Broadcom is designing custom AI chips for Google, Meta, OpenAI and Anthropic at the same time — and holds $73 billion of orders it has not shipped

Every figure in this article is drawn from Broadcom's fiscal 2026 quarterly results and earnings calls, or from CNBC and industry reporting as attributed. Broadcom's fiscal year ends in early November. Sections marked as analysis are identified as such.

The most consequential fact about the artificial-intelligence buildout is not that Nvidia dominates it. It is that every company large enough to object to Nvidia's pricing is now designing its own chip instead — and that a strikingly high proportion of them are designing it with Broadcom.

Google's TPUs. Meta's MTIA accelerators. A custom inference engine for OpenAI, confirmed by Hock Tan as the company's sixth major hyperscale customer and worth more than $10 billion, entering mass production late this year. And Anthropic, whose multi-gigawatt TPU expansion this year was announced jointly with Google and Broadcom. Four of the most important AI companies in the world, and one common supplier.

The numbers

AI revenue reached $8.4 billion in the first quarter, up 106 per cent, with custom accelerator revenue up 140 per cent. Management guided the second quarter to about $10.7 billion of AI revenue, roughly 140 per cent growth. Infrastructure software — largely VMware, whose Cloud Foundation 9.0 release has driven the conversion of perpetual licences to subscriptions — contributed $6.8 billion.

The forward book is the part that has repriced the stock. Broadcom disclosed an AI-related backlog of about $73 billion, of which roughly $53 billion is custom accelerators. On that basis Hock Tan told investors he has "line of sight to achieve AI revenue from chips in excess of $100 billion in 2027."

What Broadcom actually sells

It is worth being precise, because "custom AI chip" understates the arrangement. Broadcom does not sell a product to Google or Meta the way Nvidia does. It provides the design methodology, the intellectual property blocks — high-speed interconnect, memory interfaces, packaging — and the engineering to turn a customer's architecture into manufacturable silicon at a leading-edge foundry. The customer owns the architecture. Broadcom owns the ability to build it.

The second business, and the reason the first one is defensible, is networking. A single accelerator is useless; a hundred thousand of them wired together is a computer, and the wiring is Broadcom's oldest franchise. Its Ethernet switching silicon and optical components are what connect the racks, and every custom accelerator programme drags networking content along with it.

That pairing is what distinguishes Broadcom from a contract designer. A customer that hires it to build an accelerator ends up buying the switching silicon, the optical digital signal processors and the interconnect that make a cluster of those accelerators function as one machine — components where Broadcom has spent twenty-five years and where competition is thin. The accelerator wins the deal; the network earns the margin.

Analysis: the arms dealer position

Broadcom occupies a position that is close to structurally ideal. It does not need to win the AI race; it needs the race to continue. If Google's TPUs take share from Nvidia, Broadcom benefits. If OpenAI's inference chip works, Broadcom benefits. If they all fail and everyone goes back to buying Nvidia, Broadcom still sells the networking that connects the Nvidia racks together.

That is a materially better risk profile than any of its customers have, and the market has recognised it. The one thing that would genuinely hurt is a slowdown in the buildout itself — not a change in who wins it.

This is the same position Broadcom has occupied through three previous technology cycles. It sold the chips inside broadband modems, then inside smartphones, then inside data-centre networking, and in each case it was indifferent to which brand won the consumer-facing fight. The company has been assembled, deliberately and by acquisition, to sit one layer below wherever the argument is happening.

There is a real margin question underneath, however, and it is the standard objection to this business. Custom silicon carries lower gross margins than Broadcom's networking and software products, because a large share of the value flows to the foundry and the customer keeps the architecture. Revenue growing at 140 per cent in the lowest-margin part of the mix is excellent for the top line and dilutive to the percentage. Investors should expect Broadcom's blended margin to fall while its profits rise, and should not confuse the two.

Broadcom's own framing has been consistent about this and unusually candid for a chip company: management has repeatedly told investors to judge it on gross profit dollars rather than gross margin percentage while the accelerator mix rises. That is the right measure and it is also the one that flatters this particular transition, which is worth noting when reading the guidance.

Analysis: the concentration nobody mentions

The counterpart to Broadcom's advantage is that its customer list is extremely short. Six hyperscale customers, four of them publicly identified, account for the great majority of a $73 billion backlog. These are companies with the engineering depth to bring more of the work in-house over time, and every one of them is explicitly trying to reduce dependence on external suppliers — which is the entire reason they came to Broadcom in the first place.

There is a version of the next five years in which Google, having built five generations of TPU with Broadcom, decides it has learned enough. Nothing in the current numbers suggests it is happening. The logic of the relationship, though, points in that direction, and it is the same logic that just cost Qualcomm half its Apple revenue.

The counter-argument is that chip design has become harder rather than easier. Advanced packaging, high-bandwidth memory integration and leading-edge process nodes require capability that costs billions to maintain and is used a handful of times a year — precisely the sort of thing a company prefers to rent. Apple could bring modems in-house because it ships hundreds of millions of units. A hyperscaler building tens of thousands of accelerators has weaker arithmetic on its side.

The VMware business is the hedge against all of this, and it is an unglamorous, deeply unpopular, extremely effective one. Broadcom bought VMware and raised prices sharply, converting perpetual licences into subscriptions and losing some customers in the process. It produced $6.8 billion of high-margin, recurring, non-cyclical revenue in a quarter — the ballast under a semiconductor business whose growth rate is currently triple digits and will not always be.

Hock Tan's $100 billion claim is the number to hold him to. It is specific, it is dated, and it rests on a backlog the company has disclosed rather than a market forecast it has commissioned. Very few AI projections have that property. If it lands, Broadcom will have built more of the AI era's silicon than any company except the one everybody talks about.

The wider point is about how this buildout will be remembered. The public story of the AI boom is a story about models and the company that sells the graphics processors to train them. The industrial story is that four rivals who agree on almost nothing have each concluded they need their own silicon, and have each walked into the same building in San Jose to have it made.