US semiconductors and the Google TPU supply chain Report Interpretation

Overview

This US semiconductor-sector note assesses how Marvell’s new Google custom-silicon partnership affects Marvell, Broadcom and the distribution of value within Google’s TPU ecosystem. Bank of America argues that the partnership broadens Marvell’s upside but is largely neutral to Broadcom because Broadcom should retain the core compute die and other high-value system content.

Core views

Marvell announced a warrant structure with Google that ties a custom-silicon partnership to up to $120bn of cumulative sales to Google through FY33, in exchange for up to about 60mn Marvell shares. Bank of America interprets the agreement as an expansion of Marvell’s XPU-attach opportunity rather than a transfer of the main TPU compute die. It therefore reiterates Marvell as its top connectivity pick, citing possible upside to forecasts and greater confidence in its $15-plus CY30 EPS power. The report also stresses that XPU-attach programs can be fragmented, carry lower content value and face changes in design partners. The institution separates the companies’ roles in the TPU system. Broadcom supplies the core ASIC and advanced-package design, HBM integration, 100G/200G high-speed SerDes, and the networking fabric including switching and PHY connectivity for XPU clusters. Marvell’s expected contribution is in custom-silicon programs attached to the TPU ecosystem, such as AI inference accelerators, storage controllers, NICs, memory-interface controllers and near-memory compute. Because the detailed allocation remains unclear, the report bases its conclusion on the types of content Marvell has highlighted across customer engagements and expects Broadcom to retain the core compute die. Bank of America estimates that consensus Google capex for CY27-31E will total roughly $1.5-2.0tn+, with about 25-30% directed to TPU systems. This produces a TPU-system TAM of more than $400-600bn, compared with a historical 20-25% share of capex, as servers, compute and HBM represent a larger mix. Its exhibit uses a $1.711tn consensus capex case, $1.069tn for server/hardware, and a $470bn TPU TAM. The report expects Broadcom to retain about 55-60%+ of that market, equal to roughly $250-350bn+ of content over the next five years, because its main-die, HBM/packaging, SerDes and networking content represents the highest bill-of-materials value and is locked in through CY31. Marvell could take 10-20%, or about $50-100bn+, toward the implied $120bn opportunity by around CY32 if the warrant fully vests; MediaTek could take 20-30% in cost-optimized variants. In the report’s view, the three-partner structure chiefly redistributes lower-value inference and attach content rather than Broadcom’s high-value die content. For Broadcom’s FQ3 report on September 2, the institution expects strong near-term TPU demand and believes management can reiterate a robust custom-silicon trajectory. It highlights new Meta and OpenAI ramps, each around 1GW and above $10bn, beginning in CY27 alongside the Google TPU base. Consensus already models about $120bn of Broadcom FY27 AI sales versus management’s stated expectation of more than $100bn, but the report argues silicon content can rise with each generation as system complexity increases, analogous to Nvidia GPU content. It acknowledges that Marvell competition could weigh on Broadcom’s valuation multiple even as absolute dollar opportunity grows. Still, it calls Broadcom’s roughly 21x CY27E and 16x CY28E P/E, or about 0.5x PEG, attractive given durable AI growth and TPU-content leadership through CY31. Broadcom’s $530 price objective is based on 30x CY27E P/E, within its historical 11x-41x range and supported by double-digit EPS growth, semiconductor-leading profitability, free-cash-flow generation and returns. Marvell’s $365 price objective uses 31x CY30E EPS power of $15+, including stock-based compensation, discounted back two years. The multiple is above its 26x historical median but within its 14x-47x historical range; Bank of America attributes the support to improving visibility on major customer ASIC projects, a broader AI portfolio and potential AEC, CPO and scale-up-switch share gains.

Analysis framework

The report estimates Google’s prospective TPU-system market from consensus capex and the portion expected to be spent on TPU systems, then allocates that opportunity among suppliers according to their system-content roles. It combines this supply-chain analysis with customer agreements, expected AI ramps, earnings estimates and P/E and PEG valuation frameworks for Broadcom and Marvell.

Methodology notes

• Industry AnalysisUpstream-Midstream-Downstream Transmission

  TPU-system content allocation across core compute, packaging, networking and XPU-attach components

  The report identifies which supplier provides each portion of Google’s TPU ecosystem, then uses the relative value of those components to estimate the likely market share and revenue opportunity for Broadcom, Marvell and MediaTek.

• Industry AnalysisSupply-demand framework

  Google capex-to-TPU TAM build

  Google’s projected aggregate capex and assumed TPU-system spending share are used to derive the CY27-31E TPU-system total addressable market.

• Valuation methodsP/E and PEG Valuation

  Forward P/E and PEG valuation

  Broadcom’s price objective is based on 30x CY27E P/E and its valuation is discussed at roughly 21x CY27E, 16x CY28E and 0.5x PEG; Marvell’s objective applies 31x CY30E EPS power discounted back two years.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• Broadcom (AVGO)

  Core TPU compute, packaging, HBM integration, SerDes and networking supplier to Google’s TPU ecosystem.

  Strengths

  High-value content, Google long-term agreement through CY31, projected 55-60%+ TPU TAM share, durable AI growth and strong profitability/FCF generation.

  Weaknesses

  Competition from Marvell may pressure the valuation multiple despite continued dollar growth.

  Comparison

  Expected to retain the high-value core die while Marvell focuses on XPU-attach content and MediaTek on cost-optimized variants.

  Risks

  Semiconductor-cycle and AI-theme sensitivity, Apple and Google design-out risk, networking and other market competition, acquisition and integration risk, and $60bn net debt.

• Marvell Technology (MRVL)

  Potential supplier of TPU ecosystem XPU-attach content under a new Google custom-silicon partnership.

  Strengths

  Potential up to $120bn Google sales opportunity, improving visibility on major ASIC projects, and broader AI connectivity, switching and compute portfolio.

  Weaknesses

  XPU-attach programs are fragmented, lower-content opportunities and may change design partners.

  Comparison

  Expected to take 10-20% of TPU TAM versus Broadcom’s 55-60%+ core-content share.

  Risks

  Reduced visibility on key ASIC projects, including next-generation 3nm/2nm programs at AWS and Microsoft; AI-compute competition; and cyclicality in legacy storage, enterprise networking and carrier markets.

Key data

• Marvell potential Google sales opportunityUp to $120bn through FY33Linked to the warrant structure and full vesting potential.
• Google aggregate capex$1.5-2.0tn+Consensus expectation for CY27-31E.
• Google TPU-system TAM$400-600bn+Assumes roughly 25-30% of CY27-31E capex is spent on TPU systems.
• Broadcom TPU TAM value share55-60%+; approximately $250-350bn+Expected CY27-31E share based on core compute, HBM/packaging, SerDes and networking content.
• Marvell TPU TAM value share10-20%; approximately $50-100bn+Expected opportunity from XPU-attach programs.
• Broadcom FY27 AI sales consensusApproximately $120bnCompared with management’s stated expectation of more than $100bn.
• Broadcom price objective$530Based on 30x CY27E P/E.
• Marvell price objective$365Based on 31x CY30E EPS power of $15+ and discounted back two years.

Impact & implications

Bank of America sees a larger Google TPU ecosystem as creating room for both companies: Marvell gains potential XPU-attach revenue and improved ASIC-project visibility, while Broadcom’s established position in the core die and other high-value system content remains intact through CY31. The report nevertheless notes that Marvell’s entry can create a valuation overhang for Broadcom and that attach-content programs are less concentrated and potentially less durable.

Risks

• Broadcom faces semiconductor-cycle and AI-theme sensitivity, potential Apple and Google design-outs, competition across networking and other markets, acquisition and integration risk, and $60bn of net debt.
• Marvell’s XPU-attach programs may be fragmented, lower in content value and subject to design-partner changes.
• Marvell could lose visibility on key custom ASIC projects, face intensifying AI-compute competition, or encounter cyclicality in storage, enterprise networking and carrier markets.

What to watch

• Broadcom’s FQ3 results on September 2 and management’s commentary on TPU demand and custom-silicon growth.
• The timing and scale of Meta and OpenAI custom-silicon ramps expected to begin in CY27.
• Whether Marvell’s Google partnership remains focused on XPU-attach content or expands toward core TPU compute content.
• Progress toward the Google warrant’s implied sales milestones and visibility on major Marvell ASIC projects.