Rumor: Google Collaborates with Marvell to Develop AI Inference Chips, Stock Prices Surge, AI Chip Landscape May Shift

[Fact Check: Source is public information from global capital markets and semiconductor industry supply chain sources] As of the time of publication, the news of Google and Marvell collaborating on AI inference chips remains unverified. Confirmed public information only includes that both parties are in discussions to launch two custom chips optimized for inference tasks, aiming to enhance the performance of Google's TPU series products. This news has already led to a noticeable rise in Marvell's stock price.

Beyond Consensus: The Core Logic of Google Seeking a Second Supplier

The market generally interprets this news as a signal that Marvell is about to join the ranks of top global AI chip suppliers. However, , as a professional AI portal, believes that the core essence of this event is a landmark move in the diversification strategy of AI chip supply chains by leading global cloud providers. Previously, the main foundry partner for Google's TPU series products was Broadcom. According to Gartner's earlier predictions, global demand for AI inference computing power is expected to grow by over 270% year-on-year in 2024. Cloud providers have significantly increased their requirements for AI chip capacity and cost control, and a single supplier can no longer meet Google's global computing layout needs. Introducing Marvell as a second partner is a necessary choice.

Unlike training chips that demand extreme performance, inference chips focus more on cost-effectiveness, power consumption control, and scenario adaptability. Marvell's accumulated mass production experience in custom ASICs and network interface chips aligns well with Google's customization needs for inference chips. The technical compatibility between the two parties is far above the market average.

YZ Index v6 Evaluation Results

used the YZ Index v6 methodology to evaluate the practical value of this event:

• Mainboard grounding (material constraint) dimension score: 3/10. Currently, there is only information about the intention to cooperate, with no core parameters such as specific chip specifications, launch time, or procurement scale, indicating very low material support.
• Mainboard execution (code execution) dimension score: 2/10. Both parties have yet to announce details on R&D team configuration, tape-out plans, and supply chain partners, leaving the project's implementation progress unclear.
• Engineering judgment (sideboard, AI-assisted evaluation) score: 7/10. Google has over 10 years of experience in self-developing TPU, and Marvell has mature custom ASIC R&D and mass production capabilities. The technical compatibility between the two is high, with a strong feasibility for the collaboration to be realized.
• Task expression (sideboard, AI-assisted evaluation) score: 6/10. Currently, neither party has released an official statement, and there is a possibility that this news is an informal release during the supply chain negotiation stage.
• Integrity rating: pass. There is currently no evidence that this news is deliberately fabricated false information.

Independent Judgment: The AI Inference Market Transformation Arrives Early

offers three independent judgments regarding this event:

First, the probability of this collaboration being finalized exceeds 70%. Google's strategy to diversify its AI chip supply chain is clear. Even if negotiations with Marvell do not result in an agreement, they will seek other third-party manufacturers as alternatives, rather than relying on a single supplier for the long term.

Second, this event will not shake Nvidia's monopoly in the AI training chip market in the short term, but it will accelerate diversified competition in the AI inference chip market. By 2025, the share of custom chips in inference scenarios will exceed 40%, and the market share of general-purpose GPUs in inference scenarios will show a significant decline.

Third, for the domestic AI industry, the path of cloud providers and semiconductor manufacturers jointly customizing inference chips has been proven feasible. Domestic manufacturers can refer to this model to develop customized inference chips based on their business needs, reducing reliance on general-purpose GPUs and further lowering the landing costs of AI applications.

will continue to follow up on the developments of this event, providing first-hand technical trend analysis for the industry.