Why everyone from OpenAI to SpaceX is building their own chips (and turning up the heat on Nvidia)

The TechCrunch piece captures a well-telegraphed but now accelerating trend: the largest AI spenders are actively de-risking their dependence on Nvidia by commissioning custom silicon. OpenAI's 'Jalapeño' inference chip, co-developed with Broadcom, is the latest high-profile example, joining Google's TPUs, Apple's Neural Engine lineage, and SpaceX's in-house compute ambitions. These are not fringe experiments — they represent the very customers who account for an outsized slice of Nvidia's GPU revenue.

Nvidia's numbers remain extraordinary — $215.9B in revenue, up 65.5% YoY, with 71.1% gross margins. That kind of financial profile reflects a monopoly premium baked in. The risk is not that Nvidia loses the market tomorrow, but that as custom ASICs absorb incremental inference workloads, Nvidia's growth rate decelerates faster than consensus models anticipate. The bear case is a valuation re-rating, not a revenue collapse.

Broadcom is the clearest near-term beneficiary. At $63.9B in revenue (+23.9% YoY) and 67.8% gross margins, AVGO already generates significant AI custom chip revenue through its XPU/ASIC co-development partnerships. OpenAI's Jalapeño deal deepens that pipeline, and each new hyperscaler customer adds recurring, sticky revenue tied to multi-year chip roadmaps.

The bull/bear tension for NVDA centers on timing: Nvidia's CUDA ecosystem, software moat, and Blackwell/Rubin roadmap give it a multi-year lead in training workloads, where custom chips remain impractical for most players. But inference — the higher-volume, cost-sensitive workload — is exactly where custom ASICs are most competitive, and inference is where the next wave of AI spending is headed.

Watch for: any acceleration in hyperscaler capex commentary that explicitly mentions reduced GPU allocation per dollar spent; Broadcom's next earnings for AI revenue segment growth; and whether OpenAI's Jalapeño timeline slips (chip tape-outs frequently do), which would extend Nvidia's runway.