Power Caps the AI Buildout — the bottleneck moved from chips to megawatts

The binding constraint on AI compute has shifted from silicon to electricity and upstream gas deliverability, and — unlike a chip shortage — it is slow to relieve. Data centers add ~125 GW of US load 2026-2030 into a ~2,600 GW interconnection queue with 5+ year waits, while the fast bypass (behind-the-meter gas) is capped by turbines sold out through 2030. A chip-centric consensus underprices how long this gates the buildout — the persistence shape the desk trades, applied to power.

Queue position, not generation capacity, gates new AI load — binding the datacenter-ai industry specifically, not 'energy' generally VALID

• Datacenter Ai0.99 strong
  The demand node exists and is growing; not in dispute.
• Electricity supplies Datacenter Ai1.00 strong
  Physical necessity.
• Grid interconnection queue position0.85 strong
  2,600 GW
  ~2,600 GW queue and 5+yr median waits are LBNL-sourced.
  why

  Discount for queue GW overstating real intent — speculative and duplicate applications are counted.

~125 GW of new load meets a ~2,600 GW queue with 5+ year median waits, administratively rationed rather than priced, so capital cannot shorten it. Stated at INDUSTRY scale deliberately: the same constraint is `none` for talen-energy and vistra, which own already-interconnected generation and are made more valuable by the scarcity.

The fast bypass — behind-the-meter gas — is itself supply-capped, and it binds a DIFFERENT set of names VALID

• Heavy-duty gas turbine delivery slots0.90 strong
  116 GW
  Three OEMs, books reported sold out to 2030.
  why

  GEV Q2-26 confirms: backlog+reservations 100->116 GW, taking reservations for 2031. Discount because slot allocation is negotiated, not a published queue.

• Gas Turbines (grid + behind-the-meter)0.99 strong
  The technology exists and is the stated bypass.
• Gas Turbines (grid + behind-the-meter) supplies Electricity1.00 strong
  Physical.

~101 GW of on-site gas is announced against OEM order books sold out to 2030. The pairing is the point: crusoe is MILD on the queue and SEVERE on turbines because it sited off-grid to dodge the queue; coreweave is the reverse.

Token demand converts to megawatt demand close to one-for-one, so the demand forecast and the physical constraint are the same question VALID

• Energy per token served (site level)0.50 moderate
  2.88x1.14x1.01x2.22x3.16x1.43x
  MLPerf Inference series, llama2-70b-99.9 Server, best-in-round tokens/sec per accelerator: 3,732 -> 10,756 -> 12,305 -> 12,390 across v4.0 to v5.1.
  why

  That is 2.88x at the Blackwell step, then 1.14x and 1.01x. THE GAIN IS A ONE-TIME HARDWARE STEP, NOT A CONTINUING TREND: between generations efficiency is flat and pass-through to megawatts runs close to 1:1, discounting only at refresh points. The accelerator joules per token behind it — H100 0.1800, H200 0.1693, B200 0.0813 — show energy per token improving 2.22x in one generation, on throughput up 3.16x against board power up 1.43x. Held at 0.50 because the single-generation comparison argues lower and the flat inter-generation series argues higher.

• Power Caps Ai Buildout — signal 2026-08-030.75 strong
  Tau scaling is a TIME law and not an ENERGY law, conceded in the tau paper's own open-challenges section.
  why

  A concession against interest from a source arguing the opposite case.

• Aggregate inference token demand0.85 strong
  Token demand rising fast, held in the tokenomics layer.

Token demand passes through to megawatt demand at a DISCOUNTED rate set by the hardware refresh cycle, not at parity. Energy per token improved 2.22x from H100 to B200, so the strong one-for-one form does not hold. But throughput rose 3.16x over that same step against board power up 1.43x, so absolute draw per accelerator ROSE and efficiency did not keep pace with throughput. This is the leg that connects the rest of the thesis: the other three conclusions argue megawatts are gated and none of them says where the megawatts come from, which left token growth and grid scarcity sitting in the model as unrelated facts. Energy per token is the bridge — at roughly flat efficiency between generations, every trillion tokens of new demand is another 229 MWh that has to clear an interconnection queue. AND rather than OR, because all three must hold: the conversion must be roughly right, efficiency must not outrun throughput, and demand must actually grow. THE FALSIFIER IS SHARP AND CHEAP TO TEST: if energy per token falls as fast as throughput rises, tokens and megawatts decouple and this leg dies. Nobody publishes that series, which is why building it is the highest-value open item in the layer.

Power deliverability gates the AI buildout through AT LEAST ONE of the three physical constraints VALID

• Grid interconnection queue position0.85 strong
  Queue binds grid-connected siting.
• Heavy-duty gas turbine delivery slots0.90 strong
  Turbine slots bind behind-the-meter siting.
• Interconnection queue throughput and withdrawal rate (MISO)0.75 strong
  79.6%217.1 GW92.6 GW361.3 GW
  MISO withdraws 79.6% of every interconnection application that reaches a decision — 217.1 GW active against 92.6 GW energised and 361.3 GW withdrawn.
  why

  A headline gigawatts-in-queue figure therefore counts a large majority of applications that will never energise, so queue size OVERSTATES incoming supply and the constraint is tighter than the raw number suggests. Held at 0.75 rather than higher because this is a GENERATOR queue — supply asking to connect — so it evidences that new capacity clears slowly, not that datacenters specifically cannot connect. Measured 2026-08-04.

• Large power transformer lead times0.80 strong
  48-60mo lead times, backlogs to 2030, announced relief not before 2027-28.
  why

  Discount because the >half-of-2026-delayed figure is trade-press estimate.

OR, not AND, because these are SUBSTITUTABLE paths to the same outcome — an operator dodging the queue takes on turbine risk, and one dodging both still needs transformers. The thesis does not require all three to bind; it requires that a given project cannot route around all of them. Composing three independent 0.80-0.90 paths gives a materially higher number than any single one, which is exactly why the thesis is more robust than any one constraint.