Nvidia’s AI Boom Shifts Spotlight Toward Broadcom and Marvell’s Custom Silicon

Nvidia’s GPUs kicked off the modern AI infrastructure buildout, but the next phase of investment may rely more on “custom silicon” and the networking needed to scale it, according to a new market-focused argument from The Motley Fool published Monday, June 8, 2026. The thesis is that when AI systems mature, companies will spend less to train models on general-purpose accelerators and more to run inference and specific workloads using chips tailored for particular data center designs.

The AI “custom silicon” shift is already visible in partnerships that plug third-party compute into Nvidia’s rack-scale ecosystem. On NVIDIA’s newsroom site, the company said it and Marvell Technology (MRVL) announced a strategic partnership through NVIDIA NVLink Fusion that connects Marvell into NVIDIA’s AI factory and AI-RAN ecosystem. NVIDIA also said it invested $2 billion in Marvell as part of the deal.

Under NVLink Fusion, NVIDIA said the rack-scale platform is designed to let customers develop semi-custom AI infrastructure using Nvidia’s NVLink ecosystem. The companies outlined a division of labor: Marvell will provide custom XPUs and NVLink Fusion-compatible scale-up networking, while NVIDIA will provide supporting building blocks such as Vera CPU, ConnectX network interface cards, BlueField data processing units, NVLink interconnect, and Spectrum-X switches. NVIDIA also tied the collaboration to silicon photonics and to telecom infrastructure efforts via Aerial AI-RAN for 5G and 6G.

Marvell, meanwhile, reported momentum tied to its data center business. In its first-quarter fiscal 2027 update, covering a period ended May 2, 2026, Marvell said it delivered record first-quarter revenue of $2.418 billion, up 28% year over year, and guided second-quarter revenue to $2.7 billion at the midpoint. The company said growth would continue accelerating through fiscal 2027 driven by strength in data center solutions.

Broadcom (AVGO) is the other name highlighted in Monday’s piece, with the argument centered on Broadcom’s role in custom AI accelerators and the networking that moves data inside large AI clusters. Broadcom’s own filings show it is working directly with hyperscalers on bespoke silicon and supporting infrastructure. In an April 2026 Form 8-K, Broadcom and Google said they entered a long-term agreement for Broadcom to develop and supply custom Tensor Processing Units, or TPUs, for Google’s future TPU generations, along with a separate supply assurance agreement for networking and other components for Google’s next-generation AI racks through up to 2031.

Broadcom also described a buildout pathway that connects its custom-chip work to AI capacity commitments. The same Form 8-K said Broadcom, Google and Anthropic expanded their collaboration so that, beginning in 2027, Anthropic will access through Broadcom approximately 3.5 gigawatts of TPU-based AI compute capacity tied to Anthropic’s multi-gigawatt commitments. That arrangement is important to the customization story because it suggests demand for AI-specific silicon is moving beyond single-model deployments toward multi-year capacity plans.

Financially, Broadcom’s AI semiconductor growth has been steep enough to reinforce the market narrative. In its June 3, 2026 earnings release for the second quarter of fiscal 2026, Broadcom said semiconductor revenue from AI was $10.8 billion, up 143% year over year, driven by demand for custom AI accelerators and AI networking. The company said overall Q2 revenue rose 48% year over year to $22.2 billion, citing operating leverage alongside the AI-driven semiconductor ramp.

Still, the “next phase” framing depends on assumptions that may not hold uniformly across the industry. The Motley Fool article presents a market-view argument, not a quantified forecast of how quickly AI workloads will shift from Nvidia’s GPUs to ASIC-style accelerators like XPUs and TPUs. It also does not disclose any new contractual share targets, pricing commitments, or performance benchmarks beyond citing existing partnerships and recent results. How broadly customers adopt custom silicon, and whether Nvidia remains dominant on training and higher-margin system-level platforms, remain key uncertainties for both names.

Looking ahead, investors may watch whether partnerships like Marvell’s NVLink Fusion integration and Broadcom’s TPU and rack component agreements translate into sustained booking visibility and longer-cycle wins at hyperscalers. Additional datapoints to track include new evidence of throughput and power advantages from custom accelerators in production systems, and whether AI networking demand continues to grow at the same pace as compute demand in AI factories. Nvidia’s own recent financial messaging emphasized accelerating AI factory buildouts, underscoring that infrastructure spend is still rising, even as supply chains diversify into custom silicon.