Higher Ground

Read the full article (PDF)open_in_new

Where the money goes in the AI economy, and where it is heading

“Bustin’ out, and I’ll break you out, ’cause I’m sailin’ on,
Just, uh, sailin’ on, sailin’ on to higher ground.”
– Red Hot Chili Peppers

Executive summary

In our first two papers, we looked at how artificial intelligence (AI) is reshaping work and why the productivity payoff is taking time to show up in the data. This third paper follows the money.

If AI really is the next great technology wave, the key question for investors is simple: Who will actually make the money? As with any chain, some links look strong today, but not all of them will stay that way.

The AI economy can be viewed as a five-layer stack.

At the bottom sit chips and hardware. Then comes cloud infrastructure. In the middle are foundation models. Above that is orchestration and middleware. At the top are the applications customers actually use and pay for. Right now, most of the profits are concentrated near the bottom. NVIDIA generated more than US $215 billion in revenue in fiscal year 2026 and earned gross margins above 70 %. At the same time, according to data available in early July, the four largest hyperscalers are on track for about US $725 billion in capital spending in 2026, roughly 77 % more than in 2025.

But history suggests that these early winners do not always keep the crown.

In past technology cycles, from mainframes to personal computers to the internet, mobile, and cloud, the richest profit pools eventually moved closer to the customer.

That is already starting to happen in AI. Open-source models have improved quickly, narrowing the quality gap. And some of the biggest players in the industry are helping that process along.

Our view is that the next big value pool is likely to emerge in orchestration, the layer that makes AI useful inside real business workflows, and in what Satya Nadella, CEO of Microsoft, recently called "token capital," the proprietary AI capability firms build for themselves. In fact, this has already started, and the trend will accelerate in the coming months.

One key implication is that today’s extraordinary margins at the bottom of the stack should be treated as cyclical, not permanent.

The AI supply chain: Where the money sits today

The AI stack has five layers.

At the base is the chips and hardware layer. That includes NVIDIA GPUs, AMD accelerators, and custom chips built by the hyperscalers themselves, such as Google’s TPU, Amazon’s Trainium, and Microsoft’s Maia. It also includes the less glamorous but essential supporting pieces: memory, networking, power, and cooling.

Above that sits cloud infrastructure, led by AWS, Microsoft Azure, and Google Cloud, with newer players such as CoreWeave and Oracle trying to carve out a niche.

In the middle are the model developers, such as OpenAI, Anthropic, Google DeepMind, Meta, Mistral, DeepSeek, and xAI.

Above them is orchestration and middleware, which includes the tools that connect models to data, route tasks, evaluate outputs, and automate workflows.

At the top is the application layer, the software that embeds AI into a workflow that a customer is willing to pay for.

The concentration of profits at the bottom of the stack is hard to miss. NVIDIA’s data centre revenue reached US $75.2 billion in the first quarter of fiscal year 2027 alone, with gross margins of 75 %.

Such economics are rare at any scale. TrendForce expects global cloud provider capital spending to hit a record US $830 billion in 2026, with roughly three quarters going to physical infrastructure and about one quarter to chips.

Meanwhile, the economics look much weaker further up the stack. According to the consensus of market participants, OpenAI’s losses could reach roughly US $14 billion in 2026 against annualized revenue of about US $20 billion.

In other words, we are at a juncture where the companies supplying the picks and shovels are earning exceptional returns, while many of the companies closest to the end user are still trying to close the gap between growth and profitability.

Profits sailing on to higher grounds: What history tells us

If the history of computing teaches one lesson, it is this: Over time, profit pools tend to move up the stack, away from raw infrastructure and toward the customer relationship. Though not an iron law, this pattern has been the dominant one in every major platform shift of the last half century.

The mainframe and the PC

In the mainframe era of the 1970s and early 1980s, IBM was the king of the hill. Gross margins on large systems could reach 70 %, and mainframes represented roughly 42% of revenue but about 60% of profits. IBM controlled about 60% to 70% of the global market. Then came the personal computer.

In an effort to move quickly, IBM opened the PC architecture and outsourced key parts of the value chain, namely the processor and the operating system.

That decision changed where the money went. Microsoft supplied MS-DOS, and later Windows, the operating system layer that sat between the machine and the user, ran the software people relied on, and gave developers a common platform to build for. In this sense, Microsoft was closer to the customer than IBM’s hardware division: it controlled the everyday interface, the application ecosystem, and the standard that made one PC feel compatible with another.

According to a classic study of this market by HarvardBusiness School, Intel and Microsoft together earned more than US $15 billion in net profits from PCs by 2004. Meanwhile, Dell, HP, and IBM combined earned only about US $2.5 billion. The lesson is simple: When the hardware becomes interchangeable, profits migrate to the firms that control the standard.

The mobile carriers: Building the road, losing the toll

The smartphone era offers an even more familiar example. Telecom carriers poured hundreds of billions of dollars into 3G, 4G, and 5G networks. AT&T alone planned about US $22 billion in capital investment in 2025, compared to US $17.5 billion to US $18.5 billion for Verizon.

Globally, the mobile industry generates roughly US $1 trillion in annual revenue. Yet the stocks have largely gone nowhere over a quarter century.

Why? Because the real economic surplus from the smartphone revolution was not captured by the network owners, but rather by Apple and Google, the firms that owned the platform and the customer relationship. The carriers built the road, but they did not collect the toll. They became utility pipes. That is a useful cautionary tale for AI investors today.

The cloud layer and the manufacturer squeeze

Cloud computing tells the same story, but from another angle. AWS reported operating margins of roughly 35% in late 2025. By comparison, server manufacturers such as Dell operate with full-year operating margins closer to 6%. The process unfolded gradually.

As companies moved computing workloads from their own data centres to AWS, Azure, and Google Cloud, they stopped buying as many servers directly and instead rented computing power, storage, databases, and software tools from the hyperscalers. Those cloud providers, in turn, bought servers in enormous volumes, designed more of their own infrastructure, standardized the hardware, and pushed suppliers into a lower-margin assembly role. In practical terms, the customer relationship moved from the manufacturer selling boxes to the cloud platform selling flexible capacity and higher-value services.

Then, on top of the cloud, software companies built applications such as Salesforce, Snowflake, Datadog, and ServiceNow that captured value the cloud providers helped create but did not fully own. At each step, the same principle shows up again: The link closest to the customer often ends up with the best economics.

Commoditize your complement

A useful way to frame this comes from an old technology rule of thumb: If you can make the complementary product cheap, demand for your core product rises. In plain English, if you make someone else’s layer cheaper, your own layer becomes more valuable. That is exactly what we are seeing in AI.

NVIDIA released open Nemotron models because the goal is not to dominate the model business, but rather to keep demand for GPUs strong. Meta open sources Llama because it does not want to pay large rents to outside model providers that could squeeze its advertising margins. Google promotes Gemma for a similar reason: It is trying to make the model layer less scarce and less expensive, just as Android helped weaken the economic power of the mobile operating system.

None of these companies makes its real money from selling models. They make money from chips, cloud capacity, advertising, or software licenses. That matters. When the deepest pocketed players want the model itself to become abundant and inexpensive, pure-play model labs face a difficult problem. Their product is the very thing everyone else wants to turn into a commodity.

The model layer : commoditization in real time

If profit pools move when lower layers get commoditized, then AI is already moving faster than many expected. The gap between the best closed-weights model and the best open-weights alternative¹ is not very large and has not widened in recent years, according to benchmarks to measure model quality. Even if models like OpenAI and Anthropic continue to retain an edge, especially for the most complicated tasks, open-source models continue to improve at a similar pace. The practical implication is straightforward: for many real world uses, the best open models are now good enough.

Quality is important, but so are costs. And here, the trend is striking. The cost gap is no longer marginal.

For many enterprise workloads, open-weight models such as DeepSeek can be ten to fifty times cheaper than premium proprietary models from OpenAI or Anthropic. This changes the economic decision. If the proprietary model is only modestly better, the burden of proof shifts, and the quality premium must justify a very large cost premium.

If that is true, then distribution matters more than raw model quality. And distribution belongs, in large part, to the hyperscalers. One of the defining corporate shifts of early 2026 was the end of Microsoft’s exclusive distribution rights for OpenAI’s models on April 27.

The timing matters because Anthropic, which made Claude available across all three major clouds from the start, was able to gain traction precisely because customers could adopt it wherever they already operated. The broader lesson is familiar: the labs are increasingly becoming suppliers, while the hyperscalers are becoming the channels.

In a fight between the very best product and the most convenient distribution, distribution often wins. Netscape learned that the hard way in the browser wars of the late 1990s. Although its browser helped define the early web, Microsoft bundled Internet Explorer directly into Windows, putting it in front of users by default.

Even if many users preferred Netscape, convenience and pre-installation mattered more than marginal product quality, and the browser market quickly shifted toward the platform that already controlled the customer relationship.

That leaves model labs with a structural risk that looks uncomfortably familiar. They could end up like the mobile carriers: strategically important, capital intensive, and yet unable to capture the bulk of the surplus because they do not own the customer. It is not the only possible outcome.

A lab could still build direct distribution, switching costs, and a durable lead. But history suggests that distributors usually have the edge.

The orchestration layer : the new bottleneck

If the model itself is becoming a more interchangeable input, then the strategic high ground moves to the layer that determines how models are actually used. That is the orchestration and middleware layer. This is where firms connect models to internal data, set guardrails, route tasks, evaluate outputs, monitor quality, and make the whole system reliable enough for business use. It may not be glamorous, but this turns AI from a demo to an operating system for work.

A good way to think about it is through the factory analogy from our previous paper titled Learn to Fly. Replacing a steam engine with an electric motor did not unlock the full productivity benefit until factories were redesigned around the new technology. AI is similar, with the redesign happening in the orchestration layer.

It is where the company-specific logic sits: the routing rules, the retrieval systems, the validation steps, and the escalation paths that reflect how your business actually works. It also allows firms to create the flexibility to switch model vendors without losing the capability they have built. That is the practical expression of co-invention, the hidden investment cycle that must happen before the productivity payoff shows up in the numbers.

This layer is maturing quickly. The Model Context Protocol, launched by Anthropic in late 2024 as an open standard for connecting AI systems to outside data, reportedly gained support from OpenAI, Google, Microsoft, and AWS within a year, with governance later moving to the Linux Foundation. LangChain, a leading open-source orchestration framework, reported 90 million monthly downloads and adoption by 35% of the Fortune 500 at the time of its October 2025 Series B funding round. Its State of Agent Engineering survey found that 57% of respondents already had agents in production, that using multiple models at once had become normal, and that the main barrier was not cost or raw model power, but quality assurance.

That last point is crucial.

The issue is no longer whether a model can technically perform a task, but whether the surrounding system is reliable enough to trust it.

In fact, Gartner projects that more than 40% of agentic AI projects will be cancelled by 2027, mainly because of weak evaluation, weak governance, and friction at the integration level. In other words, the bottleneck is increasingly not the brain, but the plumbing. Firms that get this layer right are going beyond using AI, building an advantage that can endure.

Conclusion

The AI chain runs from silicon to cloud to model to workflow to customer. Today, the strongest links in margin terms are near the bottom. But history suggests that this strength will not stay there forever. Over time, value is likely to move upward.

Four markers are worth watching.

1. Model commoditization

If open source keeps closing the quality gap, pricing power at the model layer will weaken and value will continue to migrate upward.

2. Distribution

If the labs fail to build direct user relationships and meaningful switching costs, the hyperscalers may reduce them to utility suppliers.

3. Orchestration maturity

When the main question changes from “Which model should we use?” to “How do we build reliable systems around models?”, it signals that the value pool has moved to middleware.

4. Intangible accumulation

The firms that can switch the generalist model under the hood without losing their institutional AI capability are the ones turning AI spending into an asset.

A greater risk than AI failing is that investors misread where the durable value really sits and mistake today’s cash flows for tomorrow’s moat. The margins at the bottom of the stack are real. But the lessons of the mainframe, the mobile carrier, and the server manufacturer are clear: infrastructure margins tend to compress once the complementary layer becomes cheap enough.

The chain may hold together, but the value does not stay still.