TL;DR — Broadcom’s AI chip backlog reached $73 billion, built almost entirely on custom silicon for Google, Meta, Anthropic and OpenAI, the same companies trying to rely less on Nvidia. Diversifying away from Nvidia doesn’t reduce chip supplier concentration, it just adds Broadcom to the list.
Broadcom’s AI chip backlog stands at $73 billion, covering roughly the next 18 months of committed orders. Virtually all of it is coming from the same handful of companies that also happen to be Nvidia’s biggest customers. Google, Meta, Anthropic, OpenAI, and two hyperscalers Broadcom won’t name are paying Broadcom to design them a chip that isn’t a GPU — because the whole point is not needing as many of Nvidia’s.
A Beat That Didn’t Feel Like One
Broadcom’s fiscal second-quarter revenue hit $22.19 billion, up 48% year over year, with AI semiconductor revenue reaching $10.8 billion, up 143%. The stock fell anyway. The problem wasn’t the quarter — it was the guide. Third-quarter revenue guidance of $29.4 billion, up 84% year over year, and AI semiconductor revenue guidance of $16.0 billion, up more than 200%, both beat consensus. But Broadcom reiterated its full-year AI revenue target of $56 billion rather than raising it, even though first-half AI revenue already ran close to $19 billion and the third-quarter guide implies serious acceleration from there. Investors had wanted a raise. Without one, the market read $56 billion as a ceiling management is managing toward, not a floor it expects to blow past.
The backlog is the number that matters more than either the quarter or the guide. CEO Hock Tan has told investors Broadcom is carrying $73 billion in AI-related backlog over the next 18 months, with more than 80% of it expected to convert within the next four quarters. That backlog is what’s behind Tan’s public target of more than $100 billion in AI chip revenue by 2027 — a number that would roughly double the company’s current AI run rate.
Not a GPU Business
Broadcom isn’t selling merchant silicon the way Nvidia does. Its AI business runs on two pillars: custom AI accelerators — XPUs and ASICs that hyperscalers co-design with Broadcom to their own specifications — and the high-speed networking silicon that ties thousands of those chips together into a single training cluster. A custom chip takes years to design and is built for one company’s specific workload. It’s a slower, stickier business than shipping a general-purpose GPU, and once a hyperscaler commits to a design generation, switching away from Broadcom mid-cycle is expensive.
Six Names, One Motive
Broadcom’s named custom-silicon customer list grew from five to six this year, and the new addition tells you what’s driving the whole business: OpenAI, Nvidia’s single largest customer, is now also building its first custom XPU with Broadcom, targeting roughly 1 gigawatt of capacity in 2027. Google remains the longest-standing partner, with seven generations of co-designed TPU silicon behind it, including the current Ironwood generation. Meta uses Broadcom-built MTIA accelerators. Anthropic is scaling from 1 gigawatt of Google TPU capacity in 2026 to 3 gigawatts in 2027 — compute it rents through Google Cloud, built on the same TPU silicon Broadcom co-designs. Two more hyperscalers remain unnamed in Broadcom’s disclosures.
None of these companies are abandoning Nvidia. They’re building a second option, because paying one supplier for the majority of the world’s AI compute is a concentration risk every one of them is trying to manage — the same dynamic we’ve covered from Nvidia’s side of the supply chain. Custom silicon doesn’t replace GPU purchases so much as it caps how much leverage Nvidia has over price and allocation in the next negotiation.
The Margin Tradeoff
Broadcom is accepting a real cost for that backlog. The mix shift toward full AI systems and custom silicon carries a lower gross margin percentage than Broadcom’s legacy semiconductor and software businesses. But the dollar volume moving through that lower-margin business is now large enough to drive outsized growth in absolute operating profit, which is why Tan keeps leaning into it rather than defending the margin line.
The risk this creates isn’t smaller than Nvidia’s concentration problem — it’s a different shape. Instead of the industry depending on one GPU supplier, it now depends on Broadcom’s design queue, TSMC’s fabrication capacity, and a six-customer backlog where losing even one name would blow a hole in the 2027 target. Diversifying away from Nvidia doesn’t lower the industry’s dependency on a small number of chip suppliers. It just adds Broadcom to the list.
Sources: Broadcom Inc. fiscal second-quarter 2026 earnings release and call (reported June 3, 2026); company AI backlog and revenue-target commentary from CEO Hock Tan; industry reporting on named custom-silicon customers. Third Pole Markets holds no position in Broadcom, Nvidia, Alphabet, Meta, OpenAI or Anthropic as of publication. This is not investment advice — see our About page for our full disclosure policy.
Broadcom’s own disclosures are in its investor relations filings and 10-K filings on SEC EDGAR.
