Thesis

The demand for intelligence has no natural ceiling.
Intelligence begets intelligence. The practical limit is not a fixed market size, but the creativity of humans and AI systems applying it.

Directional begins with one observable direction: AI capability, usage, model-lab revenue, token demand, and compute infrastructure are all bending from gradual to near-vertical. The firm invests in the second-, third-, and fourth-order effects of that acceleration.

If you think AI is overhyped, you do not have to take it from us. Read what the people building and underwriting it are saying.

One direction

The same line keeps showing up.

Model capability, time horizon, user adoption, and frontier-lab revenue are not the same business metric. They should not naturally produce the same chart. Right now they do. That is what Directional means: not a precise endpoint forecast, but a compounding direction visible across the AI stack.

Model capability
Task horizon
AI users
Lab revenue

Every curve is exponential. The direction shows up everywhere.

The mechanism

The flywheel keeps the curves coupled.

Better models create more use cases. More use cases create more usage. More usage underwrites more hardware. Better hardware enables better models. The flywheel is what turns separate hockey sticks into one reinforcing system.

Bigger and better models

Capability · Context · Modality

More use cases

Coding · Agents · Research

More usage

Consumer · Developer · Enterprise

More AI cloud revenue

Backlog · ARR · Tokens

More AI capex

GPUs · Datacenters · Power

Bigger and better compute

Advanced packaging · EUV · Optics

What the loop forces

Hardware must scale in quality and quantity.

Capability Better hardware

Bandwidth, density, cooling, interconnect, and cluster scale.

Capacity More hardware

Memory, accelerators, racks, power, data centers, and equipment.

Investment work Find the narrow gates

Where forced change meets scarce capacity or technical control.

Value migration

We are early because the stack is still base-heavy.

Hardware came first because the stack had to be built: chips, memory, networking, data centers, power, cooling.

A developed AI market inverts. Revenue migrates up to the applications that monetize intelligence directly. That migration is beginning, and it is rapid: Anthropic has roughly 10× its revenue every year since 2023, with year-end 2026 run-rate estimated near $100 billion.

  • Hardware. Training clusters, chips, memory, fabs, power, and cooling. Last year's gains were here.
  • Infrastructure. The clouds and data centers that turn trained models into inference capacity. Google Cloud grew 63% in Q1 2026.
  • Applications. The products built on top. Anthropic scaled from roughly $9 billion at the end of 2025 to roughly $44 billion as of April 2026.
Current AI stack Capability buildout Applications Infrastructure Hardware

Most visible value so far came from making frontier models possible.

Developed AI market Intelligence monetized Applications Infrastructure Hardware

As intelligence is used, value migrates toward products and workflows.

Directional filters

Three filters. We don't need to stretch beyond them.

Atoms over bits

AI makes software abundant. Atoms stay scarce. We default to the physical.

Narrow gates

Some companies hold decades of accumulated knowledge that no competitor can recreate. ASML is the canonical case.

Pace of the curve

AI compounds at multiples per year. The use cases keep widening. We are not settling for a 12-15% IRR.

Scale check

We are early.

Frontier AI lab revenue should reach roughly $200 billion annualized by year-end 2026. That is just over 1% of US white-collar compensation, under 1% of global white-collar compensation, and a fraction of a percent of global GDP.

The larger question is not whether AI revenue can be bigger; it is what ratio of economic activity intelligence ultimately captures. White-collar productivity is only the first denominator. Drug discovery, materials science, iterative model improvement, research, and other invention-heavy domains expand the pie rather than merely lowering the cost of existing work. The bars above should be read as distance from the beginning of infinity, not as a ceiling.