Google is in discussions with Marvell Technology to develop a new set of custom AI chips, a move that reflects a broader shift among hyperscalers toward tighter control over the infrastructure underpinning artificial intelligence.

According to The Information, the collaboration would focus on two components: a memory processing unit designed to work alongside Google’s tensor processing units, and a new TPU optimized specifically for running AI models. While the discussions have not been formally confirmed, the direction aligns with Google’s long-running strategy of building vertically integrated AI systems that span hardware, software, and cloud delivery.

The technical emphasis is notable because, as AI models scale, the limiting factor is no longer just compute capacity but how efficiently data can be moved between memory and processors. Training and inference workloads are increasingly constrained by bandwidth, latency, and energy consumption tied to memory access. A dedicated memory processing unit suggests Google is targeting this bottleneck directly, attempting to reduce the “data movement tax” that has become one of the most expensive aspects of modern AI systems.

This is where the competitive dynamics sharpen. Nvidia has built its dominance not only on powerful GPUs but on an integrated architecture that tightly couples compute, memory, and software through its CUDA ecosystem. That integration has created high switching costs, effectively locking developers and enterprises into Nvidia’s stack.

Google’s response is to replicate that level of integration on its own terms. TPUs, originally designed for internal workloads such as search and advertising, have evolved into a core pillar of its cloud offering. By extending the architecture with specialized memory components, Google is attempting to optimize the full system rather than individual chips, a strategy that could yield efficiency gains at scale.

The economic logic is equally important. AI infrastructure is capital-intensive, with hyperscalers committing tens of billions of dollars to data centers, networking, and compute hardware. Relying solely on third-party suppliers like Nvidia exposes companies to pricing pressure and supply constraints. Custom silicon offers a way to reduce unit costs over time while tailoring performance to specific workloads.

For Google, TPU adoption has already become a meaningful contributor to cloud revenue growth, as it seeks to demonstrate that its heavy AI investments can translate into commercial returns. Offering differentiated hardware through its cloud platform allows it to compete more directly with rivals, particularly in attracting enterprise customers running large-scale AI workloads.

Marvell’s role in this equation is to bridge design and production. Known for its expertise in custom silicon and data infrastructure, Marvell has positioned itself as a key partner for companies seeking to build specialized chips without owning fabrication facilities. Its involvement suggests that Google is leveraging external manufacturing and design capabilities to accelerate development cycles while maintaining control over architecture.

The reported timeline—finalizing the memory processing unit design as early as next year, before moving to test production—indicates a relatively aggressive schedule, though such timelines are often subject to delays tied to fabrication, validation, and yield optimization.

The broader industry context reinforces the significance of the move. Major technology firms are increasingly designing their own chips, leading to a fragmentation of the AI hardware landscape. Instead of a single dominant supplier, the market is evolving toward multiple specialized architectures optimized for different use cases—training, inference, edge deployment, and real-time processing.

This fragmentation carries both opportunity and risk as it enables innovation at the system level, with companies optimizing hardware for specific applications. It also complicates the software ecosystem, potentially creating compatibility challenges and increasing the burden on developers to adapt workloads across different platforms.

The rise of custom silicon among Nvidia’s largest customers represents a structural challenge, even as demand for its GPUs remains strong. Each chip designed in-house by a hyperscaler reduces long-term dependence on external suppliers, even if those suppliers remain critical in the near term.

For Google, success will depend on more than hardware performance as it must ensure that its TPUs and associated chips integrate seamlessly with widely used AI frameworks, offer competitive pricing, and deliver consistent reliability at scale. Without that, even technically superior designs may struggle to gain traction beyond Google’s internal ecosystem.

There is also a strategic timing element. As the AI market matures, the focus is shifting from experimentation to efficiency. Enterprises are becoming more sensitive to the cost of running AI workloads, particularly as usage scales. Chips that can deliver comparable performance at lower cost or with better energy efficiency will have a clear advantage.

In that sense, Google’s reported plans are not just about catching up with Nvidia, but about anticipating the next phase of competition—one defined less by raw compute power and more by cost efficiency, system optimization, and total cost of ownership.

The discussions with Marvell remain at an early stage, and both companies have declined to comment publicly. But the trajectory is clear. Control over AI infrastructure is becoming as important as the models themselves.

In the coming years, the companies that can design, build, and operate their own silicon, while integrating it into scalable cloud platforms, are likely to define the competitive hierarchy of the AI industry. Google’s latest move suggests it is intent on securing that position before the window narrows.

Fears that agentic artificial intelligence could dismantle the traditional software industry are prompting a strategic response from incumbents, with Microsoft advancing a framework that would fold AI agents into the same commercial logic that has underpinned enterprise software for decades.

The argument, articulated by Microsoft Executive Vice President Rajesh Jha, is that AI systems will not eliminate software demand but redefine who, or what, counts as a user. Speaking at a recent conference, Jha described a near-term scenario in which AI agents operate inside corporate environments as digital workers, each with its own identity, credentials, and access rights.

“All of those embodied agents are seat opportunities,” he said, invoking the industry’s core pricing unit: the per-user license.

Under Microsoft’s model, an organization deploying AI at scale could see its licensing footprint expand rather than contract. Ten employees supervising multiple agents would not reduce demand for software seats; it could increase demand for them. Each agent, performing discrete tasks across systems, would require authorized access, audit trails, and integration into enterprise identity frameworks.

This framing is designed to counter a competing narrative gaining traction among analysts and technologists. In that view, large language models and autonomous agents act as intermediaries that sit above traditional software stacks, executing tasks without requiring users to engage directly with multiple applications. If that model holds, the value of many SaaS products could be compressed, with AI interfaces becoming the primary layer of interaction.

Microsoft’s counter-position is rooted in infrastructure realities. Enterprise systems are governed by strict controls around identity, permissions, and compliance. Even highly autonomous agents must authenticate, access data through approved channels, and leave verifiable records of activity. These requirements create a structural argument for preserving licensing frameworks, even as the nature of “users” evolves.

There is also a financial imperative. The global software industry, dominated by subscription-based SaaS models, depends heavily on predictable, per-seat revenue streams. A shift toward fewer human users interacting with software would, under traditional assumptions, threaten that model. Microsoft is attempting to extend its revenue base into the automation layer itself by redefining agents as licensable entities.

However, this approach introduces new tensions. Nenad Milicevic of AlixPartners argues that agentic AI may push enterprises in the opposite direction, toward minimizing the number of active “users” altogether. As automation scales, a smaller group of human supervisors could manage increasingly complex workflows, reducing the need for widespread software access.

In such a scenario, the logic of per-seat licensing begins to strain. If a single AI agent can perform the work of several employees, charging per “seat” may not align with perceived value. Vendors could respond by increasing prices for machine-based operators or introducing new tiers of licensing, but that carries competitive risk. Enterprises may favor providers that adopt usage-based or outcome-based pricing models better suited to automated environments.

The debate points to a deeper question about how value is measured in an AI-driven enterprise. Traditional software pricing assumes a linear relationship between users and output. Agentic systems break that link. One agent can operate continuously, scale tasks dynamically, and interface with multiple systems simultaneously. This creates ambiguity around what constitutes fair pricing: access, activity, or results.

There is also a dimension tied to platform control. If AI agents become the primary interface through which work is executed, the layer that orchestrates those agents could capture disproportionate value. Microsoft’s broader AI strategy, including its investments in enterprise copilots and cloud infrastructure, suggests it is positioning itself not just as a software vendor but as the operating environment for these agents. Maintaining licensing control within that stack would reinforce its ecosystem advantage.

However, the emergence of more open and interoperable AI systems could challenge that dominance. If enterprises can deploy agents that move seamlessly across different software environments, the switching costs that have historically protected large vendors may weaken. That would shift bargaining power toward customers, particularly large enterprises capable of negotiating bespoke licensing arrangements.

Operational considerations further complicate the picture. Treating AI agents as employees requires organizations to rethink identity management, cybersecurity, and governance frameworks. Each agent would need defined permissions, monitoring protocols, and accountability structures. These requirements could reinforce the role of established enterprise software providers, which already offer integrated solutions for managing users and access.

For now, the agentic AI economy remains in an early, largely experimental phase. Most deployments are limited in scope, and the economics of large-scale automation are still being tested. The contrasting perspectives from Jha and Milicevic reflect an industry attempting to map out its future before the underlying dynamics fully materialize.

What is emerging, however, is a clear divide, where incumbents like Microsoft are working to adapt existing revenue models to a world of machine-driven work, effectively extending the concept of a “user” to include AI. Critics believe that the same technology could erode those models, forcing a transition toward more flexible and potentially less lucrative pricing structures.