The commoditization of AI models and compute

By Tyler Frawley, CFA

It seems to be the prevailing view that the companies building the models and owning the compute will ultimately capture most of the value. OpenAI, Anthropic, xAI and others are broadly viewed as the ultimate beneficiaries. Meanwhile, hyperscalers are committing hundreds of billions of dollars to build the infrastructure required to support them. We think the market has largely accepted the idea that the owners of the models and compute will be the primary beneficiaries of the AI revolution.

This is an assumption that deserves a closer examination.

AI increasingly looks, to us, like a technology that will create tremendous value for society, but what if the underlying economics evolve differently than many investors currently expect? It is possible the underlying economics are pointing toward a future where both models and compute become increasingly commoditized, and where the real successes end up being the businesses that successfully implement AI rather than the companies trying to sell it.

The shift to variable costs from fixed costs

One of the most important shifts over the past year has been pricing dynamics. For a long time, companies viewed AI much like traditional enterprise software. The company paid a fixed monthly fee, employees gained access and costs were relatively predictable. That is changing quickly as pricing moves toward usage-based models.

Today, most AI services are priced on consumption. Every prompt, workflow, agent and inference consumes tokens. At that point, AI stops looking like software with fixed costs and starts behaving more like a utility where usage directly drives expense.

We are already beginning to see the implications. Uber has been one of the clearest examples, with technology news site The Information suggesting AI usage scaled quickly enough to consume what had been expected to be a full-year budget in just a few months. What appeared manageable during experimentation became a much larger operating expense once adoption scaled across the organization.

Most companies are still early in their adoption curves, but history suggests they will eventually respond the same way they respond to any rapidly rising cost structure. They will look for efficiencies, impose controls, build out more infrastructure themselves and begin searching for lower-cost alternatives to better control the economics.

Controlling the economics

As AI becomes a core business capability, enterprises gain incentive to control more of the economics. Some may choose to build infrastructure themselves. Others may increasingly optimize around lower-cost models as capabilities converge.

A recent unconfirmed report from Axios that Microsoft has explored using lower-cost DeepSeek models to support portions of Copilot illustrates this trend. Whether DeepSeek ultimately becomes material to Microsoft’s long-term strategy is less important than the fact that cost is now being actively optimized alongside capability.

Coinbase has taken this idea a step further. CEO Brian Armstrong has stated on X that the company increasingly routes AI workflows to lower-cost models when they are sufficient for the task, reserving frontier models for only the most demanding use cases. According to Armstrong, that approach has helped Coinbase keep AI spending roughly flat even as token usage has continued to rise.

Taken together, these examples suggest to us that a broader structural shift may be underway. As model capabilities converge, enterprises are likely to optimize for economics as much as performance. The most advanced model will not necessarily be the default choice. Increasingly, it becomes a specialized tool for a narrower subset of high-value tasks. In that environment, model providers may find themselves competing not only on capability, but increasingly on price and cost efficiency.

The challenge is that many model providers are implicitly relying on strong pricing power to justify their long-term economics. The industry has been built on the expectation that near-term losses can eventually be converted into high-margin revenue. That requires sustained willingness from customers to pay premium prices even as alternatives improve and internal capabilities become more sophisticated.

If providers raise prices aggressively, customers push back or migrate. If they compete on price, it becomes a race to the bottom, and elevated profitability becomes difficult to achieve.

When compute looks like a commodity

The same logic applies to hyperscalers aggressively building capacity to sell externally. There is no question that AI demand is real and expanding rapidly. We think the key question is whether supplying compute at scale will remain a structurally attractive business over time.

The economics may begin to resemble those of traditional commodity markets.

Oil, natural gas and electricity are all important. Entire economies depend on them. But importance alone does not guarantee strong economics or durable profitability. When supply expands aggressively and is increasingly interchangeable, competitive forces tend to compress returns over time.

A recent report from Bloomberg that Meta has explored selling excess compute capacity to external customers may also prove instructive. While the move, which has not been confirmed by Meta, could reflect an effort to better monetize existing infrastructure, this also raises the possibility that some hyperscalers may be building ahead of internal demand assumptions. If excess capacity increasingly needs to be sold into the market, competition for workloads intensifies, and pricing power across the industry becomes more contested.

We think every hyperscaler is currently investing aggressively to ensure future demand is met. That level of investment may be entirely rational in a growing market. But if compute becomes abundant, models converge in capability, and enterprises increasingly internalize AI workflows, the industry could begin to shift toward something closer to infrastructure and away from software-like economics.

Infrastructure can be large, essential, and durable, but it rarely supports sustained premium valuation multiples once competition normalizes and capital intensity remains elevated.

Where the value may accrue

This all suggests to us that the real economic value from AI may ultimately accrue elsewhere. Early real-world evidence is already emerging across a range of industries.

In banking, JPMorgan Chase estimates productivity gains among software engineers using internal AI coding tools could unlock over $1 billion in potential organizational value, according to reporting by Reuters. In machinery, Deere and Company stated that its AI-powered See & Spray technology has reduced herbicide usage by nearly 60 percent, materially lowering input costs for customers while improving precision. In Healthcare, HCA Healthcare has announced the deployment of AI across documentation, staffing and revenue cycle management to reduce administrative burden and improve operational efficiency.

The common thread: value accrues not from selling models or renting compute, but from using AI to improve productivity, reduce costs, enhance customer experience and make better decisions. These companies don’t necessarily need the most advanced models or infrastructure – they just need to apply AI where it matters.

Ultimately, AI may resemble electricity more than software. The value isn’t captured primarily by the producer, but by the systems built on top of it.

If that occurs, the companies who capture the lion’s share of the economics may not be the ones selling AI. They may be the companies using it effectively.

---

---