JPMorgan Research Note Recap: Broadcom's AI Custom Chip Orders Are Backlogged; TPU v9 On Track

By Rita

Chaoxiang Editor’s Note

On July 7, JPMorgan hosted an investor meeting with Broadcom management, led by CEO Hock Tan. The meeting delivered a dense set of key updates: Google’s TPU v9 roadmap remains on schedule; Apple’s ASIC partnership has been extended through 2031 and now includes new AI accelerator projects; OpenAI’s next-generation chip is nearing tape-out; and demand for the Tomahawk 6 switch chip continues to outstrip supply. JPMorgan maintains its Overweight rating on Broadcom, identifying it as the world’s second-largest AI semiconductor supplier, the largest custom ASIC provider, and the top network semiconductor vendor.

Broadcom management offered an assessment: the rapid growth of inference workloads is accelerating the adoption of custom XPUs. Among leading large model developers, the shipment ratio of XPUs to GPUs could approach 50:50 next year. This indicates that custom chips are transitioning from a 'supporting role' to one 'on par with GPUs.'

Customers are requesting accelerated tape-out schedules for next-generation ASICs and pursuing more chip design variants. The evolution of inference workloads—from single-step inference to inference chains and now to AI agents—has significantly increased the strategic value of workload-optimized custom chips. Cost-per-token competition is becoming the new battleground in AI compute.

Broadcom’s first-generation XPU for OpenAI, 'Jalapeno,' went from design to tape-out in just nine months, validating Broadcom’s execution capabilities in complex multi-die designs, advanced packaging, and high-speed SerDes. The next-generation OpenAI chip will tape out 'very soon.' Management specifically emphasized that these cutting-edge XPUs are 'pushing the limits' across multiple dimensions—including bandwidth, SerDes, power consumption, and packaging—with their complexity often underestimated by the market.

Regarding the competitive landscape for custom chips, management gave a clear view on the threat posed by hyperscalers’ in-house (COT) chip development: modern XPUs are no longer simple single-die designs; they increasingly adopt multi-die configurations, advanced packaging, and high-speed SerDes, presenting significant technical hurdles for internal teams in achieving volume production. Six-die and eight-die configurations have become industry-standard setups—precisely where Broadcom’s core competitive moat lies.

There had been prior market speculation that Google’s next-generation TPU project might be delayed or canceled. Broadcom management explicitly responded that the roadmap is 'fully on track,' with the 400G SerDes silicon already developed and operational. Broadcom also confirmed that under its five-year agreement with Google, TPU-related revenue is growing year-over-year and Broadcom maintains a majority share—this was a key precondition when signing the deal and directly addresses prior market concerns about 'share erosion.'

On the Apple front, Broadcom has collaborated with Apple for over a decade across multiple ASIC categories, including touch controllers and wireless charging. The latest agreement extends this design partnership through 2031, covering new design sockets (which JPMorgan believes include AI/data center ASICs), along with a radio frequency supply agreement. Broadcom’s wafer fab expansion in Fort Collins, Colorado, further underscores the company’s confidence in these long-term opportunities. The significance of this agreement goes beyond revenue visibility—it also serves as a strong endorsement of Broadcom’s technical prowess, given that Apple, one of the world’s most demanding chip customers, continues to select Broadcom as its ASIC partner.

Broadcom management does not believe AI compute capacity will face oversupply in 2027 or 2028. Tomahawk 6 switch chips are already sold out, primarily to leading large model developers. When allocating Tomahawk and Jericho switching products, Broadcom prioritizes the needs of its XPU customers.

Broadcom’s competitive moat in AI networking is deeper than the market realizes. Tomahawk 6’s pricing power stems from two sources: first, the technological leadership of the switch chip itself; second, deep integration with custom chip customers. This 'compute + networking' strategy is creating a virtuous flywheel: success in custom compute drives networking share, while networking scarcity further strengthens partnerships with XPU customers. Once customers adopt Broadcom’s custom chips, they naturally favor Broadcom’s networking solutions to build a complete AI compute stack—resulting in extremely high switching costs.

On the cost side, Broadcom stated it will maintain stable ASIC gross margins through technological leadership and will not allow rising costs for wafers, substrates, HBM, and other components to erode profitability. Management emphasized they will keep ASIC gross margins within the target range 'under any circumstances.' Technological leadership is the fundamental source of pricing power. Management also explicitly refuted claims that customers could procure HBM independently, stressing that HBM is an inseparable part of XPU design and must be tightly integrated with the core logic die to ensure performance—customers cannot decouple HBM from Broadcom’s design and source it separately.

Broadcom’s recent management meeting delivered a highly focused message: demand for custom AI chips is expanding from 'Google alone' to 'multiple top-tier customers.' Google, Apple, and OpenAI are all accelerating their initiatives simultaneously, each representing multi-year, long-term collaborations. Apple’s involvement is particularly noteworthy—previously focused primarily on in-house AI chip development, Apple is now procuring AI accelerator ASICs from Broadcom, suggesting a strategic shift in its AI compute approach from 'fully in-house' to 'partially outsourced.' This marks a significant structural expansion signal for the total addressable market of the custom chip industry.

Broadcom’s current valuation narrative is shifting from that of a 'semiconductor cyclical stock' to a 'core supplier of AI infrastructure.' Signals such as the sell-out of Tomahawk 6, a queue of ASIC customers, and long-term agreements extended through 2031 collectively point to a level of demand visibility far exceeding that of any prior cycle.

Key validation points will come in the second half of the year. The production ramp-up pace of TPU v9, tape-out progress of OpenAI’s next-generation chip, and the specific timeline for Apple’s AI accelerator project will be critical variables shaping Broadcom’s growth trajectory beyond 2027.

Disclaimer

This article is a synthesis and interpretation by Chaoxiang Research of a third-party brokerage research report (JPMorgan, July 7, 2026). The ratings, price targets, earnings forecasts, and related judgments cited herein reflect the views of the brokerage’s analysts and represent solely the position of their affiliated institution. They do not reflect the views of Chaoxiang Research and do not constitute investment advice.

The market involves risks, and investment decisions should be made independently. This article should not be used as a basis for buying or selling any securities.

62K Views