A recent headline caught my attention in a way that most AI news doesn’t. Google, it said, had signed a major deal to lease access to its custom AI chips to Meta for training large artificial intelligence models. On the surface, this was just another big tech company securing more computing power in the global AI arms race. I’ve previously questioned the bubble of circular investment in AI infrastructure under the Magnificent 7. The Google/Meta announcement sounded like more of the same. But one word in the headline kept nagging at me.
Rent.
Leasing is common when an asset retains value over time. Automobiles, airplanes, and heavy construction equipment are leased because the owner expects the asset to remain usable after the initial contract expires. A leased aircraft can last thirty years and you can make more money leasing it for a longer period of time than selling it once. A bulldozer can work on dozens of projects and extract money from many clients before retiring. However, AI chips behave very differently.
Modern accelerator chips cost tens of thousands of dollars each, but their technological relevance may fade in a few years as new generations deliver dramatic gains in speed and efficiency. In extreme training environments, where chips are overclocked day and night to maximize performance, some hardware life cycles are measured in months before replacement is needed.
That raises an obvious question. Why lease access to something that is both rapidly depreciating and physically worn out?
I believe the answer reveals something deeper than a commercial deal between two tech giants. It reveals the emerging economic structure of the data economy, where the good purchased is no longer hardware or software, but something more abstract and powerful.
It is the ability to produce intelligence.
What exactly do you lease?
At first glance, Meta appears to be leasing Google’s AI chips. In reality, what Meta buys is something completely different. Meta leases AI computing capacity.
In other words, the company doesn’t buy any hardware at all. It buys guaranteed access to massive amounts of computing output over time. The chips themselves are simply the specialized machines that produce that output. This distinction may sound subtle, but economically it represents a fundamental shift.
For most of computer history, organizations owned their machines. Servers, storage systems and networking equipment were located in corporate data centers and, like any other capital asset, were depreciated over many years.
Cloud computing changed that model by separating the ownership and maintenance of infrastructure from the consumption of computers. Instead of buying servers, companies rented parts of a massive shared infrastructure from providers like Amazon, Microsoft and Google. The cloud provided the infrastructure on which companies could run software.
Artificial intelligence pushes that model one step further. Training modern AI systems requires such enormous amounts of specialized computing power that the underlying hardware is increasingly abstracted away. What customers ultimately care about is not the chip in the rack, but the amount of computing output delivered over time. More specifically, AI systems are not an infrastructure to run software, but an infrastructure to deliver computational results. This is a big difference from cloud computing and part of the reason why there are so many concerns about an upcoming SaaSpocalypse.
Why the model works for Meta
For Meta, leasing intelligence capacity instead of buying hardware solves several problems at once.
First, it reduces the risk of technological obsolescence. AI accelerators are evolving quickly, and a cluster purchased today may look inefficient in a few years. By leasing AI computing capacity, Meta can avoid tying up billions of dollars in hardware that could soon lag behind the next generation of chips.
Second, leasing helps diversify supply in a market where chip availability remains limited. Meta is already investing heavily in Nvidia GPUs, exploring AMD alternatives and developing its own custom silicon. Adding Google’s TPU infrastructure creates a new source of computing capacity, freeing the company’s AI development pipeline from a single hardware vendor.
Third, the model provides flexibility as AI platforms evolve. Different hardware architectures require different software ecosystems, and the dominant training model frameworks continue to change. By leasing rather than owning AI compute, Meta retains the freedom to adapt its infrastructure strategy as the technology landscape changes.
In short, leasing allows Meta to focus on what matters most: building AI models, while outsourcing the complexities of hardware lifecycle management.
Why the model works even better for Google
If the model helps Meta reduce risk, it helps Google capture something even more valuable.
Control over the infrastructure.
For years, Google designed its Tensor Processing Units, or TPUs, primarily for internal use. These chips powered the company’s proprietary machine learning workloads in search, advertising, and recommendation systems. But with artificial intelligence accelerating across the industry, Google recognizes that TPUs could become the foundation of a much larger business.
Instead of selling hardware like traditional semiconductor companies do, Google could sell access to computer outputs. This strategy moves the company up the value chain. Rather than extracting margin from a single hardware sale, Google can generate revenue continuously as long as customers run workloads on the infrastructure. The company manages the data centers, the network fabric, the software stack and the chips themselves.
By owning the entire stack, Google can operate something closer to a computer utility than a traditional technology product.
And utilities are generally very good companies. The utility market that AWS conquered in the cloud era may be where Google has an early lead in the early data economy.
History repeats itself
There is a historical pattern here that reflects previous industrial transitions. During the industrial age, factories required enormous amounts of power. In many cases, their demand was so high that they built their own power plants rather than being completely dependent on outside suppliers. Steel mills, aluminum smelters, and chemical plants often generated electricity on site because the scale of their operations justified it.
Over time, centralized utilities emerged that could more efficiently deliver power across entire regions.
A similar dynamic appears to be unfolding in the AI economy.
Major tech companies like Google, Microsoft and Amazon require extraordinary amounts of computing power to power their own platforms. To meet that demand, they’ve built massive infrastructure systems consisting of specialized data centers filled with custom silicon, optimized networks, and dedicated cooling and power systems. Once these internal “computer factories” exist, it makes economic sense to sell excess capacity to others.
This is exactly how utilities have historically developed: high internal demand created the infrastructure, and the infrastructure eventually expanded into a broader market.
From factories to the cloud to intelligence
Seen through that lens, the Google-Meta deal represents another step in a longer economic evolution. Over the past century, the dominant infrastructure platforms of each era have shifted, along with the resources they produce.
Economic era → Core infrastructure → Output
Industrial Age → Factories → Physical Goods
Information Age → Cloud Platforms → Software
Data economy → AI infrastructure → Intelligence
Factories produced the goods that fueled industrial growth. Cloud computing provided the software that defined the Internet economy. Now massive AI infrastructure systems are emerging to generate something new: machine intelligence itself.
The companies that operate these systems may ultimately function much like the utilities that powered previous economic revolutions. They will provide the infrastructure that supports much of the economy.
The financialization of intelligence
Once intelligence is recognized as a utility, it will likely attract the same financial structures that support other forms of infrastructure. Power plants, pipelines and telecommunications networks are often financed through long-term investment vehicles designed to generate stable revenue streams. Investors finance the construction of physical assets, operators manage them, and customers purchase the output over time.
The AI infrastructure is starting to resemble that model. Investors finance massive data centers, operators deploy specialized chips and networks, and customers sign contracts to purchase computing capacity for training and model running. The underlying value is no longer just hardware. It is the ability to produce intelligence on a large scale.
How will the semiconductor industry respond?
This shift also raises an interesting question for traditional semiconductor companies. If the future of AI computing is increasingly determined by vertically integrated infrastructure providers that design their own chips and sell compute as a service, the economics of the hardware sector will undoubtedly change.
Companies like Nvidia and AMD have dominated the AI hardware market by selling increasingly powerful accelerators to customers in the technology sector. China is catching up, but Huawei, Cambricon and Biren Technology are quickly catching up, at least in terms of technological capabilities, if not yet in terms of sheer market share.
As hyperscalers build their own silicon and operate large computing facilities, the balance of power may gradually shift toward those infrastructure platforms. There are currently three layers in the stack.
Low Example Companies Role
Semiconductor suppliers NVIDIA, AMD, Intel Design AI chips
Infrastructure providers Google, Microsoft, Amazon Build and operate AI computing tools
AI developers OpenAI, Anthropic, Meta Build on top of infrastructure
Not entirely surprising, Nvidia has recently started investing heavily in software ecosystems, AI frameworks, and even model development. Moving up the stack could become essential as the industry evolves (and may warrant a follow-up article in this column in the coming weeks).
Connecting to the intelligence network
For now, the Google-Meta deal offers a glimpse of what the next infrastructure layer of the digital economy could look like. Companies no longer just buy chips. They connect to the larger global network of computing power capable of producing intelligence on demand.
In the data economy, intelligence itself could become the most valuable utility of all.
