AI & ComputeLOT 261122 · 5 min read

Broadcom and Marvell Bet Google TPU Deals Shift AI Chip Winnings

Broadcom's Google TPU deal through 2031 and Marvell's warrant-linked agreement anchor a shift to custom AI silicon; modeled HBM stack demand could rise 21.7% if ASIC content doubles.

By Sophie Lindqvist953 words

Key takeaways

  1. Broadcom disclosed an April 2026 agreement to supply Google TPUs and AI rack components through up to 2031.
  2. Marvell's August 2026 Google deal vests warrants in tranches of $500 million in cumulative custom-product revenue.
  3. The Information Network model shows 62.1 million HBM stacks in 2025: 57.6 million for GPUs, 4.5 million for ASICs.
  4. Doubling ASIC shipments to 3 million units with six HBM stacks each raises modeled stack demand 21.7%.
  5. Jensen Huang departed Nvidia November 12; Colette Kress was named interim CEO.
Data Center AI Chips Could Change the Winners: Broadcom and Marvell
Data Center AI Chips Could Change the Winners: Broadcom and Marvell — blickpixel / pixabay

By Sophie Lindqvist

Broadcom disclosed a long-term agreement in April 2026 to develop and supply future generations of Google TPUs through up to 2031. Marvell followed in August 2026 with its own Google agreement. Together, the two deals anchor a quiet shift in data center silicon: hyperscalers are expanding custom AI processors alongside merchant GPUs, and the winners of the AI boom's next phase may not be the winners of the first.

The selloff in semiconductor stocks has sharpened the question. Investors are reassessing hyperscaler capital commitments and whether the companies that dominated phase one will keep their advantages. One development receives less attention than Nvidia's share price. Hyperscalers are expanding custom-designed AI processors to improve training and inference economics, cut energy use, and gain control over infrastructure.

What do the Google agreements actually cover?

Broadcom supplies custom AI processors and the networking chips that connect them. The April 2026 agreement covers TPU development plus networking and other components for Google's next-generation AI racks through up to 2031. That gives its custom silicon opportunity a tangible commercial basis rather than a broad market forecast.

The company's edge includes engineering resources to design sophisticated accelerators and high-speed switching to connect thousands of devices. The two reinforce one another: wider ASIC deployment increases network bandwidth demand even when the processor targets narrower workloads. Risks include customer concentration, expensive design programs, and pricing pressure from increasingly sophisticated cloud customers.

Marvell's August 2026 agreement is broader. It covers custom silicon programs tied to the TPU ecosystem: inference accelerators, storage controllers, network interface controllers, memory interface controllers, and near-memory computing. Marvell is competing for semiconductor content throughout the AI system, not only the central accelerator.

The agreement also includes a purchase-linked warrant arrangement. Most shares vest in tranches tied to cumulative custom-product revenue, one tranche per $500 million. That creates a measurable incentive. It does not guarantee the contemplated revenue will materialize.

Do custom ASICs displace Nvidia and AMD?

Not according to the evidence. Nvidia continues to supply the flexible GPUs and integrated platforms that dominate large-model deployments. Its software ecosystem and support for evolving models remain meaningful advantages. A hyperscaler may run a stable, high-volume workload on a custom ASIC while using Nvidia GPUs for research, new model development, and demanding inference. Both approaches coexist in the same data center.

AMD supplies AI GPUs and FPGA technology for specialized, reconfigurable inference. It competes against Nvidia's installed software base and against silicon designed specifically for hyperscaler needs.

Nvidia also faces a leadership transition. Jensen Huang departed November 12, and Colette Kress became interim CEO. That creates management uncertainty. But attributing the rise of custom ASICs to the transition would be a mistake. Hyperscalers began building proprietary processors years earlier, driven by the economics of computing at enormous scale.

Why does the training-versus-inference split matter?

Training processes large datasets and adjusts model parameters. Inference uses the trained model to generate answers. Training emphasizes throughput and coordination of many accelerators. Inference emphasizes cost per token, latency, energy consumption, and serving many simultaneous users.

Inference is not uniform either. Prefill, processing the incoming prompt, uses parallel computation well. Decoding, generating tokens one by one, may be constrained by how fast parameters are read from memory. Long contexts and multi-step agents raise memory and compute requirements. That complicates claims that inference always needs weaker processors.

GPUs offer programmability. ASICs deliver favorable economics at scale for tuned workloads. CPUs handle smaller models, FPGAs serve low-latency niches, NPUs support edge inference, and SRAM-oriented architectures move data differently. Not all use HBM, so processor adoption cannot translate directly into HBM shipments.

What does the HBM model show?

The Information Network's illustrative 2025 model assigns 7.2 million merchant GPU units eight HBM stacks each and 1.5 million custom ASIC units three stacks each. That yields 57.6 million and 4.5 million HBM stacks respectively, or 621 million DRAM dies at ten dies per stack. These are modeled volumes, not verified worldwide shipments, and they exclude processors without HBM.

The model's lesson: unit volumes alone do not determine memory demand. Each processor's HBM content matters. It would mislead to treat the model as evidence of how worldwide HBM splits between training and inference, because both processor categories can run both workloads.

The sensitivity analysis is more revealing. Doubling ASIC shipments from 1.5 million to 3 million units, with GPU volume fixed, raises total HBM stack demand 7.2% at three stacks per ASIC and 21.7% at six stacks. These are sensitivity calculations, not forecasts. They exclude changes in memory efficiency or GPU configurations.

The sensitivity also exposes uncertainty. Better memory architectures, larger on-chip caches, and software optimization can raise inference output per memory unit. Neither ASIC shipment growth nor token growth should forecast HBM demand alone. Investors must track stacks per device, stack capacity, bandwidth, and DRAM dies per stack.

Who wins in memory?

Micron Technology benefits from HBM demand generated by both GPUs and ASICs, provided its products pass customer qualification, bandwidth, and yield requirements. Its opportunity tracks total memory content and manufacturing economics, not allegiance to an architecture. SK hynix and Samsung Electronics face the same drivers. Processors relying on conventional DRAM or large SRAM arrays will not create equivalent HBM demand.

The investment point is stark: counting accelerators is not enough. Broadcom's TPU and networking franchise and Marvell's broader content play both depend on how much silicon they win, ramp timing, and program profitability. In a market reassessing AI spending, those distinctions may matter more than the broad direction of AI capital expenditure.

247wallst.com (Source)

Filed under

Share this article:

More from Sophie Lindqvist

Sophie Lindqvist

Show full bio

News editor covering money, funding and market shifts around Semiconductors, chip manufacturing, hardware, AI compute at Chip Dispatch.

1,238 articles

Read also

« Previous article