infrastructure…Diversifying supply chains with an inference focus

AMD said on the 24th (local time) that it had signed a 6GW-scale AI infrastructure supply agreement with Meta. Following the news, voices across the industry said reliance on Nvidia was beginning to disperse. Meta CEO Mark Zuckerberg said the same day, “This is an important step for Meta in diversifying its computing resources,” adding, “We expect AMD to be a key partner over the coming years.”

Meta signed a deal on the 17th with Nvidia to secure supplies of GPUs and CPUs in the millions of units. It has also discussed supply options for Google’s AI chips, Tensor Processing Units (TPUs), and is reportedly developing its own chips. Zuckerberg’s comments, made alongside the AMD deal, were interpreted as a signal that Meta intends to deliberately reduce concentration on any single vendor.

Matt Britzman, senior equity analyst at Hargreaves Lansdown, said, “Meta is taking every necessary step to secure supply and reduce dependence on a single supplier, so that chip shortages don’t derail the expansion of its AI business.”

Another reason Meta is embracing AMD is that the agreement is focused on “inference.” The competitive landscape is no longer confined to a “GPU war” centered on training AI models. Inference infrastructure delivered to users worldwide has emerged as a core pillar of data-center investment. The inference-optimized MI450 chip is set to compete with Nvidia’s next-generation Vera Rubin platform.

Zuckerberg also emphasized that Meta formed a long-term partnership with AMD “to deploy efficient inference compute and deliver personal superintelligence.” Meta is combining hardware supplied by partners with its “Meta Training and Inference Accelerator (MTIA)” program.

China highlights an “AMD chase” of Nvidia…Deal may reshape bargaining dynamics

Chinese media described the agreement as “AMD’s chase of Nvidia.” Zhitung Caijing said, “This partnership means AMD will now compete head-on with industry leader Nvidia,” adding, “It shows that global hardware spending is accelerating even as some investors worry about an AI investment bubble.”

Some in the industry interpret the Chinese media’s framing of the deal as a “chase” as an attempt “to sound the alarm that domestic companies could fall behind in the AI infrastructure race.”

For Meta, the deal could serve as a foothold to strengthen its bargaining power. In OpenAI’s case, it was reported that merely signing a TPU contract with Google enabled it to negotiate Nvidia GPU prices down by 30%. Meta, too, could be in a more favorable position in price talks with Nvidia after the AMD agreement.

“Chips + equity” structure raises concerns over AMD’s ability to generate its own demand

With this deal, competition in AI accelerators is viewed as expanding beyond chip performance into a structure combining customer-tailored “system supply capability + software ecosystem + finance.” From a technology perspective, analysts say the competitive focus has further advanced from training to large-scale inference.

On the other hand, concerns have been raised that a supply model combining “chip supply + equity” could be problematic. Britzman said, “For AMD, this investment shows confidence in its next-generation AI hardware, but giving up a 10% stake suggests AMD may be struggling to generate demand on its own.”

There is also speculation that margins could hinge on Taiwan Semiconductor Manufacturing Co. (TSMC)’s production capacity. Techzine, an IT outlet, said, “Both AMD and Nvidia manufacture chips at TSMC, so the Taiwanese chipmaker’s capacity is the decisive factor,” adding, “If supply is constrained, margins will remain elevated.”

Meta, based on its own estimates, projected that capital expenditures (CAPEX) including AI infrastructure investment could reach as much as $135 billion this year. AMD CEO Lisa Su said, “This multi-year, multi-generation collaboration spanning Instinct GPUs, EPYC CPUs, and rack-scale AI systems aligns our roadmaps to deliver high-performance, energy-efficient infrastructure optimized for Meta’s workloads, accelerating one of the industry’s largest AI deployments and positioning AMD at the center of global AI buildouts.”

Kim Dae-young, Hankyung.com reporter kdy@hankyung.com