Reference
A plain guide to the companies behind modern computing: who designs each major class of chip, what they make, and what it powers.
Forty different company names can show up across a single AI server, a smartphone, and a data center switch, and most of them never appear on anything a buyer holds in their hand. This is a plain guide to those names: who makes what, organized by the job the chip does.
Makes the H100, H200, B200, and the newly launched Vera Rubin platform, along with its own Grace and Vera central processing units. These chips train and run AI models at the largest scale, in data centers operated by every major cloud provider.
Makes the Instinct MI300, MI325, MI350, and MI450 accelerators, sold as part of its Helios server rack alongside its own EPYC "Venice" central processing unit. Used for the same AI training and inference work as Nvidia's chips, and increasingly chosen as a second supplier by companies like Meta and OpenAI.
Makes the Gaudi accelerator line and is developing Falcon Shores. Used for AI inference, though Intel's larger near-term role in this story is manufacturing chips for other companies, covered further down.
Makes the Ascend 910C and 950PR accelerators. Used for AI training and inference inside China, where it has become the leading alternative to Nvidia hardware following export restrictions on Western chips.
A Chinese chip designer working with Huawei on AI hardware. Used as part of China's effort to build AI infrastructure that does not depend on Nvidia's CUDA software ecosystem.
Makes the Tensor Processing Unit, now in its seventh generation, known as Ironwood. Used to train and run Google's own AI models, and rented out to other companies, including Anthropic, that need large-scale compute.
Makes the Trainium and Inferentia chips through its Annapurna Labs unit. Used inside Amazon Web Services for both Amazon's own AI workloads and customer-facing cloud instances; Amazon has said it has deployed more than a million Trainium chips.
Makes the Maia AI accelerator and the Cobalt general-purpose central processing unit. Used inside Azure data centers, with Maia targeting AI inference and Cobalt reducing Microsoft's reliance on Intel and AMD for ordinary cloud workloads.
Makes the MTIA accelerator. Used internally for the recommendation and ranking systems that power its apps, with newer generations aimed at AI training as well.
Reportedly developing its own accelerator program, under the name Titan, with Broadcom as a design partner. Intended, if it ships as planned, to reduce OpenAI's reliance on Nvidia hardware for running its models.
Anthropic does not make its own chip. It runs large training and inference deployments on Google's and Amazon's hardware instead, at a scale that puts it in similar infrastructure conversations as the hyperscalers above.
Co-designs custom chips for Google, Meta, Microsoft, and reportedly OpenAI, and separately makes its own Tomahawk and Trident networking switch chips. Used both as the engineering partner behind most of the hyperscaler chips listed above, and as the company most AI clusters rely on to move data between accelerators.
Co-designs Amazon's Trainium and shares design work on Microsoft's Maia, and makes its own Teralynx switch silicon. Plays the same two roles as Broadcom, on a smaller scale, and is its closest competitor in networking.
Makes the Silicon One chip family, including the newly launched G300. Used for routing and switching inside AI clusters, as a newer entrant competing directly with Broadcom and Marvell.
Manufactures chips designed by nearly every company named above, on its most advanced process nodes. Used as the default fabrication plant for any chip that needs to compete on performance; its available capacity is the industry's most closely watched constraint.
Manufactures chips at advanced process nodes as well, and recently won orders from Nvidia, Tesla, AMD, and BYD. Used as a second source by companies looking to reduce their dependence on a single foundry.
Reopened Intel's leading-edge manufacturing to outside customers for the first time in the company's history. Used by a growing list of external customers as its 18A process matures.
China's leading domestic foundry. Used mainly to manufacture Huawei's Ascend chips, under capacity allocation set by the Chinese government.
Makes high bandwidth memory, the stacked memory chips paired with AI accelerators, and holds the largest share of that market. Used by Nvidia and other accelerator makers as a primary memory supplier.
Makes both high bandwidth memory and standard dynamic random access memory. Used across AI servers and consumer electronics alike, and recovering share in high bandwidth memory after a slower start.
Makes high bandwidth memory and standard memory, and recently exited parts of the consumer memory business to prioritize data center customers. Used by a growing list of AI accelerator makers as a third qualified memory supplier.
Licenses chip designs to nearly every mobile chipmaker and a growing number of server chipmakers, and recently began selling its own chip directly for the first time. Its architecture underlies most of the custom central processing units in this piece, including Amazon's Graviton, Microsoft's Cobalt, and Google's Axion.
Re-entering the server market with a new rack design built to connect into Nvidia's NVLink Fusion interconnect, after a previous retreat from that business. A new option in a server central processing unit market that had narrowed to two real suppliers.
Intel and AMD remain the two long-standing default central processing unit suppliers for enterprise servers, through Intel's Xeon 6 line and AMD's EPYC line, both described above.
Makes the Snapdragon line. Used in premium and mid-range Android phones across nearly every major brand.
Makes the Dimensity and Helio lines. The highest-volume chip supplier in the smartphone market, especially strong in budget and mid-range devices.
Makes the A-series chip used exclusively in its own iPhones. Used only inside Apple's own devices, which makes its chip market share a direct reflection of iPhone sales.
Makes the Exynos line, including the first 2-nanometer mobile chip to reach mass production. Used increasingly inside Samsung's own Galaxy phones in place of Qualcomm silicon.
Makes budget-tier chips including the T7250. Used primarily in lower-cost devices from brands like Redmi.
That covers the full set of names worth knowing to understand what is running inside the technology an organization buys, trains on, or carries in its employees' pockets: who designs it, who builds it, and what job it does.
