Economics / Complements and tariffs
The Price of Making It Real
A million tokens became 285 times cheaper in under two years. An electrical connection did not. Cheap intelligence enlarges what an economy can attempt; retained gains depend on the complements that turn it into useful work — and on whether a tariff helps build them or guards someone’s margin.
By Stratifi Research
November 12, 2026
A million tokens became dramatically cheaper. An electrical connection did not become a million tokens.
Epoch’s price series puts the cheapest access meeting a fixed MMLU benchmark floor at $20 per million tokens on November 30, 2022, and $0.07 on October 3, 2024. That is a 285.7-fold fall in 673 days. These are list prices for a fixed capability floor, not the cost of useful work. Even read that narrowly, it is an extraordinary change in the price of something reproducible. [1]
The cheapest benchmark-floor access fell 285.7-fold in 673 days
The cheapest listed API price for a model scoring at least 64.8 on MMLU fell from $20 (November 30, 2022) to $0.07 (October 3, 2024) per million tokens: 285.7 times lower over 673 days. These are list prices for a fixed capability floor, in nominal dollars per million tokens on a log scale. They are not the cost of useful work, a provider’s cost or a general price index. Dots are record lows only; nothing is implied between them. GPQA Diamond is a different benchmark and has its own panel.
Units: USD per million tokens (log scale). Frequency: irregular record-low price points (not a daily series). Source: Epoch AI price-trend data.Now ask what a buyer is trying to finish: a component design, a fixed fault, a document reliable enough to sign. The token bill covers part of that. Review, retries, integration and the cost of a wrong answer are still on the invoice.
The calculation has become inexpensive. Making it count remains an economic activity.
Cheap intelligence enlarges what an economy can attempt, but retained gains depend on expanding the complements that turn it into useful work—a test that tariffs can either help meet or make harder.
The thesis this piece pursues.
So the question is not what becomes expensive when intelligence becomes cheap. It is what becomes decisive. A complement can grow more important while competition pushes its price down. But abundance does not spread because a price list changed.
The last necessary thing
A project needs a combination of inputs. More reasoning does not substitute for power delivered to the right place or equipment actually on site. The limiting input sets the timing of the whole investment, even when it is a small share of its value.
Power is the clearest case. In Lawrence Berkeley National Laboratory’s latest data, the median time from interconnection request to commercial operation for the 294 generators completed in 2025 was about 61 months. That figure measures completed generating projects, development time included. It is not how long a data center waits to connect its load. [2]
Generators now take about five years from request to operation
Median months from interconnection request to commercial operation for generators completed in each year: 60.8 months in 2025 (n = 294) in the 2026 data vintage, about five years. This measures completed generating projects, development time included; it is not how long a data center waits to connect its load. The shaded band is the middle half of completed projects (25th–75th percentile), a spread of projects rather than a confidence interval. Units: months; lower panel counts projects.
Units: Months; Projects (count). Frequency: annual, by commercial-operation year. Source: LBNL interconnection queue data file (2026).Look at what is in the queue. Active natural-gas capacity in the Berkeley Lab data went from 79 gigawatts at the end of 2023 to 136 a year later and 253 at the end of 2025—more than three times as much in two years. Solar and storage are still larger queue categories, and both shrank from their 2023 peaks. Gas is where the recent surge is. [2]
Gas tripled in the queue in two years; storage shown separately
Active interconnection requests in gigawatts of proposed nameplate capacity (GW, not GWh), as of each year-end in the 2026 edition. Gas, highlighted, went from 79.21 GW (end-2023) to 136.22 (end-2024) and 252.83 (end-2025), 3.19 times in two years. Solar and storage remain larger and both shrank from their 2023 peaks. Storage has its own panel and is never added to generation. Not drawn: other, nuclear and coal, at 52.8, 10.4 and 3.7 GW at end-2025 (also their 2014–2025 peaks). A request is an option on power, not installed or firm supply; earlier report editions are not spliced.
Units: GW nameplate requested. Frequency: annual year-end snapshots. Source: LBNL interconnection queue data file (2026).But what does a queue entry become? Of the projects that entered interconnection queues from 2000 through 2020, 13% of capacity and 19% of projects were operational at the end of 2025. Three-quarters of the capacity had withdrawn. The rest was still active or suspended, so those percentages are not final failure rates. A queue is an option on power, not energized power. [2]
Most queued capacity never reached operation
Status at the end of 2025 of every project that entered an interconnection queue from 2000 through 2020, as a percentage of the cohort’s capacity (MW) or project count. 13.14% of capacity and 19.12% of projects were operational; 75.09% of capacity had withdrawn. Active and suspended projects are still unresolved, so these shares are not final failure rates, and the chart assigns no cause to withdrawals.
Units: Share of cohort (%). Frequency: single cohort snapshot (status at end-2025). Source: LBNL interconnection queue data file (2026).Then how long does a data center wait? There is no single national answer, but the load side has its own queue. ERCOT’s March 13, 2026 status report counts 238.6 gigawatts of cumulative large-load requests projected through 2030. Of that, 5.8 gigawatts was observed energized. Roughly 140 gigawatts of new submissions were pending and not yet in the table. [3]
Texas large-load requests dwarf what has been energized
ERCOT’s large-load requests by status, cumulative by the year the load is projected to connect, in gigawatts, as reported on March 13, 2026. By 2030 the table sums to 238.6 GW; 5.8 GW was observed energized. Roughly 140 GW of new submissions were pending and are not in the table, so they are not drawn. This is a load queue, not a generator queue, and it is not a waiting time. The figures are dated March 2026, not current November totals.
Units: GW (cumulative). Frequency: single dated snapshot, cumulative by projected year. Source: ERCOT March 2026 TAC report.In Virginia, Dominion reports 54.9 gigawatts of data-center capacity under contract, up from 16.5 gigawatts in mid-2023—contracted demand at several stages of agreement, not load served and not a waiting time. [4]
On the demand side, a Berkeley Lab model estimates US data-center electricity use at 192 terawatt-hours in 2024 and projects 649 in its 2030 reference case; both are model output, not metered load. [5]
A lower API price tells us neither the energy used per accepted task nor how many such tasks people will want.
What about the electricity bill itself? The national average commercial price rose about 27% in nominal terms from 2020 to 2025 and less than 2% after inflation. Whatever scarcity there is, the national average retail price is not where it shows. [6]
It shows in capacity markets. PJM’s 2027/28 capacity auction, reported December 17, 2025, cleared every zone at the cap of $333 per megawatt-day and fell 6.5 gigawatts short of its reliability requirement. The 2028/29 auction, reported July 14, 2026, cleared every zone at the cap of $325 and fell 6.8 gigawatts short. These are capacity prices set under a collar, not energy prices or rent. They do say the auction could not buy the reliability it was designed to buy. [7]
PJM capacity auctions cleared at the cap and still fell short
RTO clearing prices in nominal dollars per megawatt-day of unforced capacity (UCAP). The 2027/28 auction cleared every zone at the $333.44 cap and fell 6,516.6 MW short of the reliability requirement; the 2028/29 auction cleared all four zones at the $325.00 cap and fell 6,831.3 MW short. The earlier 2026/27 auction, also at its cap, is shown for context. These are capacity prices set under a collar, not energy prices, household bills or a measure of rent. PJM’s no-collar simulations are not drawn.
Units: USD per MW-day (UCAP); MW (UCAP). Frequency: annual capacity auctions (delivery years). Source: PJM 2027/2028 BRA report; PJM 2028/2029 BRA report.The cost of a constraint often arrives as waiting rather than a higher quote. A cheap input delivered too late is expensive. The unit that matters is a completed use at acceptable quality and time.
Scarcity is not a profit statement
The tempting next step is to name the owners of the bottleneck as the winners. That skips most of the economics. Take two complements: accelerated chips and regulated grid capacity.
Nvidia’s FY2026 10-K reports data-center revenue of $194 billion out of $216 billion total, for the year ended January 25, 2026, and a company-wide GAAP gross margin of 71.1%. Its quarterly filing puts the margin at 75% for the quarter ended in July. Neither filing reports a gross margin for the data-center business alone. [8]
The spending on the other side of the boom is just as large. Amazon, Alphabet and Meta spent about $391 billion in cash capital expenditure in the first nine months of 2026, and Microsoft spent $116 billion in its fiscal year ended in June—all-purpose capex, not AI-only spending. [9]
Is a 71% gross margin a rent, or the return on research, design risk and a production lead competitors have not matched? A gross margin is an accounting measure before operating costs, not a cost of capital or a counterfactual price. The filing establishes the margin, not the rent.
The grid looks different. Virginia’s State Corporation Commission set Dominion’s allowed return on equity at 9.8% in its November 2025 order, below the 10.4% the company requested. The same proceeding created a separate rate class for very large loads, with minimum charges. That 9.8% applies company-wide; it is not a premium on data-center investment. [10]
So here are two complements of the same build-out. One earns a market margin during a spending boom. The other earns a return a commission sets in public. Which one looks like a rent, and which looks like cost recovery?
A market-set margin during a boom is a return on a lead that competition can erode, while a commission-set return is cost recovery by construction—a view that is wrong if Nvidia’s company-wide gross margin holds above about 70% through the end of 2027 despite new entrants.
History offers a caution about reading either one too quickly. Paul David’s 1990 argument about the electric dynamo is that a general-purpose technology pays off only after the factories and practices around it are rebuilt, which takes decades. He drew the parallel to the computer era’s productivity disappointment, the Solow paradox. [11]
The fiber build-out of the late 1990s is the other half of the lesson. Mark Doms, writing in the San Francisco Fed’s Economic Review in 2004 and citing KMI Research, put US fiber route miles at about 143,000 in 1996 and 320,000 in 2000. Traffic grew far more slowly than the builders’ claims, and utilization stayed low. [12]
Where are we now? Information and communication technology investment reached a record 5.4% of GDP in the third quarter of 2026, against 4.5% at the 2000 peak and 4.0% in 2019. That comparison is descriptive. It does not measure a dynamo-style payoff, and the payoff remains a historical argument here rather than a demonstrated regularity. [13]
A scarcity can be local, temporary or institutionally made, and it can induce the expansion that ends it. The scarcity story can be right about what has to be built and wrong about who keeps the premium.
What is the guardrail guarding?
A tariff changes the terms on which imports compete with domestic production. Whether that helps convert cheap intelligence into useful work depends first on the purpose.
Is the objective a capability whose learning benefits the first producer cannot capture? A dependable source of an essential input if foreign supply is cut off? Wages and taxable income kept at home? Or simply more output from the protected industry?
These are not interchangeable. The positive case is real: a producer will not finance a costly early period if know-how leaks to competitors, and a country can value continuity of supply that no import quote prices. But a tariff on equipment needed to deploy cheap intelligence supports the equipment maker while raising costs for every domestic user of it.
The same economy is upstream in one argument and downstream in the next. Nineteenth-century tinplate makes that unusually clear.
Tinplate and the price of an earlier beginning
The United States established its tinplate industry behind the McKinley tariff, whose tinplate provision took effect in July 1891. Douglas Irwin’s analysis gives that protection real credit for timing: his entry model suggests the industry would otherwise have arrived about a decade later. [14]
But tinplate makers needed iron and steel, and those inputs were themselves protected. Irwin also stresses cheaper inputs and knowledge transferred from Britain. And earlier was not the same as worthwhile. His illustrative welfare calculations for the studied tariff found that the gains did not cover the costs. The factory proves that production happened. It does not prove that the timing was worth what users paid. [14]
The other nineteenth-century cases point in different directions, which is the point. Joseph Davis and Irwin’s model puts antebellum pig-iron output roughly 30–40% above a no-protection counterfactual, a modeled supply response rather than an observed alternate history, and finds foreign price movements mattered more than duties. Irwin and Peter Temin conclude that established American cotton-textile producers could have survived without protection by the 1830s. Réka Juhász finds that French regions more sheltered from British competition during the Napoleonic blockade adopted mechanized spinning faster—a wartime French shock, not an American tariff. More output, survival and adoption are different results, and none proves that tariffs caused the American rise. [15] [16] [17]
The modern record has the same double edge. The US International Trade Commission’s 2023 study of Section 232 in 2018–2021 modeled domestic steel quantity 1.9% higher and aluminum 3.6% higher, with downstream output 0.6% lower. Those are model estimates for that period, not a current result. [18]
The modern tinplate test
All tariff rates in this section are as of November 8, 2026.
Now run the tinplate question on the equipment that sits between cheap intelligence and energized power. The Section 232 metals duties were raised to 50%, transformers and their parts were added as steel derivatives, and a later restructuring moved the duties to full customs value and sorted products into annexes. [19]
Grain-oriented electrical steel, the specialized core steel in transformers, sits in Annex I-A at 50% of full value. Large liquid-filled transformers above 10,000 kVA, the category that covers most large power transformers, sit in Annex III at 15% all-in until the end of 2027. Smaller liquid-filled units and transformer cores and laminations sit in Annex I-B at 25%. The common claim that large power transformers face 50% is not true. [20]
So the input is taxed at 50% and the finished good at 15%. Who is protected?
The domestic side is thin. Department of Energy sources through 2024 identify a single US producer of grain-oriented electrical steel, Cleveland-Cliffs. In its June 2024 report, the National Infrastructure Advisory Council put waits for large transformers at two to four years, with average lead times near 120 weeks against about 50 three years earlier. The Commerce Department’s 2020 Section 232 transformer report, as cited by the Energy Department in 2022, found imports supplied 82% of large power transformers of 100 MVA and above by unit count in 2019. There is no primary update since for the lead time, the import share or the producer count. [21] [22] [23]
Effective protection makes the problem precise. With fixed input coefficients, effective protection of domestic value added is the output tariff minus the input share times the input tariff, divided by one minus the input share. The same Commerce 2020 survey, as cited by DOE in 2022, puts electrical steel at about 25% of transformer manufacturing cost. [21]
A 50% steel tariff can leave transformer assembly barely protected or taxed
| Case | Input share | Steel tariff | Output tariff | Effective protection |
|---|---|---|---|---|
| Large liquid-filled transformer, sourced share | 0.25 | 50% | 15% | +3.33% |
| Smaller liquid-filled transformer, sourced share | 0.25 | 50% | 25% | +16.67% |
| Large liquid-filled transformer, illustrative share | 0.35 | 50% | 15% | −3.85% |
The table is the manuscript’s three cases, under the tariff schedule as of November 8, 2026. The curves apply the same formula across steel input shares from 0.10 to 0.45, with the steel tariff at 50% and output tariffs of 15% (large liquid-filled transformers, Annex III) and 25% (smaller units, Annex I-B); dots are the rows of the source sensitivity grid. Effective protection is in percent of domestic value added. Below the labelled zero line, domestic assembly is taxed relative to imports. This is an illustration under stated assumptions (fixed coefficients, full pass-through, one tariffed input; copper and most-favored-nation duties ignored), not a measured effective rate.
Units: Effective protection (%). Frequency: not a time series (formula). Source: DOE Electric Grid Supply Chain Review (2022); Annexes I-A, I-B, II, III, IV (April 2026).The table is a construction, not a measurement. It assumes fixed coefficients, full pass-through and a single tariffed input, and it ignores copper and most-favored-nation duties. At the sourced share, the headline 15% on a large transformer protects domestic assembly by about 3%. At a modestly higher steel share, effective protection turns negative: domestic assembly is taxed relative to imports. That is the tinplate problem with a 2026 date on it.
Buying domestic steel does not escape it. Even if an assembler buys Cleveland-Cliffs electrical steel, the 50% tariff lets the domestic price rise toward import parity, so the input tax lands whether or not a ton crosses the border. The protected party is the steel producer, and the assembler pays.
Is the steel tariff what is driving transformer prices, then? The price data say no, at least not recently. The producer price index for transformers rose 78% from December 2019 to December 2025. Over the last three of those years, transformers rose 11% and switchgear 27% while steel mill products fell 1%. Copper rose 41% while energy wire rose only 6%. For decades, transformer prices followed steel prices within months; after 2022 the two parted ways. The recent squeeze reads as demand-led—grid and data-center orders—rather than a steel-tariff cost push, although asymmetric pass-through cannot be ruled out. An exploratory correlation scan tested these links; none passed its filters. [24]
Transformer prices kept rising after steel prices stopped
Five nominal price series on one axis, each rebased so its 2019 average = 100. Transformers rose 78.0% from December 2019 to December 2025. From December 2022 to December 2025, transformers rose 11.1% and switchgear 26.8% while steel mill products fell 1.0%; copper rose 40.9% and energy wire 6.4%. Dotted guides mark December 2019, 2022 and 2025. This is a descriptive comparison, not evidence of tariff causation, a current steel-tariff cost push or monopoly rent. Copper is the IMF world price.
Units: Index, 2019 average = 100 (nominal). Frequency: monthly, not seasonally adjusted. Source: FRED: PPI transformers (PCU335311335311); FRED: PPI switchgear (PCU335313335313); FRED: PPI energy wire (PCU335929335929); FRED: PPI steel mill products (WPU1017); FRED: IMF copper price (PCOPPUSDM).That leaves a narrower and sharper question. If transformer prices are set by an order backlog, a 50% duty on the core steel does not show up as the current driver of transformer prices. It shows up in the costs of any domestic plant trying to add capacity, which competes with imported transformers paying 15% all-in. Is the guardrail building the plant, or taxing it?
The same pattern runs through the other inputs. Copper’s Section 232 duty is 50% on semi-finished copper and intensive derivatives, with cathode, scrap and ore excluded. The semiconductor action puts 25% on a narrow band of advanced chips and exempts imports for US data centers, among other uses. [25] [26]
One boundary matters. A chip and a transformer cross a border as goods. Remote inference, model subscriptions and software rights do not; they raise different questions of classification, valuation and collection. A goods-tariff argument cannot simply be renamed a tax on digital labor.
A tax collector is not an industrial strategy
Tariffs also have a fiscal job, and its record is easy to misread. Customs supplied about 95% of federal receipts, excluding borrowing, in 1860 and 25% in 1865, not because customs fell—customs dollars rose—but because other receipts exploded with wartime internal taxes. Today the statutory rate, the rate importers owe at entry, gross cash collections and net cash after refunds are four different tariff rates answering four different questions. The lesson is old. A tariff that succeeds as protection shrinks its own base. A tariff that succeeds as revenue needs imports to keep coming. A tariff the courts reverse turns revenue into a liability. [27] [28]
What does the interval produce?
Temporary protection promises both intervention and eventual discipline. Annex III even carries an end date: December 31, 2027. But higher prices do not become an investment in capability because they expire.
What happens during the interval is the economics. Do costs converge? Do more producers enter, including at the input stage? Do transformer lead times fall? Is supply less exposed to interruption, or does the protected domestic assembler still depend on the same imported core steel?
A resilience benefit can be real even when the protected product stays more expensive. Insurance has a price. The test is whether continuity improves enough to justify it.
Retained gains need the same scrutiny. Higher producer revenue does not say how much is offset by higher costs to utilities, data-center builders and households.
The strongest objection to protection is not that domestic capability has no value. It is that the state can preserve an incumbent’s price advantage instead of building the capability that would make protection unnecessary. The strongest defense is not that every factory deserves shelter. It is that some learning and resilience are valuable precisely because the first private investor does not capture all their benefits.
Neither proposition supplies a verdict in advance.
Cheap intelligence makes the distinction more important. Scarce complements delay its gains, collect part of them or expand fast enough to spread them. Policy changes which of those paths is open.
The optimistic case is not a world in which nothing has a cost. It is one in which more useful work becomes economical because the remaining constraints keep loosening.
Perhaps the question to carry forward is not who will own the last scarce thing. It is whether we are building enough of it to stop calling it scarce—and whether a guardrail helps that happen or gives someone a reason to prevent it.
Sources
1
Epoch AI, LLM inference price trends: cheapest listed API price at or above fixed MMLU and GPQA Diamond score floors, data through February 2025 (retrieved March 12, 2025).
2
Lawrence Berkeley National Laboratory, Queued Up: 2026 Edition and the accompanying interconnection queue data file (May 2026).
3
ERCOT, Large Load Interconnection status report to the Technical Advisory Committee, March 13, 2026.
4
Dominion Energy, earnings and investor materials on contracted data-center capacity (54.9 GW; 16.5 GW in mid-2023).
5
Lawrence Berkeley National Laboratory, 2024 United States Data Center Energy Usage Report (December 2024).
6
US Energy Information Administration, Electric Power Monthly: average retail price of electricity, commercial sector, 2020–2025; deflated with the CPI-U.
7
PJM Interconnection, Base Residual Auction reports for the 2027/2028 (December 17, 2025) and 2028/2029 (July 14, 2026) delivery years.
8
Nvidia, Form 10-K for the fiscal year ended January 25, 2026, and Form 10-Q for the quarter ended July 26, 2026.
9
Amazon, Alphabet and Meta quarterly filings covering the first nine months of 2026; Microsoft Form 10-K for the fiscal year ended June 30, 2026.
10
Virginia State Corporation Commission, final order in the Dominion Energy Virginia biennial review, November 2025.
11
Paul A. David, “The Dynamo and the Computer: An Historical Perspective on the Modern Productivity Paradox,” American Economic Review 80(2), 1990.
12
Mark Doms, “The Boom and Bust in Information Technology Investment,” FRBSF Economic Review, 2004, citing KMI Research for fiber route miles.
13
Bureau of Economic Analysis, national income and product accounts: private fixed investment in information processing equipment and software as a share of GDP, third-quarter 2026 release.
14
Douglas A. Irwin, “Did Late-Nineteenth-Century U.S. Tariffs Promote Infant Industries? Evidence from the Tinplate Industry,” Journal of Economic History 60(2), 2000.
15
Joseph H. Davis and Douglas A. Irwin, “The Antebellum U.S. Iron Industry: Domestic Production and Foreign Competition,” Explorations in Economic History 45, 2008.
16
Douglas A. Irwin and Peter Temin, “The Antebellum Tariff on Cotton Textiles Revisited,” Journal of Economic History 61(3), 2001.
17
Réka Juhász, “Temporary Protection and Technology Adoption: Evidence from the Napoleonic Blockade,” American Economic Review 108(11), 2018.
18
US International Trade Commission, Economic Impact of Section 232 and 301 Tariffs on U.S. Industries, Publication 5405, March 2023.
19
Presidential proclamations adjusting imports of steel and aluminum under Section 232 (2025–2026): the increase to 50%, the inclusion of transformers and parts as derivatives, and the restructuring to full customs value with annexes.
20
Section 232 steel and aluminum derivative annexes I-A, I-B, II, III and IV (April 2026): tariff treatment of grain-oriented electrical steel, liquid-filled transformers, cores and laminations.
21
US Department of Energy, Electric Grid Supply Chain Review: Large Power Transformers and High Voltage Direct Current Systems (February 2022), citing the Commerce Department’s 2020 Section 232 transformer survey.
22
National Infrastructure Advisory Council, Addressing the Critical Shortage of Power Transformers to Ensure Reliability of the U.S. Grid (June 2024).
23
US Department of Commerce, Bureau of Industry and Security, Section 232 investigation report on transformers and transformer components (2020).
24
FRED / BLS producer price indexes: transformers (PCU335311335311), switchgear (PCU335313335313), energy wire (PCU335929335929), steel mill products (WPU1017); IMF global copper price (PCOPPUSDM). Monthly, not seasonally adjusted, retrieved November 8, 2026.
25
Presidential proclamation adjusting imports of copper under Section 232 (2025): 50% on semi-finished copper and copper-intensive derivatives; cathode, scrap and ore excluded.
26
Section 232 action on semiconductors (2026): 25% on a specified band of advanced chips, with exemptions including imports for US data centers.
27
US Treasury and Census Bureau historical statistics on federal receipts by source, 1860 and 1865 (customs versus internal revenue).
28
Treasury Monthly Treasury Statement and US Customs and Border Protection collections data: statutory rates, duties owed at entry, gross cash collections and net receipts after refunds.