Broadcom reported $16.7 billion of AI semiconductor revenue in its fiscal third quarter of 2026, up 221% year on year, and still expects fiscal 2027 AI revenue "in excess of $100 billion." Its Q4 guidance of $21.7 billion would put fiscal 2026 AI revenue at roughly $58 billion, by CNBC's calculation. Most of that comes from custom AI chips designed with the biggest AI companies, which Broadcom says now include Google, Meta, Anthropic and OpenAI among six core customers.
What a "custom AI chip" is
Instead of buying general-purpose GPUs, a company designs an accelerator (often called an ASIC or XPU) tuned for its own models and workloads. Broadcom provides the design expertise, intellectual property and manufacturing relationships to turn that into real silicon. Google's TPUs are the best-known example.
Why the big labs want their own
- Cost: a chip built for your workload can deliver more AI per dollar than a general-purpose GPU.
- Supply: less dependence on a single GPU vendor during shortages.
- Control: hardware and software designed together, tuned for specific models.
Broadcom has cited visibility into deployments of more than 10 gigawatts for Anthropic, more than 5 gigawatts for OpenAI and 3 gigawatts for Meta. Numbers like that show how large the AI build-out has become.
What it means for developers and businesses
- More competition in AI hardware pushes inference prices down.
- Models get optimised for different chips, so the same model may perform or cost differently across providers.
- Abstraction wins. Build AI features against a stable interface, and let providers compete underneath.
Some secondary coverage quotes different figures, so I've used Broadcom's September results and statements as reported in mainstream coverage.
Building on a competitive AI market
At NextGen Fusion, our AI development company in Lucknow, we design AI features to switch providers easily, so clients benefit as competition drives prices down.
