Edward Conard

Top Ten New York Times Bestselling Author

  • “A full-throated defense of economic dynamism.” - The Wall Street Journal
  • “…challenges misconceptions that distort our economic debates.” - Arthur Brooks, President of the American Enterprise Institute
  • “…a very valuable contribution.” - Larry Summers, former Secretary of the Treasury and director of the National Economic Council, president emeritus, Harvard University
  • “…a comprehensive explanation of the modern economy.” - Julian Robertson, Founder, Tiger Management
  • “…reminds us that inequality sends a signal of what society lacks most, in America’s case, entrepreneurship and risk taking.” - Lawrence Lindsey, CEO, The Lindsey Group, former Director of the National Economic Council
  • “Unintended Consequences is full of substance, it is one of the must-read books of the year, and once I finish it I will be giving it a second read through right away.” - Tyler Cowen, Professor, George Mason University
  • “…a must-read for serious students of economic policy.” - Glenn Hubbard, Dean, Columbia Business School, and former Chairman of the Council of Economic Advisers
  • “…a very valuable contribution.” - Larry Summers, former Secretary of the Treasury and director of the National Economic Council, president emeritus, Harvard University
  • “…reminds us that inequality sends a signal of what society lacks most, in America’s case, entrepreneurship and risk taking.” - Lawrence Lindsey, CEO, The Lindsey Group, former Director of the National Economic Council
  • “…a must-read for serious students of economic policy.” - Glenn Hubbard, Dean, Columbia Business School, and former Chairman of the Council of Economic Advisers
  • “Unintended Consequences should be read by anyone who takes for granted the superiority of progressive taxation and has not thought carefully about the trade-offs involved.” - The New Republic
  • “…challenges misconceptions that distort our economic debates.” - Arthur Brooks, President of the American Enterprise Institute
Upside of Inequality Oxford Unintended Consequences

Macro Roundup


Wednesday, December 3, 2025

AI Summary. U.S. real GDP grew at a 1.1% annualized rate over the first half of 2025, but excluding booming AI-related investment output contracted 0.1%, with non-AI investment weakening. The OECD identifies a correction in equity markets lifted by expectations of high AI returns as a key downside risk to its projections.

OECD Staff Organisation for Economic Co-operation and Development
Date Posted:

OECD estimates that US real GDP grew at an annual rate of 1.1% in the first half of 2025, but net of AI-related investment contracted slightly. OECD projects overall US real GDP growth of 2% in 2025 and 1.7% in 2026.

Is AI investment the only thing driving US growth?

Core argument: Real GDP growth in the first half of 2025 was 1.1% annualised, resilient but uneven despite rapidly cooling job growth and numerous headwinds, according to the OECD Economic Outlook.

Despite rapidly cooling job growth and numerous headwinds, real GDP growth through the first half of 2025, at 1.1% (annualised), was fairly resilient, if uneven. Excluding AI-related investments, which continued to boom, GDP contracted by 0.1%. Non-AI-related investment weakened. A key downside risk to the projection is a correction to equity markets that have been buoyed by the hopes of high returns to investment in AI, although new advances in AI could boost growth in the years ahead.

Takeaways by Macro Roundup® AI

  1. Real GDP growth in the first half of 2025 was 1.1% annualised, resilient but uneven despite rapidly cooling job growth and numerous headwinds, according to the OECD Economic Outlook.
  2. Excluding AI-related investments, which continued to boom, GDP contracted by 0.1%, with non-AI-related investment weakening, the OECD Economic Outlook reports.
  3. The OECD identifies a correction in equity markets buoyed by hopes of high returns to AI investment as a key downside risk to its projection, while noting that new AI advances could boost growth in the years ahead.

Related Articles:

  • How America’s AI Boom Is Squeezing The Rest Of The Economy — About a sixth of America's 2% real GDP growth over the past year came from investment in computer and communications equipment and data centres, and one estimate puts the artificial intelligence boom's total contribution to growth at 40%. Average electricity bills rose 7%, attributed in part to data-centre demand, while housebuilding and non-AI business investment slumped and real consumption flatlined.
  • Honey, AI Capex Keeps Eating … Everything — The author estimates that AI-related capital spending contributed between 0.6% and 1.3% of a quarter's 3% U.S. GDP growth, based on information technology equipment and intellectual property investment data. Google, Amazon, Meta, and Microsoft alone spent about $69bn in the quarter, roughly half of all information technology equipment spending on an annualized basis.
  • SF FedViews: Solid Growth, but Small Cracks Have Emerged — Since the second half of 2023, four-quarter US productivity growth has averaged nearly 2.5%, well above the pre-pandemic trend of about 1% to 1.25%, supporting real GDP growth. It remains unclear whether this reflects a higher productivity regime fostered by rising AI use or a temporary post-pandemic reorganization surge.

AI Summary. US public debt has risen sharply while household and financial sector indebtedness has fallen, a trend in place since the 2008 crisis. Because long-term forward yields in traditional safe havens such as Japan and Germany have risen substantially, the author argues Treasuries' relative attractiveness as a safe haven explains falling US long-term yields.

Robin Brooks Robin Brooks Substack
Date Posted:

Ruling out private deleveraging as old news and Fed cuts as irrelevant at the long end, Brooks concludes that the reason Treasury yields have fallen slightly this year despite an “out of control” fiscal policy is that “the US is the best of a bad bunch.”

Why are US long-term yields falling despite rising public debt?

Core argument: US public debt has risen sharply while household and financial sector indebtedness has fallen, a pattern the author describes as a long-established trend since the 2008 crisis rather than a driver of 2025's decline in long-term yields.

The chart shows indebtedness across different sectors in the US economy. It’s clear that public debt has risen sharply (red) and that this rise has been offset by falling indebtedness in the household (blue) and financial (orange) sectors. The reason I don’t think private deleveraging is behind this year’s fall in long-term yields is that this is a long-established trend since the 2008 crisis. There’s just no way a long-term trend like this wasn’t priced long ago, so this just isn’t a good explanation for why yields are down in 2025. That leaves Fed rate cuts and “the best of a bad bunch” as explanations for falling yields. As I’ve noted in the past, looking at 10y20y forward yields removes cyclical noise from the short end of the yield curve, stripping out rate cutting (or hiking) expectations. As a result, these yields provide a “cleaner” read of what’s going on with long-term yields. What’s notable is [that] traditional safe havens like Japan and Germany have seen big increases in long-term yields. This makes Treasuries an attractive safe haven on a relative basis and - in my opinion - is the main reason US Treasury yields are down this year. The US is the best of a bad bunch and that isn’t going to change any time soon.

Takeaways by Macro Roundup® AI

  1. US public debt has risen sharply while household and financial sector indebtedness has fallen, a pattern the author describes as a long-established trend since the 2008 crisis rather than a driver of 2025's decline in long-term yields.
  2. The author argues that rising long-term yields in traditional safe havens such as Japan and Germany make Treasuries relatively attractive, which he identifies as the main reason US Treasury yields fell in 2025.
  3. The author uses 10y20y forward yields to strip out rate-cut and rate-hike expectations from the short end of the curve, arguing they give a cleaner read on long-term yields.

Related Articles:

  • Our Thoughts on Large US Deficits and Their Impact on Bond Yields — Total borrowing in the US is about average despite high government deficits, as high public borrowing has contributed to the need for higher interest rates that have depressed private sector borrowing. Risk premiums on long-dated debt look about average, with no market signs that economic or debt limits are currently being hit.
  • The Race Between Asset Supply and Asset Demand — The authors estimate that demand for assets has outpaced supply by roughly 400% of GDP versus 100%, making rising asset demand the leading contributor to falling interest rates. They project that debt sustainability requires fiscal consolidation exceeding 10% of GDP, though debt could reach 250% of GDP without raising interest rates.
  • Across The Rich World, Fiscal Crises Loom — An analysis estimates most rich countries could keep debt stable as a share of their economies with only small primary surpluses—Britain's required surplus is 0.3% of GDP—even if all debt were refinanced immediately at current rates, but the tightening needed to stabilize debt ratios exceeds 2% of GDP in Britain and America and 3% in France.

AI Summary. Nonbank financial institutions' holdings of financial assets rose 74 percentage points relative to global GDP between 2008 and 2023, versus 17 points for banks. The Bank for International Settlements warns that rising government bond issuance alongside banks' post-crisis retrenchment has widened a gap in dealers' capacity to intermediate these markets.

Martin Wolf Financial Times
Date Posted:

BIS’s Hernández de Cos warns that rapidly rising sovereign debt-holding by non-bank financial institutions is creating new risk, citing highly leveraged hedge fund holdings and “dollar funding rollover risks related to [pensions’] use of FX derivatives.”

Is the shift of government bond holdings to nonbanks a risk?

Core argument: Non-bank financial institutions' holdings of financial assets rose 74 percentage points relative to global GDP between 2008 and 2023, compared with a rise of 17 percentage points for banks.

Rising public debt is one concern. Another is how it is being financed. The ratio of NBFI holdings of financial assets to global GDP rose 74pp between 2008 and 2023, while that of banks rose by only 17pp. Hernández de Cos [of the BIS] warns that the combination of rising government bond issuance with the post-financial crisis retrenchment of banks created a steadily growing gap between the supply of government bonds and the assets of bank dealers needed to support the capacity of these crucial markets to intermediate. NBFIs are a heterogeneous group. A vital distinction is between “real money” investors, such as pension funds and insurers, and leveraged speculators, notably hedge funds. The former group of long-term private investors has greatly increased its holdings of government bonds, from an amount worth 82% of global GDP in 2008 to one of 135% in 2023. Meanwhile, money market funds and hedge funds have also increased the value of their holdings from 13% to 18% of global GDP over these years. Many of these NBFIs also need to hedge foreign currency risks, given the rise in their cross-border holdings. This has caused a huge jump in their reliance on foreign currency swaps.

Takeaways by Macro Roundup® AI

  1. Non-bank financial institutions' holdings of financial assets rose 74 percentage points relative to global GDP between 2008 and 2023, compared with a rise of 17 percentage points for banks.
  2. "Real money" investors such as pension funds and insurers increased government bond holdings from 82% of global GDP in 2008 to 135% in 2023, while money market funds and hedge funds went from 13% to 18%.
  3. Hernández de Cos of the BIS warns that rising government bond issuance combined with post-financial-crisis bank retrenchment has created a steadily growing gap between the supply of government bonds and the bank dealer assets needed to intermediate these markets.

Related Articles:

  • Data Update 4 for 2025: Interest Rates, Inflation and Central Banks! — Interest rates are ultimately driven by macroeconomic fundamentals, with an "intrinsic risk-free rate" defined as the sum of the inflation rate and real GDP growth tracking the observed 10-year Treasury bond yield closely. On this view, central banks' influence comes from preferential access to data on those fundamentals, their capacity to alter them, and their credibility.
  • Investors Now See Some Corporate Bonds as Safer Bets Than Government Debt — Average net debt relative to earnings for about 2,500 companies in the MSCI ACWI global stock index improved to 1.74x in the first half of this year from 2.53x a decade earlier, indicating stronger corporate balance sheets. Meanwhile, the International Monetary Fund projects government debt across major economies will rise annually through 2030 to about 137% of GDP.
  • Why Microsoft Has Lower Borrowing Costs Than the U.S. — Two AAA-rated Microsoft bonds trade at yields slightly below equivalent U.S. Treasurys, an unusual negative spread given that companies are more likely to default than a government able to tax and that corporate bonds are harder to sell. Demand for high-quality corporate debt has compressed the AAA corporate spread to 0.3 percentage point over Treasurys.

AI Summary. The authors estimate that when Spain's radical-right party Vox was absent from the ballot, the right bloc lost 2.21 percentage points, with the mainstream conservative Popular Party capturing 82.9% of the Vox vote. Blank and spoiled protest votes rose 67%, and the two main left-wing parties also gained.

Jesús Fernández-Villaverde and Carlos Sanz Journal of Public Economics
Date Posted:

A last-minute, single-district dropout of the radical right (RR) Vox party in Spain’s 2023 election saw the classical right capturing 82.9% of the RR vote. Protest votes increased 67% to ~1.5% of the total; left parties’ share of the vote rose 1.4pp.

Where do radical-right voters go when their party isn't on the ballot?

Core argument: Using a synthetic difference-in-differences model on Spanish census tracts where Vox fielded no list, the authors estimate the right bloc would have gained an additional 2.21 percentage points of the electoral roll had the radical right run.

Due to infighting about the order of the candidates within the list, several of Vox’s candidates [announced] they were no longer running a few hours before the filing deadline [with no time for Vox to present the complete slate required to have a party list]. We estimate the effects of the RR’s absence on electoral outcomes using a synthetic difference-in-differences model. The identifying assumption is that, aside from Vox’s absence, no other shock affected electoral behavior in the treated census tracts relative to the control group. Our results imply that the right bloc would have gained an additional 2.21pp over the electoral roll had the RR run. The CR only captured a fraction of the RR vote: assuming no cross-bloc transfers, our estimates indicate that the main CR party in Spain, the Popular Party (PP), attracted 82.9% of the Vox vote. A second finding is that the RR’s absence from the ballot increases protest voting (i.e., blank and null votes). Protest votes rise by 67% (from 0.91% to 1.52% of the census) in our base line specification. This suggests a substantial share of expressive, rather than strategic, voting. A third finding is that left-wing parties capture a portion of the RR vote. The vote shares for the two main left-wing parties increase by 0.96 and 0.42 percentage points, respectively.

Takeaways by Macro Roundup® AI

  1. Using a synthetic difference-in-differences model on Spanish census tracts where Vox fielded no list, the authors estimate the right bloc would have gained an additional 2.21 percentage points of the electoral roll had the radical right run.
  2. The authors estimate that the Popular Party, Spain's main classical-right party, attracted 82.9% of the Vox vote assuming no cross-bloc transfers, while the two main left-wing parties gained 0.96 and 0.42 percentage points of vote share.
  3. Protest voting, comprising blank and null ballots, rose 67%, from 0.91% to 1.52% of the census, in the authors' baseline specification when the radical right was absent from the ballot, suggesting a substantial share of expressive rather than strategic voting.

Related Articles:

  • Political Representation Gaps and Populism — A study comparing survey responses of citizens and parliamentarians across 27 European countries finds that policymaking reflects voters' economic attitudes well, but the average parliamentarian is about one standard deviation more culturally liberal than the national mean voter. The author finds this cultural gap spans nearly all issues, countries, and established parties, and that right-wing populists fill it.
  • Europe’s Right-Wing Populists Swerve Left on the Economy — Right-wing populist parties across Europe, including the Sweden Democrats, Austria's Freedom Party, Germany's Alternative for Germany and France's National Rally, have shifted from small-government, low-tax, free-trade platforms toward calls for higher welfare benefits, subsidies and protectionism, often alongside pro-market positions.
  • Zero-Sum Thinking and the Roots of U.S. Political Divides — Zero-sum thinking — viewing gains for one group as losses for another — is associated with greater support for liberal economic policies, stronger alignment with the Democratic Party, and weaker alignment with the Republican Party. The authors also find it linked to believing the QAnon conspiracy theory holds some truth and to sympathy for the attack on the U.S. Capitol.

AI Summary. Small, often invisible national and EU regulatory barriers persist within the European Union's single market for goods, such as overlapping labelling requirements that produce a 20cm label on a 12cm toy. The International Monetary Fund estimates these internal barriers act as a drag equivalent to a 44% tariff on Europe's economy.

Richard Milne and Sam Joiner Financial Times
Date Posted:

French and EU regulations require a 20cm label on a 12cm €2 toy elephant. The IMF estimates that such “small, often invisible barriers to trade, taken together,” amount to “a drag on Europe’s economy equivalent to a tariff of 44%.”

How costly are hidden regulatory barriers within Europe's single market?

Core argument: The IMF estimates that remaining small, often invisible barriers to trade within the EU single market for goods amount to a drag on Europe's economy equivalent to a tariff of 44%.

If there is one product that symbolises the absurdities of the EU’s single market in goods, it might be a cuddly elephant. The Djungelskog toy from Ikea may only be 12cm tall and cost less than €2, but it has a 20cm-long label protruding from its behind. Environmental rules from France and EU textile legislation account for much of the length. But the elongated label reflects a strategic, even a philosophical, problem for Europe as its companies battle to compete globally. The single market in goods should be the EU’s crowning achievement. Compared with free movement of people, services or capital, it is the area that has come furthest in more than 30 years since the common market was established. Yet there remain small, often invisible barriers to trade that, taken together, amount to what the IMF estimates is a drag on Europe’s economy equivalent to a tariff of 44%.

Takeaways by Macro Roundup® AI

  1. The IMF estimates that remaining small, often invisible barriers to trade within the EU single market for goods amount to a drag on Europe's economy equivalent to a tariff of 44%.
  2. Ikea's Djungelskog toy elephant, 12cm tall and costing less than €2, carries a 20cm label whose length is largely accounted for by French environmental rules and EU textile legislation.
  3. The EU's single market in goods has advanced further than free movement of people, services or capital in the more than 30 years since the common market was established.

Related Articles:

  • The Future of European Competitiveness – A Competitiveness Strategy for Europe — A European Union competitiveness report attributes the EU's labour productivity growth gap with the US largely to weaker activity in information and communications technology and large-scale digital services; excluding main ICT sectors, EU productivity was broadly at par with the US from 2000-2019. It estimates Europe's investment share must rise about 5 percentage points of GDP.
  • Forget the US — Europe has Successfully Put Tariffs on Itself — The International Monetary Fund estimates Europe's internal barriers are equivalent to tariffs of 45% on manufacturing and 110% on services, leaving trade across European Union countries at less than half the level of trade across US states. Regulation of digital services is estimated to have cut profits for small European tech firms by up to 12%.
  • Europe Regulates Its Way to Last Place — Europe's regulators tend to act on the mere possibility of harm rather than evidence of it, an approach that can throttle innovation early. The McKinsey Global Institute found that among companies with over $1bn in revenue, U.S. firms spent 80% more on research and development, had 30% higher return on capital, and grew revenue 1.3 percentage points faster.

AI Summary. Copper demand is being boosted by grid infrastructure for the green transition and by data centres for artificial intelligence, which miner Grupo México estimates require 27–33 tonnes of copper per megawatt, over twice the need of conventional data centres. BHP estimates copper used in data centres worldwide will grow sixfold by 2050.

Camilla Hodgson Financial Times
Date Posted:

Current and planned copper production will meet only 70% of 2035’s projected world demand, according to an IEA forecast. The shortfall is being driven by increased demand from electrification and global rearmament.

How much copper do AI data centers require?

Core argument: AI data centres require 27–33 tonnes of copper per megawatt, more than twice the requirement of conventional data centres, according to miner Grupo México.

Demand for copper is being boosted by the construction of grid infrastructure for the green transition and to power data centres for artificial intelligence. These need between 27 and 33 tonnes of copper per megawatt of power, according to miner Grupo México, over twice the requirement of conventional data centres. BHP, the world’s biggest mining group by market value and a shareholder in Resolution, estimated in January that the amount of copper used in data centres worldwide will grow “sixfold by 2050”. Then there is the move towards global rearmament. “A lot of copper demand is hidden,” says mining entrepreneur Robert Friedland. “Military requirements for copper are never publicly disclosed.” Analysts at Société Générale estimate that global defence spending grew by 9.4% to $2.7tn in 2024. Prices have responded, with copper hitting a series of record highs since October. The benchmark London Metal Exchange price currently stands at more than $11,000 per tonne, compared with around $8,500 two years ago.

Takeaways by Macro Roundup® AI

  1. AI data centres require 27–33 tonnes of copper per megawatt, more than twice the requirement of conventional data centres, according to miner Grupo México.
  2. BHP, the world's biggest mining group by market value, estimated in January that copper used in data centres worldwide will grow sixfold by 2050.
  3. Benchmark London Metal Exchange copper stands above $11,000 per tonne, compared with around $8,500 two years ago, having hit a series of record highs since October.

Related Articles:

  • Capital Is Making a Comeback — Tangible investment in property, factories and equipment fell from 12.5% to 8.5% of private gross domestic product, while intangible investment in intellectual property, software and process knowledge rose from about 11.5% to 16.75%. The author argues demand for tangible capital investment will now have to rise.
  • The Pursuit of Anglo American Highlights a Global Copper Problem — Mining companies' interest in acquiring Anglo American reflects the value of its copper mines in Chile and Peru, with copper futures contracts trading well above spot prices. Such acquisitions add no new copper supply, and it is simpler and cheaper to buy a rival copper miner than to build a new mine.
  • Copper Producers Warn of Lack of Mines to Meet Demand for Metal — Copper producers warn that existing and planned mines are insufficient to meet expected demand, which is projected to climb as the world electrifies and as living standards rise in India and other developing nations. Anglo American estimates Western living standards require 200-250kg of copper per person, versus a global average of 60kg.

Thursday, December 4, 2025

AI Summary. Google's Gemini 3, trained entirely on Google's own tensor processing unit chips, is the best-performing AI model across nearly all dimensions except possibly coding, leaving OpenAI without a leading model for the first time in three years. Google's vertically integrated stack from chips to data centers likely gives it the cheapest access to computing power among leading AI labs.

Greg Jensen and Jas Sekhon Bridgewater Associates
Date Posted:

Bridgewater notes that Gemini 3’s release, “trained entirely on Google’s own TPU chips,” outperforms ChatGPT in almost all tasks, implying that “Google is now the clear leader in the AI race.”

Does owning the full chip-to-model stack determine AI leadership?

Core argument: Google's Gemini 3, trained entirely on Google's own TPU chips, is the best-performing model across nearly all dimensions except possibly coding, marking the first time since GPT-3.5 launched three years ago that OpenAI lacks a leading model.

The release of Gemini 3, trained entirely on Google’s own TPU chips, confirms Google’s leading position. Across nearly all dimensions (with the possible exception of coding, as mentioned previously), Gemini 3 is the best-performing model. This marks the first time since the launch of GPT-3.5 three years ago that OpenAI doesn’t have a leading model. Looking ahead, Google has a significant advantage in training frontier models, and efficiently and cheaply servicing them—in part due to its vertically integrated ecosystem. Google has its own large, mature, and competitive hardware stack, from chips to data centers, which likely gives it the cheapest and easiest access to compute among leading AI labs. And after nearly a decade of limiting external access to its TPU chips, Google has also started exploring selling and leasing them to other AI labs and cloud providers in recent months, opening the possibility of it dominating the full AI stack, from chips to models.

Takeaways by Macro Roundup® AI

  1. Google's Gemini 3, trained entirely on Google's own TPU chips, is the best-performing model across nearly all dimensions except possibly coding, marking the first time since GPT-3.5 launched three years ago that OpenAI lacks a leading model.
  2. Google's vertically integrated stack, spanning its own chips and data centers, likely gives it the cheapest and easiest access to compute among leading AI labs for both training and servicing frontier models.
  3. After nearly a decade of limiting external access to its TPU chips, Google has begun exploring selling and leasing them to other AI labs and cloud providers, opening the possibility of dominating the full AI stack from chips to models.

Related Articles:

  • On Benchmark Games, Gemini, and Declining Returns to Scale — Composite benchmarks show Google's Gemini roughly in line with GPT-5-class models, with any gap small, and the author argues it does not lead on strict low-leakage reasoning exams. The pattern reflects growing training costs yielding modest gains concentrated in vision, video, and some coding rather than broad reasoning advances.
  • AI Capex Risk as Predictable Engineering — The 2020 scaling-laws paper fostered a belief that mixing compute, data, and parameters yields better AI models in a deterministic way, which the author argues let unusually large capital spending plans pass without investor revolt. The author contends that curve was empirical rather than physical and is now breaking, returning AI capital spending to normal economics.
  • Is AI Hitting A Wall? — Artificial intelligence gains have come from a formula of more data and more computing power, but companies have exhausted freely available internet training data and are now striking data-sharing deals with publishers and copyright holders. Training and running large models is energy-intensive, requiring vastly more Nvidia chips with each generation.

AI Summary. The author argues that frontier model pre-training is becoming uneconomic, delivering only marginal capability gains for enormous spending on chips, power, and compute. Meaningful advances have shifted to post-training, inference-time compute, and tool-augmented reasoning rather than raw model size.

Paul Kedrosky Applied Complexity
Date Posted:

Kedrosky argues that Google’s Gemini 3 suggests “scaling curves aren’t [yet] broken, but they’re exhausting and breaking.” This implies the capex super-cycle will “briefly” deepen but likely only result in marginal improvements.

Is scaling model size still economically worth the cost?

Core argument: Gemini 3 shows larger training runs still yield some capability gains, but the author argues the gains are small relative to the spending required, with pre-training's economic viability collapsing.

The takeaway shouldn't be “scaling is alive.” It should be that frontier training is becoming uneconomic, fast, and the industry is pushing ahead anyway. That means: The capex super-cycle briefly deepens: more chips, more power, more exaflop-month training runs. The economic value of scaling shrinks while the cost explodes. Advances, such as they are, have shifted to post-training, inference-time compute, and tool-augmented reasoning—not raw model size. Gemini 3 shows you can still squeeze some capability out of larger training runs. But it also shows how little you get for that prodigious spending. Pre-training's economic viability is collapsing, forcing the industry into a financially brutal and unsustainable arms race for marginal improvements.

Takeaways by Macro Roundup® AI

  1. Gemini 3 shows larger training runs still yield some capability gains, but the author argues the gains are small relative to the spending required, with pre-training's economic viability collapsing.
  2. The author contends frontier AI training is becoming uneconomic quickly while the industry pushes ahead anyway, briefly deepening the capex super-cycle through more chips, more power, and more exaflop-month training runs.
  3. Advances have shifted to post-training, inference-time compute, and tool-augmented reasoning rather than raw model size, according to the author, who describes the industry as locked in an arms race for marginal improvements.

Related Articles:

  • Google’s Gemini 3 Means AI’s “Resource Grab” Phase Is On — Google's Gemini 3, trained entirely on Google's own tensor processing unit chips, is the best-performing AI model across nearly all dimensions except possibly coding, leaving OpenAI without a leading model for the first time in three years. Google's vertically integrated stack from chips to data centers likely gives it the cheapest access to computing power among leading AI labs.
  • On Benchmark Games, Gemini, and Declining Returns to Scale — Composite benchmarks show Google's Gemini roughly in line with GPT-5-class models, with any gap small, and the author argues it does not lead on strict low-leakage reasoning exams. The pattern reflects growing training costs yielding modest gains concentrated in vision, video, and some coding rather than broad reasoning advances.
  • AI Capex – Financing The Investment Cycle — Global data center, AI infrastructure and related power buildout could cost over $5tn, requiring funding from every public capital market plus private credit, alternative capital providers and possibly governments. The largest cloud providers generate over $700bn of annual operating cash flow and reinvest roughly $500bn in capital spending, with about $300bn a year estimated to fund AI and data centers.

AI Summary. A simulation of safety-net benefits for a single mother with two children in North Carolina shows net income rising faster than earnings below $30,000 a year, then dropping abruptly as benefits phase out. She would need to earn $70,000 to exceed the net income available at $30,000.

Angela Rachidi, Matt Weidinger, Joshua Bandoch, Nic Dunn, et al. American Enterprise Institute
Date Posted:

A non-working, non-disabled mother of two in North Carolina can collect $50k in benefits. Due to benefit phase-outs, she would have to earn $70k in market income to exceed her net income from earning $30k alongside $40k of benefits.

How much must a single mother earn to surpass benefit-level income?

Core argument: A simulated single mother with two children in North Carolina would need to earn $70,000 per year to reach a higher net income than earning $30,000 per year while keeping her full safety-net benefits.

Figure 3 simulates actual safety-net benefits a single mother with two children in North Carolina would receive based on different earnings scenarios. Notably, the mother in this example generally experiences a work incentive at earnings levels below $30,000 per year—that is, her net income grows even more than her earnings alone due to the way benefits phase in. This is primarily because the EITC and CTC—and to some extent SNAP—grow with additional earnings up to a certain point. However, this work incentive shifts to a disincentive as this mother reaches earnings levels at which benefits start to phase out. The top line of Figure 3 reflects net income (i.e., total benefits plus earnings less taxes) and shows that the combined effect of phasing out all of these benefits over a similar earnings distribution results in an abrupt drop in net income as the family starts to earn more than $30,000 per year. In North Carolina, a mother receiving this package of benefits would have to earn $70,000 per year from work to reach a higher net income than if she earned $30,000 per year from work and kept her full benefits. This is a substantial work disincentive.

Takeaways by Macro Roundup® AI

  1. A simulated single mother with two children in North Carolina would need to earn $70,000 per year to reach a higher net income than earning $30,000 per year while keeping her full safety-net benefits.
  2. Below $30,000 in annual earnings, the simulated North Carolina mother's net income grows by more than her earnings alone, primarily because the EITC, the Child Tax Credit, and to some extent SNAP phase in with additional earnings.
  3. The combined phase-out of these benefits above $30,000 in earnings produces an abrupt drop in the simulated family's net income, which the analysis characterizes as a substantial work disincentive.

Related Articles:

  • Mitigating Benefits Cliffs for Low-Income Families: District of Columbia Career Mobility Action Plan as a Case Study — The authors estimate that a hypothetical low-income family with children receives $68,686 in government support at an income of $11,000 but only $22,709 at $65,000, while its tax liability rises from $842 to $12,835. The lost assistance plus higher taxes fully offsets the $54,000 income gain.
  • “Compensate the Losers?” Economic Policy and Partisan Realignment in the US — The authors argue that partisan realignment by education reflects Democrats' declining supply of "predistribution" policies such as job guarantees, higher minimum wages, protectionism, and stronger unions, which less-educated Americans favor while more-educated Americans prefer taxes and transfers. They estimate changing views of the parties' economic policies explain roughly half the shift.
  • Welfare Is What’s Eating the Budget — The U.S. redistributes 29.4% of its gross domestic product through transfers and taxes, a larger share than any developed country except France at 30.1%. Counting transfers as income and taxes as income lost and adjusting for household size, average households in the bottom, second, and middle income fifths have roughly equal incomes despite large differences in work effort.

AI Summary. Long-dated forward yields such as 10-year and 20-year forwards are near their highs even as the Federal Reserve cuts rates, with the gap between those forwards and 10-year yields widening, which the author reads as a worsening buyers' strike for longer-dated government debt. Italy, France, the UK, Japan and Germany all show elevated long-term forward yields.

Robin Brooks Robin Brooks Substack
Date Posted:

Citing high 10y10y and 10y20y forward yields across advanced economies, Brooks notes the failure of traditional safe havens – particularly German and Japanese bonds, suggesting the “early stages of a global debt crisis that’s building and intensifying.”

Are investors backing away from long-dated government debt worldwide?

Core argument: Federal Reserve rate cuts are not pulling down long-term yields: 10y10y and 10y20y forward yields sit near their highs, and the gap between both measures and the 10-year yield has widened.

Fed cuts aren’t pulling down long-term yields: if you look at the long end of the yield curve properly by stripping out front-end yields, things look very worrying. 10y10y and 10y20y forward yields are near their highs and the gap of both metrics with 10-year yield has grown. The buyers’ strike for longer-dated Treasury debt looks like it’s getting worse. There’s many idiosyncratic trouble spots: a key feature of the recent rise in long-term yields is that there’s many fiscally distressed countries that run into trouble at different points and for idiosyncratic reasons. Italy, France and the UK are all part of this dynamic and have very elevated 10y20y forward yields. Traditional safe havens are failing: in the past, Japan and Germany would have been safe harbors in this kind of environment, but these days they’re at the heart of the rise in long-term yields.

Takeaways by Macro Roundup® AI

  1. Federal Reserve rate cuts are not pulling down long-term yields: 10y10y and 10y20y forward yields sit near their highs, and the gap between both measures and the 10-year yield has widened.
  2. Italy, France and the UK all show very elevated 10y20y forward yields, part of a dynamic in which many fiscally distressed countries run into trouble at different points and for idiosyncratic reasons.
  3. Japan and Germany, which in the past would have served as safe harbors, are at the heart of the rise in long-term yields, while the buyers' strike for longer-dated Treasury debt appears to be worsening.

Related Articles:

  • Across The Rich World, Fiscal Crises Loom — An analysis estimates most rich countries could keep debt stable as a share of their economies with only small primary surpluses—Britain's required surplus is 0.3% of GDP—even if all debt were refinanced immediately at current rates, but the tightening needed to stabilize debt ratios exceeds 2% of GDP in Britain and America and 3% in France.
  • The Race Between Asset Supply and Asset Demand — The authors estimate that demand for assets has outpaced supply by roughly 400% of GDP versus 100%, making rising asset demand the leading contributor to falling interest rates. They project that debt sustainability requires fiscal consolidation exceeding 10% of GDP, though debt could reach 250% of GDP without raising interest rates.
  • Our Thoughts on Large US Deficits and Their Impact on Bond Yields — Total borrowing in the US is about average despite high government deficits, as high public borrowing has contributed to the need for higher interest rates that have depressed private sector borrowing. Risk premiums on long-dated debt look about average, with no market signs that economic or debt limits are currently being hit.

AI Summary. France is reviving voluntary military service as a rite of passage for young people, starting with a 10-month program aimed at recruiting 3,000 people, mostly age 18 or 19, to supplement 200,000 active troops, scaling to 50,000 a year by 2035. In a "major crisis," service would become mandatory for people with certain skills if Parliament authorizes a draft.

Stacy Meichtry and Noemie Bisserbe Wall Street Journal
Date Posted:

The French government is resurrecting a framework for large-scale mobilization in order to prepare for a war with Russia “that many believe is only three to four years away.” The program will start with 3,000 18–19-year-old volunteers this summer.

Is Europe rebuilding voluntary military service to prepare for mass mobilization?

Core argument: French President Emmanuel Macron announced a 10-month military service starting next summer, recruiting 3,000 people largely aged 18 or 19 to supplement France's 200,000 active troops, with recruitment rising to 50,000 a year by 2035.

French President Emmanuel Macron is resurrecting the concept of military service as a rite of passage for young people, part of a broader push to put Europe on a war footing to counter an expansionist Russia. Macron and other European leaders are taking a gradual approach, laying the groundwork for mass mobilization with recruitment systems they can rapidly scale up in a moment of crisis. Macron said a new 10-month service would begin next summer with the goal of recruiting 3,000 people—largely age 18 or 19—to supplement the country’s 200,000 active troops. Recruitment would gradually increase to 50,000 a year by 2035. In the event of a “major crisis,” military service would become mandatory for people with certain skills, Macron said, if Parliament authorizes such a draft.

Takeaways by Macro Roundup® AI

  1. French President Emmanuel Macron announced a 10-month military service starting next summer, recruiting 3,000 people largely aged 18 or 19 to supplement France's 200,000 active troops, with recruitment rising to 50,000 a year by 2035.
  2. Macron said military service would become mandatory for people with certain skills in the event of a "major crisis," provided Parliament authorizes such a draft.
  3. Macron and other European leaders are taking a gradual approach to putting Europe on a war footing against an expansionist Russia, laying the groundwork for mass mobilization with recruitment systems they can rapidly scale up in a crisis.

Related Articles:

  • German Army Struggles to Get Gen Z Recruits ‘Ready for War’ — Germany's armed forces commissioner warned the country is not moving closer to its target of 203,000 active troops by 2031, as the military slightly shrank amid high dropout rates, with a quarter of new recruits leaving within six months. Conscientious objector applications reached 2,998, up from 200 in 2021.
  • The Intellectual Spoils of War? Defense RD, Productivity and International Spillovers — The authors estimate that a 10% increase in government-financed R&D, largely defense, is associated with a 4.3% rise in privately funded R&D across OECD countries, indicating "crowding in" rather than "crowding out." They also find international spillovers raise private R&D abroad, though the overall effect on US productivity growth is modest.
  • Germany Backs Donald Trump Goal For Nato To Spend 5% Of GDP On Defence — Germany has backed the US demand that NATO allies spend 5% of GDP on defence, up from the alliance's current 2% target. Berlin signalled it would meet the goal via the NATO secretary-general's proposal of 3.5% of GDP on hard military spending plus 1.5% on related areas such as infrastructure and cyber security.

AI Summary. Europe, at roughly 500m people and a $26tn economy with world-class aerospace, shipbuilding, pharmaceutical, vehicle and weapons manufacturers, has the material basis to be a military superpower. National militaries' preference for home-country suppliers has left subscale defense firms, and common procurement could raise output and lower unit costs.

Matt Klein The Overshoot
Date Posted:

Klein argues that better coordination and mobilization of current resources could meaningfully increase European military capacity. He suggests $200B in Norwegian windfall profits should be used to support broader European efforts against Russia.

Can common defense procurement turn Europe into a military superpower?

Core argument: Europe, defined as the EU plus Norway, Switzerland, and the U.K., has roughly 500 million people versus about 340 million in the U.S. and a $26 trillion economy versus China's $21 trillion; if material considerations were all that mattered, it would be a military superpower.

Europe [the European Union plus Norway, Switzerland, and the U.K.] has a far larger population than the U.S. (~500 million vs. ~340 million), and an economy far larger than China’s ($26 trillion vs. $21 trillion). It also has a large and sophisticated manufacturing sector, with world-beating producers in aerospace, shipbuilding, pharmaceuticals, machine tools, motor vehicles, and weapons. If material considerations were all that mattered, Europe would be a military superpower. The trick is to make sure that the additional spending translates to greater quantities, rather than higher prices. The challenge is that Europe’s national militaries have had a tendency to favor their home country’s producers, with the result that the continent is filled with subscale defense companies that are often slow and unproductive. Common defense procurement for a continental army should lead to higher output and lower costs—a few large companies handling large orders should have better unit economics than hundreds of artisanal manufacturers—but it would require Europe’s national defense elites to change their perspective.

Takeaways by Macro Roundup® AI

  1. Europe, defined as the EU plus Norway, Switzerland, and the U.K., has roughly 500 million people versus about 340 million in the U.S. and a $26 trillion economy versus China's $21 trillion; if material considerations were all that mattered, it would be a military superpower.
  2. Europe's manufacturing base includes world-beating producers in aerospace, shipbuilding, pharmaceuticals, machine tools, motor vehicles, and weapons, but national militaries' tendency to favor home-country producers has left the continent with subscale defense companies that are often slow and unproductive.
  3. Common defense procurement for a continental army should raise output and lower costs, since a few large companies handling large orders should have better unit economics than hundreds of artisanal manufacturers, though it would require Europe's national defense elites to change their perspective.

Related Articles:

  • Will Higher Defence Spending Boost The European Economy? — Defense spending typically has a fiscal multiplier below one, meaning each euro spent raises GDP by less, so it is not a short-term growth strategy. One estimate holds that extra military spending worth 1% of GDP could lift output up to 2% long term if concentrated on research and development.
  • Why Every Country Needs To Master The Electric Tech Stack — Drones are the foundation of hard power in the early 21st century, so a country unable to produce them domestically is at a disadvantage. As electric motors overtake combustion engines as the core technology that makes machines move, the same supply chains and production processes used for electronics can also produce cars, drones, robots, and appliances.
  • Behind NATO’s 2 Percent: Measuring the True Scope of Alliance Defense Investments and the NATO Defense Deficit — NATO's collective defense spending averaged 1.72% of GDP over roughly two and a half decades, below the alliance's agreed 2% threshold, leaving a cumulative shortfall the authors estimate at over $2tn that was not invested in force structure, modernization, and readiness. The authors argue closing the gap requires years of sustained investment beyond current commitments.