Broadcom: Arming the CustomersNarrow moat

The most serious competitive threat to Nvidia is not a rival GPU. It is a design partner that lets Nvidia's largest customers stop being customers. Broadcom co-designs custom accelerators — XPUs — with the companies buying the most AI compute on earth: six named customers, a Google TPU relationship running seven generations since 2014, and a ten-gigawatt OpenAI programme with first deployments targeted for the second half of 2026¹.

The economics favour the challenger in one specific place. A custom part designed for one company's known workload can beat a general-purpose GPU on performance per watt and on cost, because it carries none of the flexibility that makes Nvidia's chips useful to everyone else. That advantage is largest in inference and at enormous scale — exactly the conditions the hyperscalers operate in — and the forecasts reflect it: custom ASIC shipments are projected to grow several times faster than merchant GPUs². The revenue is already arriving: Broadcom's AI semiconductor sales reached $16.7 billion in its latest quarter, up 221% in a year, against 117% growth in Nvidia's data-center business³.

Two things blunt it. Custom silicon takes years and costs hundreds of millions before the first chip works, so it is available only to the handful of buyers with the scale to amortise it — and those same buyers keep placing enormous Nvidia orders anyway, because a custom part is useless for workloads it was not designed for. Watch the ASIC share of AI accelerator shipments, and watch whether any customer's custom programme actually reduces its Nvidia purchases rather than growing alongside them. So far, remarkably, none has.