The Stack: Which Layer of the AI Trade Is Binding Right Now
A dated status board for nine layers of the AI trade — what is binding now, what moved it, and what would flip each row.
Binding right now: L1 Semiconductors and L2 Power. Finished accelerators are gated by advanced packaging and high-bandwidth memory arriving matched, while new halls wait on interconnect and transformer lead times. Cooling and optics tighten with them. The software layer is the only one currently loosening.
As of: 14 September 2026. Reviewed weekly. Rows change only when evidence moves, not on a schedule.
Current state
Status is a constraint view, not a recommendation.
| Layer | Status | Since | What moved it |
|---|---|---|---|
| L1 Semiconductors | Tightening | 12 Sep 2026 | Packaging and memory must match |
| L2 Power & energy | Tightening | 12 Sep 2026 | Interconnect queues delay energisation |
| L3 Cooling & thermal | Tightening | 12 Sep 2026 | Rack density passed air-cooling limits |
| L4 Data-centre REITs | Stable | 12 Sep 2026 | Commencement waits on power |
| L5 Neoclouds | Stable | 12 Sep 2026 | Expansion gated upstream |
| L6 Networking & optical | Tightening | 12 Sep 2026 | Optics per rack rising faster than supply |
| L7 Hyperscalers | Stable | 12 Sep 2026 | Capex guidance has not been cut |
| L8 Servers & hardware | Stable | 12 Sep 2026 | Waits on modules and watts, not assembly |
| L9 Software & applications | Loosening | 12 Sep 2026 | App revenue growing faster than infra cost |
On the dates. This board began tracking on 12 September 2026, so every row carries that date as its starting point. A Since date changes only when that row's status actually changes — it is a status-change date, not a publish date.
What would flip each row
The column that matters most. A constraint with a nearby, plausible flip condition is a very different proposition from one without.
| Layer | Would flip if |
|---|---|
| L1 Semiconductors | A second supplier qualifies on a leading programme |
| L2 Power & energy | Transformer lead times and queue positions both clear |
| L3 Cooling & thermal | Rack density plateaus, or liquid capacity catches up |
| L4 Data-centre REITs | Powered-on megawatts and lease commencements converge |
| L5 Neoclouds | Realised rate per GPU-hour falls while fleets grow |
| L6 Networking & optical | Optical lead times shorten while cluster sizes plateau |
| L7 Hyperscalers | A major buyer guides capex down |
| L8 Servers & hardware | Module supply frees and assembly becomes the queue |
| L9 Software & applications | Inference cost per token stops falling |
Two of these carry extra nuance worth stating in full. L1 also flips if lead
times shorten for buyers outside the top tier, which is the more honest
indicator than any capacity announcement. L9 also flips if infrastructure rent
reaccelerates, which would tighten it from the cost side rather than the
revenue side.
What changed recently
| Date | Layer | Change | Trigger |
|---|---|---|---|
| 12 Sep 2026 | All | Board created | Baseline state recorded across all nine layers |
One entry, because the board is new. Every future row change is logged here with its trigger. If a status changes and no line appears above, the board is not being maintained — judge it on that.
How to read this board
- Find the tightening layer, or the one that just flipped. That is where pricing power currently sits.
- Open its Explained page for the mechanism, the roster and the falsification conditions.
- Read the neighbours. Rent frequently sits one step upstream or downstream of the layer that binds, not on it.
- Check the flip condition before acting on a status.
- Treat every name as exposure, not advice.
The layer that binds is not always the layer that earns. Packaging binds; memory earns. Interconnect binds; transformer and turbine suppliers earn. That gap is the entire premise of this site.
Which link binds first
Money follows the scarce step. When packaging and HBM are full, finished GPUs wait. When interconnect queues stretch, buildings wait on watts. When rack density jumps, liquid cooling stops being optional.
Two gates are worth reading together rather than separately, because they sit in different countries and bind at different times:
- HBM and the memory gate — concentrated in Korea
- CoWoS and the packaging gate — concentrated in Taiwan
A slot without stacks is idle. Stacks without a slot are inventory. Either failure looks identical from the outside — accelerators are late — which is why attributing a delay to the wrong gate is the most common reading error on L1.
The nine layers
| Code | Layer | Function | Open |
|---|---|---|---|
| L1 | Semiconductors | Design, foundry, packaging, HBM, equipment | Layer page |
| L2 | Power & energy | Generation, grid gear, interconnect, on-site watts | Layer page |
| L3 | Cooling & thermal | Liquid cooling, CDUs, cold plates as rack kW rises | Layer page |
| L4 | Data-centre REITs | Landlords: land, power contracts, leases | Layer page |
| L5 | Neoclouds | Rented GPU and specialist compute fleets | Layer page |
| L6 | Networking & optical | Switches, optics, fibre that scale with clusters | Layer page |
| L7 | Hyperscalers | Buyers that set capex for the whole stack | Layer page |
| L8 | Servers & hardware | OEM and ODM systems that fill the rack | Layer page |
| L9 | Software & applications | Apps and platforms that must earn a return | Layer page |
What would change the whole view
Not a single row, but the board's entire framing:
- Hyperscaler capex guidance falls materially. Every layer below is downstream of that number. A cut resets all nine rows, not one.
- Both physical gates clear at once. If memory and packaging both loosen in the same period, pricing power moves downstream to whoever rents the compute, and this board's centre of gravity moves from L1 to L5.
- Power stops being the long pole. If interconnect queues clear faster than halls are built, L2 loosens and the constraint returns to silicon.
- Inference economics break. If cost per token stops falling while application revenue flattens, L9 flips from loosening to tightening and the demand case underneath every other layer weakens.
Frequently asked questions
Which layer of the AI supply chain is most constrained right now?
Semiconductors (L1) and power (L2). L1 is gated by advanced packaging and high-bandwidth memory needing to arrive matched; L2 by interconnect queues and transformer lead times pushing energisation past hall completion.
What do Tightening, Stable and Loosening mean?
They describe capacity against demand for that layer. Tightening means the constraint is binding and pricing power is moving toward that step. Stable means balanced for now. Loosening means supply or alternatives are easing the bind.
How often is this board updated?
It is reviewed weekly, but rows change only when evidence moves — filings, capacity commentary, lead times, or a clear demand shift. A Since date is a status-change date, not a publish date.
Why do all rows currently share one date?
Because the board began tracking on 12 September 2026. That is the honest starting point rather than a set of invented change dates. Each row's date will move independently as statuses actually change.
Is this investment advice?
No. It is an educational constraint map. Not investment advice, no buy lists, and every layer page states what would change its view.
Where are the long explainers?
On the Explained desk, and on each layer page. This board is the index of current state; the layer pages carry the durable mechanism.
Not investment advice. Do your own research.