The Megawatt Is the Unit of Account

Capex per accelerator is the number the market models. Under a binding grid constraint the correct denominator is capex per megawatt, and the two diverge sharply: Vera Rubin costs 97% more per rack than GB300 but only 24% more per gigawatt. Every efficiency claim, depreciation schedule and rack-density decision reads differently under the second denominator, and most published analysis uses the first.

Grid access, not generation, is the binding physical input VALID

• Grid interconnection queue position0.85 strong
  CATEGORY ERROR CORRECTED 2026-08-05
  50 MW
  This cited MISO generator-queue withdrawal rates as evidence that queue position gates LARGE LOAD.
  why

  It is a generation and storage queue; LBNL states large loads are in separate queues not covered by it, so the withdrawal statistic measured speculative GENERATORS abandoning projects and said nothing about a datacenter obtaining load service. The better evidence is more recent and points the same way: on 2026-06-18 FERC issued Section 206 show-cause orders against all six RTOs — MISO is Docket EL26-70 — preliminarily finding the tariff unjust and unreasonable precisely BECAUSE it lacks large-load interconnection provisions, for loads above 50 MW on lines above 69 kV. The strongest support for power access binding is that as of mid-2026 there was no functioning large-load tariff at all.

• Interconnection queue throughput and withdrawal rate (MISO)0.75 strong
  410 GW198 GW
  TIMING, not withdrawal rates.
  why

  Large-load connection runs roughly 4 years nationally and about 7 in Northern Virginia, against datacenter COD of 18 months to 3 years — the mismatch IS the constraint. the ERCOT large-load queue is around 410 GW with 198 GW applied in Q1 2026 alone. NOTE ON INDEPENDENCE: this premise and the one above are close to the same proposition measured twice, so multiplying them understates the conjunction; the product rule assumes independence and these are not independent.

• Heavy-duty gas turbine delivery slots0.80 strong
  The fast bypass around the queue is behind-the-meter gas, and turbine slots are themselves supply-capped, so the escape route has its own queue.

Both the primary route and the bypass are queued, and the queue is the only input in this chain with a multi-year lead time. That is what makes the megawatt scarce in a way the chip is not: silicon supply responds to price within a cycle and interconnection does not.

Per-megawatt capex tells a different story than per-rack capex, and only one of them is decision-relevant VALID

• Ai Capex Cycle — signal 2026-08-040.65 moderate
  97%24%26.6%30.8%11%37%
  Wells Fargo NVL72 BOM: Vera Rubin at $8.34M per rack against GB300 at $4.23M, a 97% increase.
  why

  Per GW of IT power, $61,895M against $50,101M — only 24%. Memory rises from 26.6% to 30.8% of BOM; networking falls 11% per GW even as it rises 37% per rack.

• NVIDIA Vera Rubin VR200 (R200)0.70 moderate
  ~2,000W per chip against GB200's ~1,200W.
  why

  The rack costs more because it also consumes more power, which is exactly why the denominator matters.

• Energy per token served (site level)0.70 moderate
  0.000199 Wh
  0.000199 Wh per token, composite grade medium, every input sourced.
  why

  This is the coefficient that converts a token forecast into a megawatt requirement.

THE SAME BILL OF MATERIALS SUPPORTS TWO OPPOSITE HEADLINES. Per rack, Rubin looks like brutal cost inflation at +97%. Per gigawatt it is +24%, because the rack also does far more work per unit of grid connection. Under a binding power constraint the second is the decision-relevant number and almost nobody quotes it. The networking line makes the point sharpest: up 37% per rack and DOWN 11% per gigawatt, which reverses the sign of the conclusion depending on denominator.

Efficiency in the last conversion stage is capex avoidance wearing an opex costume VALID

• Package shoreline (perimeter and cross-section contention between power, memory and data)0.70 moderate
  Power escape and optical I/O compete for fixed package perimeter, and vertical power delivery is what frees it.
  why

  The engineering is usually justified on loss, and loss is not why it pays.

• Multilayer ceramic capacitor (high-capacitance, AI server)0.65 moderate
  Lateral delivery requires a perimeter capacitor sea because loop inductance prevents the regulator answering AI-workload di/dt.
  why

  Collapsing the path from ~40mm to ~2mm removes most of it.

• Vertiv Holdings Co0.60 moderate
  The power-conversion chain is where a percentage point of end-to-end efficiency is recovered, and it is a named investable layer rather than an engineering abstraction.

If interconnect is fixed at 300 MW for three years — and it is, because queue positions are multi-year — then a 3% end-to-end power-delivery improvement is ~9 MW of compute that did not require a new facility. At roughly $10M per MW of AI-ready facility that is on the order of $90M of avoided capex, recurring, from a change that appears in no marketing deck. This is why power-delivery engineering is getting attention disproportionate to how dull it sounds.

Therefore the cascade defence for depreciation weakens, and useful life is a power question rather than an accounting one VALID

• conclusion:the-megawatt-is-the-unit-of-account#10.51 moderate
  CHAINED AT ITS OWN COMPOSED VALUE
  corrected 2026-08-05.
  why

  This entered at 0.80 while conclusion 1 composes 0.5100 — citing a 0.51 conclusion at 0.80 IS the restatement the annotation claimed to avoid, and it is the framework exempting itself from its own rule at the point the rule bites hardest. Chained legs now carry the referenced conclusion's output, enforced by tools/rigor.py.

• conclusion:the-megawatt-is-the-unit-of-account#20.32 weak
  Same correction: entered at 0.65 against conclusion 2's composed 0.3185.
  why

  A depreciation argument denominated in chips does not reach this conclusion at all, but the dependency cannot be cited more confidently than the thing it depends on.

• Assumed GPU useful life (depreciation schedule)0.60 moderate
  The cascade defence — training on new silicon, inference on older, batch on the tail — REQUIRES SPARE POWER TO CASCADE INTO.
  why

  Hyperscalers use 5-6 year schedules; Amazon shortened a subset to five while Meta extended in the same window, so the industry does not agree with itself.

• CoreWeave, Inc.0.65 moderate
  A financed operator carries the mismatch directly: a GPU cluster has a 4-5 year economic life against a 15-year datacenter lease, and the debt schedule does not care which is binding.

THE INVESTABLE CLAIM AND THE HONEST LIMIT TOGETHER. Under a binding grid constraint a chip's economic life ends when the watts it occupies are worth more running newer silicon, which is strictly shorter than its physical life — the cascade defence implicitly assumes the pre-2023 world where power was abundant. That makes the power framing tighter than the accounting framing. BUT TWO THINGS CUT THE OTHER WAY AND THIS DESK HOLDS BOTH: H100 rental stabilising at $2.85-3.50/hour after a ~70% fall is evidence of a functioning secondary market, and free cash flow is identical whichever schedule is used, so this is an earnings-presentation question rather than a solvency one UNLESS the accounting distortion is itself driving overinvestment. TREAT AS UNRESOLVED. The observable that settles it is cohort-level utilisation disclosure and whether older-vintage rental floors hold.