Alphabet discusses two new AI chips with Marvell to speed up inference in Google services

Alphabet is discussing with Marvell the development of two specialized AI chips for inference — the stage when an already trained model processes real user requests. If the project materializes, Google will get hardware tailored not for laboratory records, but for fast and massive response in interactive services.

Which chips are being discussed

At the center of the discussion are two different components. The first is a memory module that should accelerate data transfer within the AI system. For inference, this is critical: even a powerful accelerator loses value if the model hits memory bandwidth constraints, delays in data exchange, and slow data delivery. In large language models, these bottlenecks often determine how quickly and cheaply the system can serve the daily flow of user requests.

The second component is an updated version of TPU, Google's proprietary tensor processors, oriented toward interactive applications. We're talking about scenarios where the user expects an answer immediately: search, AI assistants, text generation, interface suggestions. This choice demonstrates a shift in priorities: value is increasingly created not at the moment of model training, but at the moment of its continuous operation across millions of brief requests, where stable latency, predictability, and cost per response matter.

Why Alphabet needs Marvell

For Alphabet, a partnership with Marvell looks pragmatic. Marvell is strong in data center infrastructure and specialized microchips, while Google has been developing its own TPU line for many years. Instead of doing everything entirely in-house, Alphabet can split responsibility: retain control over the accelerator architecture and bring in a partner for those parts of the system where memory, interconnects, packaging, and fast data exchange between components are particularly important.

The deal also has economic logic. Inference is rapidly becoming one of the most expensive layers of AI infrastructure, because it runs on constant user traffic. The more precisely the hardware is tuned to a specific type of load, the easier it is to scale the service without explosive cost growth. For Google, this is a chance to reduce the cost per response and better control its own technology stack; for Marvell, it's a way to strengthen its position in the overheated AI accelerator market.

• A separate memory module can reduce bottlenecks when working with large models
• An updated TPU is better suited for dialogue and other interactive scenarios
• Specialized chips typically deliver more efficiency per watt and per dollar
• For Google, this is a way to strengthen control over its own AI infrastructure
• For Marvell, this is entry deeper into the market for AI acceleration

Where the effect will be noticed

If negotiations move to real shipments, the effect will first be noticed not by users, but by teams responsible for Google's infrastructure. Inference is not a one-time model run, but continuous operation under load, when the system must quickly respond to a huge number of requests in succession. In such a mode, what matters is not just computation itself, but predictable response time, energy efficiency, and absence of bottlenecks when transferring data between memory and accelerator.

In practice, this means more fine-tuning of hardware for specific products. The market is gradually moving away from the idea of a single universal chip that works equally well for all AI tasks. Instead, companies assemble separate combinations of components for training, inference, and interactive modes. The negotiations between Alphabet and Marvell fit well into this trend: those who will win are not those with more GPUs, but those who more precisely design the entire execution chain for the actual load.

What it means

The negotiations between Alphabet and Marvell show that the new race in AI is unfolding not only around models, but also around specialized hardware for everyday services. Whoever can deliver a fast answer cheaper and more reliably will gain the advantage.