Google is in discussions with Marvell Technology about developing custom chips designed specifically for AI inference, according to a report that surfaced in early 2025 and has gained renewed attention as both companies ramp up their semiconductor strategies. The potential partnership would pair Google’s massive AI deployment needs with Marvell’s growing custom silicon business, creating a new hardware track separate from the Tensor Processing Units that Google already builds with help from Broadcom.
Neither company has publicly confirmed the talks. Google declined to comment when the reports first circulated, and Marvell has not addressed the matter in its public earnings calls or press releases through early 2026. But the outlines of the reported deal align with moves both companies have been making independently, and the strategic logic is strong enough that analysts and competitors are paying close attention.
Why Google would look beyond its current chip partners
Google has been designing its own AI accelerators longer than almost any other cloud company. The first TPU debuted internally in 2015, and the program has evolved through multiple generations, with the TPU v6e (Trillium) announced in 2024. Broadcom has served as Google’s primary design and manufacturing partner for TPUs, a relationship worth billions of dollars annually and one that has helped make Broadcom the dominant force in custom AI chip design for hyperscalers.
But relying heavily on a single external partner carries risk. Broadcom also builds custom silicon for Meta, ByteDance, and other large buyers, which means Google competes for engineering bandwidth and manufacturing priority. Adding Marvell as a second design partner would give Google more leverage in negotiations, more flexibility in scheduling tape-outs at foundries like TSMC, and a hedge against any disruption in the Broadcom relationship.
There is also a technical argument. Training and inference are fundamentally different workloads. Training requires massive floating-point throughput and enormous memory bandwidth to update model weights across thousands of chips simultaneously. Inference, by contrast, involves running a finished model against live data, prioritizing low latency, energy efficiency, and cost per query. Google’s TPUs have historically been optimized for training, though recent generations have improved inference performance. A chip designed from the ground up for inference, co-developed with a partner whose engineering team brings a different design philosophy, could unlock efficiencies that a general-purpose accelerator cannot match.
Marvell’s push into custom AI silicon
Marvell has been building toward this kind of opportunity for years. The company, historically known for storage controllers and networking chips, pivoted aggressively toward custom silicon under CEO Matt Murphy, who took the helm in 2016. A string of acquisitions, including the $10 billion purchase of Inphi in 2021 for its high-speed interconnect technology, reshaped Marvell into a company capable of designing complex, application-specific processors for data center customers.
That strategy has already produced results with other hyperscalers. Marvell has publicly disclosed custom chip programs with Amazon Web Services, where it contributes to networking and custom compute silicon, and the company has described a pipeline of five major custom chip programs across its cloud customer base. In its fiscal year 2025 earnings commentary, Marvell projected that its custom silicon revenue would grow substantially as these programs moved from design to volume production.
Marvell’s SEC filings, including its annual 10-K reports, reflect this expansion through a growing list of subsidiaries and design entities. While these regulatory documents do not name specific customers for specific chip programs, they show a corporate structure that has scaled to support multiple parallel custom silicon engagements, exactly the kind of infrastructure needed to take on a project as demanding as a Google inference accelerator.
The competitive landscape around inference chips
Google and Marvell would not be operating in a vacuum. The inference chip market has become one of the most contested segments in semiconductors, driven by the explosion of generative AI applications that require constant, cost-effective model execution at scale.
NVIDIA remains the dominant supplier, with its Blackwell architecture and H200 GPUs offering strong inference performance alongside their training capabilities. But NVIDIA’s chips are expensive, power-hungry, and in chronic short supply, which has pushed every major cloud provider to explore alternatives. Amazon has its Inferentia and Trainium chips, designed in-house by its Annapurna Labs division. Microsoft has developed the Maia 100 accelerator. Meta is building custom inference silicon as well.
Startups have also entered the race. Groq has attracted attention with its Language Processing Unit, which delivers extremely fast inference for large language models. Cerebras, SambaNova, and d-Matrix are pursuing different architectural approaches to the same problem. The common thread is a bet that inference workloads are large enough and distinct enough to justify purpose-built hardware rather than repurposed training chips.
For Google, which serves billions of queries per day across Search, YouTube, Gmail, and its cloud platform, even small improvements in inference cost or latency translate into enormous savings. A custom chip that reduces the cost per token for Gemini model inference by a few percentage points could save hundreds of millions of dollars annually at Google’s scale.
What the reported deal would mean for the chip industry
If Google formalizes a partnership with Marvell, the ripple effects would extend well beyond the two companies. Broadcom, which has built a significant portion of its recent growth on custom AI chip revenue from hyperscalers, would face a direct competitive challenge from Marvell in one of its most important accounts. Broadcom’s stock has been buoyed by investor confidence in its hyperscaler relationships, and any sign that Google is diversifying away could pressure that valuation.
For Marvell, landing Google as a custom chip customer would validate its strategic pivot and could accelerate its revenue growth in a segment where design wins tend to lock in multi-year production commitments. Custom chips take two to three years to move from initial design to volume manufacturing, so a deal signed in 2025 or 2026 would likely produce revenue starting in 2027 or 2028.
The broader industry trend is clear: hyperscale cloud companies are no longer content to buy off-the-shelf processors for their most critical workloads. They want chips tailored to their specific software stacks, power budgets, and performance targets. That shift is creating a booming market for custom silicon design partners, and Marvell and Broadcom are the two companies best positioned to serve it. A Google-Marvell deal would confirm that this market is large enough to support serious competition between them.
What to watch for next
The most concrete signals will come from Marvell’s quarterly earnings calls and SEC filings, where the company is required to disclose material customer relationships and revenue concentration. If Google becomes a 10% or greater revenue customer, Marvell would need to identify it by name in its financial statements. Short of that threshold, investors and analysts will be parsing commentary about the custom silicon pipeline for hints about new program ramps.
Google’s own hardware announcements matter too. The company typically reveals new chip generations at its Cloud Next or I/O conferences, and any mention of an inference-specific accelerator built with an external partner would confirm the Marvell reports. Google’s infrastructure blog and research publications could also offer early technical clues about a new chip architecture entering its data centers.
For now, the reported talks remain unconfirmed, and the evidence is circumstantial rather than definitive. But the strategic fit is compelling, the market opportunity is enormous, and both companies have been building toward exactly this kind of collaboration. Whether the deal materializes as reported or takes a different form, the underlying dynamic is real: the race to build better, cheaper, more efficient AI inference hardware is intensifying, and Google is not standing still.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
