✦ The XPU Roster & the $60-90B PrizeNarrow moat

Broadcom (AVGO) — the future bets

Every custom chip a hyperscaler builds against Nvidia deepens its dependence on Broadcom — ~70% of a market guided to $60–90 billion.

Custom accelerators were Broadcom's decade-long apprenticeship with Google's TPU; they are now a franchise with a roster: Google formalized a long-term TPU agreement through 2031 in April 2026, Meta and Anthropic ramp their own programs, OpenAI signed its ten gigawatts, and Apple joined in 2026 — with Broadcom holding an estimated 70% of the custom-accelerator design market1. Management sizes the serviceable market for XPUs and their networking at $60-90 billion in fiscal 2027, with each major customer planning million-XPU clusters2.

The custom-silicon rosterGoogleTPUs — contracted through 2031MetaMTIA generationsAnthropiccustom-accelerator programOpenAI10GW, from H2 2026Applejoined 2026~70% of custom-accelerator design; $60–90B FY2027 SAM; million-XPU clusters planned.
Five names deep and one loyalty: every chip a hyperscaler designs against Nvidia is a toll paid to Broadcom.

The business model is the moat's best trick: the hyperscaler owns the chip design and absorbs the manufacturing capex; Broadcom supplies the hard parts — serdes, packaging, interconnect IP — and books the toll. Every custom chip a customer builds against Nvidia deepens their dependence on Broadcom. The risk is the roster's other side: five-ish customers this size can insource (Google has tried before), squeeze price, or simply pause a generation. Watch the customer count — each addition is billions — and watch Marvell and the customers' own silicon teams, the only two forces that can erode a 70% share.

Moat trajectory: Widening

Widening a customer at a time: Google formalized through 2031, Apple arrived in 2026, and the roster now spans five names deep at ~70% design share. Rosters this concentrated can shrink as fast as they grow — but this year every move was an addition.

The number that tests this moat
Third-party estimate
Named XPU customers
5 (Google, Meta, Anthropic, OpenAI, Apple)

Each name is worth billions and each addition widens a ~70% share of custom-accelerator design, against a market management sizes at $60–90B for fiscal 2027. Concentration cuts back: five customers this large can insource, squeeze, or pause a generation — the roster count each quarter is the cleanest read on the franchise.

Source: Custom-ASIC market analysis (third-party) ↗
References
  1. Third-party estimateRoster: Google (TPU deal through 2031), Meta, Anthropic, OpenAI, Apple; ~70% of custom-accelerator design.
    Tom's Hardware — the custom AI ASIC state of play (May 2026): Broadcom XPU customers incl. Google (long-term TPU agreement through 2031, Apr 2026), Meta, Anthropic, OpenAI and Apple; ~70% of the custom-accelerator design market; $60–90B FY2027 SAM guided; million-XPU clusters planned per customer — 2025–2026 · publ. May 2026 · source ↗
  2. Third-party estimateSAM guided at $60–90B for FY2027; million-XPU clusters planned per major customer.
    Tom's Hardware — the custom AI ASIC state of play (May 2026): Broadcom XPU customers incl. Google (long-term TPU agreement through 2031, Apr 2026), Meta, Anthropic, OpenAI and Apple; ~70% of the custom-accelerator design market; $60–90B FY2027 SAM guided; million-XPU clusters planned per customer — 2025–2026 · publ. May 2026 · source ↗
Sources
Generated September 22, 2026