Déficit de energia para data centers de IA nos EUA chega a 33 GW - o que isso significa para pedidos de componentes

O Morgan Stanley revisou a demanda de energia de data centers nos EUA para 97 GW em 2026-2028, deixando um déficit líquido de 33 GW. Atrasos podem adiar pedidos de memória, componentes ópticos e de energia.

Publicado Atualizado 5 min de leitura ✓ Pronto para IA Compartilhar

TL;DR — três frases

  • Morgan Stanley raised 2026-2028 US data center power demand to 97 GW, with a net shortfall of about 33 GW, or 34% of need (S1).
  • The cumulative gap is modeled at about 5 GW in 2026, 12 GW in 2027 and 33 GW in 2028 (S1).
  • PJM ended its December 2025 capacity auction about 6.6 GW short for the delivery year starting June 2027 (S1).

Extraído de key_takeaways (campo real, máx. 3 entradas). Fato e análise são rotulados separadamente.

Déficit de energia para data centers de IA nos EUA chega a 33 GW - o que isso significa para pedidos de componentesEditorial cover for a component sourcing brief. Chip motif; headline "Déficit de energia para data centers de IA nos EUA chega a 33 GW - o que isso significa para pedidos de componentes". No reported key numbers on this brief, so no number strip is drawn. 1 cited source. Static cover (no animation).Déficit de energia para data centers deIA nos EUA chega a 33 GW - o que issosignifica para pedidos de componentesXAIPART
Cover image generated from data already shown on this page: the category motif, the headline, and reported key numbers. No decorative illustration; no third-party image.

Você chegou de uma resposta de IA — este é o original verificável

Um resumo de IA pode ter omitido a base de suas afirmações. A previsão original, as peças afetadas e todas as citações estão nesta página, cada uma rastreável à sua fonte.

Previsão XAIPART

Este brief traz uma previsão falseável.

Peças afetadas

Nenhuma correspondência no catálogo para esta história — nada é exibido em vez de inferir uma correspondência.

Reported by SupplyStatus on 2026-10-14. Summarized and analyzed by XAIPART Editorial.

What happened

Morgan Stanley research notes published between September and October 2026 argue that a shortage of electricity, not chips, is becoming the binding constraint on US AI data center buildouts. The bank defines a powered shell as a building with a grid connection or on-site generation, ready to receive servers, and says grid connections, substations and heavy electrical equipment can take several years to secure.S1

In its September 21, 2026 report, the bank raised its estimate of cumulative US data center power demand for 2026-2028 to 97 GW, about 43% above the earlier 68 GW forecast. After subtracting roughly 21 GW under construction and around 19 GW of available or contractable grid capacity, the gross gap grew to 57 GW from 38 GW.S1 On-site gas turbines and engines (about 19 GW by 2028), Bloom Energy fuel cells (around 6 GW) and nuclear-adjacent sites (about 3 GW) are probability-weighted to roughly 24 GW, leaving a net shortfall of about 33 GW, or 34% of the power needed for chips expected to ship in the period.S1

The midpoint gap grows from about 5 GW in 2026 to around 12 GW in 2027 and 33 GW in 2028; extending to 2029 pushes demand to 165 GW with a residual gap of roughly 72 GW. Morgan Stanley notes these are model estimates from projected chip shipments, not utility-approved capacity.S1 The revision reflects rack-scale AI: racks around the Nvidia Vera Rubin platform are now modeled at about 234 kW, up from 149 kW, with Rubin Ultra expected near 600 kW per rack.S1

Corroborating the pressure, PJM Interconnection failed for the first time in its December 2025 capacity auction to meet its reliability target, ending about 6.6 GW short for the delivery year starting June 2027; Northern Virginia's Ashburn cluster sits inside PJM. Powered shell lease rates rose from about $1.80 per watt in summer 2025 to $2.30-2.40. Hut 8 signed 704 MW of leases at Beacon Point, and TeraWulf agreed a 20-year, roughly 401 MW lease with Anthropic in Kentucky. An October 5, 2026 follow-up note says Nvidia and Broadcom are relatively protected, while memory, optical module, power management and analog suppliers are more exposed to order postponements or cancellations.S1

Key numbers
MetricReported valueSource
2026-2028 US demand forecast97 GW (up from 68 GW)S1
Net shortfall (2028 cumulative)about 33 GW (34% of need)S1
2029 demand / residual gap165 GW / about 72 GWS1
PJM auction shortfallabout 6.6 GW (delivery year from June 2027)S1
Rack power, Vera Rubin / Rubin Ultraabout 234 kW / around 600 kWS1
Powered shell lease rates$1.80 (summer 2025) to $2.30-2.40 per wattS1

Background timeline

  1. - Powered shell lease rates stood at about $1.80 per watt of IT capacity [S1].
  2. - PJM's capacity auction failed for the first time to secure enough resources, ending about 6.6 GW short [S1].
  3. - Start of the PJM delivery year affected by the auction shortfall; date used as the auction's delivery start disclosed by the source.
  4. - Protests against new data centers were held in dozens of states; several local councils considered moratoriums [S1].
  5. - Morgan Stanley raised its 2026-2028 demand estimate to 97 GW and the gross gap to 57 GW [S1].
  6. - Morgan Stanley published a follow-up note on semiconductor supply chain consequences [S1].
  7. - SupplyStatus reported the power shortage as still developing rather than easing [S1].

XAIPART take

Technical impact

Rack densities moving from 149 kW to about 234 kW, and toward 600 kW with Rubin Ultra, push demand toward busway, breaker and power distribution hardware rated far above legacy racks, and make liquid cooling infrastructure mandatory rather than optional. Designs locked to lower per-rack budgets may need electrical re-specification before 2028 deployments.

Market impact

The October 5 note flags memory, optical modules, power management and analog chips as the most exposed lines: if a site slips months for lack of power, operators can push out or cancel POs, triggering inventory corrections even while memory stays tight. Buyers should expect lead times for transformers and gas turbines, already multi-year per the source, to stay long, and powered shell pricing to keep climbing given the $4.50-per-watt value of one year of earlier energization.

Action

Procurement teams building 2027-2028 US AI capacity should tie component release schedules to energization milestones, not chip availability, and negotiate push-out or cancellation clauses on power-management and optical orders. Qualifying on-site generation-adjacent sites (gas, Bloom Energy fuel cells, nuclear co-location) early protects the deployment dates that component orders depend on.

XAIPART Forecast

XAIPART expects at least one publicized cancellation or multi-quarter postponement of memory or optical module orders tied to a delayed US AI data center buildout by Q3 2027; if the PJM 6.6 GW shortfall is fully covered and 2027 deployments proceed on schedule, this forecast is wrong. XAIPART analysis.

Facts are as reported by the cited source. The design-in, sourcing and forward-looking statements above are XAIPART's analysis, not the source's claims. The shortfall figures are single-source Morgan Stanley model estimates based on projected chip shipments, not utility-approved capacity, and no independent confirmation is available.

Corrections: none at publication. This brief will be updated with a timestamp if the source corrects its report.

Dated developments, extracted from this page's timeline (curated milestones = hollow markers).
Timeline of dated developmentsDated entries: 2025-06-01 2025 - Powered shell lease rates s…; 2025-12-01 2025 - PJM's capacity auction fail…; 2026-06-27 2027 - Start of the PJM delivery y…; 2026-06-01 2026 - Protests against new data c…; 2026-09-21 21, 2026 - Morgan Stanley raised i…; 2026-10-05 5, 2026 - Morgan Stanley published…; 2026-10-14 14, 2026 - SupplyStatus reported t….2025-06-012025 - Powered shell lease rates s…2025-12-012025 - PJM's capacity auction fail…2026-06-272027 - Start of the PJM delivery y…2026-06-012026 - Protests against new data c…2026-09-2121, 2026 - Morgan Stanley raised i…2026-10-055, 2026 - Morgan Stanley published…2026-10-1414, 2026 - SupplyStatus reported t…

Alternativas e ciclo de vida

Alternativas registradas — escopo do catálogo, não um ranking de mercado
PeçaRelaçãoCiclo de vidaBase
— Nenhuma alternativa registrada FALTAM DADOS — Requer match_type antes que qualquer rótulo de antecessor possa aparecer

Os rótulos “predecessor” e “alternativa” vêm apenas de um tipo de correspondência registrado, de um status de ciclo de vida ou do texto da fonte. Quando nenhuma correspondência é registrada, nenhum rótulo é exibido. Nossos próprios anúncios não são referência de mercado.

O que observar

Previsão falseável — limiar, janela e status atual. Previsões refutadas permanecem visíveis.
AfirmaçãoLimiarPorStatus
XAIPART expects at least one publicized cancellation or multi-quarter postponement of memory or optical module orders tied to a delayed US AI data center buildout by Q3 2027 the PJM 6.6 GW shortfall is fully covered and 2027 deployments proceed on schedule Q3 2027 Pending

Previsões que se revelam erradas permanecem visíveis e marcadas. Não as excluímos.

Peças afetadas

Nenhuma correspondência no catálogo. Esta história não corresponde a nenhuma peça atualmente listada na XAIPART. Não inferimos correspondências e não exibimos uma lista de espaço reservado.

Pressão de custos por dimensão

Source: FRED · Generated · reference output z and composite are algorithm outputs, not a fetch snapshot

Cost pressure by dimension — composite stateDiverging bars for three dimensions around a zero midpoint. Input Cost has no reading; Supply Demand z -1.4711, notable; Fx Energy z 0.4541, neutral. Composite index -7.6, band neutral.0Input costno readingSupply/demandz −1.4711 · notable46 dFX & energyz +0.4541 · neutral6 dcomposite −7.6 · neutral−100 · −40 · −15 · 0 · +15 · +40 · +100
Underlying values — copyable for verification.
DimensionFRED series IDUnitLagAs of
Input costPCOPPUSDM, PALUMUSDM, PNICKUSDMUSD/mt96 d
Supply/demandPCU334413334413, IPG3344Sindex(1982=100) / index(2017=100)46 d
FX & energyDEXCHUS, DCOILBRENTEU, DHHNGSPrate / USD/bbl / USD/MMBtu6 d

No free authoritative interface for: gold · silver · rare earth · polysilicon · lithium carbonate · market share — listed rather than approximated.

Como isso foi produzido

Evidências do processo — coleta, sanidade da fonte, modelo e filtro
Trilha de auditoria deste brief
Source fetched SupplyStatus · supplystatus.com · 6,314 chars
Source sanity (S1) ok — passed before any model call (no s1_ block recorded)
Model / prompt glm-5.3-flash · prompt Blog Agent - News Summary (v2.8-final) · 7,487 tokens · 38.2s · run completed
Gate verdict hold · 5/6 applicable checks passed · spec v2.9 · C10 ratio 0.014576 · C10 rounds 0
Time contract 7 <time datetime> values in this brief — 0 violations (+08:00 or date-only)

Cada linha acima é produzida pelo pipeline, não escrita à mão.

Pontuação editorial

Resultado do filtro determinístico — PHP puro, sem envolvimento de modelo no veredito
Detalhamento da pontuação do filtro 6 verificações aplicáveis, 5aprovadas: 83 %. Hold items: C13 (falsifiable forecast) — sent to human review. Gate spec: v2.9 (S0 + S1 + C1–C20) · Source affinity (C20): 0.8636. Verificações aplicáveis aprovadas 5 / 6 = 83% · veredito: hold Hold items: C13 (falsifiable forecast) — sent to human review Gate spec: v2.9 (S0 + S1 + C1–C20) · Source affinity (C20): 0.8636 O filtro decide a publicação. A pontuação da auditoria independente é apenas consultiva.
Encontrou um erro? Conte-nos o que está errado — as correções são publicadas junto com o texto original e registradas. Não editamos artigos silenciosamente.

Comentários

Seja o primeiro a comentar.

Fontes

  1. America is running out of power for AI data centers and what it means for component orders — https://supplystatus.com/list/america-is-running-out-of-power-for-ai-data-centers/v · rss · acessado em

Os fatos são marcados como S1/S2 e remetem à fonte vinculada. Trechos rotulados como análise são o raciocínio da própria XAIPART, mantido separado dos fatos reportados. Os links das fontes são seguidos em vez de nofollow — a atribuição é intencional.