Loading reports … Fetching the current index of long-form research.
Reports
Loading reports … Fetching the current index of long-form research.
Power in the United States is consolidating faster than its institutions can convert it, and conversion decides the year.
Power in the United States is consolidating on a handful of balance sheets, budget lines and agencies faster than its institutions can convert it, and conversion decides the year. The United States enters mid-2026 with record commitments from Washington and four American firms, and the question is whether it can turn announced capital, compute, capacity and security commitments into operating plant, power and fielded systems inside the planning horizon. On the public record the grid sets that pace, and the agencies that run the public programs lost staff faster than the government as a whole. China, which already converts at scale in manufacturing, is the benchmark. This edition measures the United States and makes no claim about how fast Chinese or European institutions convert.
The United States holds more capital, compute and security commitment than its institutions can yet convert. China's industrial base is the benchmark it is measured against.
The thesis of this outlook is that power in the United States is consolidating on a handful of balance sheets, budget lines and agencies faster than its institutions can convert it, and that conversion decides the year: whether announced capital, compute, capacity and security commitments become operating plant, power and fielded systems inside the planning horizon. Four propositions test it, and Part V states the reading that would falsify each. Capital and compute are concentrating on four American balance sheets. The grid sets how fast that capital becomes capacity. The interest bill compresses the fiscal room for everything else. The federal agencies that administer the conversion programs lost staff faster than the government as a whole, a risk to conversion that the public record does not yet measure.
The United States is the best-measured pole, because interconnection queues, award and staffing records, budget projections, and company guidance are public. The other poles can be measured only in part. For China, reachable sources show industrial scale and private AI investment; they show nothing comparable on how fast the state converts commitments or how its institutions govern the firms that hold them. For Europe, they show the rise in military spending and nothing on how fast it becomes fielded capability. The thesis is therefore stated for the United States, and the report makes no claim about Chinese or European institutions beyond what the data below show.
China is the benchmark because it already converts at scale. It produced 28.0 percent of world manufacturing value added in 2024, $4.65 trillion of a $16.62 trillion world total, on the World Bank's series. Its share was 8.5 percent in 2004, rose past the United States in 2010, and peaked at 30.0 percent in 2021 before settling near 28 percent as output in dollar terms held level. China's 2024 output was 2.9 times that of Japan and Germany combined. The US share fell from 21.9 percent in 2004 to 15.5 percent in 2021, the last year the World Bank reports for the United States (Exhibit 1).1
The US response has been to commit capital at scale. Through January 31, 2025, the CHIPS Program Office had made 20 awards worth up to $33.7 billion in direct funding and up to $5.5 billion in loans, including $7.865 billion to Intel, $6.6 billion to TSMC, and $4.745 billion to Samsung, whose final award came in below the $6.4 billion first announced. Boston Consulting Group and the Semiconductor Industry Association projected in May 2024 that the program and the private investment around it would triple US fab capacity by 2032, lifting the US share of world capacity from 10 percent in 2022 to 14 percent, and its share of the most advanced logic, below 10 nanometers, from zero to 28 percent.2
Trade has rerouted without shrinking. China's recorded share of US goods imports fell from 21.6 percent in 2017 to 13.4 percent in 2024, while its share of world goods exports kept rising, and much of what replaced it from Vietnam, India and the rest of Southeast Asia embeds Chinese inputs that US customs records do not observe.3 The dollar has moved less than the headlines suggest. Its share of allocated official reserves stood at 56.92 percent in the third quarter of 2025, near the 56.32 percent of the second quarter, the lowest since 1995, while its role in foreign-exchange turnover, commodity invoicing and international debt is essentially unchanged.4 The reserve dollar is eroding at the margin and the transactional dollar holds, so the American position rests on what the country can build more than on what it can finance.
The CHIPS figures are the strongest evidence that conversion can work, and Part IV takes them as the case against this outlook. They are also projections of plant that is still being built by agencies that lost staff in 2025. A committed award is a promise conditioned on construction milestones; the report found no public disbursement total for the incentives program dated before its cutoff, so the share of the $33.7 billion already paid out is a reading this outlook cannot supply and Part V puts on the watchlist.
World military expenditure reached $2.89 trillion in 2025, the eleventh consecutive annual rise and a 2.9 percent real increase, the smallest since 2021. Europe drove it: spending rose 14 percent to $864 billion, with Germany up 24 percent to $114 billion and Spain up 50 percent to $40.2 billion. US spending fell 7.5 percent to $954 billion, which SIPRI attributes mainly to the absence of any new Ukraine aid package in 2025; outside the United States, world spending grew 9.2 percent.5
The American dip is a one-year accounting effect, and the commitment that follows it is the largest on record. Congress approved more than $1 trillion for FY2026, a total that counts about $155 billion from the 2025 reconciliation act, money that remains available over several years. The FY2027 request released on April 3, 2026 is $1.5 trillion: a $1.15 trillion base for regular appropriations and $350 billion through reconciliation, which CSIS calculates as a 38 percent real increase over FY2026 and the highest single-year level since the Second World War.6 Defense money, like chip money and AI money, is now committed faster than any of the three has yet been converted. Defense differs from the other two in one respect: no public reading at the cutoff shows how fast budget authority becomes fielded capability, through contract awards, production lines, and units equipped. This outlook therefore measures the defense commitment by its size and makes no claim about the speed of its conversion.
The AI build-out is the largest private capital commitment in the economy, and it is held by four firms. The technical lead it buys over China is 39 points on the Arena leaderboard.
By April 30, 2026, after the first-quarter earnings round, Alphabet, Amazon, Meta, and Microsoft had guided to more than $700 billion of capital expenditure for 2026, nearly double what they spent in 2025 and about $100 billion above their guidance one quarter earlier. Alphabet raised its range to $180 to $190 billion and Meta to $125 to $145 billion; Amazon held at about $200 billion and Microsoft pointed to about $190 billion for the calendar year.7 The supplier side shows where the money lands: NVIDIA's data-center revenue was $193.7 billion in the fiscal year to January 2026, up 68 percent, with $62.3 billion in the fourth quarter alone.8
The concentration runs from the capex line to the equity market. The ten largest companies in the S&P 500 made up almost 40 percent of the index at the end of June 2025, a level S&P Dow Jones Indices says had not been seen since the mid-1960s, and the four firms above are among them. An investor in an S&P 500 index fund therefore holds the same balance sheets that carry the build-out, so a revaluation of AI capex moves both the GDP accounts and the index in the same direction.9 That shared exposure is what turns a technology cycle into a macro variable, and it is why the capex guidance round is the first row of the watchlist.
The capital lead buys a narrow technical lead. Stanford's 2026 AI Index records private AI investment of $285.9 billion in the United States and $12.4 billion in China in 2025, a ratio of about 23 to one, and notes that the Chinese figure understates the total because state guidance funds sit outside private-capital databases. On the Arena leaderboard in March 2026, the top American model scored 1,503 and the top Chinese model 1,464, a gap of 39 points.10
Three explanations compete for why so large a capital lead buys so narrow a gap, and the data here cannot rank them with confidence. Chinese laboratories may be learning efficiently from open-weight releases and from distillation of American outputs. Export controls may be leaking, through stockpiled or diverted accelerators. Or the benchmarks may be saturating, so that the measured gap understates the real one. The policy reading is the same under all three: an export-control regime built to protect a lead in compute has to be designed for a world in which the frontier does not track who owns the most of it.
For conversion, the narrow gap matters through the defense channel. Frontier capability reaches the force through a small number of laboratories whose models run on the same facilities, accelerators, and power the commercial build-out uses. A slower grid therefore slows national-security AI and commercial AI together, which is why Part III treats the grid as the shared constraint.
Capital is abundant and power is scarce. The grid connects new supply more slowly every year, and the interest bill takes a rising share of the money that would pay for the rest.
The International Energy Agency projects global data-center electricity use to more than double, from about 415 terawatt-hours in 2024, roughly 1.5 percent of world consumption, to about 945 terawatt-hours by 2030, slightly more than Japan uses today. The United States accounts for by far the largest share of the increase: its data-center demand rises by about 240 terawatt-hours by 2030, and data centers account for nearly half of all US electricity demand growth over the period.11
Demand arrives faster than the grid can connect supply. At the end of 2024, about 2,290 gigawatts of generation and storage were waiting in US interconnection queues, 12 percent less than a year earlier. Of the capacity that entered the queues from 2000 to 2019, 13 percent had come online by the end of 2024. The median project completed in 2024 took 55 months from interconnection request to commercial operation, against 36 months in 2015 and 22 in 2008 (Exhibit 2).12 That is the plainest public measure of conversion in this report: the time it takes to turn a committed power project into delivered electrons has more than doubled in sixteen years.
A 55-month median is longer than the planning horizon most capex guidance covers. A data center whose power is requested in 2026 joins a queue that, at the 2024 median, delivers it in 2030, which means the 2026 build-out runs on power already connected or on generation built beside the load. That is the mechanism behind the fractured-consolidation scenario in Part V: if grid capacity falls short in the regions that host the build-out, capex is spent on facilities that wait for power, the return on it falls, and the guidance round that follows is cut.
Net interest on the federal debt passed national defense for the first time in FY2024, at $882 billion against $874 billion, and reached $970 billion, 3.2 percent of GDP, in FY2025. The Congressional Budget Office's February 2026 baseline puts it at $1.0 trillion, 3.3 percent of GDP, in FY2026 and $2.1 trillion, 4.6 percent, in FY2036, with $16.2 trillion of interest over the decade. Interest exceeds defense in every year of the baseline and is almost double it by 2036.13
Interest follows yields, the stock of debt, and the maturity of that stock, so faster growth or a year of restraint barely moves it. Every $100 billion of additional interest is $100 billion that does not fund defense, research, the grid, or industrial policy without new borrowing, which adds to the interest. The FY2027 defense request makes the trade-off concrete: its $350 billion reconciliation tranche is one-time money outside the recurring baseline, and financing it adds to the interest line CBO already projects. Allies reading the American commitment to the 2025 NATO pledge of 5 percent of GDP by 2035 are reading it against that trajectory.
The conversion programs run through a handful of federal agencies. Those agencies lost staff faster than the government as a whole in 2025, and this outlook judges that loss the least measured risk to conversion.
The federal civilian workforce fell from 2,312,301 in December 2024 to 2,074,649 in December 2025, a net loss of about 238,000 people, or 10.3 percent. Separations rose 81 percent to 348,219, and hires fell 56 percent to 116,912; separations net of hires account for 231,307 of the loss, and the remaining 6,345 reflects flows the published figures do not itemize. The losses were deepest in the agencies that run the conversion programs: Commerce, home of the CHIPS Program Office, lost 13.8 percent of its staff; Energy, which runs the loan programs and grid funding, 15.1 percent; Treasury, which administers the IRA tax credits, 23.2 percent; and the National Science Foundation, which funds the research base, 30.3 percent (Exhibit 3).14
The mechanism from staffing to conversion runs through the steps each program requires. A CHIPS award pays out against construction milestones that Commerce staff must verify. An IRA credit is claimed through Treasury guidance and IRS processing. A loan closes after Energy staff complete diligence. Each step is a queue, and a queue served by fewer people moves more slowly unless the work is automated or dropped. The report cannot show the slowdown directly, because no agency published processing times or disbursement rates for these programs before the cutoff. What the data show is that the agencies on the critical path lost between 1.3 and 3 times the government-wide rate.
The aggregate data neither confirm nor refute an AI-driven payoff. Nonfarm business productivity grew 2.1 percent in 2025 after 3.0 percent in 2024, against 1.5 percent a year in the 2007 to 2019 cycle and 2.2 percent since 1947. The current cycle, since the end of 2019, has run at 2.1 percent. The 2024 figure is the March 2026 revision; earlier vintages put it at 2.7 percent.15 The comparison abroad sharpens the point: on the OECD's harmonized 2024 data, labor productivity grew about 0.4 percent across OECD countries and fell 0.9 percent in the euro area and 0.6 percent in Japan.16 Productivity is running faster than in the last cycle and at its long-run average, which is consistent with an early AI contribution and equally consistent with the post-pandemic recovery in capital deepening. The test is whether the rate holds above 2 percent through 2027 as the capex already spent comes into use.
The strongest reading of the same evidence is that conversion is working. Fab capacity is on course to triple by 2032 and the most advanced logic to reach 28 percent of world capacity from nothing. Four firms are spending more than $700 billion this year and NVIDIA's data-center revenue rose 68 percent, so the capex is turning into installed hardware at a pace no state program has matched. The defense request passed $1 trillion. Productivity is growing faster than in the last cycle. On this reading the binding constraints are energy and capital costs, and federal staffing is a political story with little bearing on output, since the money flows whether or not the agencies are fully staffed.
This outlook accepts half of that case. The grid is the binding constraint on the AI build-out, and Part III puts it first. Private capex converts on its own balance sheets and does not wait for an agency. The disagreement is about the public programs, where conversion runs through the agencies by statute: CHIPS milestones, IRA credits, loan closings, research grants, and permitting. There, three readings would settle it. A CHIPS disbursement total that keeps pace with construction, IRA credit processing times that hold level through 2026, and interconnection medians that stop rising would each show that the staffing loss is not binding, and the institutional proposition of this outlook would fail. Until those readings exist, the report weighs the staffing loss as a risk to the public half of conversion and the grid as the constraint on both halves.
The outlook resolves into three scenarios distinguished by which constraint binds first. A war with Iran was under way at the cutoff, so the scenarios are written around it.
The year's largest shock arrived before this outlook went to press. Israel and the United States launched strikes on Iran on February 28, 2026, and Iran closed the Strait of Hormuz to most shipping within days. A ceasefire declared on April 8 has held only in part: the strait remained restricted at the start of May, the United States imposed a blockade of Iranian ports on April 13, and Brent traded at $111 a barrel on May 5.17 The scenarios below are therefore conditioned on a war in progress. The probabilities are the author's judgment; no model produced them.
Managed concentration is the base case on one current reading and two judgments. The reading is capex guidance, which rose in the April round. The first judgment is that the strait reopens by the third quarter; it had not reopened at the cutoff, though the ceasefire, however partial, had held for a month. The second is that an equity drawdown would stay inside the AI-exposed names. The weight is 55 percent, little more than even, because the war adds a condition the base case cannot control.
Fractured consolidation carries 30 percent because its trigger is already visible in the data. The 55-month interconnection median and the regional concentration of the build-out mean that a shortfall in two or three host regions would bite within the horizon, and the market would see it first in a guidance cut. Kinetic escalation carries 15 percent. It means a war that spreads beyond the Gulf or keeps the strait closed past the third quarter; the base case already includes the war contained in the Gulf. The weight is high against the historical base rate for a great-power war in any one year, and it is set there because the year already holds one regional war, with the strait still restricted at the cutoff.
For investors, capital, compute, and energy operate as one system, and an allocation to one leg that ignores the other two carries the grid risk without being paid for it. For industrial policy, the CHIPS and IRA wave has committed the capital, and the next four years turn on the workforce, supplier base, and permitting speed that convert it. For allies, the United States remains the primary security underwriter through the decade with less fiscal room each year, which gives European governments fiscal grounds to rearm on top of the strategic ones. For export-control policy, the 39-point Arena gap means the regime has to work in a world where the frontier does not depend on American compute.
| Variable | What a move would mean | Reading | Falsifying trigger | Horizon |
|---|---|---|---|---|
| Four-firm 2027 capex guidance | Concentration continues or stalls; decides base case against fracture | Above $700bn for 2026 (Apr 30) | Below $750bn for 2027 | Q4 results, early 2027 |
| Interconnection median, request to operation | Whether the grid still binds conversion | 55 months (2024) | Below 50 months | LBNL 2026 edition |
| CBO August 2026 update, FY2036 net interest | Fiscal room for defense and industrial policy | $2.1tn, 4.6% of GDP | Below $1.8tn | Aug 2026 |
| CHIPS incentives disbursed against awards | Whether the public half of conversion is keeping pace | No total published | Disbursement keeps pace with milestones | 2026 GAO or OIG report |
| Strait of Hormuz and Brent | Base case against kinetic escalation | Restricted; Brent $111 (May 5) | Closure past Q3, or Brent sustained above $120 | Q3 2026 |
| Best Chinese model against best US model | Export-control design; frontier tracks compute or not | 39 points; Arena 1,464 vs 1,503 | Chinese model leads on Arena | 2026 H2 |
Triggers show the direction of the move that would change the view. Readings as of early May 2026.
Source: Company guidance via Sherwood News; LBNL; CBO; Commerce OIG; Reuters; Stanford HAI. Data through May 5, 2026.
The four propositions of Part I map onto the first four rows. If capex guidance holds, the grid median stops rising, the interest path stands, and the staffing loss shows up in disbursement, the thesis holds and the base case stays at 55 percent. If the grid row moves the wrong way first, weight shifts to fracture; if the strait row does, it shifts to escalation. Either way, the outlook's judgment is that power in the United States is consolidating faster than its institutions can convert it, and the readings above are where that judgment will be tested before the year is out.
At the start of May the Strait of Hormuz remained restricted, the US blockade of Iranian ports was in force, and a US effort to escort ships through the strait began on May 4 (S&P Global Commodity Insights, May 5, 2026). No hyperscaler cut its 2026 capex guidance in the April round; Alphabet and Meta raised theirs. The FY2027 defense request of $1.5 trillion awaits appropriations, and its $350 billion reconciliation tranche awaits a vehicle.
The strait condition was met and then reversed. The United States and Iran reached an interim agreement in mid-June, and US forces lifted the blockade of Iranian ports on June 18, after about two months; tanker traffic through the strait recovered over the summer.18 Attacks on shipping resumed in September, and October was the worst month on record for attacks on commercial vessels in and around the strait; Brent rose 13.6 percent in October to its highest settlement since the end of April.19 On the scenario definitions above, the year has moved from the base case toward kinetic escalation without a war beyond the Gulf. The edition's other triggers, 2027 capex guidance, the 2025 interconnection median, a CHIPS disbursement total and FY2036 net interest in a revised CBO baseline, have no new primary reading this note could confirm, and the edition does not re-forecast them.
Every figure carries a named, dated source and was re-checked in October 2026 at the edition's data cutoff, early May 2026. Primaries read directly include the WDI series, the CHIPS Inspector General report, the LBNL queue study, the BLS release, and NVIDIA's filing. Where no primary was reachable, the figure rests on a dated secondary account named in the note: SIPRI's fact sheet through Defense News and Breaking Defense, company capex guidance through Sherwood News, the AI Index through its published summaries, CBO's projections through the House Budget Committee and CRFB, federal staffing through Pew's analysis of OPM data, and the Gulf war through Al Jazeera, Barchart, Bloomberg Government, Reuters, and S&P Global. This edition retired seven exhibits recovered from an earlier draft, whose series could not be traced to a primary table, and withdrew claims the record would not support, among them a 2020 to 2025 capex series, a 25-point LMArena gap, a year-by-year net interest path, and unsourced claims about naval forging and bomber sortie rates. Investment figures are announced or committed values, which differ from capital disbursed. The November 2026 release restates the thesis for the United States, the one pole the evidence tests, and adds a dated note on events since the cutoff, drawn from the news reporting named in its notes; it does not re-forecast the scenarios.
| Dataset | Source | Frequency | Coverage | Retrieved | Used in |
|---|---|---|---|---|---|
| Manufacturing value added | World Bank WDI | Annual | World, by economy | 2024 (US 2021) | Exhibit 1; Table A1; Part I |
| Interconnection times and queue | LBNL, Queued Up 2025 | Annual | United States | End-2024 | Exhibit 2; Table A2; Part III |
| Federal civilian workforce | Pew analysis of OPM data | Annual | United States, by agency | Dec 2025 | Exhibit 3; Table A3; Part IV |
| Hyperscaler capex guidance | Company guidance via Sherwood News | Quarterly | Four firms | Apr 2026 | Part II |
| AI investment and model gap | Stanford HAI, AI Index 2026 | Annual | US and China | Mar 2026 | Part II |
| Federal net interest | CBO baseline, Feb 2026 | Annual | United States | 2036 proj. | Part III |
| Military expenditure | SIPRI | Annual | World, by country | 2025 | Part I |
| Economy | 2004 | 2010 | 2021 | 2024 |
|---|---|---|---|---|
| China | 8.5 | 18.0 | 30.0 | 28.0 |
| United States | 21.9 | 16.7 | 15.5 | n/a |
| Japan | 14.1 | 11.3 | 6.4 | 4.7 |
| Germany | 7.7 | 6.3 | 5.0 | 5.1 |
Source: World Bank WDI (NV.IND.MANF.CD), current US dollars; shares over the world aggregate. The US series ends in 2021. Behind Exhibit 1.
| Measure | Value |
|---|---|
| Median months, request to operation, built 2008 | 22 |
| Median months, built 2015 | 36 |
| Median months, built 2024 | 55 |
| Active queue, end-2024, GW | 2,290 |
| Share of 2000–2019 capacity online by end-2024, % | 13 |
Source: Lawrence Berkeley National Laboratory, Queued Up: 2025 Edition, December 2025. Behind Exhibit 2.
| Agency | Change, % |
|---|---|
| All federal civilian (2,312,301 to 2,074,649) | −10.3 |
| Commerce | −13.8 |
| Energy | −15.1 |
| Treasury | −23.2 |
| National Science Foundation | −30.3 |
Source: Pew Research Center analysis of OPM Federal Workforce Data, March 13, 2026. Behind Exhibit 3.
| Recipient | Award |
|---|---|
| Intel | 7.865 |
| TSMC | 6.6 |
| Samsung (6.4 first announced) | 4.745 |
| All 20 awards, up to | 33.7 |
Source: US Department of Commerce, Office of Inspector General, OIG-25-021-I, June 2, 2025. Cited in Part I.
Hart, Brendan. 2026. “The Power Curve 2026 Outlook.” The Power Curve, June. https://brendanhart.com/reports/2026-outlook.
