Markets
The US is where most large AI buyers and many chip designs sit. This door follows who gets paid across demand, neoclouds and power, which gate is stuck, and the dates that matter — in plain English. Education only, not investment advice.
US
Equinix booked a record $533 million of annualized gross bookings in Q3 2026 and plans more than 13,000 cabinets in Q4, its biggest delivery quarter on record. Price is rising and churn is low. The utility and the builders are paid first; the stuck step is the install.
US
Digital Realty signed a record $627 million of yearly contracts in the third quarter, and its pile of signed contracts that have not started reached $1.85 billion. New deals now take 17 months to start paying, and fewer leases started than planned. The stuck step is the handover date.
US
Microsoft on 28 October printed Azure at $33.2 billion, up 48%, and $45.2 billion of capex, short of a ~$50 billion guide. Who gets paid is Microsoft on Azure that billed. The stuck step is still handover of rented GPU-as-a-service racks.
US
Texas Instruments' data center sales rose about 95% in the third quarter, to about 13% of revenue. It has clean rooms to spare, but tools for its back-end lines are arriving late and its finished-goods cushion fell 14% this year. That is the stuck step.
US
Meta on 21 Oct: Q3 capex $34.91bn, FCF -$2.12bn. Who gets paid is server and hall builders on cheques that already cleared. The stuck step is still live megawatts.
US
A $100 billion plan for 21 data centers in rural Virginia needs 3,000 MW. The utility has to promise 299 MW by the end of 2028. The rest comes from gas plants Stack must build itself. The scarce thing is the first power.
US
NiSource on 16 Oct signed Microsoft to about 1,100 MW from 2028. AGP and GenCo are paid; the only named plant is a 600 MW battery still on paper.
US
US AI data centres are paying for their power plants before the halls exist. AEP will build $4 billion of fuel cells for Google, ordered before Google was named. The power gear makers are paid first. The halls are paid last.
US
PJM, the biggest US grid, issued its first Capacity Advisory for 12-13 October 2026. Its last auction came up 6,831 MW short, and its new fast lane has approved 2.1 GW for mid-2029. Plants already running are paid first. The stuck step is new supply, and who pays for it.
US
Micron’s 9 October 10-K shows $133.2 billion of fiscal 2026 sales and $12.75 billion of customer deposits already in the door. Who gets paid is Micron, on bits that already shipped. The stuck step is Idaho DRAM wafer output, still dated mid-calendar 2027.
US
Texas's new large-load rule took effect on 8 Oct: $50,000 of security per megawatt and a $100,000 study fee before a big data center can join a grid study. Wire companies get paid first. ERCOT's approval to switch on stays paused while it checks projects; its reports are due 10 Dec.
US
US data-center construction ran at a record $84.95 billion yearly pace in August 2026, up 73.2% on a year ago, Census data show. Builders get paid as the halls go up. The servers are not in that number. The stuck step is power: the grid side is growing far slower than the halls.
The US is where the big buyers sit and where many AI designs are born. Hyperscalers and neoclouds write the cheques that pull wafers, HBM, packaging, power equipment and data-centre halls. Accelerator designers (and increasingly custom silicon teams) define what those buyers order.
Physical scarcity often shows up one step down the map — in Taiwan packaging, Korea memory, or local power interconnection — but the demand signal starts here. Speed-to-power, transformer lead times and interconnection queues decide how fast contracted GPU fleets can actually light up. In September 2026, AI labs are also sounding out smaller 20–30 MW capacity blocks for speed to usable megawatts while gigawatt campuses stay on multi-year clocks.
Equipment makers and some memory capacity also sit on US listings even when fabs are elsewhere. For the layer map, start at Hyperscalers, Neoclouds and Semiconductors.
As at September 2026, power-ready halls and long-lead electrical gear are a binding gate for US AI infra even when chips are booked — and smaller powered MW deals are one way labs buy around that wait.
| Step | Who | How they get paid | Status (Sep 2026) |
|---|---|---|---|
| Demand / capex | Microsoft, Amazon, Google, Meta, Oracle and peers | Cloud and platform spend; multi-year capex waves | Still high; sets pull for the stack |
| Neocloud / GPU rental | CoreWeave and specialist GPU clouds | Contracted capacity and utilization | Mixed; financing and power matter |
| Accelerator design | NVIDIA, AMD, Broadcom and custom silicon teams | Chip ASP and attach | Design sits here; finish line often elsewhere |
| Small MW / speed-to-usable capacity | Powered-shell landlords, smaller DC developers, neoclouds | 20–30 MW blocks that can light sooner | Active hunt (labs sounding out; CNBC 18 Sep 2026) |
| Power-ready halls / GW campuses | Utilities, developers, landlords, equipment vendors | Interconnect, transformers, leased MW | Binding gate (long lead times; multi-year campus phases) |
Status is Second Order's read from the latest US Pulses (Speed-to-Power; AI labs small MW; CoreWeave) and related Markets pieces.
| Name | What they do | Which step |
|---|---|---|
| Microsoft / Amazon / Alphabet / Meta / Oracle | Large cloud and platform buyers | Demand / capex |
| OpenAI / Anthropic (labs) | Frontier model buyers of campus + small MW capacity | Demand / speed buy |
| NVIDIA / AMD / Broadcom | Accelerator and networking silicon design | Design |
| CoreWeave (and peers) | Specialist GPU cloud capacity | Neocloud |
| US power & electrical chain | Transformers, interconnect, power-ready sites | Power gate |
Examples of exposure roles only — not recommendations.
| Date | Event | What to look for |
|---|---|---|
| 18 Sep 2026 | CNBC — labs hunt 20–30 MW deals | Whether reported small-MW chase turns into named contracts |
| Ongoing 2026 | Hyperscaler capex commentary | Whether AI spend holds while ROI proof is still early |
| Ongoing 2026 | Transformer / interconnect lead-time prints | Whether >160-week class delays ease |
| Ongoing 2026 | ERCOT and other large-load queue updates | Whether power-ready MW clears or lengthens |
| 2028 | PORTS-Pike first ~800 MW (stated) | Whether campus energise holds vs slips |
| 2028–2032 | Camellia / Georgia Power 3.2 GW phases | Phase timing vs small-MW rental mix |
| Quarterly | Neocloud backlog vs revenue conversion | Contracts turning into billed utilization |
The US is where most large AI buyers and many chip designs sit. Their spending pulls the Asian manufacturing steps and the local power buildout.
It is how fast a site can secure interconnection, transformers and energized halls so booked GPUs can run — often slower than chip delivery in 2026.
Neoclouds sell specialist GPU capacity, often as overflow or dedicated clusters when hyperscaler capacity is full or specialised.
For speed to usable capacity. Smaller blocks at already-powered sites can host inference and deployable workloads sooner than multi-year gigawatt campuses.
No. Education only — who gets paid, which gate is stuck, and which dates can change the read.
Updated 21 Sep 2026 · Next update after the next dated small-MW contract print or major US campus / power-queue disclosure
