Broadcom explained: the business behind custom AI chips

When Meta set out its expanded custom-chip partnership with Broadcom in April 2026, the agreement covered chip development, advanced packaging and Ethernet networking. Those responsibilities explain Broadcom’s place in AI: it helps turn a customer’s computing requirements into processors and the connections around them. The collaboration is intended to span multiple chip generations through 2029.[2]

A custom processor still needs a large supporting industry. Broadcom brings reusable chip designs and interfaces, engineers who adapt them, and products that move data between computers. Alongside that semiconductor business, it sells the software that organisations use to run private clouds. Understanding both sides makes the company considerably easier to follow.[1]

What the customer customises

Broadcom’s annual report describes an application-specific chip platform that combines its intellectual property with a customer’s requirements. The building blocks include processor logic, connections to high-bandwidth memory and high-speed interfaces that send data in and out. Engineers must make these components work together inside an advanced package, then support the design through manufacturing. Broadcom calls its custom AI accelerators XPUs. The commercial product is therefore a customer-specific processor supported by a platform that can serve successive designs.[1]

Meta’s MTIA programme illustrates the division of work. Broadcom’s role extends across logic, memory integration, high-speed input/output and packaging, with its Ethernet portfolio supplying network connections. The April announcement describes an initial deployment commitment exceeding one gigawatt and plans for further expansion. That figure sizes the commitment; it does not establish that the promised computing infrastructure was already operating.[2]

Packaging is part of the engineering product

A modern accelerator can combine several pieces of silicon, or dies. Their arrangement affects how power and data travel through the package. Broadcom’s XDSiP technology combines vertical die stacking with a wider arrangement of chips and memory in the same package. Its face-to-face connections join the upper metal layers of stacked dies. The company also supplies design methods for power delivery, clock distribution and signal integrity. These are repeatable engineering capabilities that can be adapted to different processors; the customer does not have to invent every connection from the beginning.[3]

There is a useful example beyond a frontier-model developer. In February 2026, Broadcom said it had begun shipping a custom 2-nanometre compute chip for Fujitsu using this packaging platform. The company describes an architecture in which computing, memory and network input/output can scale independently. The shipment is a concrete milestone for a particular customer. It also shows why the underlying integration technology can matter commercially across more than one chip programme.[4]

Manufacture remains another dependency. Broadcom’s latest quarterly filing says TSMC produced approximately 95% of the wafers made by its contract manufacturers during the first three quarters of fiscal 2026. That denominator matters: it describes Broadcom’s outsourced wafer production. Broadcom’s design and integration role sits alongside the foundry’s physical fabrication capability.[6]

AI chips are growing inside a two-part business

The results show why describing Broadcom solely as an AI-chip supplier misses a substantial business. Infrastructure software accounted for 29.6% of quarterly revenue, calculated from the reported segment figures. Semiconductor revenue also extends beyond AI. Broadcom’s AI semiconductor measure includes custom accelerators and networking, so it cannot be read as sales of custom processors alone.[5]

The software side includes VMware Cloud Foundation, or VCF. Its AI capabilities address an operating problem: different teams may want access to the same model without each maintaining another copy. Broadcom’s September description of VCF 9.1.1 says a shared Model Runtime can host and scale models while organisations use separate namespaces. That multi-tenant model sharing was listed as generally available. The same article placed the AI Gateway among future-release capabilities, a distinction that matters when evaluating what can actually be installed.[7]

Standard Chartered provides an example of the broader customer relationship. In July, Broadcom and the bank said VCF underpinned a private-cloud architecture supporting 70% of the bank’s global infrastructure footprint. The account describes a foundation for banking applications, payments and digital services. It is evidence of enterprise infrastructure use; the percentage does not measure how much banking work AI performs.[8]

A design win creates a lasting relationship—and exposure

Custom work ties Broadcom to its customers’ product decisions. The quarterly filing describes competition for designs that consumes engineering resources before the investment is assured of a return. If a customer changes plans, a processor designed around that customer may have limited resale opportunities. Winning one generation can create an opening for the next; losing it can close that route.[6]

Sales concentration needs equally careful interpretation. One distributor accounted for 50% of Broadcom’s net revenue in the latest quarter. A distributor is a route through which products reach customers. The filing’s figure cannot identify a single AI company buying half of Broadcom’s output.[6]

The company’s appeal consequently rests on work that continues around the chip: integrating its components, adapting the next design and connecting it to a usable system. The Fujitsu shipment makes that progression tangible. For a reader following AI suppliers, the useful question is which part of this engineering relationship a new agreement actually adds—and when that part becomes a delivered product.[4]

Report an error or suggest a correction