Capability frontier
Better hardware
Models demand more bandwidth, lower latency, higher density, larger coherent clusters, and cooling architectures built for rack-scale compute.
Directional Holdings
The same curve is appearing everywhere.
AI capability, usage, lab revenue, token demand, and compute infrastructure are bending from gradual to near-vertical. Directional Holdings invests in the second-, third-, and fourth-order effects of that acceleration.
Directional Holdings is an investment firm. We invest in public equities across the AI buildout. The portfolio is concentrated, AI-only, and held with a multi-year view.
Model capability
Task horizon
AI users
Lab revenue
Every curve is exponential. The direction shows up everywhere.
The mechanism
The flywheel keeps the curve bending.
Better models unlock more use cases. More use cases drive more usage. More usage forces more and better hardware. The hardware buildout then enables the next generation of models. The shape is the thesis; the flywheel is the mechanism.
AI × Hardware flywheel
01 Better models
02 More use cases
03 More usage
04 More hardware needed
05 Better hardware developed
Adoption reality
We are still so early.
Each dot is one of 8.1 billion people, colored by the most advanced AI interaction they have ever had. Most have had none. Roughly 3.5 million are using it to build something. That is 0.04% of the world.
- 2,100 dots
- 6.8B never used AI
- 393 dots
- 1.3B free chatbot users
- 6 dots
- ~20M paying users
- 1 dot
- ~3.5M building with AI
One dot is everyone building with AI today. About 0.04% of the world.
1 dot = ~3.2M people. World population ≈ 8.1B. Estimates as of Q1 2026.
Time horizon
What AI can do for you doubles every four months.
A task that takes a human two minutes is not worth paying much for. A task that takes a human three hours is worth a lot. METR, an AI evaluation lab, measures the longest task an AI model can complete reliably, meaning it finishes the work correctly about 80% of the time. That length is doubling roughly every four months. Each doubling brings more valuable work within AI's reach.
Conviction
“We're not investing approximately $200 billion in capex in 2026 on a hunch.”
They see demand from two sides at once: as the cloud running frontier models through Bedrock, and as a major equity holder in both Anthropic (~17%, ~$33B) and OpenAI ($50B committed).
The forcing function
The flywheel demands hardware.
The loop does not merely create more software usage. It demands that the physical substrate keep accelerating: faster systems, larger clusters, denser packaging, more memory, better networks, more power, and more capacity than the market expected to build.
Capacity frontier
More hardware
Adoption demands raw volume: more memory, more accelerators, more networking gear, more data center capacity, and more electrical infrastructure.
Directional work
Find the narrow gates
Bottlenecks are not assumed. They are discovered where forced change runs through scarce capacity, irreplaceable equipment, or hard-won process knowledge.
This was the trade. The hardware required to run AI at scale was never built to absorb a step-change in demand. Fab capacity, memory, power, grid interconnect, clean-room space, and data center footprint were all going to run short. The fund bet on those constraints.
Net portfolio return
Jan 2025 to May 2026
Memory, compute, and equipment exposure
Version 2.0
The trade scales beyond hardware.
This trade continues. AI is only just beginning. Version 2.0 brings industry experts, technical advisors, specialized research, and institutional knowledge to invest in what comes next.