JPMorgan Bullish on Broadcom: TPU v9, OpenAI XPU, and Apple Collaboration Visibility

JPMorgan hosted an investor meeting with Broadcom (AVGO-US) management on Tuesday, July 7, with CEO Hock Tan in attendance. On July 8, JPMorgan released a research report maintaining Broadcom's 'Overweight' rating and positioning the company as the world's second-largest AI semiconductor supplier, the largest custom chip (ASIC) supplier, and the largest networking semiconductor supplier.

Regarding recent market rumors of delays or cancellations in Google's next-generation TPU project, Broadcom management confirmed that the TPU v9 roadmap is 'fully on track.' Key details include:

- 400G SerDes silicon has been completed and is operational - TPU v9 uses 2nm process technology, integrating 4 compute dies, 16 HBM stacks, and 400Gbps SerDes interfaces - Mass production remains on schedule for 2028, with no delays or cancellations - Broadcom signed a five-year agreement with Google in March 2026, covering TPU generations from v8 to v11, securing future demand across four generations - Broadcom retains the majority share of TPU revenue—a key contractual term that directly counters market concerns about margin erosion - Broadcom maintains a technological lead of over 18 months compared to Google's in-house chip design team

Broadcom's collaboration with Apple, spanning over a decade and covering touch controllers and wireless charging ASICs, has been extended to 2031. New design projects—potentially including AI and data center ASICs—have been added, along with a radio frequency supply agreement. Management views the extension as a dual win: enhancing revenue visibility and serving as a technical endorsement, given Apple is 'one of the most demanding chip customers.' The expansion of Broadcom's wafer fab in Fort Collins, Colorado, is cited as evidence of long-term demand confidence.

Broadcom's first-generation XPU with OpenAI, codenamed 'Jalapeno,' went from design to tape-out in just nine months—validating Broadcom's execution in complex multi-die design, advanced packaging, and high-speed SerDes. The next-generation OpenAI chip is expected to tape-out 'soon.' Management emphasized that cutting-edge XPUs are pushing the limits of bandwidth, SerDes, power efficiency, and packaging—complexities often underestimated by the market.

Broadcom's Tomahawk 6 switch chip is sold out, with core customers being leading large model builders. When allocating Tomahawk and Jericho switch products, Broadcom prioritizes XPU customers. Management asserts that Broadcom's AI networking moat is deeper than perceived, with pricing power stemming from both technological leadership and deep collaboration with custom chip clients.

On AI compute supply and demand, Broadcom management explicitly stated they do not foresee an oversupply of AI compute capacity in 2027 or 2028. JPMorgan's broader strategy report reinforces this, citing persistent capacity bottlenecks through 2028, with meaningful new supply unlikely before then—supporting a narrative of sustained hardware shortage.

Broadcom's three structural advantages:

(1) Inference demand is reshaping compute architecture—XPUs are moving from 'supporting role' to 'co-equal with GPUs.' Management notes rapid growth in inference workloads is accelerating XPU adoption. Among leading AI model builders, XPU-to-GPU shipment ratios could approach 50:50 next year. As workloads evolve—from single inference to inference chains to AI agents—optimized custom chips gain strategic value. The per-token cost competition is becoming the new battleground in AI compute.

(2) The moat in multi-die architecture—why hyperscalers can't easily build in-house. While concerns persist that Google, Amazon, and Microsoft might design their own chips, Broadcom argues modern XPUs are no longer simple monolithic designs. Multi-die architectures, advanced packaging, and high-speed SerDes create high barriers to mass production. Six- and eight-die configurations are now standard—precisely Broadcom's core strength. Designing a chip is one thing; manufacturing millions at scale, on tight timelines, with acceptable yields, is another. Broadcom's moat lies not in 'design capability' but in 'engineering execution for complex, high-volume multi-die systems.'

(3) The 'compute + networking' flywheel effect. Custom XPU customers (Google, Apple, OpenAI) rely on Broadcom for ASIC design and high-volume manufacturing. These same customers require high-performance networking switches (Tomahawk) to interconnect tens of thousands of XPUs. Broadcom prioritizes Tomahawk allocation to XPU customers, reinforcing stickiness. The more a customer depends on Broadcom's XPU, the more likely they are to use Broadcom's networking chips—and vice versa. This cross-selling and lock-in effect creates high switching costs. Tomahawk 6's sell-out and pricing power are direct evidence of this flywheel accelerating.

This report aligns with JPMorgan's broader strategy: bullish on semiconductors, focused on custom chips, viewing recent pullbacks as buying opportunities. The core thesis—that new capacity bottlenecks will persist through 2028—strengthens the long-term AI hardware shortage narrative. JPMorgan remains cautious on 'AI cannibalization' sectors—software, business services, and media—where AI could compress pricing or margins.