Google is developing new custom chips designed to speed up how its AI models generate responses, and is in talks with Marvell Technology to help design them, according to a report published Monday by Bloomberg. The effort is aimed at the inference side of AI workloads — where trained models serve user queries — and would reduce Google's reliance on Nvidia GPUs for a growing share of its data center compute.
Bloomberg reports that the new silicon is intended to work alongside Google's existing Tensor Processing Unit (TPU) line, which the company has developed with Broadcom for more than a decade. The report does not specify a production timeline or manufacturing partner for the new chips.
Marvell Enters Google's Silicon Roadmap
Bloomberg reports that Google has been in discussions with Marvell to co-design custom AI accelerators, a move that would expand Google's silicon supplier base beyond Broadcom. Business Today and Startup Fortune both report that the partnership would focus on improving efficiency for inference workloads specifically, rather than training.
Benzinga, citing the Bloomberg report, writes that the Marvell engagement is part of Google's broader strategy to vertically integrate more of its AI stack. Marvell has not publicly confirmed the discussions, and Google has not issued a statement on the reported talks.
DeepBrief previously reported on the Google-Marvell custom AI inference chip discussions earlier on Monday, based on the same underlying Bloomberg reporting and corroborating coverage.
Inference Economics Drive The Push
The new chips target inference — the compute required to generate outputs from an already-trained model — rather than training, where model weights are learned. Inference workloads have grown as products like Google's Gemini and consumer search-integrated AI features scale to billions of queries.
Google's reported chip effort targets inference workloads, rather than the training workloads where Nvidia GPUs account for the majority of hyperscaler AI accelerator spending, according to Nvidia's quarterly filings.
Crypto Briefing reports that Google's custom silicon strategy is explicitly framed around reducing per-query compute cost as AI features become more deeply integrated into consumer products. Parameter.io writes that Alphabet shares advanced on the Bloomberg report, reflecting investor interpretation that in-house silicon could improve Google Cloud margins on AI services.
Inference-specific accelerators have become an active design target across the industry. DeepBrief previously covered Nvidia's unveiling of the Groq 3 LPU at GTC, which Nvidia positioned as an inference-optimized part, and Meta's custom chip program as part of its broader AI stack.
TPU History And Broadcom Relationship
Google first deployed its TPU in 2015 and has since released multiple generations, with Broadcom serving as the primary co-design and packaging partner, according to Bloomberg's report. The TPU has been used internally for Google Search, YouTube recommendations, and Gemini model training, and is offered externally through Google Cloud.
Bloomberg reports that the Marvell engagement does not replace the Broadcom relationship but adds a second supplier for a distinct chip line. Startup Fortune characterizes the move as Google expanding its custom silicon supplier options to diversify its supply chain. Neither Google nor Broadcom has issued a public statement on whether the reported Marvell discussions alter the scope of the existing TPU partnership.
Hyperscaler Custom Silicon Expands
Google's reported move sits within a broader pattern of hyperscalers designing their own AI accelerators. Amazon Web Services has developed the Trainium and Inferentia chip lines. Microsoft has disclosed its Maia accelerator. Meta has deployed its MTIA chips internally, as DeepBrief previously reported.
Benzinga notes that the common thread across these programs is a focus on workloads the operator runs at high volume internally — where the economics of designing custom silicon outweigh the engineering cost versus buying Nvidia parts. Nvidia, in its most recent quarterly filing, reported data center segment revenue that the company attributed largely to hyperscaler and AI-focused customer spending.
Market Reaction And Status
Parameter.io reports that Alphabet shares rose on the Bloomberg report. Crypto Briefing frames the news as relevant to investor assessments of Nvidia's long-term share of AI compute spending among the largest cloud operators.
As of publication, Google, Marvell, Broadcom, and Nvidia have not issued official statements on the reporting. The Bloomberg article is the primary source for the Marvell engagement, with Business Today, Startup Fortune, Crypto Briefing, Parameter.io, and Benzinga all citing Bloomberg's reporting in their own coverage.
