AMD in AI: Instinct GPUs, OpenAI Deal & Taalas

The other chip company

AMD, incorporated in 1969, makes processors: Ryzen chips for consumer computers, EPYC chips for servers, Radeon graphics cards, and, under the Instinct brand, the accelerators it sells for AI training and inference. Lisa Su has been chief executive since 2014, and the company has spent that decade turning a longtime distant-second-place chipmaker into Intel's closest competitor in server processors and, on AI accelerators specifically, into the most credible alternative to Nvidia. Its 2022 acquisition of Xilinx added programmable chips used in networking and defense hardware to that lineup. [2]

Atlas interpretation: AMD's relevance to AI is less about any single model and more about market structure: it is the company whose existence keeps Nvidia's AI chip dominance from being total, which matters to every lab and cloud provider that would rather not depend on one supplier for the hardware training runs and inference at scale require. [2]

A deal that ties a customer's fortunes to its supplier's stock

In October 2025, AMD and OpenAI announced a partnership to deploy six gigawatts of Instinct GPUs across multiple hardware generations, with the first gigawatt of MI450 chips arriving in the second half of 2026. AMD's own announcement describes issuing OpenAI a warrant for up to 160 million shares of AMD stock, vesting in tranches tied to deployment milestones and to AMD's share price crossing set thresholds. The structure gives OpenAI an ownership stake in its supplier rather than a conventional purchase agreement, and ties AMD's incentive to actually deliver the hardware on schedule. [3]

Buying a bet that some models hold still

In August 2026, AMD acquired Taalas, a startup that etches a specific model's weights directly into chip silicon rather than running the model on general-purpose hardware. AMD's announcement describes the deal as advancing compute for AI inference; independent coverage describes the underlying trade as giving up the flexibility to run other models on the same chip in exchange for lower inference cost, a bet that pays off only for a model stable and popular enough to be worth hard-wiring. [4][5]

Atlas interpretation: The two deals point the same direction from opposite ends of AMD's business. The OpenAI partnership is AMD selling flexible, general-purpose accelerators at a scale that requires giving its customer equity to make the commitment credible. The Taalas acquisition is AMD betting that a slice of the market wants the opposite: hardware with no flexibility at all, in exchange for running one model as cheaply as physics allows. [3][4]