10 most valuable public AI chip companies

A buyer can now get most of an AI rack from Nvidia alone. Its Vera Rubin platform is in full production, putting Nvidia's GPUs next to its own Vera CPUs, Spectrum-6 switches and BlueField-4 processors, and its inference offering also uses Groq technology under a nonexclusive licence. Data centre revenue reached $89.0 billion in the quarter to 26 July 2026, and it forecasts total revenue of $108 billion, plus or minus 2 percent, for the following quarter.

Its customers now buy in gigawatts. Nvidia says OpenAI's existing and planned commitments represent about 12 gigawatts of its compute, while AWS plans to add 2 million GPUs through the second quarter of Nvidia's fiscal 2029. Anthropic's initial commitment is up to 1 gigawatt, and Nvidia announced plans to invest up to $10 billion in the lab. To keep up, Nvidia raised its supply and capacity commitments from $119 billion to $279 billion in one quarter. Its revenue outlook assumes no data centre compute sales to China.

When Google, Meta and OpenAI design their own AI chips, Broadcom turns those designs into silicon: Google's TPUs, Meta's MTIA and OpenAI's Jalapeño. It also sells the Ethernet that connects the resulting clusters. On its September call, as reported by SDxCentral, management sized two programs:

• Anthropic: an upcoming 1 gigawatt Ironwood deployment, then 5 gigawatts of TPU v8i in fiscal 2027, with prospects for another 10 and no year given for those.
• OpenAI: 1.3 gigawatts planned for 2027, with prospects for more than 5 across Jalapeño and its successor, also without a year.

AI semiconductor revenue was $16.7 billion in the quarter to 2 August 2026. Management's outlook is about $115 billion in fiscal 2027 and $230 billion in fiscal 2028, on supply it says is secured, and Counterpoint projects it will hold about 60 percent of AI server compute ASIC design partner work in 2027. Not all of the value is chips: infrastructure software, including VMware, brought in $8.75 billion of the quarter's $29.6 billion.

Three of the biggest AI buyers have signed for AMD GPUs by the gigawatt. OpenAI's 6 gigawatt agreement and Meta's 6 gigawatt agreement span multiple generations, starting with MI450 based GPUs, and Anthropic plans up to 2 gigawatts of MI450 series GPUs in Helios racks. That makes AMD a substantial second source for all three. OpenAI and Meta each received warrants for up to 160 million AMD shares, tied to deployment or shipment milestones as well as share price, technical and commercial conditions.

Most of the execution is still ahead: the planned first deployments begin in the second half of 2026 for OpenAI and Meta and the first half of 2027 for Anthropic. Today AMD's data centre segment, Instinct GPUs and EPYC server CPUs together, generated $6.7 billion in the quarter ended 27 June 2026.

An AI server still needs a host processor beside its accelerators, whoever makes the GPUs, and even Nvidia wants Intel's: its announced $5 billion investment agreement calls for Intel to build custom x86 CPUs for Nvidia's AI platforms. Intel reported data centre and AI revenue up 59 percent to $6.3 billion in the second quarter of 2026 and showed Xeon based systems with SambaNova and Foxconn. Its factories are a separate bet, on winning business from outside chip designers; Intel's filings still list securing major external customers for 14A and later processes as a risk.

Amazon's Graviton, Google's Axion, Microsoft's Cobalt and Nvidia's Vera compete for the same AI server racks, and all four are built on Arm technology, so Arm earns licence fees and royalties from each of them. Now it sells silicon too. The AGI CPU, developed with Meta as lead partner and customer, has OpenAI and Cerebras integrating it, according to Arm, which says customer demand exceeds $2 billion across fiscal 2027 and 2028. Becoming a chip supplier widens its opportunity and puts it into some of the markets its licensees serve.

Look past the accelerator in an AI rack and Marvell is in the chips around it. Google signed a July agreement for silicon attached to its TPU systems, including network interfaces, memory interfaces and storage controllers; most of Google's associated warrant vests as its purchases grow, a smaller portion over time. Nvidia separately invested $2 billion and is working with Marvell on custom chips and networking compatible with NVLink Fusion. Its optical connectivity business gains a role as clusters grow. In custom AI server compute ASIC design itself, Counterpoint still projects about 8 percent share in 2027, well behind Broadcom.

Qualcomm already puts AI processing into phones, PCs and cars through Snapdragon; the data centre is its next market. In July it said two hyperscaler custom chip programs had purchase orders and would start generating revenue in the December quarter, and it has since named Amazon in a multigeneration AI silicon and optical connectivity collaboration. Management targets about $5 billion of data centre revenue in fiscal 2027 and more than $15 billion by fiscal 2029. For now handsets still carry its chip business, and they slipped: handset revenue fell 20 percent to $5.1 billion in the quarter ended 28 June 2026.

US export controls keep squeezing Nvidia's sales in China, and Cambricon is one of the domestic chipmakers growing into the gap. Its cloud product line, chips, boards and systems, accounts for almost all of its revenue, which more than doubled to 6.0 billion yuan in the first half of 2026. It is still a small player in its own market: IDC data reported by Reuters on 13 October put Cambricon at 5 percent of China's AI accelerator shipments in the first half, against Nvidia's 34 percent and Huawei's 28 percent.

Hygon names China's tax and customs administrations, its state banks and the three national telecom operators among its customers, according to its own investor response. What it sells them is x86 compatible server CPUs paired with DCU accelerators, for China's push to domestic computing infrastructure. Revenue rose 66.5 percent to 9.1 billion yuan in the first half of 2026; the report gives no split between CPUs and DCUs.

Cerebras builds one processor the size of a silicon wafer and rents out its speed as an inference cloud. OpenAI committed to 750 megawatts of it under a multiyear agreement, valued by Cerebras at more than $20 billion, and at the end of June remaining performance obligations stood at $25.4 billion, including some data centre pass through costs. The design avoids HBM and CoWoS packaging, which Cerebras says eases supply constraints. The contracted growth comes with concentration: three customers accounted for 76 percent of second quarter revenue.