Where the Next Trillion Dollars in AI Will Be Made

>

New technologies always start by obsessing over what is rarest. Eventually, that resource becomes common, and we stop thinking about it. We saw this happen as we moved from steam to steel, then to internet speed, and finally to the cloud.

What used to be precious simply becomes the foundation for new things. As people build on top of it, the real value shifts away from the base and toward the new applications being created. AI has finally reached this stage, and we are watching it happen right now.

Roughly a trillion dollars a year is going into AI infrastructure on a single conviction: that whoever owns the compute owns the future. It is a defensible bet, and it has built an extraordinary engine. It has also, so far, absorbed far more capital than it has returned in value. The spending and the payoff have begun to drift apart, and the distance between them is the most interesting question in the industry.

The question was never whether AI is real. Silicon and electricity settled that. The question worth answering is which direction the value travels from here. 

The technology stack exhibits a geological progression: value does not remain at the foundation but rises. It drifts toward the workload, where AI evolves from a raw capability into an integrated system that runs the business.

The First Trillion Came Out of the Ground

Begin at the bottom, because that is where the money started, and it did not start with a model. It started with electricity, and there is not enough of it.

S&P's 451 Research projects US data center demand climbing to roughly 76 gigawatts in 2026, more than doubling by 2030, and the grid cannot keep pace. Goldman Sachs has flagged an 11-gigawatt shortfall today. 

Morgan Stanley separately projects a 44-gigawatt US power deficit by 2028. Microsoft has publicly described an $80 billion backlog of Azure demand it cannot serve, gated not by buyers but by power that has been sold and cannot yet be switched on. 

CEO Satya Nadella has said openly that GPUs are sitting idle in inventory because there is not enough electricity to run them. The constraint on the most capital-rich industry in history turns out to be the speed at which a utility can energize a substation.

That is why the spending at the base is difficult to fathom. The five largest hyperscalers are guiding to roughly $725 billion of capital expenditure in 2026, nearly double the prior year, with about three quarters aimed at AI infrastructure. 

The Stargate program alone is a $500 billion, ten-gigawatt undertaking. American investor-owned utilities are planning around $1.4 trillion of capital expenditure through 2030 to serve the new load. The true bottleneck of the AI economy turned out to be civil engineering.

>

One layer up sits the company that converted all of that into the cleanest fortune of the cycle. NVIDIA closed its 2026 fiscal year at $215.9 billion in revenue, up 65% year-on year. Its data center segment alone booking $75.2 billion in the quarter ending April, up 92% in a year. 

In the first act of any platform shift, the house belongs to whoever sells the tools. For the first trillion, that verdict is already in.

The Second Trillion Is Already Moving

The companies on the steepest trajectories today are no longer selling chips. They are selling the place where enterprise data lives and enterprise work gets executed.

>

Databricks crossed a $6.9 billion annual run rate in June 2026, growing more than 80%, and is in talks to raise at a valuation of $165 to $175 billion, roughly six months after a round that priced it at $134 billion. The company has become the environment where enterprises put the data their agents will act on, and the market is paying accordingly. 

Palantir sits at the other pole of the same argument, at a market capitalization of around $300 billion on revenue that grew 85% year over year in Q1 2026, with US commercial revenue up 133%. 

Jensen Huang has publicly described ServiceNow, which crossed $95 billion in market value on a similar thesis, as "destined to be the operating system for enterprise AI agents." Neither company is a foundation lab. Neither sells compute. Both are being priced as the layer where models become systems that run a business.

CoreWeave tells the same story from a different direction. It grew from a specialist GPU provider into a contracted revenue backlog approaching $100 billion in a single year, on multi-year take-or-pay contracts from the largest AI labs. 

It did not win by being a cheaper hyperscaler. It won by being built for one workload and refusing to be general. Two layers, one lesson: capture rises as it nears the work.

Why Value Climbs

In every prior computing cycle, the scarce resource of the early years became the commodity of the later ones. Bandwidth, storage, servers, raw cloud compute. Each was once the bottleneck, each was eventually abstracted into the background, and in each case the durable economics migrated to whoever sat closest to the human's work.

AI is running the same script at higher speed. In the first quarter of 2026, roughly $242 billion flowed to AI companies, and about $188 billion of it went to just four of them. But concentration of capital is not the same as capture of value. 

At the top of the stack, a line is hardening between the durable and the disposable. On one side sit companies with proprietary data, embedded workflows, and distribution a frontier lab cannot reproduce. On the other sit the products that are, when described plainly, a thin coat of paint over someone else's model.

The clearest sign is in the software itself. Horizontal wrapper apps have watched their revenue collapse as the labs shipped the same capabilities natively. Vertical, domain-specific software has held up and, in categories where regulation and proprietary data create real switching costs, has accelerated. 

Vertical AI funding tripled to roughly $3.5 billion in 2025, with healthcare alone taking close to half. Veeva holds around 80% of its market and is embedded across clinical, regulatory, quality, and commercial workflows that a horizontal platform cannot casually replicate. Generic capability is being absorbed by the very labs that supply it. Specificity is what survives the encounter.

The Deployment Layer Is Where the Next Trillion Is Realized

Which leaves the layer the market is slowest to price, because it photographs poorly and demonstrates worse. A model that exists is worth nothing to an enterprise until it runs reliably inside the business. 

The space between those two states is the deployment layer. It is operational, unglamorous, and faintly boring. It is also where the next trillion is genuinely realized, because it is the only place where value crosses from a lab's balance sheet onto a customer's.

The evidence is quiet but consistent. Research from Forrester, Anaconda, and IDC has found that roughly 88% of enterprise AI agent pilots never reach production. The bottleneck is no longer capability. It is deployment. 

Everything downstream of that number, including governance, orchestration, cost per successful run, and the plumbing that turns a working demo into a system a Fortune 500 will actually run, is where the compounding happens.

This is the layer Lyzr is built for. Where the industry average puts fewer than a third of enterprise AI pilots into production, roughly 85% of the projects Lyzr runs reach it, most of them in the industries that punish shortcuts hardest. 

Willis Towers Watson's intelligent retirement advisor has been running in production for over a year, fully governed and compliant, and has pulled customers back from ChatGPT into an advisor WTW itself owns. The moat is not the model. It is the compliance posture, the regulated data, and the fact that the agent is already there when the customer opens the app.

The unit economics move the way the thesis predicts. The cost of putting one dollar of recurring revenue into production has fallen from $0.45 to $0.19 over the past year, cohort after cohort. That descending curve, not any one feature, is the real moat. 

The most revealing detail is what customers talk about once an agent is live, which is almost never the model. A contact center that took 16% off its average handle time. A sales team that recovered thousands of hours a year from manual research. 

An HR agent resolving the majority of employee questions with no person in the loop. The conversation is about hours returned, cycles compressed, and headcount that never had to be added.

The companies that minted the first trillion built the engine. The next belongs to those who get it installed and running where the work actually lives. It is the least glamorous layer in the stack, and steadily the most valuable.

*This content is produced in association with Lyzr. Sources include Lyzr's deployment research, S&P Global 451 Research, Goldman Sachs, Morgan Stanley, Forrester, Anaconda, and IDC.

>