Nvidia acquired Groq, the startup behind LPU inference chips, while simultaneously shipping its own Vera CPUs.1 The moves mark a pivot: Nvidia is no longer just selling GPUs — it is building a vertically integrated compute stack.
Broadcom is working on custom AI processors with both Alphabet and OpenAI.1 Arm is also part of the OpenAI collaboration. Three of the most powerful AI labs and cloud operators are now co-designing the silicon that runs their models.
Alphabet's TPU program, co-developed with Broadcom, has been running for years. The current wave differs in scale and simultaneity: multiple hyperscalers are locking in custom silicon relationships at the same time.
Tiger Global has increased stakes in Nvidia, Broadcom, and TSMC concurrently.1 The investment pattern signals confidence that custom silicon — not commodity GPU purchasing — is where AI infrastructure value accrues.
The strategic logic is straightforward. Custom chips tuned to specific model architectures deliver better performance-per-watt than general-purpose GPUs. For hyperscalers running billions of inference queries daily, that efficiency gap translates directly to cost.
Vertical integration also reduces supply chain exposure. Companies that design their own chips and manufacture through dedicated TSMC capacity are less vulnerable to GPU allocation constraints — a recurring problem during AI's rapid scaling phase.
The consolidation creates a two-tier compute market. At the top: hyperscalers with proprietary silicon and manufacturing relationships. Below: companies still dependent on commodity GPU purchasing from Nvidia or AMD.
General-purpose GPU revenue faces commoditization pressure as custom silicon deployments scale. The 18-24 month window before this pressure becomes acute gives Nvidia limited time to complete its own vertical integration before its core GPU business faces structural headwinds.1
The race to own AI's silicon stack is no longer a future scenario. It is the current state of competition.