The margin moved to custom silicon
AI revenue up 143%, guided to triple again. When a hyperscaler commissions its own accelerator, it is freezing a workload into silicon — and betting it understands it.
Broadcom reported $10.80bn in AI semiconductor revenue, up 143%, and guided next quarter to $16.0bn on custom accelerators and switching silicon sold to hyperscalers.
What a custom ASIC concedes
Commissioning your own accelerator is a multi-year bet that a workload's shape has stopped changing. That is a strong claim, and it is only economic at volumes where a few percentage points of efficiency exceed the design cost — which is itself a statement about how large these workloads have become.
Custom silicon takes the work you have already characterised. The merchant GPU keeps the work that is still moving.
Which is why this is not displacement. Nvidia's data-centre revenue exceeded $50bn in the quarter — several times Broadcom's entire AI line. Fast growth from a small base in frozen workloads is a different business from supplying the frontier.
The half nobody covers
Switching silicon matters as much as the accelerators. Past a certain cluster size the interconnect determines utilisation, and an accelerator waiting on the network is an accelerator not amortising. Selling both the compute and the fabric around it is an unusually strong position.
The same logic, further up
Memory capex at $38bn and construction contracts at ₹15,000 crore are the same move in different currencies: capital going to the layers that cannot be scaled on demand.
What this asks of you
If you are forecasting hyperscaler capex, watch whether the 200% guide lands. A guide of that size reflects design wins already booked, so a miss says more about intentions than any survey will.
And if you are choosing where to build, note what the custom-silicon trend implies: the platforms are optimising hard for the workloads they already have. That is excellent for anyone doing the same thing they are, and worth thinking about if you are not.
Science-Technology News — Broadcom: The Architect of Custom AI ASICs → · Intellectia — AI Chip Stocks August 2026: Nvidia vs AMD →