Massive datacenter buildouts continue

Last updated

Tracker summary: OpenBrain is building the biggest datacenters the world has ever seen. Annual spending on datacenters doubles to $400 billion, led by Microsoft, Google, and Amazon, and the US adds over 5 GW of AI power draw.

At a glance

  • Assessment: Confirmed
  • Confidence in assessment: 95%
  • Predicted timing: Through 2026
  • Primary source: Late 2025: The World’s Most Expensive AI, archived PDF p. 3; Early 2026: Coding Automation, archived PDF p. 6

Assessment scope: Confirmed (95%) applies to the broad datacenter buildout trend. It does not confirm the scenario’s exact cluster, spending or US AI-power quantities. Operating infrastructure, contracted capacity and planned capacity are assessed separately; the numerical checkpoints below remain unresolved.

What AI 2027 Predicted

OpenBrain is building the biggest datacenters the world has ever seen.

Source: Late 2025: The World’s Most Expensive AI: archived PDF, p. 3. Its footnote describes a distributed cluster with 2.5 million H100-equivalent GPUs, $100 billion spent and 2 GW online, with construction underway to at least double it through 2026.

Annual spending on datacenters doubles to $400 billion, led by Microsoft, Google, and Amazon, and the US adds over 5 GW of AI power draw.

Source: Early 2026: Coding Automation, footnote 16: archived PDF, p. 6, describing growth over 2025.

The scenario dashboards label annualized datacenter spending at $458B/year in April 2026 and $524B/year in August 2026 (p. 6), then $600B/year in December 2026 (p. 8). These are scenario quantities, not observed actual spending.

Tracker addition: power-grid bottlenecks are contextual indicators used here to assess the feasibility and pace of construction; they are not part of either quoted scenario claim. Planned capacity, investment commitments and grid queues do not by themselves establish the scenario’s realized spending, installed hardware or operational AI power quantities.

How We Track This

  • Compare realized, consistently scoped datacenter spending with the scenario’s $400B annual-spending statement and its $600B/year December 2026 dashboard pace; annual totals and annualized point-in-time rates are distinct.
  • Compare operational US AI power additions with the over-5-GW quantity; distinguish power actually online from planned campus capacity and connection queues.
  • Assess the large-cluster scale described on p. 3 separately from worldwide spending.
  • Track power-grid constraints as contextual feasibility evidence added by the tracker, not as an independently predicted outcome.

Broad-trend confirmation rests on actual construction and operating infrastructure, not on treating all of these quantities as a conjunctive pass test. The evidence below also includes plans and commitments, but not a complete matched series for the numerical targets. None is newly resolved by this scope clarification.

Current Evidence

Dated assessment: November 9, 2026

Retained Confirmed (95%) for the broad buildout. Crusoe reports 1 GW delivered and operational within more than 6 GW of gross contracted capacity. Meta’s October 19 engineering account describes a completed first phase that has trained models. Operating infrastructure supports the broad assessment independently of announced campuses. Those issuer descriptions do not establish a matched national power series, a global spending total or the scenario’s exact late-2025 cluster scale.

Scenario checkpointResult in the cited evidence as of November 8
Late 2025: 2.5 million H100-equivalent GPUs, $100B spent and 2 GW online in the distributed clusterUnresolved as a combined, contemporaneous cluster measurement. Later delivered capacity and contracted pipeline cannot reconstruct it.
2025: $400B annual datacenter spending and doublingSeparately unresolved on a comparable datacenter-only basis; later 2026 spending cannot fulfill the fixed 2025 total.
2025: over 5 GW of added US AI power drawSeparately unresolved; operating US AI-only additions are not interchangeable with planned campuses or contracted capacity.
April 2026: $458B/year spending paceElapsed and unresolved; the evidence does not supply a matched annualized observation.
August 2026: $524B/year spending paceElapsed and unresolved on the same basis.
December 2026: $600B/year spending paceFuture endpoint; not confirmed by current guidance.

These quantities are summarized from the archived scenario, pp. 3, 6 and 8. They remain separate from the broad-trend rating and from the tracker’s grid-bottleneck context.

The UK Treasury’s October 28 Data Centre Covenant estimates a need for 6 GW of AI data-center capacity by 2030. It announces a Ministerial Control Centre and plans to reserve or reallocate grid connections, transfer pre-2030 connection contracts and develop tightly controlled temporary-generation access for exceptionally strategic projects. Local-benefit and environmental tests qualify projects for support. Further strategic-site criteria and the temporary-power framework remain forthcoming. This is an announced delivery policy and future need estimate, not evidence of energized capacity, installed accelerators or completed global buildout. Data Centre Covenant, opening estimate and government actions

Earlier reporting documented operational buildings and continuing construction at Abilene; those dated counts are not a refreshed October operating inventory. Five more Stargate sites have been announced. Bloomberg and Reuters reported in March 2026 that Oracle and OpenAI dropped a planned ~600 MW expansion lease at Abilene. Oracle called the reports false and said the 4.5 GW agreement is on track. Reuters, March 6, 2026 (syndicated copy); Oracle statement reported March 9, 2026. The disputed expansion lease is distinct from the existing campus and the wider capacity agreement. Power grid constraints are increasingly cited as a bottleneck; this is tracker-added context, not a quoted scenario prediction.

Reuters/Bloomberg reported in May 2026 that Microsoft’s AI data center power demand may force the company to delay or abandon its 2030 goal of matching all hourly electricity use with renewable purchases. The report says expensive, energy-intensive AI data center construction is reshaping the feasibility of pre-AI climate commitments and notes that some new data centers may require multiple gigawatts of capacity. This reinforces that power availability and energy procurement are central bottlenecks in the AI buildout.

A 2026-05-27 Axios report said Microsoft, Google, Amazon, and Meta are joining an Elemental Impact initiative using data centers as testbeds for climate and energy technologies, including advanced cooling, energy storage, and low-carbon building materials. This is not a capacity announcement, but it is further evidence that AI data-center growth has become large enough for hyperscalers to coordinate around energy, emissions, and infrastructure constraints.

IEA’s 2026 energy and AI analysis reports that global data-center electricity demand grew 17% in 2025 and that reliable onsite gas generation for critical, variable data-center load can require overbuilding generation by 30% to 70% relative to demand. This strengthens the tracker’s power-bottleneck evidence for continued AI data-center buildout.

Local opposition is now a material friction point for data-center buildout. Data Center Watch estimates that $18B in US data-center projects were blocked and another $46B delayed amid opposition over the prior two years, while cautioning that data-center delays have multiple intertwined causes. The Coweta County, Georgia referendum campaign against a more-than-800-acre project shows large projects facing organized local opposition. The overall buildout trend remains confirmed, but permitting, local politics, water concerns, and power concerns may slow some announced projects.

New York imposed what the state describes as the first statewide moratorium on new hyperscale data centers. For up to one year, the Department of Environmental Conservation will withhold discretionary environmental permits that had not already been deemed complete while the state develops standards covering energy demand, water, air quality, grid effects, and ratepayer impacts. This is a binding state-level constraint on buildout speed, while the confirmed national construction trend continues.

July announcements add planned gigawatt-scale capacity and operating manufacturing capacity. AMD and Anthropic announced up to 2 GW of planned MI450 deployment, with the first gigawatt scheduled to begin deployment in the first half of 2027. NAVER, NVIDIA, and Brookfield proposed expanding a Korean AI factory beyond its current 55 MW, with 200 MW planned by 2028 and a path toward gigawatt scale. NVIDIA plans a $1B investment subject to conditions, and Brookfield entered a nonbinding term sheet for up to $9B. Wistron opened a 324,000-square-foot Fort Worth facility that produces GB300 systems and is preparing for Vera Rubin production as part of a $700M manufacturing investment. The future power figures remain plans rather than operational 2026 capacity.

New US and EU plans couple AI compute expansion to energy and sovereign-capacity policy. DOE announced a planned 1.8 GW Western Kentucky AI/HPC campus with 2 GW of new gas generation and up to 2.6 GW of batteries. The European Commission opened a tender for up to seven AI Gigafactories. The Kentucky power agreement still requires regulatory approval and is expected to complete in 2031. None of these figures count as operational 2026 capacity.

JLL’s August midyear report provides an aggregate measure of North American construction: 25 GW of H1 absorption, 66 GW under construction, and 95% of that construction pre-committed, with vacancy at 1%. Absorption and commitments measure demand rather than operating electrical load, and JLL attributes demand to hyperscalers, neoclouds, and pure-play AI companies rather than AI alone. New South Wales separately linked a 75-day assessment commitment to energy, water, environmental, community, skills, and infrastructure requirements. NSW says 19 state-significant projects worth A$50.3 billion are in its pipeline, not completed investment.

NVIDIA committed credit support for the initial 4.25 IT-GW phase of SB Energy’s PORTS-Pike campus, named OpenAI as the customer under a 20-year lease, and invested $1.5 billion in SB Energy. The campus is planned capacity rather than operational load. Pennsylvania separately removed AI data centers from Fast Track permitting and required local approval plus binding grid, generation, transparency, and community commitments for projects seeking state permits or support. Together, the actions document continued gigawatt-scale contracting and tighter infrastructure constraints.

Crusoe reported more than 6 GW of gross contracted capacity across data centers and cloud, including 1 GW delivered and operational. The distinction matters: 1 GW is the stated operating portfolio figure; the remainder is contracted pipeline. The figures strengthen the documented buildout without resolving national AI-only power or construction totals.

Meta’s October 19 engineering report says Prometheus’s first phase is complete, has trained Muse models and currently comprises hundreds of thousands of GPUs; roughly one million GPUs remains its full-scale target. NEXTDC separately disclosed suspended network works at M3 Melbourne and delayed expansion permits, with the next grid increment expected in FY28. Data-hall construction continues and affected capacity remains contracted subject to customer rights, including credits or termination. Existing supply cannot reliably support the full ramp even in delivered halls. This is a project-specific constraint, not evidence that all M3 halls are unpowered or that development has stopped industry-wide. The reports distinguish operating infrastructure, planned scale and constraints on powering contracted capacity; they do not supply a global AI-only operating-load total.

Sources:

Counterevidence & Limitations

  • Bloomberg and Reuters reported in March 2026 that Oracle and OpenAI dropped a planned ~600 MW expansion lease at Abilene. Oracle called the reports false and said the 4.5 GW agreement is on track. Reuters, March 6, 2026 (syndicated copy); Oracle statement reported March 9, 2026. The disputed expansion lease is distinct from the existing campus and the wider capacity agreement.

  • Some planned expansions have been scaled back or delayed (the disputed Abilene expansion lease; see the dated report and rebuttal below), raising questions about whether announced capacity will fully materialize

  • Power grid constraints could slow buildout significantly below planned rates; some projects face multi-year waits for grid connections

  • Not all announced capacity will necessarily be used for AI training vs. inference, and the mix matters for the scenario’s capability predictions

  • Announced ≠ built ≠ operational: there is a significant gap between press releases and actual compute online

  • If AI revenue growth disappoints, some of these capital-intensive projects may be delayed or cancelled, similar to overbuilding patterns in previous tech cycles

  • On August 3, Texas directed the PUCT and ERCOT to audit every data center advancing through ERCOT’s interconnection process before any project may proceed. Texas cited more than 474 GW of requests, over five times its record peak demand, and said about 90% of new power requests are data centers. The audit covers grid dependence, onsite generation, water, cooling, incentives, ownership, and community impacts. This binding verification gate shows how grid and local-resource constraints can delay planned capacity, while leaving the confirmed national buildout trend intact.

What Would Change Our Assessment

  • Downgrade: Multiple major projects cancelled; significant pullback in construction
  • Maintain: The buildout is clearly happening at massive scale

Update History

DateUpdate
2026-11-09Assessment reviewed. Status: Confirmed. Confidence: 95%. Retained Confirmed (95%) for the broad buildout only. Crusoe’s delivered operational capacity and Meta’s completed first phase support the broad trend. Separated operational from contracted/planned capacity and left the late-2025 cluster, 2025 spending/power and April/August 2026 quantities unresolved; December remains a future endpoint. The numerical subtargets are not jointly confirmed.
2026-11-07Correction: Corrected the reported Abilene expansion from 2 GW to about 600 MW and paired the March report with Oracle’s rebuttal, separating that lease from the wider 4.5 GW agreement. Corrected the April 13 history: the five-site announcement was September 23, 2025, posted by SoftBank September 24; nearly 7 GW includes Abilene and CoreWeave projects, and 1.5 GW applies to two SoftBank sites together, not each of five sites. Restored the scenario’s archived page-3/page-6 quantities and separated grid constraints as tracker context. Planned capacity and commitments do not establish realized spending, installed hardware or operational AI power.
2026-11-02Added the UK’s proposed grid-allocation and conditional temporary-power measures, with its 2030 capacity-need estimate kept separate from operating capacity. Implementation details remain forthcoming; status and confidence unchanged.
2026-10-26Added Meta’s completed Prometheus phase and NEXTDC’s M3 power/permitting delays. Current infrastructure, future scale and contracted power ramps remain separate; status and confidence unchanged.
2026-09-21Crusoe reported more than 6 GW contracted and 1 GW delivered and operational across its portfolio. The operating figure is material; contracted pipeline remains separate. Status and confidence unchanged.
2026-08-24NVIDIA backed an initial 4.25 IT-GW PORTS-Pike phase for an OpenAI lease, while Pennsylvania removed AI data centers from Fast Track permitting and imposed grid and local-approval requirements. These actions reinforce both the buildout and constraint mechanisms. Status and confidence remain unchanged.
2026-08-17JLL reported 25 GW of North American H1 absorption and 66 GW under construction, 95% pre-committed. NSW also linked faster data-center assessment to energy, water, community, and infrastructure requirements. These developments reinforce both the buildout and constraint mechanisms, with status and confidence unchanged.
2026-08-10Texas required an audit before any data center may advance through ERCOT’s interconnection process, citing more than 474 GW of requests and saying about 90% of new power requests are data centers. This adds a binding grid and resource constraint without changing the confirmed national buildout assessment.
2026-08-03DOE announced a planned 1.8 GW Kentucky AI/HPC campus paired with new generation and storage, while the EU opened a tender for up to seven AI Gigafactories. These plans reinforce the buildout trend, but remain conditional or pre-award and do not count as operational 2026 capacity.
2026-07-27AMD and Anthropic announced up to 2 GW of planned MI450 capacity, NAVER and partners proposed a Korean buildout toward gigawatt scale, and Wistron opened a $700M US AI-system factory. These developments reinforce the confirmed buildout trend while leaving future capacity and conditional financing uncounted as operational infrastructure.
2026-07-20New York imposed a statewide pause of up to one year on discretionary environmental permits for new hyperscale data centers while it develops energy, water, grid, and environmental standards. This adds a binding state-level constraint on buildout speed. Status and confidence remain unchanged.
2026-06-15Added evidence that local opposition is delaying some data-center projects, including a Coweta County referendum campaign and Data Center Watch’s estimate of $64B in US projects blocked or delayed amid opposition. This is a limitation on buildout speed, not counterevidence to the overall confirmed buildout trend.
2026-06-08IEA reported that global data-center electricity demand grew 17% in 2025 and that reliable onsite gas generation for variable data-center load can require 30% to 70% overbuild relative to demand. This reinforces that power infrastructure remains a central constraint for AI data-center expansion.
2026-06-01Added Axios reporting on a Microsoft/Google/Amazon/Meta data-center climate initiative. This reinforces the confirmed assessment that energy and emissions constraints are now central to the AI data-center buildout.
2026-05-11Reuters/Bloomberg reported that Microsoft’s AI data center power demand may force revision of its 2030 clean-energy target, underscoring that grid and energy constraints are materially affecting hyperscaler planning. This reinforces the existing confirmed assessment; confidence unchanged because it is already at 0.95.
2026-04-13Stargate expanding to multiple GW-scale campuses. OpenAI, Oracle, and SoftBank announced five new US Stargate sites (March 2026) with nearly 7 GW planned capacity and $400B+ investment. Abilene campus deploying 450K GB200 GPUs per IntuitionLabs. New sites can scale to 1.5 GW each. UAE Stargate expected 2026. GW-scale datacenter buildout confirmed and accelerating. Sources: OpenAI blog, IntuitionLabs
2026-03Datacenter construction at unprecedented scale globally. Power availability confirmed as primary constraint, with utilities and regulators scrambling to keep pace.
2026-01Nvidia Rubin platform announced at CES: ~5× inference performance vs. Blackwell, entering full production with H2 2026 availability. Meta guided $116–118B infrastructure for 2026. Power described as “primary supply constraint” across the industry.
2025-07Stargate ($500B commitment) and competing mega-datacenter projects announced. Power grid constraints emerging as real bottleneck in multiple US regions.
2025-05Stargate UAE: G42, OpenAI, Oracle, Nvidia, SoftBank, and Cisco announce 1 GW Abu Dhabi campus expected to open 2026 (announced May 22). International Stargate expansion confirms the buildout is geographically distributed and accelerating.