Broadcom: chips ship, software sticks

Key takeaways

• Beat and raise. Q3 revenue USD 15.95bn (+22%). AI USD 5.2bn (+63%). Q4 guide USD 17.4bn with AI at USD 6.2bn
• Strategy is simple. Custom AI chips plus Ethernet now. VMware subscriptions for staying power
• OpenAI shifts the curve. A multi-year custom-chip deal points to 2026 production and deeper visibility

---

Custom chips, compounding visibility

Broadcom reported Q3 on 4 September with record sales and another AI step-up. The headline was a new multi-billion-dollar AI order. Several outlets report the buyer is OpenAI and that the two will co-develop a custom accelerator for internal use, targeting mass production in 2026, with deal size near USD 10 billion. If that’s right, two things follow. First, a marquee customer diversifies beyond a single GPU supplier. Second, custom programs usually run across multiple chip generations, creating visibility that merchant silicon rarely offers. Broadcom has not named the customer, so keep “reported” in your notes—yet the direction looks credible given Broadcom’s Google TPU track record and the stronger AI guide.

What happened, what’s next

Print vs consensus. Beat-and-raise. Broadcom posted Q3 revenue of USD 15.95bn vs USD 15.83bn expected and earning per share USD 1.69 vs USD 1.66. AI revenue rose 63% to USD 5.2bn.

Outlook. Q4 revenue guided to USD 17.4bn (above expectations) with AI at USD 6.2bn. Read-through: accelerators keep scaling, Ethernet tight into 800G, VMware conversions steady.

Tone. Execution first. Management spoke to firm delivery windows and demand tied to booked orders. The highlight is a new multi-billion AI customer order. Press reports point to OpenAI co-developing a custom chip, with mass production in 2026.

Price action. On 4 September 2025, Broadcom closed at USD 306.10, up USD 3.65 (+1.2%). After-hours trading moved higher as the guide landed above consensus.

Valuation read. Still premium. A clean beat plus an above-street guide supports it, but the bar rises each quarter. Delivery dates and cash conversion must keep pace with the AI narrative, or the multiple does the talking first.

Positioning: engines, roads, and tolls

Broadcom supplies the engines (task-specific accelerators), lays the roads (Ethernet fabric), and runs the tolls (VMware software). Its edge is depth:

• Co-designed engines lock in customers and raise switching costs.
• Standard roads: Ethernet uses tooling and skills every data center already has.
• Toll leverage: VMware’s installed base makes up- and cross-sell easier.
  

  
  Risks:

• Customer concentration.
• Nvidia's InfiniBand strength at the elite training tier.
• Slower Cloud Foundation conversions.
• Power/cooling/optics gating 800G rollouts.
  
  
  Broadcom’s strategy: roads and engines, one map

Bigger AI models need two things: faster roads and more efficient engines. Think of a transport system with three parts working together.

• Custom accelerators are the engines. Built with top customers for a known route (e.g., Google’s TPU). They roll off the line on dated schedules, sized to demand.

• Ethernet is the road network. 400G today, 800G ramping, ecosystem ready for 1.6T. Switches, optics, power, and cooling must scale together for 800G; one weak link bottlenecks the cluster.

• VMware is traffic control and tolls. Legacy licences convert to Cloud Foundation subscriptions; then networking and security attach on top.

For buyers, fewer vendors mean a single roadmap and clearer service level agreements. For investors, contracted orders and delivery windows steady guidance. Chips drive volume now; software keeps the cash flowing.

The risks: customers can postpone projects, pour more budget back into GPUs, or delay road upgrades. Power and cooling can slow the move to faster lanes like 800G. On software, pushing clients to subscriptions can create friction. If both the roads and engines stall at once, margins take the hit.

AI: threat and opportunity

Opportunity. Every bigger model needs fast, standard networks and efficient accelerators. As inference spreads beyond a few training islands, Ethernet’s price/performance helps it win. Custom silicon broadens as more workloads stabilise. The OpenAI partnership adds a second proof point beyond Google—multi-year design, 2026 mass production targeted, and likely repeat generations if v1 lands. That is real visibility.

Threat. Digestion happens. Hyperscalers can push deliveries or tilt spend back to GPUs or in-house chips. Power, optics, and cooling can slow 800G rollouts. On software, Cloud Foundation conversions can face timing friction. If roads and engines stall at once, mix pressures gross margin and cash.

Investor watch

• Guide and cadence: Q4 bridge, AI run-rate, date-stamped delivery windows.

• AI mix: accelerators vs Ethernet; 800G momentum and any 1.6T signals.

• VMware durability: Cloud Foundation conversions, renewal tone, RPO, attach into networking/security.

• Margins and cash: semi gross margins on accelerator scale, software margins stability, free cash flow conversion vs inventory builds.
  
  

What decides the next leg

This print says the organs and arteries kept pace with the industry’s brain. Two drivers matter now: guidance credibility and AI breadth. Two risks stand out: digestion and conversion friction. If dates hold, 800G ramps cleanly, and VMware cash keeps compounding, duration extends. If schedules slip, the multiple moves first. Build the engines, pave the roads, run the tolls and the story writes itself.