Google Is Building a Four-Partner Chip Supply Chain to Challenge Nvidia in AI Inference

Why It Matters

By controlling a multi‑partner silicon stack, Google can slash inference costs at massive scale, eroding Nvidia’s market share in the most revenue‑critical AI workload. The strategy also mitigates supply‑chain risk, giving Google pricing leverage and operational independence.

Key Takeaways

• Google ships Ironwood TPU, delivering 42.5 FP8 exaflops per super‑pod
• Broadcom to build training TPU “Sunfish” on TSMC 2 nm by 2027
• MediaTek’s “Zebrafish” inference chip targets 20‑30% lower cost
• Google expects 4.3 M TPUs shipped in 2026, 35 M by 2028
• Multi‑partner supply chain reduces pricing and capacity risk for Google

Pulse Analysis

Google’s chip programme is now a coordinated ecosystem rather than a single‑vendor effort. By partnering with Broadcom for high‑end training accelerators, MediaTek for low‑cost inference silicon, Marvell for memory‑processing units, and Intel for surrounding compute, Google spreads design risk and creates internal competition that drives down prices. The Ironwood TPU, already in mass production, showcases the immediate payoff: 192 GB of HBM3E memory, 7.2 TB/s bandwidth, and a super‑pod capacity of 9,216 liquid‑cooled chips delivering 42.5 FP8 exaflops. This scale underpins Google’s AI services and sets a performance baseline for the upcoming v8 family.

The strategic focus on inference reflects a shift in AI economics. While training remains a bursty, high‑cost activity, inference runs continuously across billions of daily queries, making per‑query cost the decisive metric. MediaTek’s “Zebrafish” chip promises 20‑30% lower cost than competing solutions, directly targeting this margin pressure. As custom ASICs capture an increasing share of the AI compute market—projected to grow 45% in 2026 versus 16% for GPUs—Google’s diversified supply chain positions it to outpace Nvidia, whose GPUs excel at training but face higher unit costs for inference workloads.

Looking ahead, Cloud Next will likely unveil the v8 roadmap and concrete timelines for the Sunfish and Zebrafish designs, reinforcing Google’s narrative of supply‑chain resilience and cost leadership. With shipments slated to reach 4.3 million units in 2026 and over 35 million by 2028, the volume advantage could translate into substantial savings for Google Cloud customers and a compelling alternative to Nvidia’s ecosystem. If Marvell’s memory‑processing unit materializes, the stack will further tighten integration, cementing Google’s position as a self‑sufficient AI infrastructure provider.