If you want to understand why AI keeps getting cheaper and more capable year after year, follow the chips. Last week’s announcement between Google and Marvell Technology is one of the clearest signals yet that the hyperscalers are not just buying AI capacity, they are locking in the infrastructure to own it.

Google granted Marvell the right to purchase up to $12.2 billion of its stock, tied to a co-design agreement covering custom AI semiconductors for Google’s TPU (Tensor Processing Unit) clusters. The warrant structure is straightforward: one block of shares unlocks for every $500 million in custom chip revenue Marvell books from Google, running from Q3 fiscal 2027 through fiscal 2033. If Google hits every tranche, Marvell books roughly $120 billion in product revenue over that period.

On August 28, Marvell raised its fiscal 2028 revenue forecast to approximately $18 billion, up from $16.5 billion previously, with data center revenue expected to grow more than 60%. The Google deal is the primary driver.

What Marvell Is Actually Building

The agreement covers more than one type of chip. Marvell will co-design inference accelerators, storage controllers, network interface controllers, memory interface controllers, and near-memory computing products for Google’s TPU ecosystem.

That breadth matters. TPU clusters are not just processors, they are entire systems where memory bandwidth, networking throughput, and storage access are as important as raw compute. Google is not outsourcing one component, it is partnering on the complete silicon stack that sits behind its AI infrastructure.

CEO Matt Murphy described the arrangement as “game-changing” and said the potential revenue contribution could be “a lot larger than modeled,” a comment that points to Marvell’s deeper integration across multiple layers of Google’s AI hardware.

The Real Story: AI Infrastructure Is Being Built at Lock-In Scale

The warrant structure tells the strategic story. Google is not simply placing a large purchase order. It is creating a financial incentive for Marvell to build around Google’s roadmap, and giving itself a stake in Marvell’s success in return. This is infrastructure procurement as equity strategy.

Wedbush analyst Matt Bryson noted that the agreement fits a broader pattern where hyperscalers are developing processors tailored to specific AI workloads rather than relying solely on general-purpose chips. Critically, he observed that this opportunity may benefit multiple chip designers, rather than simply shifting share among competitors. The market for custom AI silicon is growing fast enough to accommodate new entrants.

Google has now completed what analysts are calling the “custom silicon hyperscaler sweep,” with deals across inference chips, networking, and storage from multiple semiconductor partners. The scale of commitment suggests Google believes the AI compute demand it is seeing is not a temporary peak but a durable, decades-long infrastructure buildout.

Why This Matters for Businesses That Depend on AI

Most business leaders do not spend time thinking about chip architecture. But the decisions being made in semiconductor boardrooms right now have direct implications for what AI tools will cost and what they will be capable of in 24 to 36 months.

The pattern is consistent across the industry: when hyperscalers invest heavily in custom silicon optimized for AI inference, the cost-per-query for running AI models falls. That is why OpenAI cut GPT-5.6 Luna pricing by 80 percent this year, and why Google’s Gemini Flash models are priced at a fraction of what frontier models cost eighteen months ago. Custom chips are the upstream reason frontier AI keeps getting cheaper.

For businesses currently evaluating AI tools, this trajectory matters. The models that feel expensive or computationally intensive today will almost certainly become more affordable as infrastructure investments work their way through the value chain. That is not an argument to wait, it is an argument to build capability now and plan for costs to fall.

On AI strategy: Organizations that build internal AI competency now, including the data infrastructure, processes, and team skills to deploy AI effectively, will be in a position to absorb rapidly improving capabilities as they arrive. Those that wait for AI to “settle down” risk finding themselves behind on execution when the infrastructure does catch up.

What This Means for Business

For technology buyers: Custom chip partnerships like this one reinforce that major cloud platforms are making decade-long commitments to AI infrastructure. Businesses choosing between cloud AI providers should factor in infrastructure depth, not just current model quality.

For finance and operations leaders: The economics of AI tools will continue to improve. Building a business case for AI adoption on today’s pricing is conservative, the actual cost trajectory is downward. Factor that into multi-year planning.

For teams building on AI APIs: Marvell’s deep integration into Google’s TPU stack means Google Cloud’s AI services will have consistent access to tailored compute for the foreseeable future. That reduces the platform risk of building on Google’s infrastructure relative to smaller providers.

For anyone still on the fence about AI adoption: The scale of capital commitment from companies like Google is not noise. It is the market’s clearest signal that AI is now foundational infrastructure, on the same level as cloud computing was in 2012. The question is not whether AI will be central to business operations, it is whether your organisation will be ready when it is.

---

Enterprise DNA helps businesses build the data fluency and AI capability to take advantage of these shifts. If you want to understand what the current AI infrastructure buildout means for your specific industry and operations, Omni Advisory is where that conversation starts.