Google assembles four-partner chip supply chain with Broadcom, MediaTek, Marvell to challenge Nvidia in inference

Google is scaling its custom TPU strategy aggressively, with Ironwood now generally available, millions of units planned for 2026, and a v8 roadmap targeting TSMC 2nm in late 2027. The company has split the next generation across Broadcom’s training chip "Sunfish" and MediaTek’s lower-cost inference chip "Zebrafish," while Marvell and Intel expand the broader AI infrastructure stack. The multi-partner approach strengthens Google’s cost position in AI inference and intensifies competitive pressure on Nvidia, making this a sector-relevant AI hardware development.

Analysis

This is less about a single TPU cycle and more about Google turning AI compute into a procurement architecture it controls. The key second-order effect is bargaining power: by splitting training and inference across different design partners, Google commoditizes its suppliers against each other while preserving optionality on both price and capacity. That should improve Google’s gross margin resilience even if unit costs stay elevated, because inference scaling will increasingly be internalized rather than bought at hyperscaler prices.

The most actionable implication is that Nvidia’s moat is shifting from “best accelerator” to “default rack architecture,” and Google is actively attacking the latter. As inference becomes the dominant cost center, the competitive battleground moves from FLOPS to cost-per-request and supply assurance; custom ASICs can win there even with weaker versatility. That creates a creeping share loss risk for Nvidia in cloud inference and adjacent networking attach, while strengthening the ecosystem roles of Broadcom, TSMC, and—more tactically—Marvell if it wins the memory-processing slot.

The timeline matters: near-term sentiment may overstate the revenue impact before 2027, but the capex and design-in commitments are real today. The risk to the thesis is execution failure at advanced nodes and integration complexity across multiple partners; if 2nm schedules slip, Google still has to keep buying external silicon to bridge the gap. Conversely, the upside surprise is that the million-chip deployments could pull forward software and pricing changes at Google Cloud, turning this into a margin story sooner than the market expects.

Consensus may be underestimating how asymmetric this is for Google versus the suppliers. Google gets structural cost leverage; suppliers get volume, but their economics become more dependent on one customer’s roadmap and bargaining leverage. The market may also be over-reading this as a pure Nvidia-negative when the better framing is a reallocation of AI capex toward a broader custom-silicon stack, which can support AVGO/TSM/MRVL even as it pressures NVDA mix.

Key Decisions for Investors

• Long GOOGL vs short NVDA into Cloud Next / next 3-6 months: thesis is inference cost takeout and margin expansion at Google versus gradual cloud inference share loss for Nvidia; use a 1:1 dollar-neutral pair and trim if Nvidia reaccelerates on non-Google hyperscaler demand.
• Add AVGO on pullbacks for a 6-12 month horizon: Broadcom is the cleanest beneficiary of custom accelerator proliferation and network attach; risk/reward is attractive so long as Google’s TPU ramp stays on schedule and Broadcom’s custom-AI backlog remains intact.
• Tactical long MRVL 6-9 months, but only as a smaller position: optionality around memory-processing and a second inference TPU can re-rate the stock, yet it is a negotiation trade, not a signed-win trade; size modestly and stop if no contract emerges by next year.
• Underweight / short-duration hedge on NVDA through 2027: use call spreads or a relative-value short versus GOOGL/AVGO rather than an outright short; the market may be slow to price share erosion in inference, but training demand still supports the stock near term.
• Long TSM on any node-capacity concern dips, 12-24 months: Google’s multi-partner strategy still routes through TSMC, making it the toll collector on the entire roadmap; downside is node transition risk, but the volume concentration is a net positive.