Google is in advanced talks with Marvell Technology to co-develop a new generation of custom silicon for AI inference, a move first reported by The Information that would break Broadcom’s long-running lock on Google’s TPU program.
Two chips are on the table, including a dedicated LLM inference processor. Marvell stock has ripped more than 50% year-to-date on the news and its preceding $2 billion Nvidia investment.
The deal signals a broader shift in where the money inside AI is actually going.
What Was Reported
According to The Information, Google is negotiating with Marvell on two new AI chips designed specifically for running, not training, large models. In the proposed arrangement, Marvell would act in a design-services capacity, similar to the role MediaTek already plays for Google. One of the chips under discussion is a dedicated memory processing unit. The companies are reportedly aiming to finalize designs as soon as next year before handing off for test production.
The news comes days after Google extended its long-term TPU and networking agreement with Broadcom through 2031. In other words, Google is not replacing Broadcom. It is adding a second design partner to a program that has been functionally single-sourced for years.
Why This Matters: The Economics Have Shifted
For most of the current AI cycle, the money went into training. Building frontier models required enormous clusters of GPUs running for months. That is the market Nvidia has dominated. But as models get deployed at scale, the cost center shifts. Every ChatGPT query, every Claude response, every Gemini answer is an inference call. At Google’s volume, those queries add up to a bill that dwarfs training costs over time.
Custom inference silicon is cheaper per query than general-purpose GPUs. That is the entire strategic point. The cloud providers that design their own chips, including Google with TPUs, Amazon with Trainium, and Microsoft with Maia, are trying to get off the Nvidia premium for the workloads they run constantly. Marvell is one of the firms that makes that possible.
Why Marvell, Specifically
Marvell has quietly built one of the strongest custom silicon businesses outside of Broadcom. Its data center revenue hit a record $6.1 billion in fiscal 2026, with custom silicon running at a $1.5 billion annual run rate across 18 cloud-provider design wins. Customers already include Amazon on Trainium, Microsoft on Maia, and Meta on a new data processing unit.
The setup got stronger at the end of March when Nvidia invested $2 billion in Marvell and partnered through NVLink Fusion to integrate Marvell’s custom chips and networking into Nvidia’s ecosystem. The message to hyperscalers was direct: you can buy Nvidia GPUs, custom Marvell silicon, or both, and they will work together. Barclays analyst Tom O’Malley upgraded Marvell to overweight and raised his price target from $105 to $150 on the combined momentum.
| Metric | Value | Context |
|---|---|---|
| FY2026 Data Center Revenue | $6.1 billion | Record high, driven by custom silicon |
| Custom Silicon Run Rate | $1.5 billion/year | Across 18 cloud-provider design wins |
| Nvidia Investment | $2 billion | Closed end of March 2026 |
| Stock Performance YTD | +50%+ | Roughly 30% gained in April alone |
| Barclays Price Target | $150 | Upgraded from $105 to overweight |
Broadcom Still Holds the High Ground
Despite the Marvell story, Broadcom’s position in custom AI accelerators remains dominant. The company commands more than 70% market share in the category. Its AI revenue hit $8.4 billion in its most recent quarter, up 106% year over year, with guidance of $10.7 billion next quarter and a long-term target of $100 billion in annual AI chip revenue by 2027.
The Google-Broadcom TPU extension through 2031 puts a floor under that trajectory. What Google is doing with Marvell is not a Broadcom replacement, it is a negotiating posture. Hyperscalers that depend on one chip partner, whether Nvidia or anyone else, face pricing risk, supply risk, and strategic vulnerability. Adding Marvell to the mix gives Google leverage on all three.
The Bigger Picture
Google’s chip strategy now spans four external partners (Broadcom, MediaTek, Marvell, and TSMC), an in-house design team, and a product line that covers training, inference, and general-purpose cloud compute. That complexity is deliberate. Every hyperscaler is racing to diversify silicon sources as AI workloads move from experimental to permanent line items on the P&L.
The short version for anyone tracking the AI market: training made Nvidia the most valuable company on earth. Inference is about to make a much longer list of companies extremely valuable, and Marvell just moved to the front of that list.
Stay ahead of the AI shift. Prompt Insider publishes weekly breakdowns on AEO, AI search, and the tools marketing teams need to stay visible. Subscribe to the newsletter for the next piece delivered straight to your inbox.
FAQs
What did Google and Marvell announce?
Google and Marvell have not announced a signed agreement. The Information reported that the two companies are in advanced negotiations to co-design two custom AI chips focused on inference, including a dedicated LLM inference processor and a memory processing unit.
Does this replace Broadcom’s role with Google?
No. Google extended its TPU and networking agreement with Broadcom through 2031 just days before the Marvell news broke. Marvell would be an additional design partner, not a replacement. Broadcom still holds more than 70% market share in custom AI accelerators.
Why is Google interested in inference chips specifically?
Inference is the cost of running AI models after they are trained. At Google’s query volume across Search, Gemini, and Cloud, inference spend scales with usage and becomes the dominant cost over time. Custom inference silicon is cheaper per query than general-purpose GPUs, which is why hyperscalers are racing to design their own.
How has Marvell stock reacted?
Marvell is up more than 50% year-to-date, with roughly 30% of that gain coming in April alone after the Nvidia investment and the Google talks. Barclays raised its price target from $105 to $150 and upgraded the stock to overweight.
What does this mean for Nvidia?
Nvidia remains the dominant supplier of AI compute, especially for training. The Marvell partnership announced at the end of March actually strengthens Nvidia’s position by integrating Marvell’s custom chips into the Nvidia ecosystem through NVLink Fusion. The longer-term risk is that hyperscalers keep shifting inference workloads to custom silicon, which compresses the share of AI compute spend Nvidia captures.