Skip to content
l0gIndependent analysisEnglish edition
Start here Support l0g
ENFR

// analysis

TSMC follows chip demand all the way to the power grid

Illustration for the analysis: TSMC follows chip demand all the way to the power grid

Data center power schedules are entering TSMC’s industrial decisions. An investigation into contracts, forecasts and the way infrastructure risk can reach semiconductor factories.

20 min read#TSMC#Taiwan#Asia#Artificial intelligence#Data centers#Electricity#Semiconductors#Risk
dated revision: October 21, 2026French originalprimary sourcesno tracker
// reading pathAsia in the global financial systemFollow the rates, currencies, credit and capital flows around this analysis.

The company manufacturing chips for the AI industry is taking a closer look at the buildings that will eventually house them. On 15 October 2026, TSMC chief executive C. C. Wei said the company was sending its own teams to check construction progress and power supply at US data centers. He described the feedback as favorable and assessed demand as still exceeding TSMC’s supply. The revealing detail is how far this work reaches: decisions about semiconductor factories now draw on the physical readiness of infrastructure used by its customers’ customers. 1

A processor order can have a budget behind it, a signed contract and a real need for computing capacity, while still depending on a grid connection whose schedule needs to be secured. Businesses throughout the chain incur costs before the whole installation becomes operational. A delay can then travel through delivery dates, cash requirements and, eventually, subsequent equipment orders. Following that process gives a more useful view of the AI investment cycle: an industrial forecast is only as executable as the projects, connections and supply agreements on which it rests.

Public documents reviewed through 21 October make it possible to reconstruct this mechanism. They also provide meaningful reasons for caution about a uniformly pessimistic interpretation. Some commitments have financial protection, buyers have very different resources, and TSMC continues to invest. The relevant questions concern which orders can turn into deployed capacity, when that can happen and who bears the cost of a change in schedule.

Looking beyond the immediate customer

A foundry such as Taiwan-based TSMC manufactures circuits designed by its customers. In the semiconductor supply chain, those components must subsequently be incorporated into equipment and installed in a working computing facility. The manufacturer and the cloud operator therefore earn revenue at different stages, under their respective contracts. A chip may already have been sold while the facility intended to use it is still being prepared. 4

That separation creates a reason to investigate demand beyond the immediate buyer. A customer may reserve additional components to secure its expansion, while a cloud operator anticipates strong demand for services. Each decision can make sense in its own setting. Adding them together still requires scrutiny: projects may target overlapping customers, depend on shared infrastructure or assume opening dates that cannot all be reconciled with the necessary construction work. These are possibilities to investigate at project level, rather than conclusions that follow from the total value of announcements.

The checks discussed in October build on a process TSMC had already described on 16 July. It was then examining data center projects and server racks to assess where the equipment would be deployed. That history matters: the company is extending a form of demand verification it had already explained. The public material reviewed does not identify the sites visited in October or provide the inspection reports. The favorable assessment of those visits remains management’s account. 2 1

From silicon to billable computingDiagram: TSMC manufactures circuits, suppliers assemble GPUs and systems, cloud operators deploy them and deliver computing. The electricity grid supplies the site. TSMC says it checks US construction and power. Conditional risk arrows show possible upstream order delays after persistent site delays; compatible reallocation can cushion the effect.l0g / 0121.10.2026From silicon to billable computingDeliveries, power and demand checksTSMC checks sites and powerTSMCCircuitsSupplierGPUCloudServersServiceBillableElectricity gridSite powerIf site delays persistInventories and orders may shiftupstream under the contracts.Other demand: reallocation possible if compatible.Sources:S01S02S04
From silicon to billable computingDiagram: TSMC manufactures circuits, suppliers assemble GPUs and systems, cloud operators deploy them and deliver computing. The electricity grid supplies the site. TSMC says it checks US construction and power. Conditional risk arrows show possible upstream order delays after persistent site delays; compatible reallocation can cushion the effect.l0g / 0121.10.2026From silicon to billablecomputingDeliveries, power and demand checksTSMC checks sites and powerTSMCCircuitsSupplierGPUCloudServersServiceBillableGridSite powerIf delays persist:orders mayshift upstreamReallocation possible if compatibleSources:S01S02S04
01 / From silicon to billable computing GPUs are graphics processors suited to parallel computing. Physical deliveries, demand checks and the possible transmission of delay follow different routes. TSMC describes its checks and favourable feedback; the risk loop represents a conditional mechanism. 1, 2, 4

Electricity forecasts also begin with chips

The latest national assessment identified in this investigation is Berkeley Lab’s United States Data Center Energy Usage Report: 2025 Update, published on 18 June 2026. The year in the title describes the update, while the publication took place the following year. The researchers estimate that US data centers consumed 192 terawatt-hours in 2024. Their reference scenario reaches 649 TWh in 2030, with the scenarios combining extreme assumptions spanning 521 to 843 TWh that year. 5

These figures measure energy consumed over a year. Gigawatts quoted for a site describe power: the amount it can draw at a given moment. Converting between the two requires assumptions about operating hours, equipment load and the facility’s additional needs. Location introduces another distinction. National electricity consumption can grow while a particular project encounters a transmission or connection constraint, far from resources available elsewhere.

The Berkeley report is especially useful for its method. It starts partly with server and accelerator shipments, then applies assumptions about power requirements, utilization, equipment lifetime and supporting infrastructure. The authors explain that their scenarios leave out major disruptions associated with factors such as grid power availability, regulation and supply chains. Even the historical estimates are reconstructed from equipment data, rather than from a comprehensive census of electricity meters. 5

This matters when interpreting financial forecasts. An electricity projection built from chip shipments estimates the energy requirements associated with an equipment deployment path. Independent scrutiny of its feasibility calls for different evidence: a location, a connection, a credible completion date and financing. TSMC’s reported visits seek information of that kind. This comparison is our interpretation of the documents; there is no public indication that TSMC uses the Berkeley model.

From servers to projected electricityLinear scale from 0 to 900 TWh per year. The 2024 estimate is 192 TWh. For 2030, the reference case is 649 TWh, with compounded-assumption scenarios from 521 to 843 TWh. The model starts with equipment shipments, then power, utilisation, lifetime and facility overhead. Major disruptions in grid power availability are not modeled.l0g / 0221.10.2026From servers to projected electricityUnited States · data centers · TWh per year2024Estimate192 TWh2030Scenarios649 TWhreference5218430200400600800900IT shipmentsServersstorage,networkInstalled fleetPower,lifetime and useElectricityIT × PUEMajor grid disruptions are not modeled.Sources:S05
From servers to projected electricityLinear scale from 0 to 900 TWh per year. The 2024 estimate is 192 TWh. For 2030, the reference case is 649 TWh, with compounded-assumption scenarios from 521 to 843 TWh. The model starts with equipment shipments, then power, utilisation, lifetime and facility overhead. Major disruptions in grid power availability are not modeled.l0g / 0221.10.2026From servers to projectedelectricityUS data centers · TWh/year2024192 TWh2030649 TWhreference5218430300600900ITshippedFleetW · lifeuseEnergyIT × PUEMajor grid disruptions are notmodeled.Sources:S05
02 / From server shipments to TWh The 521–843 TWh range is a sensitivity exercise combining extreme scenario assumptions. Equipment shipments feed the projection; project delivery conditions require separate local scrutiny. PUE is total facility energy divided by IT equipment energy, over the same period and scope. 5

Dominion puts megawatt requests through several filters

Dominion’s presentation to PJM on 29 September shows how a utility handles the pipeline. For contracts executed as of July 2026, it distinguishes electric service agreements (ESA), construction letters of authorization (CLOA) and engineering letters of authorization (ELOA). The three categories account for 11,997, 9,356 and 32,473 megawatts, respectively. Their sum, 53,826 MW, brings together commitments at different stages and with different schedules. 6

The engineering authorizations make the filtering visible. Of the associated 32,473 MW, Dominion includes 13,429 MW in its forecast from 2032 onward and excludes 19,044 MW. The model then considers energization dates and the gradual increase in load. Dominion reports an average five-year ramp period and a mature demand-to-contracted-capacity ratio of 63%. Those parameters describe its electricity forecasting method; they do not measure processor activity. 6

The meaning of the headline number changes as a project advances. Engineering work prepares a possible connection. A construction authorization commits the parties to building infrastructure. A service agreement governs supply. Actual electricity demand then depends on the site’s deployment and operation. A useful grid forecast emerges from this sequence, with adjustments for timing and load, rather than from adding requested capacity alone.

PJM, which coordinates this regional electricity system, also describes how a campus enters the load forecast. Its 29 September article discusses developer commitments and the handling of duplicate requests. It receives information from electricity suppliers for its forecasts, adding another stage of collection and verification between a project’s initial request and the system-wide outlook. 7

The gap between initial requests and the retained forecast reflects the maturation and screening of projects along the way. A property or power option can be abandoned while the computing demand it was intended to serve moves elsewhere. For the chip manufacturer, the deployment’s destination and timing remain central, beyond the fate of one address.

Dominion screens the projectsThree zero-based bars show 11,997 MW under ESAs, 9,356 MW under CLOAs and 32,473 MW under ELOAs. ELOAs split into 13,429 MW included in the model and 19,044 MW excluded. The model then applies energisation dates, an average five-year ramp and a 63% demand-to-contract-capacity ratio. Firm is Dominion’s classification.l0g / 0321.10.2026Dominion screens the projectsContract capacity · July 2026 · MWESAFirm11,997CLOAFirm9,356ELOA32,473010,00020,00030,00013,429 MW included19,044 MW excludedPowerdate5 yearsramp63%demandIncluded ELOAs from 2032; Firm is a model classification.Sources:S06
Dominion screens the projectsThree zero-based bars show 11,997 MW under ESAs, 9,356 MW under CLOAs and 32,473 MW under ELOAs. ELOAs split into 13,429 MW included in the model and 19,044 MW excluded. The model then applies energisation dates, an average five-year ramp and a 63% demand-to-contract-capacity ratio. Firm is Dominion’s classification.l0g / 0321.10.2026Dominion screens theprojectsContract capacity · July 2026 · MWESAFirm11,997CLOAFirm9,356ELOA32,473010,00020,00030,00013,429 MW included19,044 MW excludedIncluded ELOAs: from 2032Powerdate5 yearsramp63%demandMature demand / capacitySources:S06
03 / Dominion screens projects before forecasting Contract capacities cover existing loads and future energizations at different stages. Included ELOAs enter the model from 2032 onwards. Timing and load ramp then help construct the demand forecast. 6

Contracts determine where the cost lands

Power requests come with financial protections. In its May 2026 report, Dominion describes an initial $250,000 engineering payment, with further payments possible. The construction stage triggers deposits and, under the arrangement it describes, reimbursement of costs incurred if the project is cancelled before energization. The utility presents a sequence in which customer obligations increase as the development becomes more concrete. 8

Virginia’s regulator also confirms stricter contractual requirements for Dominion’s large-load customers in the state. New customers contracting for service from 1 January 2027 will face a minimum fourteen-year term, together with collateral requirements depending on credit quality. A long-term payment obligation helps the utility recover its costs even when the electricity consumption originally expected takes longer to materialize. 9

Understanding risk transmission therefore requires identifying the beneficiary of each protection. Reimbursement for construction can protect the utility while leaving a developer exposed to land, a building or equipment whose revenue arrives later. A guarantee can reassure a lender by moving an obligation to another company. Across the supply chain, exposure is distributed among balance sheets and contracts; the economic cost still depends on whether assets can be used or redeployed.

Contracts for reserved grid capacity give substance to headline gigawatt announcements. A request carrying an expensive exit obligation provides different information from an engineering-stage request. Its actual force must be established from the relevant agreement: a forecasting model’s classification of a project as “firm” first describes the category assigned by the model’s author.

Reserved capacity still has to become an available service

CoreWeave’s disclosures illustrate the distinction between stages of development. At 30 June 2026, it reported 1.5 GW of active power and approximately 3.7 GW of contracted power. Its revenue backlog, an indicator of future revenue under customer contracts, stood at about $104 billion, subject to delivery and service-availability requirements. These figures were published on 11 August. 10

The comparison is useful when its definitions are preserved. Active power describes an operational position under the company’s metric; contracted power concerns capacity covered by its agreements. Identifying delayed projects requires additional information, site by site. The financial issue lies in the service condition: the promised capacity has to become available before it can support the expected revenue under the contracts.

On 17 September, CoreWeave subsequently reported approximately 4.2 GW of contracted power as of 11 August. That release did not provide an active-power figure for the same date. The illustration therefore keeps the two June observations together and presents the later update separately. 11

CoreWeave: power and deliveryAt 30 June 2026, CoreWeave reports 1.5 GW of active power and approximately 3.7 GW of contracted power, shown as separate bars on a common scale. The 17 September release reports approximately 4.2 GW contracted as of 11 August, without specifying active power as of 11 August. The approximately USD104 billion revenue backlog at 30 June remains subject to delivery and service-availability requirements.l0g / 0421.10.2026CoreWeave: power and delivery30 Jun 2026 · reported power · GWActive1.5 GWContracted~3.7 GW01234~4.2 GW contracted · 11 Aug 2026Active power on 11 August: unspecified30 Jun 2026 backlog → performance → revenue~104USD billionConditionsDelivered, availableRevenueAs fulfilledSources:S10S11
CoreWeave: power and deliveryAt 30 June 2026, CoreWeave reports 1.5 GW of active power and approximately 3.7 GW of contracted power, shown as separate bars on a common scale. The 17 September release reports approximately 4.2 GW contracted as of 11 August, without specifying active power as of 11 August. The approximately USD104 billion revenue backlog at 30 June remains subject to delivery and service-availability requirements.l0g / 0421.10.2026CoreWeave: power anddelivery30 Jun 2026 · power (GW)Active1.5 GWContracted~3.7 GW01234~4.2 GW · 11 Aug 2026Contracted; active unspecified~104USD bn30.06DeliveryavailableRevenueasfulfilledConditional backlog conversionSources:S10S11
04 / CoreWeave, from contract to delivered service The two power measures retain their definitions and date. Attributing their gap to delayed capacity requires project-level evidence. Revenue backlog combines remaining performance obligations and other estimated future contract revenue; conversion depends on execution. 10, 11

A recent example shows how power schedules can change a customer’s decision. In October 2025, Nscale announced a 50 MW project at Loughton in the United Kingdom, intended to host roughly 23,000 GB300 GPUs from the first quarter of 2027. 14 Reuters reported on 19 October 2026 that Microsoft had withdrawn after grid access slipped into the 2030s; Microsoft and Nscale say their other agreements remain in place. 15

Loughton is a British project, whereas TSMC’s reported visits concern the United States. It illustrates a local decision and the possibility of reorganizing deployment. The exit agreement and the grid operator’s revised schedule would be needed to allocate the costs in detail.

How a delay can reach the foundry’s investment decisions

NVIDIA explicitly identifies the relevant link in its quarterly report for the period ended 26 July 2026. The company says a lack of available data center infrastructure or customer financing can delay purchases of new architectures. In the same document, it describes supply commitments made to meet demand over several years. Sales schedules and industrial obligations can consequently move out of alignment. 12

If a site falls behind schedule, its operator may seek another location or renegotiate delivery dates. If the supplier can allocate equipment to another buyer, aggregate component demand may hold up. If inventories accumulate or future requirements are revised, the timing of new orders can change. A problem must therefore pass through several commercial decisions before it reaches TSMC, with opportunities for absorption along the way.

TSMC’s investment program explains why this question matters financially. The company maintains a worldwide 2026 capital expenditure budget of $60–64 billion and reports $40.64 billion spent from January through September. The annual budget was raised in July. These figures cover the group’s overall investment, extending beyond AI applications and US factories. 3 2

The annual report explains why commercial visibility and contractual protection can differ. Customers generally place product orders with limited advance notice, although some make payments to reserve capacity. The treatment of those payments depends on the agreements reached. The report also describes a substantial fixed-cost base and the way factory utilization changes the allocation of those costs across units produced. 4

That relationship gives on-site verification its economic purpose. A factory requires decisions taken well in advance, while its future markets continue to evolve. If new capacity becomes available when orders shift to later dates, fewer units must absorb the same fixed costs. Adjustments may involve reallocating capacity between customers, changing future equipment purchases or altering the production ramp. Their scale depends on the length of the delay, the manufacturing processes involved and the availability of alternative demand.

The buffers matter as much as the bottleneck

Computing buyers have very different resources with which to absorb a delay. For the twelve months ended 30 June 2026, Microsoft reported $182.9 billion of operating cash flow and $115.9 billion of cash additions to property and equipment. Both figures cover the entire group. The second is a cash outflow measure that excludes newly entered finance-lease commitments. They demonstrate substantial funding generated by existing businesses, while leaving the economics of each individual project open for examination. 13

Location and timing create further opportunities to adapt. Subject to technical, contractual and data-location constraints, some computing work can move between sites or be deployed in stages. Grid work already underway and financial guarantees may also make a promised service date more credible. Delays are harder to absorb when several locations are affected together or an operator has little financing headroom. These are conditions to examine in individual cases; the sector contains a wide range of business models and exposures.

The strongest counterpoint is embedded in the statement that prompted this investigation: TSMC considers the feedback from its checks favorable and its assessed demand above supply. The exercise can help determine where to accelerate investment as well as where to avoid a mismatch in timing. The public sources reviewed establish no amount of TSMC orders or revenue lost because of US power constraints. 1

Following demand through to the service date

A useful reading of the evidence brings together the commercial commitment, the site’s stage of development, its power-availability date and the financial capacity of the parties. An order figure becomes more informative when it can be linked to those conditions. A connection schedule becomes more credible when the construction work and payment obligations behind it are identifiable.

Tracking the risk over time will require comparisons on consistent terms: contracted and active power at the same date for an operator, service dates for the same set of projects, revisions to investment plans, and comments on inventories or capacity utilization. A setback at one location may represent a move elsewhere; strong demand can coexist with delivery deferrals. The evidence has to be reconciled across those possibilities.

The public record available on 21 October shows that infrastructure constraints have become part of the industry’s demand assessment. Companies themselves document the channels through which those constraints can affect finances, while their aggregate effect on TSMC’s future sales remains unknown. Inspecting construction sites adds information that budgets and global forecasts struggle to provide: whether a specific project can turn its equipment into an available service on the date it has promised.

Methods and limits

Evidence cutoff: 21 October 2026. TSMC’s statements are examined alongside publications from Berkeley, Dominion, PJM, Virginia’s regulator and corporate financial disclosures. Management assessments, projections and observed conditions retain their attribution and dates. Transmission from a site delay to future chip orders is a conditional mechanism; its aggregate effect on TSMC remains unknown.

Figure 1 distinguishes deliveries, power supply, demand checks and the potential propagation of a delay. Figure 2 compares estimated annual electricity use in 2024 with the 2030 reference case and the range produced by combining extreme assumptions. Berkeley’s projection uses equipment shipments among its inputs and requires scrutiny against local conditions for project delivery. TWh measure annual energy; GW measure power.

Figure 3 retains Dominion’s three contract categories as of July 2026: 11,997 + 9,356 + 32,473 = 53,826 MW. Engineering authorizations are screened as follows: 32,473 − 19,044 = 13,429 MW included from 2032 onwards. The intermediate total of 34,782 MW precedes timing and ramp adjustments in the model. The 63% mature load ratio relates electricity demand to contract capacity.

Figure 4 places CoreWeave’s active and contracted power metrics side by side as of 30 June 2026. Their arithmetic gap describes two differently defined indicators; attributing it to delays requires project-level evidence. Contracted power updated to 11 August is shown separately. Revenue backlog estimates future contract revenue conditional on delivery and service availability. The British Loughton case rests on Nscale’s announcement and Reuters reporting; detailed contracts and site inspection reports remain unavailable.

The opening image is a conceptual illustration generated with imagegen. It combines a silicon wafer, a substation and a data center, without depicting an identified TSMC facility.

Sources and supporting documents

  1. S01 · TSMC / LSEG · Event: 2026-10-15 · Accessed 2026-10-21TSMC Q3 2026 Earnings Call: edited transcript
    Relevant passages, scope and access

    Q3 2026: 2026-07-01 / 2026-09-30; full-year 2026 budget

    p. 4: budget; pp. 5–6 and 16: site checks, power and management's findings; p. 17: AI/HPC scope.

    Call dated 15 October; exact file publication date unverified. Management reports positive findings, without public site-level inspection reports.

    Management statement; edited transcript subject to LSEG rights.

    Complete document downloaded and examined for research.

    Related link
    https://investor.tsmc.com/english/quarterly-results/2026/q3

  2. S02 · TSMC / LSEG · Event: 2026-07-16 · Accessed 2026-10-21TSMC Q2 2026 Earnings Call: edited transcript
    Relevant passages, scope and access

    Q2 2026: 2026-04-01 / 2026-06-30; multi-year outlook

    p. 4: 2026 budget; pp. 11–12: project checks and aggregation of customer forecasts.

    Checks were already discussed in July. An August archive path does not establish an exact file publication date.

    Management statement; LSEG rights.

    Complete document downloaded and examined for research.

    Related link
    https://investor.tsmc.com/english/quarterly-results/2026/q2

  3. S03 · TSMC · Published: 2026-10-15 · Accessed 2026-10-21TSMC 3Q26 Management Report
    Relevant passages, scope and access

    Q1–Q3 2026 and January–September 2026 cumulative spending

    p. 5, table V. Capital Expenditures, US dollars in billions.

    Actual worldwide consolidated spending; the remaining $19.36–23.36 billion is l0g arithmetic.

    Issuer's unaudited quarterly financial release.

    Complete document downloaded and examined for research.

    Related link
    https://investor.tsmc.com/english/quarterly-results/2026/q3

  4. S04 · TSMC / SEC · SEC filing: 2026-04-16 · Accessed 2026-10-21TSMC 2025 Form 20-F
    Relevant passages, scope and access

    Fiscal year 2025; balance sheet at 2025-12-31

    p. 8 / PDF 14: risks; p. 17 / PDF 23: orders and advances; p. 28 / PDF 34: fixed costs; note 5, F-21 / PDF 116: depreciation.

    Annual filing; published terms do not describe every customer contract. Revenue geography primarily follows customer headquarters.

    Issuer's risk disclosures; audited annual financial statements and management's forward-looking discussion.

    Complete document downloaded and examined for research.

    Related link
    https://investor.tsmc.com/english/annual-reports

    SEC filing dated 16 April 2026; the general annual-reports page displays 17 April. The EDGAR index and report signature establish the filing date. SEC EDGAR

  5. S05 · Lawrence Berkeley National Laboratory / US Department of Energy · Published: 2026-06-18 · Accessed 2026-10-21United States Data Center Energy Usage Report: 2025 Update
    Relevant passages, scope and access

    2024 historical estimate; projections through 2030; shipment data available through Q3 2025

    pp. 6–7 / PDF 8–9: methods and limits; pp. 8–9 / PDF 10–11: figures; table 2, p. 9; pp. 9–10: energy and power.

    Equipment-based model; major local power-availability constraints are not modeled. The combined range is a sensitivity exercise, without assigned probabilities.

    DOE-funded public research; industry assumptions are documented.

    Complete document downloaded and examined for research.

    Related link
    https://eta-publications.lbl.gov/publications/united-states-data-center-energy-2025

    Related link
    https://escholarship.org/uc/item/33m6w3x0

    Related link
    https://doi.org/10.71468/P1RP4F

  6. S06 · Dominion Energy / PJM · Published: 2026-09-29 · Accessed 2026-10-21Dominion Energy LSE 20-Year Data Center Forecast
    Relevant passages, scope and access

    Contracts as of July 2026; billing peak as of August 2026; 2027–2047 forecast

    p. 2: billing peak; p. 5: ramp assumptions; p. 6: ESA/CLOA/ELOA and screening; p. 10: forecast table.

    MW describe contract stages. 'Firm' is a model classification; the calculated 34,782 MW is not final forecast demand.

    Utility with an interest in development and cost allocation; reported data, not independently audited by this investigation.

    Complete document downloaded and examined for research.

  7. S07 · PJM Inside Lines · Published: 2026-09-29 · Accessed 2026-10-21PJM Kicks Off Annual Process of Vetting Large Load Requests
    Relevant passages, scope and access

    Process leading to the 2027 forecast expected in January 2027

    Paragraphs on twelve utilities, contracts, duplicates, Charles River Associates, and the absence of a PJM load-interconnection queue.

    Documents project vetting; does not publish a national volume of duplicate requests.

    Operator's description of its own forecasting process.

    Article read on the web; direct download failed (HTTP 503).

  8. S08 · Dominion Energy · Published: 2026-05 · Accessed 2026-10-21Meeting the Demands for Large-Load Customers with Reliable, Affordable, and Increasingly Clean Energy
    Relevant passages, scope and access

    Contract process described in May 2026; GS-5 effective from 2027-01-01

    pp. 4–5: ELOA, CLOA, ESA; pp. 6–7: term, minimums and the new GS-5 class.

    Explains contractual stages and financial safeguards. Publication dated to the month; regulatory provisions cross-checked with S09.

    Dominion's presentation of its own safeguards, with scope limited to the mechanisms described.

    Complete document downloaded and examined for research.

  9. S09 · Virginia State Corporation Commission · Published: 2026-02-24 · Accessed 2026-10-21SCC Data Center Initiatives
    Relevant passages, scope and access

    Approved measures; some apply to contracts entered into from 2027-01-01

    p. 2: fourteen-year commitments, minimum transmission/distribution charges, collateral and exceptions; publication-date footer.

    Corroborates safeguards. Minimum T&D billing uses 85% of contract demand (specified in S08); dates and exceptions apply.

    Regulator's institutional summary; not a substitute for all orders and contracts.

    PDF read on the web; direct binary download failed (HTTP 403).

    Related link
    https://www.scc.virginia.gov/about-the-scc/scc-facts/

  10. S10 · CoreWeave · Published: 2026-08-11 · Accessed 2026-10-21CoreWeave Reports Strong Second Quarter 2026 Results
    Relevant passages, scope and access

    Quarter ended and position as of 2026-06-30

    Additional Second Quarter 2026 Financial Highlights, note 1; Continued Rapid Scaling of Purpose-Built AI Infrastructure.

    Active power, contracted power and backlog are distinct metrics; backlog remains conditional on delivery and service availability.

    Issuer-defined metrics and disclosures.

    Complete document downloaded and examined for research.

  11. S11 · CoreWeave · Published: 2026-09-17 · Accessed 2026-10-21CoreWeave Continues to Contract New Compute Capacity at Higher Prices
    Relevant passages, scope and access

    Contracted power as of 2026-08-11; comments on Q3 2026

    Paragraph reporting approximately 4.2 GW of contracted power as of 11 August 2026.

    Measurement date differs from publication date; this release supplies no matching active-power measure.

    Issuer's commercial update.

    Complete document downloaded and examined for research.

  12. S12 · NVIDIA / SEC · Published: 2026-08-26 · Accessed 2026-10-21NVIDIA Form 10-Q for the quarter ended July 26, 2026
    Relevant passages, scope and access

    Quarter ended 2026-07-26 (Q2 of fiscal year 2027)

    pp. 25–26: infrastructure, purchase deferrals and manufacturing commitments; filing date verified in the SEC index.

    Disclosed prospective risk; establishes no TSMC loss amount. PDF compilation metadata is not the filing date.

    Issuer's risk and commitment disclosures.

    Complete document downloaded and examined for research.

    Related link
    https://www.sec.gov/Archives/edgar/data/1045810/000104581026000075/nvda-20260726.htm

    Related link
    https://www.sec.gov/Archives/edgar/data/1045810/000104581026000075/0001045810-26-000075-index.htm

  13. S13 · Microsoft · Published: 2026-07-29 · Accessed 2026-10-21Microsoft FY 2026 Q4 earnings release
    Relevant passages, scope and access

    Fiscal year from 2025-07-01 to 2026-06-30

    Cash Flows Statements, Twelve Months Ended June 30, 2026 column; operating cash flow and property/equipment additions.

    Group-wide consolidated figures; cash property/equipment spending excludes newly contracted finance leases.

    Cash-flow table in the release, labeled unaudited.

    Complete document downloaded and examined for research.

  14. S14 · Nscale · Published: 2025-10-15 · Accessed 2026-10-21Nscale–Microsoft AI infrastructure announcement, including Loughton
    Relevant passages, scope and access

    Announced project; planned service from Q1 2027

    Section on the United Kingdom and the Loughton site.

    Specifications and schedule announced in 2025; the site is in Britain and the capacity is not reported as operational.

    Project operator's commercial announcement.

    Complete document downloaded and examined for research.

  15. S15 · Reuters · Wire report dated: 2026-10-19 · Accessed 2026-10-21Reuters: Loughton report, 19 October 2026
    Relevant passages, scope and access

    News report dated 19 October 2026 in the republication accessed.

    Reuters report dated 19 October 2026, by Paul Sandle.

    Journalistic reference used in the articles; additional primary documents were not obtained in this collection.

    Journalistic source; reporting is distinguished from original project documents.

    Complete wire report read in an Investing.com republication of Reuters. Direct access to the Reuters URL failed during this verification. The original link’s first publication date remains unconfirmed; the 18 in its URL is part of its address.

This analysis is not investment advice.

// cite this analysis

l0g, “TSMC follows chip demand all the way to the power grid”, l0g.fr, published October 21, 2026, updated October 21, 2026, https://l0g.fr/en/analysis/tsmc-chip-demand-data-center-power-risk/

$ cd ../analysis