NEW YORK, June 22, 2026, 12:02 EDT
- Broadcom fell in Monday trading even after JPMorgan pushed back on reports of trouble in its Google TPU roadmap.
- The bigger issue is no longer just custom chips. It is whether Broadcom can lock in full AI racks, networking gear and financing at gigawatt scale.
- The main risk is customer concentration: the same Google, Meta, Anthropic and OpenAI demand that gives Broadcom visibility also gives those buyers leverage.
Broadcom shares slipped near midday on Monday as investors weighed JPMorgan’s defense of the chipmaker’s Google AI program against a broader reset in high-valuation AI infrastructure stocks. The stock traded at $396.72, down $14.63, with its latest trade around 11:47 a.m. in New York, market data showed.
The move matters because the market is now testing whether Broadcom’s AI story is a quarter-to-quarter chip cycle or a longer contract cycle. JPMorgan analysts Harlan Sur and Mayur Ramdhani told clients to look past reports that Broadcom and Google had delayed or canceled a next-generation TPU v9 2-nanometer program, saying the team remained on track with “NO delays; NO cancellations,” Benzinga reported on Monday. Benzinga
The argument comes after a hard June pullback. Barron’s reported last week that Broadcom had fallen 11% in June after guidance that investors viewed as not strong enough for an AI stock priced for near-perfect execution, while JPMorgan kept an Overweight rating and a $580 target.
Broadcom’s latest numbers explain why the debate has sharpened. The company said fiscal second-quarter AI semiconductor revenue rose 143% from a year earlier to $10.8 billion and forecast $16.0 billion in AI semiconductor revenue for the third quarter, growth of more than 200%. Chief Executive Hock Tan said the gain came from custom AI accelerators and AI networking.
An ASIC, or application-specific integrated circuit, is a chip built for a narrow job rather than many jobs. That is the appeal for large AI users: Nvidia’s GPUs remain the flexible industry standard, but repeat workloads such as inference — when an AI model answers a user query — can favor custom chips built to cut cost and power use.
The undercovered catalyst is in the contract plumbing. Broadcom’s April SEC filing did not just say it would develop and supply Google’s future tensor processing units, or TPUs. It also disclosed a supply-assurance agreement for networking and other components in Google’s next-generation AI racks through up to 2031.
That shifts the story from chip design to rack economics. In a large AI data center, the processor is only part of the bill; the switching fabric, interconnect, power envelope and delivery schedule decide whether a buyer can turn capital spending into usable compute. Broadcom’s June 9 platform with Apollo and Blackstone adds another layer: a $35 billion initial tranche for more than 1 gigawatt of compute, with a structure aimed at more than 20 gigawatts through 2028.
Tan framed the platform as a way to meet AI demand with “speed and certainty.” Apollo President Jim Zelter called it a “bold, collaborative model,” language that matters because the financing is becoming as important as the chip spec. Broadcom Inc.
Anthropic is a live example of that mechanism. The AI company said in April it had signed with Google and Broadcom for multiple gigawatts of next-generation TPU capacity starting in 2027, and CFO Krishna Rao called it Anthropic’s “most significant compute commitment to date.” Anthropic said its run-rate revenue had topped $30 billion, up from about $9 billion at the end of 2025. Anthropic
Peers are not standing still. Nvidia still dominates general-purpose AI accelerators, while Marvell is chasing the same custom-chip budget. Reuters reported this month that Broadcom and Marvell help cloud companies such as Amazon and Google design custom chips that generate tens of billions of dollars in revenue and offer an alternative to Nvidia hardware.
There is also an emerging design-tool challenge. Architect Labs said last week it had raised $24 million to use AI to speed custom chip design, a process Reuters said can take roughly two years and cost hundreds of millions of dollars. That will not displace Broadcom near term, but it points to the pressure customers want to put on ASIC lead times and supplier margins.
The risk is straightforward. If Google or another hyperscaler spreads orders across more suppliers, Broadcom’s visibility could narrow even if AI demand stays high. Reuters reported after Broadcom’s June results that Marvell was making inroads with hyperscale customers and that Broadcom’s shares dropped after Tan left the $100 billion 2027 AI chip sales forecast unchanged; Ben Bajarin of Creative Strategies said, “they just didn’t raise it.” Reuters
For now, the market is treating Broadcom less like a traditional chip supplier and more like an AI capacity broker: part silicon designer, part networking vendor, part beneficiary of private-credit financing. That is a powerful position if the Google TPU roadmap lands on schedule. It is also a concentrated one if the buyers decide the best way to cut AI costs is to make Broadcom compete harder for every rack.
