Every training-cluster generation is gated by one input at a time. For most of the last two years, that input was leading-edge logic — wafer starts and CoWoS-class packaging slots at the handful of fabs capable of producing frontier accelerators. That constraint is easing. The next one is memory.

Why HBM, not wafers

High-bandwidth memory bandwidth, not compute throughput, is what determines effective utilization on the largest training runs. As model and cluster sizes grow, the ratio of memory bandwidth to compute needed to keep accelerators fed has been rising, not falling — every generation ships with more HBM stacks per package, and every generation sells out earlier.

What we're tracking: HBM4 qualification timelines across the two qualified suppliers, allocation commitments to the largest buyers, and the capital intensity of the advanced packaging step that turns memory dies and logic into a finished package. Each of these moves independently and each can bind before the others.

What it means for positioning

• Memory suppliers with qualified HBM4 capacity capture disproportionate pricing power in 2026.
• Packaging capacity, not wafer capacity, becomes the next headline shortage.
• Accelerator vendors without secured HBM allocation face shipment risk regardless of wafer supply.

We size the compute sleeve of the book around this sequencing — long the capacity owners, sized down on names whose roadmap assumes memory supply that isn't contracted yet.

Ask any hyperscaler what limits their build plan today and the answer is rarely silicon. It's interconnection — the queue of projects waiting for a grid operator to confirm they can be safely connected without destabilizing the network.

A unit-by-unit accounting

We track queued data-centre load against confirmed, dispatchable generation additions across three grid operators with meaningful hyperscale exposure. The gap between requested interconnection capacity and confirmed generation has widened every quarter we've measured it — new load is arriving faster than new dispatchable generation can be permitted, financed, and built.

What we're tracking: interconnection queue depth and study timelines by region, gas turbine order backlogs (now extending years past historical norms), and the utilities and independent power producers positioned to monetize the gap through behind-the-meter generation and long-duration contracts.

What it means for positioning

• Turbine and transformer manufacturers with visible order backlogs are a multi-year long, not a cyclical trade.
• Utilities able to fast-track large-load interconnection capture pricing power disproportionate to their regulated returns elsewhere.
• Compute buildout guidance that doesn't reference secured power is a schedule risk, not a certainty.

Software priced per seat assumes the cost of doing the work scales with headcount. Once inference is cheap enough to do the work instead of assisting a human doing it, that assumption inverts — the vendor's pricing model is now taxing a headcount that no longer needs to grow.

Mapping the inversion

We rank software categories by two variables: how much of the delivered value is currently produced by a human operator inside the tool (versus the software itself), and how sensitive that category's task is to model capability rather than domain-specific tooling. Categories high on both are exposed first.

What we're tracking: categories where the labour-arbitrage margin — the gap between what the vendor charges per seat and the fully-loaded cost of the human work being augmented — is largest, since that gap is what a capability-driven entrant competes away first.

What it means for positioning

• The short leg of the book targets vendors whose per-seat pricing is unusually exposed to this inversion.
• The long leg looks for vendors already re-pricing toward usage or outcome-based models ahead of the inversion.
• This is a multi-year repricing, not a single earnings event — position sizing reflects that horizon.