✦ The Future BetsNarrow moat

Broadcom (AVGO) — the future bets

Broadcom's futures arrive as purchase orders — ten OpenAI gigawatts, five XPU customers, Ethernet at every scale — tolls on other people's AI ambitions.

Broadcom's future bets are unusually easy to price, because most of them arrive as purchase orders. The custom-accelerator business the company spent a decade building for Google quietly became the industry's default alternative to Nvidia — and in late 2025 it landed the order that reframed the company: OpenAI, co-developing and deploying ten gigawatts of its own accelerators on Broadcom silicon and Ethernet, first racks in the second half of 2026, complete by end of 20291.

Committed for the bets, at 2 Aug 2026 ($B)RPO, total$179.2BPurchase commitments$126.8BRPO due within 12 months~$44.8BBroadcom Form 10-Q, quarter ended 2 August 2026
Customers have committed $179 billion to Broadcom, and Broadcom has committed $127 billion to its own suppliers.

The roster behind that headline is the real story: Google (now contracted through 2031), Meta, Anthropic, OpenAI and — newest — Apple, with Broadcom holding roughly 70% of the custom-accelerator design market and guiding to a $60-90 billion serviceable market in fiscal 2027 alone2. Beside it runs the networking wager: Tomahawk 6 shipping at 102.4 terabits, Tomahawk Ultra chasing NVLink's scale-up role, and Jericho4 wiring million-XPU clusters across data centers3 — Ethernet as the open answer to Nvidia's private fabrics at every scale. The quietest bet is software: VMware Cloud Foundation 9.1, pitched as the private-cloud home for production AI, with Broadcom's own survey showing private-cloud inference (56%) now ahead of public (41%)4.

Notice the shape: every bet sells a toll on someone else's AI ambitions — the customer designs the chip, owns the capex, takes the model risk, and Broadcom clips the fee on silicon, switches, and virtualization. AI revenue is already $16.7 billion a quarter, growing 221%5, so these are less options than extensions — the question the price already assumes answered is duration. Watch three things: OpenAI's racks actually landing on schedule, the XPU roster holding as hyperscalers' appetites shift, and whether the VMware repricing goodwill cost shows up as private-AI defections.

Moat trajectory: Widening

Widening on orders, not narratives: OpenAI's 10GW signed, Google extended through 2031, Apple joined the XPU roster, Tomahawk 6 reached production volume, and AI revenue compounds at 143%. The bets are extensions of franchises already billing — the rare futures portfolio with a backlog.

The number that tests this moat
Reported
The headline order's size
10GW (OpenAI, H2 2026–2029)

The portfolio's center of gravity is a single purchase order: ten gigawatts of OpenAI-designed accelerators on Broadcom silicon and Ethernet. Around it, a five-name XPU roster at ~70% design share and a networking lineup shipping in volume. Watch delivery against the calendar — announced gigawatts and energized gigawatts differ by tens of billions of dollars.

Source: OpenAI–Broadcom collaboration announcement ↗
✦ Future bets — beyond today's moat
✦ FUTURE BET 01
The OpenAI Deal & the 10-Gigawatt Order
Ten gigawatts promised by 2029 in the currency of 2025's enthusiasm — the order that reframed Broadcom, priced before a rack has landed.
✦ FUTURE BET 02
The XPU Roster & the $60-90B Prize
Every custom chip a hyperscaler builds against Nvidia deepens its dependence on Broadcom — ~70% of a market guided to $60–90 billion.
✦ FUTURE BET 03
Ethernet Everywhere: Tomahawk, Thor & Jericho
Broadcom bets AI clusters wire like the internet did — open Ethernet from 250 nanoseconds to a million XPUs, against Nvidia's private fabrics.
✦ FUTURE BET 04
VMware Private AI & the Software Second Act
The toll road adds an AI lane — VCF pitches the private cloud as AI's production home, to customers still sore from the repricing.
References
  1. ReportedThe custom-accelerator business the company spent a decade building for Google quietly became the industry's default alternative to Nvidia — and in late 2025 it landed the order that reframed the company: OpenAI, co-developing and deploying ten gigawatts of its own accelerators on Broadcom silicon and Ethernet, first racks in the second half of 2026, complete by end of 2029.
    OpenAI–Broadcom strategic collaboration (Oct 13, 2025): co-develop and deploy 10GW of OpenAI-designed AI accelerators; racks scaled entirely with Broadcom Ethernet; first deployments H2 2026, completion by end-2029 — Announced Oct 2025; deployments H2 2026–2029 · publ. Oct 13, 2025 · source ↗
  2. Third-party estimateThe roster behind that headline is the real story: Google (now contracted through 2031), Meta, Anthropic, OpenAI and — newest — Apple, with Broadcom holding roughly 70% of the custom-accelerator design market and guiding to a $60-90 billion serviceable market in fiscal 2027 alone.
    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 ↗
  3. ReportedBeside it runs the networking wager: Tomahawk 6 shipping at 102.4 terabits, Tomahawk Ultra chasing NVLink's scale-up role, and Jericho4 wiring million-XPU clusters across data centers — Ethernet as the open answer to Nvidia's private fabrics at every scale.
    Broadcom — AI-infrastructure networking lineup (OFC 2026): Tomahawk 6 (102.4 Tbps, production volume; 128K-XPU fabrics in two tiers), Tomahawk Ultra (~250ns, scale-up alternative to NVLink; systems through H2 2026), Thor Ultra 800G NICs — 2025–2026 · publ. 2026 · source ↗
  4. ReportedThe quietest bet is software: VMware Cloud Foundation 9.1, pitched as the private-cloud home for production AI, with Broadcom's own survey showing private-cloud inference (56%) now ahead of public (41%).
    Broadcom — VMware Cloud Foundation 9.1 (May 5, 2026): private-AI services standard (model store/runtime, agent builder, vector DB); Private Cloud Outlook: 56% run/plan production inference in private cloud vs 41% public (−15% YoY); 9 of the top-10 Fortune 500 committed to VCF — May 2026 · publ. May 5, 2026 · source ↗
  5. ReportedAI revenue is already $16.7 billion a quarter, growing 221%, so these are less options than extensions — the question the price already assumes answered is duration.
    Broadcom Inc., Q3 FY2026 results (Form 8-K exhibit 99.1) - revenue $29,591M, up 86%; AI semiconductor revenue $16.7 billion, up 221% year over year and 54% quarter over quarter; Q4 FY2026 guidance of approximately $34.8 billion of revenue and $21.7 billion of AI semiconductor revenue (+236%); free cash flow $13,665M, 46% of revenue; semiconductor solutions revenue $20,839M and infrastructure software $8,752M — Quarter ended 2 August 2026 · publ. 2026-09-02 · source ↗
Sources
Generated September 22, 2026