We may be asking the wrong question about AI spending

The scale of the artificial intelligence investment boom is becoming difficult to comprehend. Microsoft, Amazon, Alphabet and Meta are building data centres, securing power, ordering chips and expanding networks at a pace rarely seen in corporate history. Long-term estimates now run into the trillions of dollars. It is reasonable to ask whether the world will ever spend enough on AI to justify it.

Early in the AI cycle, the clearest opportunity appeared to be in the picks and shovels: the companies supplying the chips, networking equipment and manufacturing capacity needed to build the new infrastructure. 

Demand was visible, supply was constrained, and the beneficiaries were relatively easy to identify. However, things have since shifted.

The economics of foundation labs remain less clear. Companies such as OpenAI and Anthropic are building extraordinary technology and have demonstrated extraordinary technical progress. What they have not yet demonstrated is what the mature economics of that business ultimately look like.

I feel differently about the hyperscalers. Companies like Microsoft, Amazon, and Alphabet are not merely financing a new AI industry. They are deploying a new production input across some of the world's largest and most profitable existing businesses. These companies already serve billions of consumers and hundreds of millions of businesses globally. That distinction changes how we should think about the economics of AI.

Not all AI spending has the same economics

Much of the debate treats AI capital expenditure as a single pool of investment with a single required source of return. This is a mistake.  

The foundation labs, chip suppliers and hyperscalers are making fundamentally different economic bets.

A foundation lab primarily needs to earn a return by selling access to intelligence. Its economics depend on model usage, pricing, differentiation and the cost of training and running inference. A chip company earns its return by selling the wafers and racks used to build the infrastructure. A hyperscaler, on the other hand, has many more options.

Microsoft can sell AI services through Azure, but it can also use the same infrastructure to improve Microsoft 365, GitHub, Security, Dynamics and Fabric. It can deploy AI internally to write software faster, automate support, improve cybersecurity and raise employee productivity. The result is not just new revenue, but a more efficient and valuable business. 

Amazon can sell AI capacity through AWS while also using it across retail, advertising, logistics, customer service and internal software development. Alphabet can deploy it through Cloud while improving Search, YouTube, advertising and its own operations. Rather than selling cloud infrastructure or enterprise software, Meta uses AI to improve content recommendations and engagement tools for its 3+ billion users across Facebook, Instagram, WhatsApp, and Threads. The infrastructure may be the same. The paths to monetisation are not. 

AI is reinforcing the cloud

Amazon’s latest quarter offered one of the clearest explanations yet of how these economics will work. AWS is no longer describing AI and traditional cloud computing as separate businesses. Management said growth in AI was helping drive growth in the core AWS business because AI workloads require much more than specialised accelerators. 

They also consume CPUs, storage, databases, networking and security services. As customers build AI applications, they tend to increase their use of the rest of the cloud platform as well.

Amazon’s chief financial officer made the same point more directly: as customers invest in AI, AWS sees a corresponding increase in core consumption. The relationship should strengthen as more AI workloads move from experimentation into full-scale production. This is important because it suggests AI is not cannibalising the cloud. It is increasing cloud intensity. Every agent needs somewhere to run. It needs access to data, memory, identity, governance, security and tools. 

It may require specialised chips for some tasks and conventional computing for others. As a result, AI revenue and core cloud revenue become increasingly intertwined. The return on AI infrastructure therefore does not need to come only from the direct sale of AI services. It can also appear as higher demand for the existing platform.

The ecosystem is the economic unit

Much of the debate about AI spending assumes that intelligence is the product being sold. I think that's the wrong way to think about it. Intelligence is increasingly becoming an input. Nobody asks how much revenue Microsoft’s electricity consumption generates. The reason is obvious. Electricity isn't the product. It's an input into producing software, running data centres and serving customers. 

What much of the debate around AI spending misses is that intelligence may be evolving in much the same way. We continue to ask whether companies earn an adequate return by selling AI itself. But the larger opportunity may be using AI to make existing businesses more valuable. 

We've already seen this dynamic play out at Meta. AI didn't create an entirely new business. It made Meta's existing advertising engine substantially more valuable by improving recommendations, engagement and ad targeting.

Microsoft illustrates the same dynamic in a different way. Let's say an employee of a large retailer wants to understand why sales have slowed in Victoria. Rather than opening Outlook, Teams, Dynamics, SharePoint and half a dozen dashboards, they simply ask Copilot.

Within seconds, Copilot retrieves customer records from Dynamics, analyses financial data stored in Fabric, searches for documents in SharePoint, reviews recent Teams meetings and Outlook emails, checks permissions in Entra, and applies governance policies in Purview. Azure provides the compute that orchestrates the entire interaction.

To the employee, it feels as if they are being asked a single question. Behind the scenes, however, that interaction has increased the value of almost every part of Microsoft's ecosystem. The customer stores more data in Fabric, relies more heavily on Azure, embeds Microsoft's security and governance tools more deeply into its organisation and becomes increasingly dependent on the workflows that connect them all.

The spending is large, but it is not all irrevocable

The trillions of dollars discussed in forecasts can sound like a binding commitment. They are not. 

One of the more interesting insights from Amazon's latest earnings call was management's description of the capital cycle. Data centres require years to plan and can remain productive for decades. Servers and networking equipment, however, have much shorter lifespans and are typically purchased only a few months before deployment. This distinction matters.

It means Amazon has considerable visibility into customer demand before committing a significant proportion of its AI infrastructure spending. Jassy's message was simple: if the demand isn't there, Amazon won't spend the capital.

That doesn't eliminate the risk of overinvestment. Power still needs to be secured, data centres must be planned years in advance, and some capacity will inevitably prove unnecessary. But investors shouldn't treat every long-term capital expenditure forecast as money that has already left the building. The spending is, to a meaningful extent, contingent on demand.

Ironically, if AI adoption ultimately proves slower than expected, disciplined capital allocation may become a source of strength for the hyperscalers rather than a weakness.

The model may be less important than the platform

Amazon’s discussion of foundation models also reinforced something I've been thinking about for some time: the value is likely to migrate up the stack. Jassy argued that there will not be one model that rules the world. Instead, customers will use a mix of proprietary and open models, with capabilities continually leapfrogging one another.  

That is an important observation. If model capabilities converge, the model itself may become less important than the environment in which it is deployed. 

Enterprises still need identity, permissions, security, data access, monitoring, workflow integration and accountability. 

They need the AI to work seamlessly with systems they already use and within the controls they have already established. Those requirements don't disappear as models improve. If anything, they become more important. That is why I increasingly think the platform may prove more valuable than the model itself.

The companies that own the customer relationship, workflow, data, and distribution are likely to capture a disproportionate share of the economic value, regardless of which frontier model leads the benchmarks at any given time. That is one reason I am more constructive on the hyperscalers and enterprise application companies than on the businesses selling the original picks and shovels.

A different way to think about the return

None of this proves that the current AI investment boom will earn attractive returns. The possibility of overinvestment is real. Foundation labs may consume enormous amounts of capital without building durable economics. Enterprise adoption may take longer than expected. Prices may fall faster than costs. Some data centres will almost certainly be built in the wrong places or at the wrong time.

Those are real risks, and they remain one of the most important areas for investors to watch. But treating all AI spending as a single homogeneous bet obscures more than it reveals. The foundation labs must prove that intelligence itself can be sold profitably.

The hyperscalers have a different opportunity. They can sell AI infrastructure, deploy it internally, strengthen existing platforms and build entirely new applications on top. AI can generate new revenue while simultaneously making some of the world's largest businesses more valuable.

That may prove to be the most important distinction of all. We continue to debate whether AI will become a multitrillion-dollar industry. Perhaps the bigger opportunity is that AI becomes an indispensable input for existing industries.