A kWhBill experiment
Can Texas Power the AI Buildout?
KWhBill tracks what power costs. This page asks whether the power will be there.
One page of Texas power-availability facts, drawn from the generation interconnection queue and the large-load interconnection queue — two different queues that are not the same thing. Every number below has a source and a cutoff date.
The queue in numbers
Source: Lawrence Berkeley National Laboratory, Queued Up 2026 Edition (data through end of 2025)
- 408 GW active. ERCOT's interconnection queue holds 1,527 projects seeking to connect — generation plus storage — second only to the non-ISO West (567 GW). Nationally, 8,244 projects totaling 2,061 GW (1,312 GW generation, 749 GW storage) are actively seeking grid interconnection.
- 119 GW with draft or executed interconnection agreements.ERCOT has more queue capacity at this stage than any other region (3 GW additional suspended). Nationally, 549 GW sits at this stage but has not reached commercial operation — including 256 GW solar, 161 GW storage, 76 GW wind, and 45 GW gas. A signed agreement is not a built power plant.
- 24% completion in ERCOT — the highest of any region. Of interconnection requests submitted from 2000–2020, 24% reached commercial operation in ERCOT (by number of requests), ahead of ISO-NE (18%) and well ahead of CAISO (8%). Nationally, just 13% of queued capacity from that cohort had come online by end of 2025; 75% had been withdrawn.
- Texas is ≥20% of the national queue in every fuel.The state map shows Texas holding at least one-fifth of active U.S. queue capacity in each of gas (253 GW national, +86% in 2025), solar (773 GW), storage (749 GW), and wind (220 GW).
- Slow pipeline. The median project that reached commercial operation in 2025 spent 61 months — over five years — from interconnection request to operation (national figure; ERCOT runs faster than most regions).
Queue data were compiled, cleaned, and aggregated by LBNL in partnership with GridTracker; the dataset is licensed CC BY 4.0, used here with attribution to Lawrence Berkeley National Laboratory and GridTracker.
The other queue: large loads
ERCOT also runs a separate large-load interconnection process for loads of 75 MW and above — data centers, bitcoin miners, hydrogen. This is a different queue from the generation queue above, and the two cannot be netted against each other.
- ≈200 GW screened in Batch Zero's first classification round(industry reporting, September 2026): 362 large-load requests totaling 191.8 GW classified, with 204 projects / 66.4 GW conditionally counted as base load. Nothing here is final energization approval.
- ≈474 GW of total large-load requests were cited in the Texas governor's August 3, 2026 directive — roughly five times ERCOT's all-time peak demand of ~86 GW, with ~90% reportedly from data centers.
- New large-load interconnections are paused pending a verification audit ordered by the governor; ERCOT did not issue its planned August 2026 Batch Zero classifications.
Load-side figures are drawn from September 2026 industry reporting on ERCOT publications (trade press), not from a primary LBNL dataset — treat them as directionally indicative, not audited.
Delivered, but not yet powered: “dark GPUs”
The queues above measure generation waiting to connect and loads waiting for approval. There is a later stage this page did not previously name: hardware that has arrived but cannot yet be switched on, because the site that will host it does not yet have the power to run it.
The industry shorthand for this stage is“dark GPUs”, by analogy to dark fiber: fiber laid but not lit. Here it means hardware delivered but not yet powered — not hardware that is broken, and not a Texas total.
Some commentary uses the phrase more broadly for GPUs that are powered but underused because of system design or idle time. This page does not use it that way. Conflating the two would turn a power-readiness problem into a utilization problem, and the remedies are different: one needs upstream power, the other needs scheduling or design fixes.
- Amazon, reported October 2026. Reporting based on internal emails and a meeting recording obtained by Business Insider said AWS was temporarily storing Nvidia GB300 racks in intermediate warehouses because target data centers had room for more racks but not enough available power to turn them on. Internal projections cited in the reporting put receipts at about $5 billion in August 2026 and about $3 billion a month thereafter — a receipt projection, not a measure of deployed capacity. Amazon disputed that any team lacked a deployment path and described staging near deployment sites as standard practice. Treat the episode as reported, not audited: it shows the stage exists, not how large it is.
- Nvidia DSX POD, announced October 21, 2026. A reference design for complete Vera Rubin AI factories built by partners inside modular containers — seven NVL72 racks, 504 GPUs per pod — integrated, cabled, liquid-cooled, and validated in the factory before delivery. Nvidia says onsite deployment at a prepared site can fall from about six months to 14 days, with partner deliveries planned from early 2027. That 14-day figure is explicitly conditional: it applies once site utilities are ready. The design compresses integration; it does not create upstream power. Nvidia's own description leaves upstream generation, conversion, and regional electrical design to partners.
Put together, the two cases split "time to first token" into stages that can move independently: hardware delivery, onsite integration, power readiness, and construction labour. Factory integration can now run in parallel with site preparation instead of after it — which makes power readiness, not integration, the stage most likely to be left standing alone at the end. That is why this page tracks the generation queue and the large-load queue separately, and why neither total, on its own, is a measure of power actually available to a new site.
DSX POD is an announced reference design with deliveries planned, not operating evidence. No figure in this section is a Texas total, and nothing here claims that delivered-but-unpowered hardware moves retail prices.
What it means for site selection
Three observations, no procurement advice: the 408 GW generation queue is overwhelmingly requests, not firm capacity — history says roughly three in four requests never build. The 119 GW with signed agreements is the nearest-term serious pipeline, but even signed capacity withdraws at material rates (41% of nationally IA-signed capacity from 2000–2022 had withdrawn by end of 2025). The large-load queue is measured in a separate process with its own rules and a current pause — do not read generation-queue totals as supply available to new data centers.
What we're watching next
This page is a briefing, not a news feed — but three dated items would change what it says, so they are tracked here rather than left implicit:
- December 10, 2026 — ERCOT's verification-audit report.Responses to ERCOT's September 14 impact questionnaire (Batch Zero loads and medium loads, with notarized attestation) were due October 12, 2026. The governor-ordered audit of data centers in the process — grid and water use, on-site generation, ownership — is expected to report around December 10, 2026, after which the Batch Zero study resumes. If the audit slips, the study timeline slips with it.
- November 1, 2026 — first Batch re-study filings (target).October stakeholder workshops (3a on Batch scope, 3b on planning scope) feed Revision Requests covering re-study of loads with unfulfilled ramp, later-Batch inclusion for uncommitted loads, and Batch One inclusion for loads excluded from Batch Zero. These are filing targets, not filed rules.
- Early 2027 — first DSX POD partner deliveries.The section above records the announcement; the test is whether partner delivery dates hold, and whether "14 days at a prepared site" survives contact with real sites.
Dates above are trade-press-reported targets and expectations, not ERCOT publications — treat them as a watch list, not a schedule. This section was added October 25, 2026.
Evidence & limits
- Price: Texas residential average 15.91¢/kWh, U.S. 18.16¢/kWh — EIA retail-sales snapshot for August 2026, as published on the Texas electricity rates page.
- Generation queue: LBNL Queued Up 2026 Edition — data through December 2025; Excel data file updated May 2026; report published June 2026.
- Regional labeling: LBNL reports at the ERCOT-region level. ERCOT is the Texas grid operator's footprint — it is not exactly the state of Texas, and this page labels ERCOT-measured figures as ERCOT, not Texas.
- Two queues, two stories: the generation queue (LBNL) and the large-load queue (ERCOT Batch Zero) are independent processes. Comparing the two totals is descriptive, not predictive.
- No proven causality: a growing queue does not mechanically mean rising retail prices — or lower ones. This page does not make that claim.
- Delivered, but not yet powered: Business Insider reporting, October 2026, based on internal emails and a meeting recording (Amazon's response included); Nvidia DSX POD announcement, GTC Berlin, October 21, 2026, with partner deliveries planned from early 2027. Neither source is a Texas total.
- “Dark GPUs” term: the shorthand above comes from October 2026 Business Insider reporting on delivered hardware waiting for power, as propagated in secondary reporting; the dark-fiber analogy is the reporting's, not this page's invention. This page uses the narrow meaning only — delivered but not yet powered.
- Not advice: an information briefing from public sources — not legal, investment, or energy-procurement advice.
Why this page exists
This page is an experiment: before building anything on the supply side, we publish one page and see whether anyone reads it. If there is real search demand, it grows into a section or a site. If not, this page stays a one-page archive. That's the whole plan.
Texas power availability, as it changes.
A kWhBill experiment. Data through December 2025 unless noted.