inthesingularity.ai
Wire · Data centers · Pipeline

W.094Bernstein counts 586 GW of planned North American data centers and rates 199 GW credible, 34%. Development stage accounts for half of the 387 GW cut; state politics for 7%

A Bernstein note reported this week counts 586 GW of planned data center capacity in North America, up from 411 GW in July by Zhitong Finance’s account, and rates 199 GW of it credible, 34%. The other 387 GW is removed in four steps: 198 GW for development stage, meaning little or no construction progress, 113 GW for the tenant’s credit, 48 GW for the developer and 28 GW for state political risk. We have not seen the note; the figures are as Chinese financial media reported them.

Rated credible, of 586 GW planned
199 GW
34%, probability-weighted; 33% in July and 35% in August
Removed for development stage
198 GW
51% of the 387 GW cut
Removed for state political risk
28 GW
7% of the cut; tenant credit removes 4.0 times as much
Texas, credible of planned
~43 of 150
GW, about 29%; 21 GW is under construction

What the note is reported to say

The 199 GW is an expected value, not a list of approved projects. According to Wallstreetcn and Cailian Press, the team led by Madison Rezaei takes the Aterio database of 8,858 North American projects, which are under construction, announced or only holding land, and multiplies each one by probabilities for its stage, its tenant, its developer and its state. That description implies capacity already operating is outside the count. The total in October was 586 GW; after weighting, 199 GW remains. Zhitong Finance dates the note November 10.

Stage does the most work. Wallstreetcn reports the base probabilities as 90–98% for a project under construction, depending on the developer, 50% for one announced with a third-party developer and 25% for one that only holds land. The stage discount removes 198 GW, 51% of the 387 GW cut. Tenant credit removes 113 GW and the developer’s strength and record 48 GW. State political risk removes 28 GW, 7% of the cut and 4.8% of the pipeline.

Cailian Press sets the note against fears, after the midterms, that state pauses would stop construction. Democrats won the governorships of Texas and Georgia in the November 3 elections, and Reuters reported that Democrats who campaigned in part on a moratorium on data center development won offices. Texas’s governor-elect wants construction paused until the legislature acts, Cailian Press reported. Goldman Sachs estimated that temporary state halts could affect about 45% of planned US capacity expansion, according to Reuters. Bernstein’s arithmetic puts the political deduction at about a quarter of the tenant deduction, and the model treats projects already under construction as exempt.

From 586 GW planned to 199 GW credible

Planned North American data center capacity in October 2026 (faint grey), the four amounts Bernstein’s model removes (pink) and what it rates credible (strong grey), in gigawatts, as Wallstreetcn and Cailian Press reported them. The steps are the model’s attribution of a probability weighting, not counts of cancelled projects. The reports do not say whether a gigawatt here is IT load or power drawn from the grid.

0150300450600Plannedunder construction, announced or land heldPlanned North American data center capacity, October 2026: 586 GW586Development stagelittle or no construction progressRemoved for development stage: 198 GW, 51% of the cut−198Tenant creditwho has leased it, if anyoneRemoved for tenant credit: 113 GW, 29% of the cut−113Developer strength and recordpower, capital, leases, executionRemoved for developer strength and record: 48 GW, 12% of the cut−48State political riskregulatory risk by stateRemoved for state political risk: 28 GW, 7% of the cut−28Credibleprobability-weighted, 34% of plannedRated credible: 199 GW, 34.0% of 586 GW199
0150300450600PlannedOctoberPlanned North American data center capacity, October 2026: 586 GW586StageRemoved for development stage: 198 GW, 51% of the cut−198Tenant creditRemoved for tenant credit: 113 GW, 29% of the cut−113DeveloperRemoved for developer strength and record: 48 GW, 12% of the cut−48State politicsRemoved for state political risk: 28 GW, 7% of the cut−28Credible34% of plannedRated credible: 199 GW, 34.0% of 586 GW199
The deductions
StepGWShare of the cutShare of planned
Planned586100%
Less: development stage−19851%33.8%
Less: tenant credit−11329%19.3%
Less: developer strength and record−4812%8.2%
Less: state political risk−287%4.8%
Credible19934.0%

Most of what is under construction counts; a quarter of what is only announced does

Planned capacity by stage, in gigawatts, split into the part rated credible (strong grey) and the rest (faint grey), as Zhitong Finance reported. The two stages add to 584 GW planned and 198 GW credible, against totals of 586 and 199; the reports do not explain the difference, which may be rounding.

Rated credibleNot rated credible
0125250375500Under constructionUnder construction: 75 GW rated credibleUnder construction: 12 GW not rated credible75 of 87 GW, 86%Announced, not startedAnnounced, not started: 123 GW rated credibleAnnounced, not started: 374 GW not rated credible123 of 497 GW, 25%
0125250375500Under constructionUnder construction: 75 GW rated credibleUnder construction: 12 GW not rated credible75 of 87Announcednot startedAnnounced, not started: 123 GW rated credibleAnnounced, not started: 374 GW not rated credible123 of 497

Who is behind the credible part

A little over a third of the plans account for three-quarters of what is credible. Hyperscalers building for themselves and established operators that lease are 37% of planned capacity and about 76% of the credible total, Cailian Press reported. Zhitong puts Amazon, Google, Meta and Microsoft at about 78 GW of credible capacity between them, 39%: about 33 GW for Amazon, 20 GW for Google, 15 GW for Meta, 10 GW for Microsoft. Wallstreetcn puts hyperscalers’ own building at about 82 GW, 41%.

More than half of the plans come from developers the model barely credits. Developers that are not hyperscalers, colocation operators, former crypto miners or neoclouds have announced 308 GW, 53% of the pipeline, of which 33 GW is rated credible, 11%. Developers scoring between 0 and 1 on Bernstein’s scale of 0 to 5 have announced 142 GW, and 7 GW of it survives. Wallstreetcn names O’Leary Ventures, with 15.0 GW planned and 489 MW credited, and Joule Capital Partners, with 11.7 GW and 1.1 GW. At the other end, QTS is credited with 11.5 GW of 14.3 GW.

The tenant matters more than the state. A lease to Google, Amazon, Microsoft, Meta or Nvidia carries no discount in the model. A project leased only to OpenAI or to Groq is halved, and one leased to CoreWeave, Nebius or Lambda is multiplied by 0.45, according to Wallstreetcn. Oracle’s multiplier was lowered to 0.75 from 0.80. The state multipliers have a floor of 0.70, which Texas, Pennsylvania, Virginia and New York are at.

Three-quarters of hyperscalers’ own plans are rated credible, a ninth of other developers’

The share of each builder type’s planned capacity that Bernstein rates credible (strong grey), as Zhitong Finance reported it, each figure given as approximate. The reports give planned gigawatts only for “other developers” (308 GW, of which 33 GW credible), so the bars are rates, not amounts.

0%25%50%75%100%Hyperscalers building their ownHyperscalers building their own: about 75% of planned capacity rated credibleHyperscalers building their own: about 25% not rated credible~75%Specialist colocation operatorsSpecialist colocation operators: about 62% of planned capacity rated credibleSpecialist colocation operators: about 38% not rated credible~62%NeocloudsNeoclouds: about 29% of planned capacity rated credibleNeoclouds: about 71% not rated credible~29%Former crypto minersFormer crypto miners: about 24% of planned capacity rated credibleFormer crypto miners: about 76% not rated credible~24%Other developersOther developers: about 11% of planned capacity rated credibleOther developers: about 89% not rated credible~11%
0%25%50%75%100%Hyperscaler self-buildHyperscalers building their own: about 75% of planned capacity rated credibleHyperscalers building their own: about 25% not rated credible~75%Colocation operatorsSpecialist colocation operators: about 62% of planned capacity rated credibleSpecialist colocation operators: about 38% not rated credible~62%NeocloudsNeoclouds: about 29% of planned capacity rated credibleNeoclouds: about 71% not rated credible~29%Former crypto minersFormer crypto miners: about 24% of planned capacity rated credibleFormer crypto miners: about 76% not rated credible~24%Other developersOther developers: about 11% of planned capacity rated credibleOther developers: about 89% not rated credible~11%
Tenant multipliers and named developers
TenantMultiplierBefore
Google, Amazon, Microsoft, Meta, Nvidia1.00
Microsoft with OpenAI0.95
Oracle with OpenAI (Stargate)0.800.85
Jane Street, Bloomberg0.80
Nscale with Microsoft0.80
Oracle0.750.80
xAI0.60
OpenAI alone0.50
Groq0.50
CoreWeave, Nebius, Lambda0.45
Nscale alone0.40
DeveloperPlanned (GW)Credible (GW)Share
QTS14.311.580%
STACK11.17.769%
Vantage9.07.280%
O’Leary Ventures15.00.53.3%
Joule Capital Partners11.71.19.4%

The pipeline grew 43% between July and October; the credible share stayed near a third

Planned capacity (faint grey) and the part rated credible (strong grey) in the three editions of Bernstein’s analysis, in gigawatts, as Zhitong Finance reported them. The label over each strong grey bar gives the credible gigawatts and their share of planned. The months are those of the pipeline data; the notes were published later.

PlannedRated credible
0200400600July: 411 GW planned411July: 136 GW credible, 33.1% of planned136 · 33%JulyAugust: 488 GW planned488August: 170 GW credible, 34.8% of planned170 · 35%AugustOctober: 586 GW planned586October: 199 GW credible, 34.0% of planned199 · 34%October
0200400600July: 411 GW planned411July: 136 GW credible, 33.1% of planned136 · 33%JulyAugust: 488 GW planned488August: 170 GW credible, 34.8% of planned170 · 35%AugustOctober: 586 GW planned586October: 199 GW credible, 34.0% of planned199 · 34%October

How large 199 GW is

The credible part alone is about three times what runs today. Wallstreetcn reports Bernstein as saying 199 GW would be 3.4 times current operating capacity, which implies a base of about 58.5 GW, built over five to ten years; the note gives no year-by-year schedule in the accounts we read. 451 Research put grid power supplied to US data centers at about 64.4 GW in 2025, on which 199 GW is 3.1 times and 586 GW is 9.1 times. The 75 GW credible and under construction is by itself more than that base.

The part that will not be built is priced at trillions. Wallstreetcn cites construction costs of $7–15 billion a gigawatt before chips. On 387 GW that is $2.7T to $5.8T of construction that the model does not expect to happen.

In Texas the discount falls on the near years. The database holds about 150 GW in Texas and the model credits about 43 GW, about 29% (Zhitong gives 43 to 44 GW); the state multiplier alone removes about 12 GW, twice Virginia’s loss. Of 11 GW announced for 2027 and not yet started, 0.3 GW is credited; of 18 GW for 2028, about 5 GW; of 39 GW for 2029, about 8 GW. ERCOT’s large-load requests were about 474 GW, according to the governor’s August letter, 3.2 times the Texas capacity in Bernstein’s database; interconnection requests are a different measure from a project database.

Other counts are of the same order. Wood Mackenzie said in August that US utilities and grid operators would realistically commit to about 28% of the 1,066 GW requested of them. CBRE counted 7.48 GW under construction in the eight primary markets in the first half; Bernstein’s 87 GW under construction covers every market and builder, including hyperscalers’ own sites, so the two are not the same measure. Plans with on-site power grew to 231 GW from 185 GW in August; 47 GW of that is rated credible, 20%.

What this changes on inthesingularity.ai

  • The Texas data center risk pack sizes exposure to a pause. Lead: set its exposed capacity beside Bernstein’s Texas figures, about 43 GW credible of 150 GW, 21 GW under construction and treated as exempt, and about 12 GW removed by the state multiplier.
  • The compute balance pack takes a supply path in gigawatts. Lead: a tracker row for Bernstein’s planned and credible totals by edition (411 / 136, 488 / 170, 586 / 199), and for the under-construction figure, 87 GW.
  • W.091 set lab contracts against cloud backlog. Lead: the tenant multipliers (OpenAI alone 0.50; CoreWeave, Nebius, Lambda 0.45; Oracle 0.75) are a second analyst’s view of the same concentration, and could be applied to the neocloud backlog table there.

Sources and method

ItemFigureSource
Bernstein586 GW planned in October, from 411 GW in July; 199 GW credible, about 34%; 87 GW under construction, 75 GW of it credible (86%); about 497 GW announced and not started, 123 GW credible; earlier editions 411 / 136 and 488 / 170; credibility by builder type; the four largest hyperscalers 78 GW (Amazon about 33, Google about 20, Meta about 15, Microsoft about 10); behind-the-meter 231 GW planned and 47 GW credible; note published Nov 10Zhitong Finance, via Sina, Nov 11 (in Chinese)
BernsteinThe four deductions (198, 113, 48, 28 GW); Aterio database of 8,858 projects; stage probabilities, developer scores and tenant and state multipliers; “other” developers 308 GW planned and 33 GW credible; hyperscaler self-build about 82 GW; named developers; Texas 150 GW planned, about 43 GW credible, by target year; 68% of credible capacity in 13 flagged states; 3.4 times capacity operating today; $7–15B per GW excluding chipsWallstreetcn, via NetEase, Nov 12 (in Chinese)
BernsteinThe four deductions; 37% of planned and 76% of credible capacity with hyperscalers and established operators; halving for projects leased only to OpenAI or Groq; analyst Madison Rezaei; other banks’ estimates after the midtermsBigGo Finance, Nov 12, citing Eastmoney (a Cailian Press article)
BernsteinThe same figures in Chinese, set against fears after the election of state pauses; “expected to be built”; the Texas governor-elect wants construction paused until the legislature actsCailian Press, via Eastmoney, Nov 12 (in Chinese)
BernsteinFirst edition (July data): title; 33% credible, 135 GW of a 492 GW dataset; a 5–10 year build; no stand on timingZeroHedge, August
BernsteinEarlier note, Nov 6: the election left about 17 GW of newly exposed credible capacity vulnerable to stricter rules; Bernstein forecast a 2–4% effect on credible capacity nationwideBenzinga, via KuCoin, Nov 6
Goldman SachsTemporary state halts could affect about 45% of planned US data center capacity; Democrats who campaigned in part on a data center moratorium won offices, including the Texas and Georgia governorships; Core Scientific fell 11.6% and Iren 12.5% on Nov 4Reuters, via MarketScreener, Nov 4
Installed baseGrid power supplied to US data centers about 64.4 GW in 2025, up 25%S&P Global, 451 Research, May 6
ERCOTAbout 474 GW of large-load requests, about 90% data centers, per the governor’s August letter; a moratorium on new data center interconnections called for on Aug 3Utility Dive, Aug 21
CBRE7,481 MW under construction in the eight primary North American markets in the first half of 2026, more than 80% preleased; vacancy 1.4%CBRE, Aug 27
Wood Mackenzie1,066 GW of requests to US utilities and grid operators, which would realistically commit to about 28%Capacity, Aug 25

We have not seen Bernstein’s note. Every Bernstein figure here is from media accounts, most of them Chinese-language, read in translation; where a figure appears in only one account the text names it. Accounts differ on scope (most say North America, some the US; Wallstreetcn cites Canadian projects) and none of this week’s accounts defines the gigawatt used; Investing.com, reporting the 488 GW edition on September 14, called it a US pipeline of nameplate capacity. “Credible” is Zhitong’s and Wallstreetcn’s word for the probability-weighted total; Cailian Press says “expected to be built”. The 198 GW is the stage factor’s part of the cut, which also trims projects under construction; projects not yet started lose 497 − 123 = 374 GW across all four factors. Reconciliation: 198 + 113 + 48 + 28 = 387; 586 − 387 = 199; 199 ÷ 586 = 34.0%. Shares of the cut: 198 ÷ 387 = 51.2%; 113 ÷ 387 = 29.2%; 48 ÷ 387 = 12.4%; 28 ÷ 387 = 7.2%. 113 ÷ 28 = 4.04; 28 ÷ 586 = 4.8%. By stage: 75 ÷ 87 = 86.2%; 123 ÷ 497 = 24.7%; 87 + 497 = 584 and 75 + 123 = 198, against 586 and 199. Editions: 136 ÷ 411 = 33.1%; 170 ÷ 488 = 34.8%; 199 ÷ 586 = 34.0%; 586 ÷ 411 = +42.6%. ZeroHedge’s account of the first note quotes a dataset of 492 GW and 135 GW credible (33%); Zhitong gives 411 GW and 136 GW for July, and the accounts do not reconcile the two. Builders: 33 + 20 + 15 + 10 = 78; 78 ÷ 199 = 39.2%; 82 ÷ 199 = 41.2%; 308 ÷ 586 = 52.6%; 33 ÷ 308 = 10.7%; 7 ÷ 142 = 4.9%. Scale: 199 ÷ 3.4 = 58.5; 199 ÷ 64.4 = 3.09; 586 ÷ 64.4 = 9.10. The 451 Research figure is grid power to US data centers and Bernstein’s base is its own count for North America, so the two multiples differ. Cost: 387 × 7 = 2,709 and 387 × 15 = 5,805, $ billions; Wallstreetcn rounds these to $3–6 trillion. Texas: 43 ÷ 150 = 28.7%; 474 ÷ 150 = 3.16. On-site power: 47 ÷ 231 = 20.3%. A separate Bernstein note of November 6 estimated the election’s effect directly; it is listed in the sources and its figures are not mixed with these.

Wire reports are written within hours of the news, on a six-hour schedule. Each lists its sources and shows the arithmetic behind its figures. Not investment advice.