FactIQ
Browse documentation

What happens when the AI boom runs out of money?

Ben Thompson told Invest Like The Best that money — not compute — may be the AI boom’s binding constraint. We checked that argument against public data: 23 claims from the conversation, 6 FactIQ and primary-source findings, every number traceable.

On this page

Source: Ben Thompson on Invest Like The Best — “What Happens When the AI Boom Runs Out of Money”, published 18 August 2026. Research retrieved 21 August 2026. This page paraphrases the discussion; it does not reproduce the transcript.

Start here

The money hasn’t run out — but whoever carries the bill keeps changing. Four companies spent $408.91B on capex in their latest annual periods, nearly triple their 2023 total, and Alphabet turned to equity despite generating $73.26B of free cash flow. The data confirms Thompson’s financing pressure. His bust forecast is the one part no dataset can confirm.

The spending is real

Amazon, Alphabet, Meta, and Microsoft reported $408.91B of combined capex — 2.9× their 2023 total. Not all of it is AI, but the acceleration is unmistakable.

The funders keep changing

Free cash flow gave way to debt, then equity, and NVIDIA now aims to mobilize over $500B from outside investors. Each step of Thompson’s sequence checks out.

The bottlenecks are real too

Taiwan supplied 30.7% of direct U.S. chip imports in 2025, and data-center servers already draw 7% of commercial electricity. Real constraints — but neither proves a bust.

The short answer

Thompson’s core worry is a timing mismatch: infrastructure arrives years before the revenue meant to pay for it. The audit bears out the mismatch — spending has outrun any single company’s cash, supply runs through Taiwan, and power demand is climbing. What the data cannot show is the ending: whether funding dries up, when, or how hard.
23
claims from the conversation

1 supported, 8 partly supported, and 14 not verified against equivalent evidence.

29
evidence entries

Each entry names its source, period, unit, dataset ID, and calculation — including the plausible-looking proxies we rejected, and why.

Who is funding the boom

The spending surge is measurable. The predicted moment when the money stops is not.

What the video says

At 8:37, Thompson describes working “down the capital curve”: free cash flow first, then debt (“it took, like, a year”), now equity. At 31:22, the episode cites roughly $800B of capex in 2026 heading toward $1.3T in 2027, and asks whether revenue shows up before the money does.

What the data shows

Combined capex at Amazon, Alphabet, Meta, and Microsoft climbed from $140.14B in 2023 to $408.91B in their latest annual periods — 2.9×. Alphabet generated $73.26B of free cash flow in 2025, then announced $80B of equity offerings in June 2026 anyway. Across the economy, U.S. private investment in information-processing equipment and software is up 46.1% since early 2023.

What we make of it

Every step of the widening funding base checks out. The forecasts do not: the episode never says which companies or assets make up the $800B, and consolidated capex bundles GPUs with warehouses and office buildings. Confirmed: pressure on the financing model. Unconfirmed: the bust.

The chip bottleneck

Taiwan’s role in U.S. chip supply is real and large. Trade statistics just can’t prove the stronger claim that TSMC stands alone at the leading edge.

What the video says

The conversation opens with the risk that overwhelming U.S. AI superiority would make destroying TSMC rational for China. At 40:30, Thompson argues one company operates at the leading edge. At 46:42, he says acute scarcity is what finally pushed customers toward Intel.

What the data shows

Taiwan supplied 30.7% of direct U.S. integrated-circuit import value in 2025 ($12.94B of $42.20B); China supplied 3.0%. Smartphone imports show real diversification elsewhere: China’s share fell from 75.8% in 2023 to 45.2% in 2025 while India’s rose from 8.4% to 42.3%. Intel’s revenue and operating cash flow remain 33.1% and 67.1% below their 2021 levels.

What we make of it

Customs records cannot see Apple by name, the fab a chip came from, or chips routed through third countries — so “only TSMC” stays untested. Intel’s decline explains why a second foundry would matter. It does not yet show that scarcity has saved one.

Business models

Where Thompson quotes accounting figures, he mostly checks out. Where he turns a figure into a moat, the data pushes back.

What the video says

At 57:50: Microsoft generates about $20B of free cash flow a quarter and pays a $10B dividend. Meta runs a near-costless content engine, Reality Labs has consumed “hundred-some billion dollars,” and at 1:15:45, NVIDIA has kept its margins despite competitors.

What the data shows

Microsoft’s quarterly free cash flow was indeed $19.64B — but dividends paid were $6.76B, not $10B. Advertising was 97.6% of Meta’s 2025 revenue. Reality Labs accumulated $83.58B of operating losses over 2020–2025, which is not the same as spending on Oculus. NVIDIA’s gross margin slipped from 75.0% to 71.1%. AWS reached $128.73B of revenue at a 35.4% operating margin.

What we make of it

The broad shapes hold: Microsoft mints cash, Meta lives on advertising, NVIDIA’s product economics remain extraordinary, and AWS is huge and profitable. Three specifics get corrected: the dividend size, losses-versus-spending, and “maintained” margins that actually fell 3.9 points.

What survives a bust

Power is the best candidate for the boom’s lasting legacy — provided the plants get finished, sited well, and still have customers afterward.

What the video says

At 1:22:24: the United States brought more power online than expected. At 1:23:49: if the boom breaks, abundant power will be its legacy — the way fiber outlived dot-com and rail outlived the railroads.

What the data shows

U.S. net generation rose from 4,009.8 TWh in 2020 to 4,429.5 TWh in 2025, up 10.5%. The EIA estimates data-center servers alone took 7% of commercial-sector electricity in 2025 and projects 22%–33% by 2050.

What we make of it

Two gaps remain. “More than expected” needs the expectation, which the episode never states. And national annual totals say nothing about the constraint that matters locally — whether a specific substation in Northern Virginia or Texas is ready. A post-bust asset also has to be completed and reusable. None of that is observable yet.

Evidence ledger

Open any row for its source, period, unit, dataset ID, and calculation. All FactIQ research was read-only, retrieved 21 August 2026.
Supported

Equivalent evidence confirms the statement or retrieval as described.

Partly supported

Evidence confirms a defined part but not the full measure, period, entity, or causal claim.

Not verified

No equivalent public dataset tests the statement. This is not a ruling that the statement is false.

Capital

10 entries · 3 supported · 2 partly supported · 5 not verified

Partly supportedAI infrastructure moved from free-cash-flow funding through debt and then to equity, with returns not yet sufficient to recycle the system back to free cash flow.

Public filings confirm very large cash-funded capex and Alphabet’s 2026 equity raise, but FactIQ cannot establish one common financing sequence for the whole industry or whether AI revenue is insufficient to fund it.

Source
SEC company financials via FactIQ; Alphabet 2026 SEC filings for the equity event
Period
2023 through June 2026
Units
US$ billions
Dataset / series IDs
AMZN_capex_consolidated_annual, GOOG_capex_consolidated_annual, META_capex_consolidated_annual, MSFT_capex_consolidated_annual, SEC CIK 0001652044, June 2026 offering filings
Calculation / method
Capex and financing events were checked separately; no industry-wide causal sequence was inferred.
Partly supportedNVIDIA is assembling a $500 billion financing effort that reaches pension and insurance capital.

NVIDIA announced partnerships intended to mobilize more than $500B of third-party capital over time. The announcement names large alternative-asset managers and notes insurance businesses, but does not allocate a pension-fund amount. It is a target, not capital already raised or spent.

Source
NVIDIA investor-relations announcement, 10 August 2026
Period
Announced 10 August 2026; mobilization horizon not fixed
Units
More than US$500B target
Dataset / series IDs
NVIDIA: AI Compute Infrastructure Financing Platforms announcement (2026-08-10)
Calculation / method
No aggregation; the amount is the company’s announced target. It must not be presented as deployed capital.
Not verifiedThe AI buildout is in the ballpark of, or may exceed, the railroad boom as a share of GDP.

FactIQ has modern BEA investment and rail-industry series but no harmonized 19th-century railroad-capital series with the same boundary, price basis, and GDP denominator. The comparison cannot be reproduced without substituting non-equivalent measures.

Source
Coverage gap — modern BEA data do not provide a comparable 19th-century denominator
Period
Railroad boom era versus 2026 AI investment
Units
Investment as % of GDP; both boundaries unspecified
Dataset / series IDs
nipa|A679RC|T50305 (modern broader IT investment, context only)
Calculation / method
No cross-era ratio; the modern series includes all information-processing equipment and software, not AI infrastructure alone.
Not verifiedThe 1870s railroad system ran out of money before long-lived infrastructure produced its returns.

FactIQ does not contain the required 1870s issuance, default, construction, and cash-flow panel. Modern rail output confirms that rail still contributes to the economy, but it cannot verify the historical funding mechanism or causal account.

Source
Coverage gap — no 19th-century project-finance panel in FactIQ
Period
1870s and later railroad operating life
Units
Financing flows, defaults, construction, and output; definitions unspecified
Dataset / series IDs
bea.gdp_by_industry (modern rail context only)
Calculation / method
No historical calculation; current rail activity is not evidence for why 1870s financing failed.
Not verifiedBNSF generated more free cash in 2025 than See’s Candies did over its entire lifetime.

FactIQ’s Berkshire financials are consolidated and do not provide a comparable lifetime See’s cash-flow series or a BNSF free-cash-flow series on the same accounting basis.

Source
Coverage gap — subsidiary-level, lifetime cash-flow comparison unavailable
Period
BNSF 2025 versus See’s lifetime through 2025
Units
Free cash flow; definition not supplied
Dataset / series IDs
BRK-B_operating_cash_flow_consolidated_annual (non-equivalent consolidated context only)
Calculation / method
No calculation; consolidated Berkshire cash flow cannot be allocated to either subsidiary.
Not verifiedAI-related capex will be about $800 billion in 2026 and $1.3 trillion in 2027.

The episode does not define included companies, assets, or forecast source. FactIQ actuals show scale and acceleration, but company capex includes non-AI assets and does not verify either forward estimate.

Source
Coverage gap; SEC company actuals via FactIQ provide non-equivalent context
Period
Forecasts for 2026 and 2027; actual context uses 2023 and latest annual periods
Units
US$ capex; industry boundary unspecified
Dataset / series IDs
AMZN_capex_consolidated_annual, GOOG_capex_consolidated_annual, META_capex_consolidated_annual, MSFT_capex_consolidated_annual
Calculation / method
No forecast aggregation. The separate four-company actual-capex result is reported below and is not labeled AI-only.
SupportedAmazon, Alphabet, Meta, and Microsoft reported $408.91B of combined capex in their latest annual periods, 2.9 times the comparable 2023 total.FactIQ retrieval

This establishes the speed and size of public-company investment. It is not an AI-only measure: Amazon’s total includes logistics and Microsoft’s latest period is fiscal 2026 while the others are calendar 2025.

Source
SEC 10-K company financials via FactIQ
Period
2023 comparison; latest periods are calendar 2025 and Microsoft FY ended June 2026
Units
US$ billions of consolidated capital expenditures
Dataset / series IDs
AMZN_capex_consolidated_annual, GOOG_capex_consolidated_annual, META_capex_consolidated_annual, MSFT_capex_consolidated_annual
Calculation / method
Latest: 131.82 + 91.45 + 69.69 + 115.95 = $408.91B. 2023: 52.73 + 32.25 + 27.05 + 28.11 = $140.14B. Ratio = 2.92×.
SupportedU.S. private investment in information-processing equipment and software rose 46.1% from early 2023 to the second quarter of 2026.FactIQ retrieval

The BEA series provides a macro cross-check on acceleration but is much broader than AI infrastructure and includes software. It cannot be compared directly with the episode’s capex forecasts.

Source
U.S. Bureau of Economic Analysis NIPA via FactIQ
Period
2023 Q1 to 2026 Q2, seasonally adjusted annual rates
Units
Current US$ millions at annual rate
Dataset / series IDs
nipa|A679RC|T50305
Calculation / method
($1,612.853B − $1,103.962B) / $1,103.962B = 46.1%.
SupportedAlphabet issued equity even though Google can generate substantial cash internally.

The cash-generation premise and financing event are both documented. Alphabet reported $73.26B of 2025 free cash flow on the simple operating-cash-flow-minus-capex definition, then announced equity offerings totaling an expected $80B in June 2026.

Source
SEC 10-K data via FactIQ; Alphabet SEC free-writing prospectus dated 1 June 2026
Period
Calendar 2025 cash flow; June 2026 offering announcement
Units
US$ billions
Dataset / series IDs
GOOG_operating_cash_flow_consolidated_annual, GOOG_capex_consolidated_annual, SEC CIK 0001652044 accession 0001193125-26-251733
Calculation / method
$164.71B operating cash flow − $91.45B capex = $73.26B. The $80B offering figure is the issuer’s expected aggregate amount, not completed proceeds.
Not verifiedNVIDIA’s financing backstops and investments transfer risk to its balance sheet and amount economically to hidden price cuts.

FactIQ financial statements show commitments and investments only at reported accounting boundaries; they do not price the expected loss on the unspecified 25% backstop discussed in the episode. Calling that value a price cut is an interpretation, not a reported metric.

Source
Coverage gap — deal-level expected-loss assumptions are unavailable
Period
At publication; underlying deal not uniquely identified in the captions
Units
Expected US$ loss or effective discount
Dataset / series IDs
NVDA 2026 10-K commitments (context only)
Calculation / method
No expected-value calculation; probability of default, recovery, utilization, and contract value are not specified.

Chips

8 entries · 2 supported · 1 partly supported · 5 not verified

Not verifiedOverwhelming U.S. AI superiority could make destroying TSMC a rational response for China.

This is a geopolitical scenario, not an observed economic quantity. Trade data can measure exposure to Taiwan, but it cannot establish military intent, probability, or an optimal response.

Source
Coverage gap; U.S. Census trade data provide non-equivalent exposure context only
Period
Hypothetical future scenario; trade context is calendar 2025
Units
No defined measure for the strategic claim; contextual trade values are US$
Dataset / series IDs
us_census_hs_M_10d_8542******_5830, us_census_hs_M_10d_8542******_-
Calculation / method
No probability or game-theory calculation. The separate Taiwan import-share result is recorded below as context only.
Partly supportedApple is diversifying iPhone production to India but is not truly moving out of China.

U.S. smartphone imports show a sharp shift toward India while China remains a major supplier. The product category covers all smartphones, not Apple alone, so it supports the direction but cannot verify Apple’s supply chain.

Source
U.S. Census Bureau international trade via FactIQ
Period
2023 and 2025
Units
US$ import value; share of all U.S. smartphone imports
Dataset / series IDs
us_census_hs_M_10d_8517130000_5700, us_census_hs_M_10d_8517130000_5330, us_census_hs_M_10d_8517130000_-
Calculation / method
China: $44.79B / $59.09B = 75.8% in 2023 and $23.51B / $52.01B = 45.2% in 2025. India: 8.4% to 42.3%.
Not verifiedTSMC slowed its growth in 2023, 2024, and 2025, making today’s compute shortage worse.

FactIQ has TSMC revenue and cash-flow series, but no current, comparable wafer-capacity additions series. Revenue growth cannot substitute for physical capacity growth or prove the claimed effect on compute availability.

Source
Coverage gap — TSMC physical capacity additions are unavailable in FactIQ
Period
2023–2025
Units
Claimed growth measure not specified; likely wafer capacity
Dataset / series IDs
TSM_revenue_consolidated_annual (non-equivalent context only), TSM_operating_cash_flow_consolidated_annual (non-equivalent context only)
Calculation / method
No calculation; financial growth was rejected as a proxy for wafer capacity.
Not verifiedCurrent infrastructure spending will not become usable compute until 2028 or 2029 because fabs lead data centers.

This is a project-timing forecast. FactIQ lacks a linked ledger of fab, packaging, server, data-center, grid-connection, and delivery milestones needed to test it.

Source
Coverage gap — no linked semiconductor-to-data-center project schedule
Period
Spending around 2026; claimed delivery in 2028–2029
Units
Years to usable compute; project boundary unspecified
Dataset / series IDs
None — no comparable project milestone dataset
Calculation / method
No calculation; company capex dates do not identify the date capacity becomes operational.
Not verifiedTSMC is the only company operating at the leading edge, creating a singular concentration risk.

FactIQ trade records show material Taiwan exposure, but customs values do not identify foundry, process node, design owner, or fabrication origin after packaging. They cannot establish the “only company” claim.

Source
Coverage gap; U.S. Census integrated-circuit imports provide non-equivalent context
Period
Calendar 2025
Units
US$ import value versus process-node manufacturing share
Dataset / series IDs
us_census_hs_M_10d_8542******_5830, us_census_hs_M_10d_8542******_-
Calculation / method
No foundry-share calculation; country-of-export customs data are not a process-node capacity census.
SupportedTaiwan supplied 30.7% of the United States’ direct integrated-circuit import value in 2025; China supplied 3.0%.FactIQ retrieval

This supports material direct exposure to Taiwan and shows why “China dependence” and “Taiwan dependence” should not be collapsed into one statistic. It still understates chips routed through third countries and does not identify TSMC.

Source
U.S. Census Bureau international trade via FactIQ
Period
Calendar 2025
Units
US$ import value at HS 8542, summed once at the 10-digit product level
Dataset / series IDs
us_census_hs_M_10d_8542******_5830, us_census_hs_M_10d_8542******_5700, us_census_hs_M_10d_8542******_-
Calculation / method
Taiwan: $12.94B / $42.20B = 30.7%. China: $1.27B / $42.20B = 3.0%. Quantity series and regional aggregates were excluded.
Not verifiedAcute compute scarcity ultimately saved Intel by making customers willing to qualify another foundry.

FactIQ can show Intel’s financial deterioration, not foundry qualification decisions or the counterfactual in which scarcity did not occur. The causal “saved” verdict is not testable yet.

Source
Coverage gap; Intel SEC financials provide context only
Period
2021–2025 financial context; foundry outcome is ongoing
Units
US$ revenue and operating cash flow; no customer-qualification measure
Dataset / series IDs
INTC_revenue_consolidated_annual, INTC_operating_cash_flow_consolidated_annual
Calculation / method
No causal calculation; the separate financial trend below records the observed context.
SupportedIntel’s revenue fell 33.1% and operating cash flow fell 67.1% from 2021 to 2025.FactIQ retrieval

The deterioration explains the stakes around Intel’s foundry strategy. It does not show that scarcity caused a recovery or that a major external customer has qualified a process.

Source
Intel SEC 10-K financials via FactIQ
Period
Fiscal years ended 2021 and 2025
Units
US$ billions
Dataset / series IDs
INTC_revenue_consolidated_annual, INTC_operating_cash_flow_consolidated_annual
Calculation / method
Revenue: ($52.85B − $79.02B) / $79.02B = −33.1%. OCF: ($9.70B − $29.46B) / $29.46B = −67.1%.

Business models

8 entries · 1 supported · 4 partly supported · 3 not verified

Not verifiedChinese AI models are roughly six to nine months behind the U.S. frontier.

FactIQ has no longitudinal model-capability benchmark with a stable frontier definition, test set, and release-date adjustment. Company financials and trade data cannot test the claim.

Source
Coverage gap — no comparable model-capability benchmark in FactIQ
Period
At publication, 18 August 2026
Units
Months behind an undefined capability frontier
Dataset / series IDs
None — no model-capability benchmark dataset
Calculation / method
No calculation; the benchmark and comparison rule are unspecified.
Not verifiedOpen-weight models are not free to use because inference still has a marginal cost, and some models cost materially more per answer.

The accounting distinction is sound, but FactIQ does not carry like-for-like serving costs by model, hardware, token mix, latency, batching, or utilization. The relative cost claim cannot be tested.

Source
Coverage gap — no normalized model-serving cost dataset
Period
At publication
Units
US$ per answer or per token; workload not specified
Dataset / series IDs
None — no model-level inference-cost series
Calculation / method
No calculation; a reproducible comparison requires the same workload and serving configuration.
Partly supportedMicrosoft generated about $20 billion of free cash flow and paid a $10 billion dividend in its latest quarter.

The free-cash-flow figure is close on a simple definition, but the dividend is not. For the quarter ended June 2026, FactIQ records $55.44B operating cash flow, $35.80B capex, and $6.76B dividends paid.

Source
Microsoft SEC 10-Q/10-K financials via FactIQ
Period
Quarter ended 30 June 2026
Units
US$ billions
Dataset / series IDs
MSFT_operating_cash_flow_consolidated_quarterly, MSFT_capex_consolidated_quarterly, MSFT_dividends_paid_consolidated_quarterly
Calculation / method
$55.44B − $35.80B = $19.64B simple free cash flow; reported cash dividends paid were $6.76B, not $10B.
SupportedAWS revenue reached $128.73B in 2025, up 41.8% from 2023, with a 35.4% segment operating margin.FactIQ retrieval

This supports the video’s description of AWS as a large, profitable internal capability sold externally. It does not rank Amazon’s moat against every other technology company.

Source
Amazon SEC 10-K segment financials via FactIQ
Period
Calendar 2023 and 2025
Units
US$ billions and operating margin
Dataset / series IDs
AMZN_revenue_ex_tax_amazon_web_services_annual, AMZN_operating_income_amazon_web_services_annual
Calculation / method
Revenue growth: ($128.73B − $90.76B) / $90.76B = 41.8%. Margin: $45.61B / $128.73B = 35.4%.
Partly supportedMeta has an exceptional advertising business because it pays zero dollars for Instagram content.

FactIQ confirms that advertising supplies almost all Meta revenue, but SEC line items do not prove that content has zero economic cost. Revenue sharing, creator incentives, moderation, hosting, and product costs are different measures.

Source
Meta SEC 10-K financials via FactIQ
Period
Calendar 2025
Units
US$ revenue and share of consolidated revenue
Dataset / series IDs
META_revenue_ex_tax_advertising_annual, META_revenue_consolidated_annual
Calculation / method
$196.18B advertising revenue / $200.97B consolidated revenue = 97.6%. No zero-content-cost calculation is available.
Partly supportedMeta spent a cumulative “hundred-some billion dollars” on Oculus.

FactIQ records $83.58B of cumulative Reality Labs operating losses from 2020 through 2025. That is directionally large, but operating loss is not spending, Reality Labs is broader than Oculus, and pre-2020 amounts are not in the segment series.

Source
Meta SEC 10-K segment financials via FactIQ
Period
2020–2025
Units
US$ billions of segment operating loss
Dataset / series IDs
META_operating_income_reality_labs_annual
Calculation / method
Sum of annual Reality Labs operating income for 2020–2025 = −$83.58B. The result is not relabeled as spending.
Partly supportedNVIDIA has maintained its margins despite competitors pursuing the market.

NVIDIA’s gross margin remained very high but did not stay unchanged: it fell from 75.0% in fiscal 2025 to 71.1% in fiscal 2026. Financing risk outside product pricing is a separate claim that gross margin cannot capture.

Source
NVIDIA SEC 10-K financials via FactIQ
Period
Fiscal years ended January 2025 and January 2026
Units
Gross profit as % of revenue
Dataset / series IDs
NVDA_gross_profit_consolidated_annual, NVDA_revenue_consolidated_annual
Calculation / method
$97.86B / $130.50B = 75.0%; $153.46B / $215.94B = 71.1%, a decline of 3.9 percentage points.
Not verifiedGoogle agreed to sell roughly 20% of its TPUs to Anthropic, and Amazon is preparing to sell Trainium externally.

FactIQ company financials do not contain chip-unit contracts, external sales volumes, or a consistent denominator for “20%.” Earnings-call language about future sales would still be guidance, not observed shipments.

Source
Coverage gap — no customer-contract or custom-accelerator shipment dataset
Period
At publication and future sales horizon
Units
TPU share and Trainium unit/revenue sales; denominator unspecified
Dataset / series IDs
None — no custom-accelerator contract series
Calculation / method
No calculation; the quantity, period, and denominator behind 20% are not defined.

Power

3 entries · 1 supported · 1 partly supported · 1 not verified

Partly supportedThe United States brought substantially more power online than expected, delaying electricity as the binding AI constraint.

FactIQ confirms generation growth after 2020, but the episode does not name the prior expectation or define “brought online.” Annual generation also cannot identify whether data-center interconnections had adequate local capacity.

Source
U.S. Energy Information Administration electric-power data via FactIQ
Period
2020–2025
Units
TWh of net generation, all fuels and sectors
Dataset / series IDs
ELEC.GEN.ALL-US-99.Q
Calculation / method
Quarterly values summed by calendar year: 4,009.8 TWh in 2020 to 4,429.5 TWh in 2025, up 10.5%. No forecast-error calculation was possible.
SupportedEIA estimated data-center servers used 7% of commercial-sector electricity in 2025 and projected 22%–33% by 2050.Primary source

This independently supports the video’s focus on power as a material constraint. It is a modeled range for server electricity, not total data-center load, and it does not predict whether a financing bust occurs.

Source
U.S. Energy Information Administration, Annual Energy Outlook 2026 analysis, 19 May 2026
Period
Estimated 2025; projected 2050
Units
% of U.S. commercial-building electricity consumption
Dataset / series IDs
EIA AEO2026 Commercial Demand Model — data center server energy use
Calculation / method
No recalculation; percentages are EIA model outputs. Supporting cooling and ventilation are discussed separately and are not included in the server-only share.
Not verifiedIf the AI boom breaks, its durable legacy will be abundant power, as fiber survived the dot-com bust and rail survived the railroad bust.

FactIQ and EIA show rising generation, investment, and demand, but cannot observe the outcome of a future bust or establish that planned power assets will be completed, located where needed, and economically reusable.

Source
Coverage gap; EIA generation and forecast data provide context only
Period
Hypothetical post-bust future; context through 2027/2050 scenarios
Units
Generation, capacity, location, completion, and utilization; no single measure
Dataset / series IDs
ELEC.GEN.ALL-US-99.Q, EIA AEO2026 Commercial Demand Model
Calculation / method
No counterfactual calculation; announced or forecast capacity is not treated as completed reusable infrastructure.

Method and sources

The transcript supplied the research outline; the ledger preserves where the data ends.

Caption provenance

The video’s English captions were downloaded and de-duplicated on 21 August 2026 with stitch-youtube-vtt at commit 4b723d84b16cb176fccf0f0f34dd8b09f79c7323. The captions were used to locate topics and timestamps; the transcript is not checked in or republished. Automatic captions render “See’s” as “seized,” “Trainium” as “Tranium,” and several company names phonetically. Those errors were corrected only in paraphrase, not treated as factual discrepancies.

FactIQ research

Research used read-only access to SEC company financials, U.S. Census ten-digit HS trade series, BEA NIPA investment, and EIA electricity generation. Company capex is consolidated and not labeled AI-only. Fiscal calendars were preserved: Microsoft’s latest annual point ends June 2026; Amazon, Alphabet, and Meta end December 2025. Free cash flow is explicitly the simple calculation operating cash flow minus capex, not a company-defined non-GAAP measure. Trade was summed once at the ten-digit product level; quantity series and regional aggregates were excluded.

External primary sources