brain/
conceptstock-market

Phantom data-center load — interconnection-queue demand is heavily inflated

Notes

Phantom data-center load — interconnection-queue demand is heavily inflated

One-line summary: Announced/queued data-center demand (1,200+ proposed DCs, ~100–300 GW against a total US generation base of 1,400 GW) is largely speculative: where states add financial commitment (Texas's ~$55k/MW queue fee), "expected" demand drops by a third, and only ~75% of generation projects in queues ever get built — so load forecasts riding raw queue numbers systematically overstate the real buildout.

The insight

Data-center power-demand forecasts range from ~100 GW to 300–400 GW — up to a third of total existing US generation capacity. But the queue is congested with speculative, multi-state duplicate requests. The early natural experiments (Texas, Illinois) show that requiring skin-in-the-game (fees, proof of financial closure, permitting) shrinks the "real" number dramatically. ERCOT's own view: the notional request queue is ~3x the whole Texas system, but the paid, committed queue is "incrementally tens of gigawatts per year," which they consider manageable — and Texas isn't worried about price increases because it can build supply.

Evidence

⬆ TSMC says it audits customer demand precisely because the aggregate is not credible (2026-07-17 ingest)

The sharpest first-party statement the wiki has that supplier order books over-state real load — from the supplier best positioned to know:

  • cc-wei in 2026-07-16-earnings-tsm-q2-fy2026: "Now, remember that I believe every customer tell me the truth. Everyone. You put all the truths together, it's not the truth. We have to make some of the judgment, you know what I mean... Because all the customer are very aggressive, right? That's a CEO's job. CEO got to be aggressive. They give me the number of their demand, and I believe they try their best to tell me the truth. I put all together, all the truths together is not a truth. Mark down that word."
  • And the countermeasure he says TSMC runs — cc-wei in 2026-07-16-earnings-tsm-q2-fy2026, asked directly whether TSMC's chips could end up in customer inventory: "Actually, we are checking the AI data centers of progress, the building, the location, the demand, the rest. We checking all that to make sure that TSMC's chips will not be put in inventory."

Both directions matter, and they partly cancel. The first quote is a monopolist stating that summed customer demand is systematically overstated — the phantom-load thesis in one sentence, from the top of the supply chain. The second is TSMC saying it physically verifies data-center construction progress before committing capacity, which is a genuine control against phantom load propagating into TSMC's capex.

So the honest read is: TSMC confirms the phenomenon exists at the order-book level, and claims to filter it. That makes phantom load a bigger risk for suppliers without TSMC's visibility (memory, power equipment, gensets, grid) than for TSMC itself. It also means TSMC's $60–64B capex embeds a demand estimate that has already been haircut — Wei is telling you the raw customer numbers were higher.

⚠ Unverifiable and interested: Wei offers no detail on the audit's method or how much he haircuts, and he has every reason to assure investors his capex is disciplined.

Trading implications

  • Discount any thesis whose demand leg is a queue or announcement number. The wiki's AI-power chains cite announced-GW figures; the realized number is likely ~1/3 lower once commitment is required, on top of the "<half under construction" shortfall already tracked on ai-capex-to-power-and-materials-cascade.
  • Utility/genset/turbine order books built on committed (paid) queue positions are higher quality than those built on request pipelines — a diligence lens for GEV/PWR/VRT backlog claims.
  • Falsifier feed: state-level financial-commitment reforms (TX, IL) produce datable "real demand" prints — each one tests the demand leg of the AI-power cascade.

2026-07-18 — NY's statewide moratorium: the political-siting risk, and the water/land "myths" rebuttal (All-In)

The phantom-load demand skepticism now has a supply-side political companion: blue states blocking construction outright. Governor Kathy Hochul signed "the nation's first ever statewide moratorium on hyperscale data centers." david-sacks in 2026-07-18-podcast-all-in-podcast-can-the-ai-industry-regulate-itself-stripe-wants rebuts the stated rationales point-by-point — the read-through is that the loudest anti-DC arguments (water, land, pollution) are weak, so the binding constraint is politics/permitting, not physics:

  • david-sacks: "these data centers are a model of land use efficiency... The whole water consumption thing is largely a hoax. The modern data centers recirculate the water closed loop systems... a typical data center uses the same amount of water as two and a half In-N-Out burgers."
  • The scale-of-the-myth data (peter-diamandis/Salim in 2026-07-19-podcast-moonshots-urgent-update-ai-sputnik-moment-kimi-k3-released): every US data center on-site ≈ 17B gallons of water vs. golf-course irrigation ≈ 531B gallons (31×), and California almond farming ≈ 1T gallons (60×); Amazon warehouses occupy ~10× more US land than all data centers combined.
  • The behind-the-meter escape (source-attributed, Chamath in 2026-07-18-podcast-all-in-podcast-can-the-ai-industry-regulate-itself-stripe-wants): blocked grid interconnection pushes developers to self-generation to escape the queue and clean-air permitting — Elon "bought the company that makes all this mobile turbines" for Colossus (Memphis), and "when you act as a block, you could make the claim that it doesn't [fall under clean-air permitting]... new solutions like Bloom Energy... still fall under the personal use clean air permit." Ties to ai-power-gap-to-genset-bridge-power and bloom-energy.
  • The demand-is-real, supply-is-short counter to phantom-load (source-attributed, Chamath, same source): a PJM auction where they "needed like 7 or 8 gigawatts" and "had 156 megawatts... show up," and ~"40% of all these projects are getting mothballed and stopped" — so the committed, energizable-today segment is priced at an extreme premium. Read against the queue-inflation thesis: the notional queue is phantom, but deliverable power is genuinely scarce.
  • Influence-op overlay (david-friedberg, source-attributed): the anti-DC movement is analogized to the Russia-Today-era anti-GMO campaign; Sacks cites an OpenAI blog ("PRC-linked influence operations are targeting AI debates") and a coming congressional investigation. ⚠ Politically-charged and interested (all-in panel is pro-buildout) — recorded as a claim about why the opposition exists, not adjudicated.

Net: the moratorium is a real, datable supply-siting risk (5+ year effective pause once queues reset, per Sacks) that is distinct from — and partly opposite to — the phantom-demand thesis: headline queue demand is inflated, but blue-state siting bans plus a genuinely short deliverable-power segment make the realized buildout constrained on both ends. Beneficiaries of the behind-the-meter escape: bloom-energy, gensets/gas turbines, and permissive states (TX).

Contradictions / tensions

  • Hyperscaler capex actuals (Q1 2026 prints on ai-capex-to-power-and-materials-cascade) are real spend, not queue phantoms — the phantom-load critique bites hardest on utility load forecasts and speculative developers, less on the top-4 hyperscaler self-build.
  • Cembalest's "take the under" on consultant demand projections (already on the cascade page) is independent corroboration of the same skepticism.

Sources

Related

Referenced by
Mechanisms
Aging grid at end-of-life → replacement + hardening capex wave → T&D equipment inflation → grid picks-and-shovels demand independent of AI loadAI capex sprint → power-supply gap → grid-component + materials bottleneck → nuclear/copper/transformer beneficiary cascade
Concepts
AI safety as a vector for regulatory capture / centralizationElectricity-price driver decomposition — load growth is not the national driverGrid-enhancing technologies — ~260 GW of latent capacity in the existing grid
Entities
Alice YakeDavid FriedbergDavid SacksDoug ArentSam AltmanTSMC (Taiwan Semiconductor Manufacturing Co.)