The custom AI chip wave: who is eating into Nvidia, and how big is the Broadcom opportunity?
Google TPU, Meta MTIA, Microsoft Maia, OpenAI Titan — hyperscaler custom silicon is everywhere. In 2026 ASIC shipment growth (~+45%) runs near triple that of GPUs, and Broadcom holds roughly 60% of design partnerships. Public data plus legendary-investor positioning show what remains of Nvidia’s moat and who profits on the custom chain.
What happened: customers became competitors
Nvidia’s biggest customers are becoming its most organized competitors. Google’s TPU is generations deep; Meta has MTIA and Microsoft has Maia; even OpenAI — built almost entirely on Nvidia — is developing custom silicon with Broadcom (the Titan program, targeting 2027 deployment at over a gigawatt of compute). The logic is blunt: when inference cost is the biggest variable in an AI business, owning the silicon means owning the economics.
The data: ASIC growth near triple that of GPUs
Industry trackers put custom-ASIC AI-server shipment growth at ~44.6% YoY for 2026 versus ~16.1% for merchant GPUs, lifting ASICs to ~27.8% of AI-server shipments. Keep perspective: Nvidia still holds 90%+ of training and ~70% of the overall AI chip market — custom silicon is eating the fastest-growing slice, inference, rather than storming the training throne.
The biggest winner isn’t a hyperscaler — it’s the arms dealer, Broadcom
Hyperscaler custom silicon doesn’t mean in-house fabrication: design partnerships, IP, networking and packaging are largely outsourced. Broadcom holds roughly 60% of AI-server ASIC design partnerships, carries a ~$73B AI backlog, and guides fiscal-2026 AI semiconductor revenue toward ~$56B (roughly tripling in a year). Whether Google’s, Meta’s or OpenAI’s chips “win” or not, Broadcom collects first — the classic picks-and-shovels seat. Marvell is the second source on the same logic.
The legends have voted: AVGO enters the consensus board
This isn’t theoretical — check the Q1 2026 13Fs: Druckenmiller opened a Broadcom position in the same quarter he exited Alphabet, and Broadcom sits in Coatue’s top five. On our consensus board AVGO now appears as a standalone name, while Nvidia shows the classic split — Duan Yongping adding 91% versus Tepper and Coatue trimming. That is exactly what the Consensus Score quantifies in real time, before the narrative catches up.
How to think about it: three checkpoints
First, don’t read the custom wave as Nvidia’s obituary: the CUDA ecosystem, training dominance and rack-scale systems remain deep moats — valuation is the real debate. Second, the custom chain’s certainty ranking is roughly design partners (AVGO/MRVL) > foundry & equipment (TSM/LRCX/AMAT) > any single chip’s success. Third, respect the cycle: ASIC orders track hyperscaler capex, and if AI capex contracts the growth story reverses just as fast. Educational content, not investment advice; all figures are sourced and may change quarterly.
Sources: Tom’s Hardware — The custom AI ASIC state of play (May 2026) · TechTimes — Broadcom forecasts $56B as custom silicon demand surges · Introl — Custom silicon inflection 2026: hyperscaler ASICs vs Nvidia · InvestorPlace — The rise of custom AI chips is breaking Nvidia’s grip · HeyGoTrade — Q1 2026 13F: Druckenmiller opens AVGO