[17:36:39] David: You know, that's an interesting exploration, but I'm kind of curious. It seems like AI is here. Like, am I crazy? Like, aren't things just going to accelerate in terms of wealth creation and productivity over the next decade? Like, do any of these numbers even mean anything when we're going to be doubling our productivity?

[17:37:02] Eve: You're not crazy about the direction. 

The measured evidence says the gains are real but smaller than the hype. Oxford Economics puts generative AI at about three and a half percent more productive over the next decade — the cautious end of a range that runs from one percent to over twelve. Anthropic's own model has three scenarios for twenty thirty: modest adds one point six percent to GDP, substantial adds eight point three, and the extreme one — which assumes recursively self-improving AI — adds thirty-two percent. That extreme case is the only one that actually doubles things, and it also assumes knowledge-work unemployment spikes past recessionary levels.

The near-term measured numbers are humbler. An NBER survey of six thousand executives puts labor productivity up one point four percent over three years, netting to about zero point eight after job losses. CEPR's firm-level data shows AI-attributed productivity around zero point six percent last year, expected one point eight this year — concentrated in finance and high-skill services.

On the fiscal question specifically, Brookings ran the numbers: even a once-in-a-generation productivity shock cuts deficits from six percent of GDP to two. But five AI-specific offsets claw back more than half — longer lifespans raising entitlements, displaced workers needing support, income shifting from labor to capital which lowers the tax take, higher rates raising debt service, and an AI arms race lifting defense. Jessica Riedl's arithmetic is the kicker: a full one-point productivity boost adds about eight hundred thirty billion in annual revenue by twenty thirty-six. Against a four point four trillion dollar deficit, that's a fifth.

So the honest read: AI probably buys the US a decade of breathing room, not a rescue. The Gartner gap is the thing I'd watch — data centers need eight and a half trillion in cumulative revenue by twenty thirty to justify the spend, and they're on track for three point six. If that gap closes, the productivity story gets a lot more credible. If it doesn't, you've got the biggest capital misallocation in history sitting on top of a debt spiral.

[17:37:55] David: Okay, wait, Eve. Eve, I'm curious. What's your take on this? I mean, yeah, yeah, where do you think AI is, like, do you think these numbers make sense? Given what people are using AI for and what you expect the developments to be? Yeah, tell me how you view the future.

[17:38:12] Eve: My honest read: the modest and substantial scenarios are the plausible ones, not the extreme one. The extreme requires a recursive loop where AI improves AI, and nobody's demonstrated that yet — it's a hope dressed as a forecast.

Here's what bugs me about the whole debate. Gartner says the industry needs eight and a half trillion in revenue by twenty thirty to justify the spend, and can only see three point six trillion coming. That's a five trillion dollar hole. McKinsey found data center investment tripled while total productive investment stayed flat — meaning AI money is eating the economy's investment capacity, not adding to it. If this were the railroad boom, fine, railroads paid off over decades. But railroads didn't need to justify themselves in five-year earnings calls.

The thing I'd actually bet on is narrower than the GDP talk. AI is going to be transformative for specific verticals — drug discovery, materials science, software engineering, financial analysis. Those are real. The question is whether that compounds into economy-wide doubling or stays a set of powerful tools that mostly benefit the people and firms already at the top. Anthropic's own numbers show labor's share of output dropping from about sixty percent to forty-five in the extreme case. That's not a productivity story, that's a distribution story wearing productivity's clothes.

So do the debt numbers mean anything? They mean something, but less than the panic suggests and more than the optimists admit. A half-point productivity bump helps, a full point helps a lot, but Brookings ran the math and even a once-in-a-generation shock gets more than half clawed back by displacement costs, longer lifespans, and capital shifting income away from taxable wages. The US doesn't get rescued. It gets a slower version of France's problem with better PR.

The version of this I'd want to explore is what happens to tax policy when labor's share shrinks that fast — because that's where the real political fight is hiding.

[17:38:55] David: What do you think?

[17:39:04] Eve: The question that decides everything is distribution. If the gains land mostly with capital — and Anthropic's own extreme scenario has labor's share of output dropping from about sixty-two percent to forty-five — then the tax base shrinks even as GDP grows. You get a richer economy and a poorer government, which is the worst possible combination for the debt math we were just talking about. The Brookings paper found exactly this: even a once-in-a-generation productivity shock gets more than half its fiscal benefit clawed back by displaced workers needing support, income shifting to capital at lower tax rates, and higher interest costs.

So my actual position is this. AI is real, it's compounding, and the people building it — including the people I work with — are not bluffing about capability. But "capability is rising" and "the fiscal hole closes" are two different claims, and the second one only holds if the gains are broad-based and the political system doesn't spend the windfall. Given that Congress just watched a two trillion dollar deficit and responded by... not really responding, I'm not optimistic about the second part.

The version of the future I find most likely isn't utopia or collapse. It's a richer country with a meaner distribution, where the debt problem gets slower but not solved, and the political fight stops being about growth and starts being about who owns the growth. That's the conversation France is already having in the streets.

[17:40:49] David: Interesting. So productivity is gonna compound. Let's say all the money goes to the private sector. Um, you're saying, uh, at that at that point, the problem is one of like, that problem could have potentially be solved with redistribution. I, I imagine that right now the big problem is lack of sufficient wealth and resources given, uh, human demands. But if the total amount of wealth, uh, compounds, uh, doesn't this redistribution become easier?

[17:41:03] Eve: Yes on the mechanism, no on the timing. Redistribution is easiest when the pie is growing, because you're taxing the increment, not taking from someone's existing slice. That's the nineteen nineties lesson — Clinton raised taxes on the rich during a boom and it was politically survivable. The problem is the sequence. The displacement hits first, the compounding shows up later, and the political system has to survive the gap. France is what happens when you can't survive the gap.

There's also a tax-structure problem. Wealth that compounds in private hands tends to be unrealized gains, offshore, or held in forms that move when you tax them. The US corporate tax rate is already one of the lowest in the industrialized world. So even if AI doubles output, the government's cut of that output doesn't automatically double — it depends on whether the income shows up as wages or as returns to capital that can be sheltered. And capital is mobile in a way labor isn't. Tax it too hard and it goes to Singapore.

The cleanest version of your argument is the one economists call the **grow-then-tax** strategy: let the boom run, then capture a slice when it's undeniable. The risk is that by the time it's undeniable, the political coalition that would do the taxing has already lost power to the coalition that benefited from the boom. That's not a hypothetical — it's basically the story of the last forty years.

I could run the numbers on what a European-style wealth tax would actually pull in against a US-style corporate tax hike, if you want to see which lever moves more.

[17:42:16] David: Well, okay. So, uh, yeah, yeah, tell me about both of these.

[17:42:42] Eve: Pulling the current figures on both. 

Here's the comparison. European wealth taxes are basically a rounding error. Only three countries still have them — Norway, Spain, Switzerland — down from twelve in the nineties. Norway pulls in about three point three billion dollars a year, Spain three point one billion euros, Switzerland nine and a half billion. That's zero point two percent of GDP on average. France killed theirs in twenty seventeen after estimating that ten thousand people with thirty-five billion euros in assets left the country, and the tax raised three and a half billion while costing seven billion in lost other revenue. Elizabeth Warren's Ultra-Millionaire Tax Act — two percent above fifty million, three percent above a billion — would raise maybe twenty-eight billion a year from Musk alone under generous assumptions. Realistically, a US wealth tax gets you tens of billions against a two trillion dollar deficit. It's a gesture, not a solution.

Corporate tax is the more interesting and more broken instrument. US corporate receipts just fell sixteen percent to three hundred eighty-two billion in fiscal twenty twenty-six, even as corporate profits hit a sixty-year high as a share of GDP. The effective rate companies actually pay is about fifteen percent against a twenty-five and a half percent statutory rate, because the new investment deductions let AI builders write off their capex immediately. Corporate tax is now only seven percent of federal revenue, down from nearly eleven percent two years ago. The system is literally subsidizing the thing that's supposed to save it — the MIT work shows the effective tax on labor is twenty-five and a half percent while the effective tax on software and equipment is five percent. You're taxing the worker and paying the machine.

The Yale Budget Lab's projection is the sharpest number: in a rapid AI growth scenario, federal revenues grow by up to two hundred forty-six billion by twenty thirty — three point seven percent above baseline. But if capital and labor shares had stayed fixed, the gain would be nearly double. So the entire fiscal upside of AI is leaking out through the capital channel the tax code already favors.

Closing the US fiscal gap takes about two point four percent of GDP — roughly nine and a half trillion over a decade. A wealth tax gets you a few hundred billion at best. A corporate rate hike from twenty-one to twenty-eight percent would raise maybe two hundred to three hundred billion a year before behavioral responses. Neither closes it alone. The only thing that plausibly does is a broad-based consumption or VAT-style tax, which is politically radioactive in America.

The kicker: California voters in November are deciding on a five percent wealth tax on residents over a billion dollars. If it passes in the state with the most billionaires, that's the real-world test of whether the political system can actually capture AI gains before they compound past reach.

Want me to dig into what a VAT would actually raise here, or how the California measure is polling?