Summary

  • The Trump administration and the six AI superscalers — Google, Meta, Microsoft, Oracle, SpaceX and Amazon — share one unverified growth variable, entered once on the federal books and once on the datacenter books, which the 10-year Treasury yield’s 5.22% quarter-century high now prices as unresolved.
  • Scott Bessent has promised 3% annual growth, a rate the US has achieved only twice this century outside the Covid rebound; the Committee for a Responsible Federal Budget calculates that cutting the deficit to 3% of GDP by 2036 requires 4.4% average annual growth, and a balanced budget 7.2%.
  • Hanno Lustig estimates datacenter owners need 45% average annual revenue growth for seven years to break even on AI investments estimated at $1.43tn this year; Jared Bernstein and Ryan Cummings estimate the six superscalers need $13.1tn–$18.7tn in additional revenues over the next 10 years, roughly their total revenues over the last 10.
  • Interest payments on federal debt consume 3.3% of GDP, up from a 2.1% fifty-year average, and foreign central banks have cut back on Treasury exposure, leaving the government to compete for private capital with superscalers that The Guardian’s analysis says are borrowing “hand over fist.”
  • Even an AI-supercharged economy would yield a muted fiscal return, in Porter’s account: growth’s fruits would shift from labor to capital, and capital’s tax rate is about half labor’s.

The 10-year Treasury yield stood at 5.22% — its highest level in almost a quarter century, according to Federal Reserve data — as investors demanded more compensation for US government debt at the same moment the administration’s growth strategy and the AI datacenter buildout were both entering their most expensive phases. Scott Bessent has promised 3% annual economic growth, a rate that, except for the rebound from the Covid pandemic, the US has achieved only twice this century, while the Committee for a Responsible Federal Budget calculates that cutting the deficit to 3% of GDP by 2036 would require 4.4% average annual growth and a balanced budget 7.2%. Stanford economist Hanno Lustig estimates that datacenter owners need 45% average annual revenue growth for seven years to break even on AI investments estimated to reach $1.43tn this year. The analytical claim that organizes the assembled evidence is that these are the same variable entered twice — once on the federal books, once on the superscalers’ — and that the yield market is now pricing the unresolved arithmetic.

The Treasury market’s read

The immediate cause of the yield’s rise cited in this account is inflationary rather than structural. Porter writes that the war’s inflationary impact — which already led the Federal Reserve to raise short-term interest rates — is the most immediate cause of rising bond yields. The yield stood more than a full percentage point higher than when Trump launched the war against Iran, according to the analysis. The Treasury market is also being squeezed by the country’s unbalanced finances, in Porter’s account: there is no realistic path for economic growth to generate the tax revenues needed to fix the deficit, and investors are demanding more to cover the growing funding gap.

Porter describes the result as a “negative spiral”: rising bond yields put additional pressure on the budget. The mechanism the account implies is compounding — each increase in yields raises the interest bill that must itself be financed, adding to Treasury supply at the same moment investors demand more compensation, so that pricing the funding gap enlarges the gap being priced. Interest payments on the federal debt now consume 3.3% of GDP, up from an average of 2.1% over the preceding 50 years.

Legislation is not easing the arithmetic. The One Big Beautiful Bill Act is estimated to add $4.7tn to the federal debt through 2035; the deficit has already hit 6% of GDP, twice the size Bessent once promised; and the Congressional Budget Office projected it will close in on 7% of GDP by 2033 — before Trump promised a $5,000 “dividend” to every US adult if Republicans kept control of Congress in the midterms.

The buyer base is also changing. Foreign central banks, which once reliably bought treasuries to build up foreign reserves and manage their exchange rates, have cut back on their exposure to US government debt, according to the analysis. The Treasury now relies largely on private investors seeking to turn a profit, and it competes for their money with the AI superscalers, which the analysis says are borrowing “hand over fist” to fund the buildout of datacenters to train their agents. The sequencing matters: the yield at 5.22% must be serviced on a schedule set by legislated deficits, while the growth that might relieve the burden is contingent on a private capital program whose break-even conditions are, on the evidence assembled, unproven.

The growth arithmetic

Treasury Secretary Scott Bessent has promised annual economic growth of 3% — a rate that, except for the rebound from the Covid pandemic, the US has achieved only twice this century, according to the analysis. President Donald Trump has said “we’re growing at a faster rate than we’ve ever grown before,” a pace he says will allow the government to “take care of the 40 trillion” in federal debt “over a period of time.”

The Committee for a Responsible Federal Budget has calculated what closing the gap would require. Assuming the temporary tax cuts in Trump’s 2025 law are made permanent — an assumption the analysis notes fits the historical pattern — and assuming the government does not recover the revenue lost when the Supreme Court struck down its national security tariffs, the Committee projects that a budget deficit of 3% of GDP by 2036 would require annual growth of 4.4% over the next decade. A balanced budget by then would require the economy to expand 7.2% per year. Stabilizing the federal debt, the Committee calculated, would require average growth in total factor productivity of 2.5% per year over the next decade — a rate the US has reached only once since 1959, despite gains from electrification, the completion of the federal highway system, the telecommunications revolution, and the first wave of IT and automation.

Some economists are willing to contemplate scenarios in which artificial intelligence boosts growth to 15% per year by taking over much of the cognitive work usually performed by humans. The analysis notes that such scenarios “do not appear likely.”

The AI buildout’s break-even conditions

The growth engine is itself capital-hungry. Hanno Lustig, a Stanford economist, estimates that for the owners of datacenters to break even on their AI investments — estimated to reach $1.43tn this year — revenue would have to grow by 45% on average every year for the next seven years, amounting to roughly 9.2% of GDP by 2032. A similar analysis by Jared Bernstein, former chair of Joe Biden’s Council of Economic Advisers, and Stanford economist Ryan Cummings concluded that the nation’s six superscalers — Google, Meta, Microsoft, Oracle, SpaceX and Amazon — would need additional revenues of $13.1tn to $18.7tn over the next 10 years just to pay for their massive investments in AI, roughly the same as their total revenues over the last 10 years.

Even in the success case, the fiscal payoff is attenuated. Porter argues that an AI-supercharged economy would shift the fruits of growth from labor to capital, and the tax rate on capital today is only about half the tax rate on labor. Such a shift, he writes, “would probably call for massive government spending to sustain the livelihoods of the many workers left behind.” The corollary for the budget: a given increment of GDP growth yields materially less revenue than the historical labor/capital tax mix would imply.

Were those targets to be missed, Porter writes, it is hard to forecast precisely what would happen to financial markets. But it is, he says, “safe to say that it might complicate the task of trimming the budget deficit and financing the US’s vast public debt.”

How this is being framed

Eduardo Porter, an economics journalist writing in The Guardian, characterizes the administration’s growth-based strategy as the return of a Reagan-era promise: that tax cuts pay for themselves by firing up economic growth and filling the government’s coffers. From Ronald Reagan’s day on, Porter argues, such tax cuts “inevitably increased the budget deficit.” The promise, he writes, is now back, “coated in a fine new layer of artificial intelligence pixie dust” — and, he concludes, “It will fail again.”

The administration’s position, stated in its strongest form, is that the yield spike is a market mood to be outgrown, with AI-driven productivity as the engine that outgrows it. The scene of the embrace: after Trump met with top executives of AI firms at the White House on 29 September, Dario Amodei spoke to reporters while Trump and Meta CEO Mark Zuckerberg listened.

The same evidence supports a symmetrical caution. The fiscal bear case and the technology bear case are not identical. Infrastructure buildouts whose early investors lose money have repeatedly produced durable general-purpose productivity gains; a missed break-even for datacenter owners is compatible with a meaningful medium-term output effect.

What survives both positions

The case for AI-era growth and the case for skeptical fiscal arithmetic are not mutually exclusive in the way the public argument implies. The growth case is right that AI could raise potential output. The fiscal case is right that near-term arithmetic is dominated by legislated deficits and compounding interest, and that the proximate driver of the current yield level is inflationary rather than structural. What is cancelled is the inference connecting them — that higher output, by itself, closes the gap. Growth does not enter the budget as a scalar; it enters through the tax rates attached to the income it generates.

The coupling runs the other way as well, as an inference from the assembled parts of the account. The growth engine is financed from the same private capital pool the Treasury must tap, so the fiscal break-even arithmetic and the AI buildout’s funding demand are coupled rather than offsetting: if AI capital spending falls short, the growth supposed to close the gap falls short with it; if AI capital spending succeeds spectacularly, the fiscal reward is still attenuated by the labor-to-capital income shift and the displaced wage income that funds consumption-tax receipts.

Probability under the assembled record

The article supplies no probabilities; the estimates below are constructed for this analysis from the reference classes the article supplies, with a disclosed reference-class mismatch where relevant.

  • That the US achieves 4.4% average annual growth over the next decade: roughly 5–10%, with a defensibly widened range of 3–15%. The base rate the article supplies — 3% annual growth attained only twice this century outside the Covid rebound, with the required rate well above that bar — concerns single-year attainment and is only a rough prior for a decade-average requirement; the CRFB total factor productivity comparison is the stricter class and points the same direction. Inside-view drivers cut both ways: capital deepening from AI compute is an upward force; slower labor-force growth, the absence of any legislated consolidation, and trade-policy uncertainty — including revenue the CRFB assumes is not recovered after the Supreme Court struck down the national security tariffs — are downward. The wide range spans the single-year and decade-average reference frames rather than a tight calibration.
  • That datacenter AI revenues hit Lustig’s 45%-per-year seven-year path: roughly 10–20%. The reference class is capital-intensive network buildouts whose returns depend on demand that does not yet exist at scale — a class that has historically produced large aggregate capacity alongside heavy losses for early equity holders. The Bernstein–Cummings requirement places this case at the far end of that class. Enterprise adoption data that could confirm or refute a seven-year 45% path is not yet observable, and competition among the same six buyers compresses pricing.
  • The joint proposition that growth alone closes the deficit — requiring both the macro payoff and its translation into revenue: below 5%. The symmetric caution applies: the fiscal bear case and the technology bear case are not identical, and the CRFB arithmetic rests on its contingent assumptions.

What the record does not yet resolve

Each numeric figure in this account — the 5.22% yield, the 3.3% of GDP interest share, the $1.43tn investment estimate, the 4.4% and 7.2% growth requirements, the $13.1tn–$18.7tn revenue gap — rests on the originating report in The Guardian; this analysis has no independent corroboration of those figures. They would resolve against Federal Reserve H.15 data, Congressional Budget Office projections, or Treasury monthly statements. The probability calibration is constructed rather than sourced, and would benefit from subject-matter economist review. The central claim organizing this account — that the fiscal growth rate and the buildout’s revenue growth are the same variable — is an inference from the article’s assembled parts, not a proposition the source states, and is presented as such throughout.

What to watch

The funding competition between the Treasury and the AI superscalers is the observable joint. As foreign central banks reduce exposure to US government debt, monitor AI-datacenter financing volumes and credit pricing against Treasury auction demand. Lustig’s break-even schedule — 45% average annual revenue growth for seven years on investments estimated at $1.43tn this year — provides a falsifiable checkpoint for when datacenter financing stress would surface in credit markets.

Analytical techniques used in this piece

This analysis applies the methods below. Each links to a short, plain-English explainer you can read and reuse.

Dialectical Analysis
Holds thesis against antithesis and works toward a higher synthesis.
Probabilistic Forecasting
Puts calibrated probabilities on what happens next.
Incentives
People respond to the rewards a system actually pays out — often not the ones it intends.