Intel joined an effort to make AI accelerator systems less dependent on Nvidia hardware. Nvidia then bought about 5% of Intel. The purchase came with a multigeneration product partnership, but no disclosed agreement to manufacture Nvidia chips in an Intel fab.

Key takeaways

  • Nvidia completed its purchase of 214.7 million Intel shares at $23.28 per share in December 2026, for about $5 billion.
  • If Intel’s ownership of its foundry falls below 51%, the U.S. government can buy an additional 5% of Intel at $20 per share.
  • GlobalFoundries signed a five-year, $2 billion agreement in October to supply TSMC with silicon-interposer capacity in Malta, New York.
  • Intel said in November 2026 that it had signed its first outside 18A-P customer; 18A-P entered risk production in June.
  • Intel’s Xeon 6+ Clearwater Forest, its first 18A-based data-center CPU, offers configurations with up to 288 efficiency cores.

AI infrastructure is turning semiconductor rivals into capital partners and, in some cases, capacity customers without ending their competition over chip designs. Intel is the hardest test of that bargain: Nvidia’s money and Washington’s ownership conditions can help preserve a domestic manufacturing option, but only trusted processes, outside production commitments, and competitive products can make it a credible second source.

Intel built its business around the chip, not the shared factory

Intel’s server proposition once centered on selling Xeon processors into data centers. A 2014 account framed those chips as building blocks for modernizing legacy facilities: Intel designed the processor, made it, and sold the resulting product. Customers judged the chip they received. They did not have to decide whether to entrust their own designs to Intel’s production line.

AI systems widened the question. In 2024, Intel joined Google, Microsoft, Meta, and others in the Ultra Accelerator Link Promoter Group, which sought a standard for connecting accelerator server chips and reducing dependence on Nvidia hardware. A useful CPU or accelerator still mattered, but so did the connections among chips that might come from different designers.

Intel appeared in enterprise-framed coverage 22.1% of the time in the later of the compared 2024–2026 periods, up from 16.7% in the earlier one. That measures what the coverage discussed, not how many customers committed designs to Intel’s fabs.

In September 2025, Nvidia and Intel announced multiple generations of jointly developed x86 products, including RTX PC systems-on-chip and custom Nvidia data-center processors. Nvidia completed its purchase of 214.7 million Intel shares at $23.28 apiece in December, spending about $5 billion. The agreement gives both companies a direct product interest: Nvidia gains an x86 partner, and Intel gains an investor and collaborator. It does not establish that Nvidia bought Intel shares to sustain an independent foundry, or that Nvidia has placed a foundry order there.

Washington can keep the foundry attached to Intel

Intel’s CFO described the U.S. government’s arrangement as a deterrent to selling the foundry. If Intel’s ownership of that unit falls below 51%, the United States has an option to buy another 5% of Intel at $20 a share. The condition attaches a consequence to an ownership decision; it does not specify what the fab must yield or who must buy its output.

Governments are taking similar financial positions elsewhere. An October plan for two GlobalFoundries plants in Malta, New York, paired up to $21 billion in backing from Japan with a planned 4% U.S. stake in GlobalFoundries. The public interest is understandable: a fab takes capital and time to establish, and a country may value capacity that a chipmaker cannot justify on near-term sales alone. The stalled GlobalFoundries–STMicro project in France is a reminder that public support does not complete a difficult factory partnership.

Share ownership can hold an asset in place through a weak commercial period. It cannot settle the manufacturing questions that an outside designer asks before handing over a chip: whether the process meets its requirements, whether the schedule holds, and whether Intel can protect a customer’s design while selling competing products of its own.

TSMC bought the kind of capacity Intel still has to sell

Financing chip capacity is easier to see when the financing turns into a purchase. In October, GlobalFoundries signed a five-year, $2 billion agreement to provide TSMC with silicon-interposer production capacity at Malta. TSMC and GlobalFoundries compete as foundries, yet TSMC committed to buy a specific manufacturing service from its peer.

An interposer is not a leading-edge processor wafer, so that contract does not show GlobalFoundries replacing TSMC’s core fabrication business. It shows where selective reintegration has a practical address: a production line, a defined component, and a buyer obligated under a multiyear agreement. As AI systems combine processors and packaging from different suppliers, a rival can supply one constrained part without becoming the designer of the whole system.

Intel reached an earlier stage of that test in November. The company said it had signed its first outside customer for 18A-P, which entered risk production in June. That is a commercial milestone, not merely a financing event. It is not yet evidence of repeat, high-volume production for a range of independent designers. June reporting said Nvidia was testing Intel technology and that Google had placed an order for future TPU manufacturing; those accounts relied on sources, rather than an announced Nvidia foundry contract or a completed Google production run.

Intel’s own products give customers another test

Intel also has to demonstrate its process in chips it designs itself. Its Xeon 6+ Clearwater Forest is the company’s first data-center CPU built on 18A, with configurations offering up to 288 efficiency cores. That gives customers a product in which to assess Intel’s design and manufacturing execution together. It cannot, by itself, answer whether Intel will serve an outside designer neutrally.

Nor has strategic backing removed ordinary product and demand risk. AMD projected roughly $17 billion in fourth-quarter 2026 revenue against Intel’s $15 billion to $16 billion outlook; reporting treated the possibility that AMD could become the largest x86 vendor as a possibility, not an accomplished change. After Intel issued weaker-than-expected guidance, its shares fell 28.58% in one day, reducing the market value of the U.S. government’s stake by roughly $12 billion. Washington could constrain a foundry sale, but it could not constrain that loss.

Frequently asked questions

Who is Intel’s first outside 18A-P customer?

Intel had not identified the customer in the information cited here. The announced milestone confirms an external customer agreement, but not the customer’s name, chip design, order size, or production timetable.

Has Nvidia committed to manufacture chips at Intel’s fabs?

No announced Nvidia foundry order is disclosed in the agreement described here. Nvidia’s commitment is an equity purchase plus a multigeneration x86 product partnership, not a stated fab-production contract.

When can the U.S. government exercise its option to buy more Intel shares?

The option is tied to Intel’s foundry ownership: it applies if Intel’s ownership of that unit drops below 51%. The piece does not specify any current trigger, expiration date, or obligation for the government to exercise it.

When will Intel’s 18A-P customer move to high-volume production?

That timing is not disclosed. Risk production began in June, but the reported customer agreement does not establish repeat, high-volume manufacturing or a volume-production schedule.

Capital investment versus contracted manufacturing capacity

RelationshipTimingDisclosed commitmentManufacturing-order status
Nvidia → IntelDecember 2026214.7 million shares at $23.28 each; about $5 billionNo Intel fab order disclosed
TSMC → GlobalFoundriesOctober 2026Five-year, $2 billion agreementContracted silicon-interposer capacity at Malta, New York
Intel → first outside 18A-P customerJune–November 202618A-P risk production began in June; first customer announced in NovemberCustomer identity, volume, and schedule not disclosed

Nvidia’s name is on Intel’s shareholder register. For an independent second source, the name that matters is the outside customer on the production schedule.