a16z Podcast · · 1h 4m
Alex Imas & Tyler Cowen | AI Is Moving Fast. Why Isn’t the Economy?
With Alex Imas, Tyler Cowen
Summary
Alex Imas and Tyler Cowen debate whether AI capability gains are translating into economic growth, and they largely agree that the real economy is moving far more slowly than the technology. Imas argues that agentic systems are improving fast, but that token price competition, flat revenue despite exploding usage (citing Ramp data), and weak productivity evidence leave willingness to pay unproven. He also says that if AI breaks the coordination limits on firm size, labour markets could shift abruptly, and that GDP may understate gains unless alternative welfare measures are used. Cowen is more sceptical of extreme revenue projections for OpenAI and Anthropic, of Waymo's real-world benefits, and of fast-takeoff science, while still calling the leading AI firms great companies and expecting AI to change the world over time. They both see institutional inertia, regulation, and local data-center opposition as brakes on the pace of adoption.
Key takeaways
- Tyler Cowen (bearish on Anthropic): Prediction markets don't see Anthropic's IPO above $2T, and its revenue projections fall short of fast-takeoff scenarios.
- Tyler Cowen (bearish on OpenAI): OpenAI's extreme revenue projections are likely false, since no one could name sectors that justify them.
- Alex Imas on US growth & recession risk: Current data implies AI keeps growth in single digits; double-digit growth needs increasing returns to ideas.
- Alex Imas on AI: overall outlook: Real economic transformation needs AI agents that coordinate well, not just more individually competent agents.
- Erik Torenberg on AI infrastructure & data centers: Compute supply is the real scarcity; short-term demand exceeds visible usage, though payoff is speculative.
What was said, by company and theme
31 views from 3 speakers. Every quote is checked against the episode's transcript (how).
Anthropic
Across the shows: 115 speakers in other episodes, 174 bullish and 87 bearish calls, past 30 days
Where everyone stands on Anthropic
Tyler Cowen · George Mason UniversityBearish
Prediction markets don't see Anthropic's IPO above $2T, and its revenue projections fall short of fast-takeoff scenarios.
Cowen points to prediction markets for the Anthropic IPO price, which bounce around but do not rise above $2 trillion. He also says Anthropic's own projected revenue numbers are not what one would expect under super-fast takeoff scenarios. He describes himself as underwhelmed. This is a modest bearish update on the speed of takeoff and the revenue numbers, not a rejection of the company.
“If you look at prediction markets for what the Anthropic IPO will go for, they bounce up and down, but it's not flying above 2,000,000,000,000.”at 6:05
OpenAI
Across the shows: 39 speakers in other episodes, 15 bullish and 34 bearish calls, past 30 days
Where everyone stands on OpenAI
Tyler Cowen · George Mason UniversityBearish
OpenAI's extreme revenue projections are likely false, since no one could name sectors that justify them.
Cowen says that in the Bay Area, people pushed very high revenue numbers for OpenAI and Anthropic on him. He asked which other sectors would grow fast enough to justify those numbers, and nobody answered or even tried. He concludes the extreme revenue numbers are false. He adds that this does not make him bearish on the sector overall.
“what are the other sectors that are gonna grow sufficiently rapidly to justify the extreme revenue numbers you're giving me for Anthropic and OpenAI? And no one had an answer... So I think that means the extreme revenue numbers are false.”at 7:19
Waymo
Across the shows: 6 speakers in other episodes, 6 bullish and 0 bearish calls, past 30 days
Where everyone stands on Waymo
Tyler Cowen · George Mason UniversityBearish
Waymo's contribution to human welfare is tiny; it is fun, but mostly unavailable outside a few cities.
Cowen argues the driverless car has not delivered much. Its biggest benefit is entertainment value. It is not better than an Uber, and is a little worse for him because it obeys the speed limit. Most of America still does not have access, and Europe is far behind. He concludes that the transition is still far from done.
“Like the contribution of driverless cars to actual human welfare is tiny... Most of America still doesn't have it. Europe, forget about it.”at 22:52
NVIDIA
Across the shows: 147 speakers in other episodes, 139 bullish and 112 bearish calls, past 30 days
Where everyone stands on NVIDIA
Tyler Cowen · George Mason UniversityUncertain
Nvidia's leadership turmoil weakens its ability to act as buyer or lender of last resort in a macro shock.
Cowen says the system is robust, but the biggest change would be if something went wrong at the macro level and Nvidia had to step in as buyer or lender of last resort. A weakened or departed founder-type CEO would make that much less feasible than it would have been without this episode. He thinks this wears off in the medium term, but it could slow things down for a while. He treats the structural forces as ultimately dominant.
“if something goes wrong at the macro level and Nvidia had to step in as a buyer or lender of last resort with a weakened or gone CEO founder type, they're a lot less able to do that”at 57:28
AI: overall outlook
Where everyone stands on AI: overall outlook
Tyler Cowen · George Mason University
Cowen remains bullish on AI: it will change the world, benefits exceed costs, but the path will be more normal.
Cowen says that the revenue concerns do not make him bearish on the sector as a whole. He calls the leading AI companies great companies that will be fine. He expects AI to change the whole world, with benefits exceeding costs. He qualifies this: the adoption path will take time and be quite a bit more normal than many people claim.
“That doesn't make me bearish on the sector as a whole. I'm still bullish.”at 7:38
Alex Imas · Google DeepMind
Real economic transformation needs AI agents that coordinate well, not just more individually competent agents.
Imas's formal paper argues that the relevant competitor to AI agents is a human organization, not an individual. Even at minimal AGI (average human ability on every dimension), economic impact stays small because organizations already beat average humans through division of labour. Current 'agentic swarms' are mostly parallel processing: one agent beats two agents splitting work, while two humans who divide labour and communicate beat one human with double the resources. If agents get better at coordinating, they can break the Coasian ceiling on firm size, reshaping firms, wages and employment.
“What you actually need for real economic transformation is not just competence... It's the ability of these agents to coordinate with one another.”at 32:16
Tyler Cowen · George Mason University
AI agents reaching 1.5% of GDP within five years would be an explosion; 10% of GDP is not soon.
Asked when agents will be 10% of GDP, Cowen says not soon. He argues that rising from near zero to 6% would count as an explosion, and that 6% is still not near. He reasons that 1.5% of GDP is a huge number in absolute terms and that he could see it within five years. He frames this as a forced guess rather than a firm call.
“I don't think it's soon. ... If it's like one and a half percent, I could see that, I don't know, within five years.”at 37:19
Alex Imas · Google DeepMind
Agent adoption is bottlenecked by human-built websites, regulation, and platforms blocking outside agents.
Even capable consumer agents hit bottlenecks. Websites are built for humans and differ from one another, so agents make errors such as booking a bad seat. Regulation is a major constraint. Companies spend billions on choice architecture to upsell humans, and agents don't respond to that psychology, so platforms have an incentive to keep outside agents out. As an example, a consumer agent cannot shop on Amazon for the user.
“Websites are designed for human beings, and they're all different. ... So these platforms have an incentive not to allow, you know, somebody else's agent to go in and shop for them.”at 38:00
Alex Imas · Google DeepMind
AI will do almost everything, since training verifiers makes nearly any task verifiable.
Imas says many people dismiss AI by saying certain tasks are not verifiable, but that is copium. Reinforcement learning with verifiable rewards in post-training means that if you train a verifier, almost anything becomes verifiable. Chess is an example where intuition-based skill fell to machines. On that logic, AI will be able to do almost everything.
“But it's going to be able to do almost everything. ... A lot of people use that as copium, like, Oh, that's not verifiable.”at 48:30
Tyler Cowen · George Mason University
Fast AI takeoff is overstated; simulation cannot substitute for experiment in most areas of science.
Cowen says Leopold Aschenbrenner's Situational Awareness is too enamored of fast takeoff. The essay assumes simulation can substitute for experiment, which produces an endless loop of accelerating science. Cowen does not think simulation substitutes for experiment very much. It will in some areas, but not broadly enough to drive the accelerating loop the essay describes.
“he's assuming simulation can substitute for experiment ... And I don't think simulation very much subs in for experiment. It will in some areas.”at 54:53
Alex Imas · Google DeepMind
AI-built structural models, such as digital twins for macro, will be a major underrated breakthrough.
Imas argues AI could build better structural models of micro and macro phenomena, like a model of the cell or digital twins for macroeconomics. Structural understanding lets you predict and create new things, unlike reduced-form estimates that break under policy changes. He thinks AI's potential to recreate that structure is underrated, and Tyler Cowen notes Raj Chetty's digital twin paper as an example.
“to the extent that you are gonna be able to put structure on things that we need to be able to use to predict, that's going to be the real big blow up, I think.”at 58:59
AI regulation & safety
Where everyone stands on AI regulation & safety
Tyler Cowen · George Mason University
Tougher AI regulation is more likely and sooner, a modest bearish factor on takeoff speed and revenue.
Cowen says a lot of data centers will still be built, but the likelihood of regulation, tougher regulation, or regulation arriving sooner has risen, including the possibility of Trump giving ground on guardrails. He calls this not a major force, but a reason to be a modestly more bearish on the sector in terms of takeoff speed and revenue.
“There's greater likelihood of regulation or somewhat tougher regulation or regulation a bit sooner”at 10:01
Alex Imas · Google DeepMind
Alignment should rely on mechanism design and incentive structures around agents, not only on agents' own preferences.
Imas contrasts two approaches. One aligns the models themselves with a constitution so their preferences match the deployer's. The other, the mechanism-design view from economics, does not depend on the agent's preferences at all. Checks and balances make the agent act in the principal's interest even if the agent is individually misaligned. He says research coming from labs, internal and external, increasingly takes this direction. He cites the Hugging Face incident as mostly human error on top of human error, exposing broken incentives for agents. Labs learned from it and built better environments, so the trend is positive despite setbacks.
“you align things, don't almost like don't even care what the preferences of these agents are... you can create a mechanism around them where the incentives are such that the agent behaves in an aligned way.”at 25:46
Tyler Cowen · George Mason University
Real AI harms come mostly from bad humans directing AIs; checks and balances, not perfect alignment, are the answer.
Cowen says as an economist he thinks in principal-agent terms. Unless conditions are very restrictive, the principal cannot control the agent well, no matter what contract is written. He says the Hugging Face incident did not update him much, since the answer was always checks and balances, as with separation of powers. He thinks alignment work matters, but the actual problems seen so far, such as hacks, mostly come from evil humans instructing AIs to do harmful things. He adds that the labs are more reasonable in practice than their public talk suggests.
“the actual problems we've seen, you know, the hacks, they're mostly coming or maybe entirely coming from evil human individuals where the AIs are doing what the humans told it to do.”at 27:36
Alex Imas · Google DeepMind
Political backlash will endogenously constrain extreme AI scenarios, so AI's economic impact will be slower than feared.
Imas argues that the governors' backtracking happens because unemployment and growth look fine, so there is no instrumental reason for a strong political reaction yet. He says that people projecting extreme scenarios miss the fact that an extreme scenario would trigger an extreme political reaction, which would constrain it. The result is back-and-forth dynamics, and future data will show things moved much slower than people expected.
“with the beginning of an extreme scenario, there will be the beginning of an extreme political economy reaction, and that will kind of endogenously constrain things.”at 41:19
AI infrastructure & data centers
Where everyone stands on AI infrastructure & data centers
Tyler Cowen · George Mason University
Most data centers will still be built, but regulatory risk will slow the buildout somewhat.
Cowen argues that the price competition worries are overplayed. He thinks a lot of data centers still will proceed. However, he expects the greater regulatory likelihood to slow things down a bit, though he says it is not a major force.
“I think a lot of data centers still will proceed, but I think it will slow things down a bit.”at 9:58
Tyler Cowen · George Mason University
Local data-center opposition will slow buildout and raise costs, but most projects will still go ahead.
Cowen expects the backlash against data centers to slow things somewhat but not stop them. He points to 50 states and many localities, some of which will say no. He sees the Texas governor already backtracking, and says projects will proceed if unions are involved and pay good wages. He cites Josh Shapiro's strong rhetoric as contrasting with a much softer reality. In his view the backlash raises costs but won't matter much in the end.
“It seems it will slow down a bit, but still basically go through. ... So it raises costs but I don't think ultimately it will matter that much.”at 40:08
Erik Torenberg · Andreessen Horowitz (a16z)
Compute supply is the real scarcity; short-term demand exceeds visible usage, though payoff is speculative.
Torenberg says the firm sees the biggest scarcity on the supply side, and that in the short term demand is bigger than it looks. The view rests on super users who say they can't get enough compute. He acknowledges the counter-argument: super users are a small share of people and companies, AI is not yet broadly diffused, and there is no measurable impact on companies yet. So the spend justified by future usage is speculative. The firm has launched a fund that invests in data centers and other infrastructure to address the supply constraint.
“we think that the biggest scarcity is on supply, and in the short term, that demand is much bigger than it seems”at 53:35
US growth & recession risk
Where everyone stands on US growth & recession risk
Alex Imas · Google DeepMind
Current data implies AI keeps growth in single digits; double-digit growth needs increasing returns to ideas.
Following the Jones & Jones framework, singularity-style growth requires increasing returns to ideas, meaning each breakthrough makes the next cheaper to find, with no real bottlenecks. Imas says the parameter can be estimated: above one gives double-digit growth, below one gives standard diminishing-returns growth. Using current data, including the meter report, he estimates it is decidedly below one. Models currently need increasingly more resources to squeeze out more intelligence. He also argues that automated sectors shrink in GDP share as their prices fall, so the bottlenecks dominate measured growth. That is why AI could show up less in GDP than in capabilities.
“and you take that data and you estimate this parameter, it is decidedly less than one.”at 12:59
Tyler Cowen · George Mason University
AI might add 3-4% growth, but 10% growth is possible only as a temporary post-recession or war rebound.
Cowen says that high growth historically happens after recessions, wars or pandemics, citing Guyana's 40% growth from a tiny oil base and US growth above 8% in the 1980s coming out of deep recession. If the US had a shock combined with an AI-driven burst giving 3-4% growth, the country could for one year reach 10%. He would bet heavily against it, and the effect would be temporary, with much of it coming from the rebound rather than from AI.
“that combined with a burst of AI driven growth that might have given us three to 4%... could we for a year get to 10%? I'll say yes. I'd bet heavily against it, and the effect would be temporary.”at 13:54
Alex Imas · Google DeepMind
GDP may understate AI-driven welfare gains from free digital goods, so alternative measures are needed.
Imas says GDP has tracked human welfare well historically, but technology creates a wedge because many products are free or nearly free. Research by Brynjolfsson, Collis and others proposes a GDP-B measure that amends GDP with survey-based estimates of the welfare gained from unpriced or cheap goods. He argues that the economy could grow at 3-5%, already high, while being truly flourishing, and that such a world would only show up with these alternative measures.
“you could see a world where you're still kind of like in this forward of, 3% to 5% GDP range... but you have a world that's actually truly flourishing. How do you pick that up?”at 21:12
AI & jobs
Where everyone stands on AI & jobs
Alex Imas · Google DeepMind
A coordination breakthrough could cause a sudden phase transition in wages and employment, leaving large unemployment pools.
Imas expects the real economy to move slowly, but says a fast transition is possible. In his paper, a coordination breakthrough could cause a phase transition in wages and employment. New sectors would still emerge, but the normal rematching of workers to new jobs, which historically took a long time, would not have time to happen. That leaves larger pools of unemployment and recursive loops where people get scarred out of the labour market. He calls this his personal worry.
“if there's a breakthrough in coordination, there is the potential for wages and employment to suddenly... There's a phase transition as far as what will happen to the economy.”at 34:25
Tyler Cowen · George Mason University
Hiring will slow in the AI transition, but an incredible number of new projects and jobs will follow; Cowen is bullish.
Cowen says that employers cannot easily tell whether a candidate will be good at AI, because the old signals, such as a degree from a good school or strong grades, are weaker. That slows hiring and may raise the natural rate of unemployment for a while. He expects this to fade as AI itself helps match people and certifications emerge. He concludes he is pretty bullish, since there will be an incredible number of new projects and jobs, and labour markets have been fine so far.
“some additional slowness in hiring people, but there'll be an incredible number of new projects, new jobs. So I... I'm pretty bullish actually.”at 36:19
Alex Imas · Google DeepMind
AI could create many valuable scarce jobs rather than concentrating all returns in capital; data will tell.
Imas accepts that AI could push economic activity toward capital, as the Piketty-style essay argues. He offers an alternative: value moves toward scarcity, so if AI creates more jobs whose scarcity makes them valuable in a rich society, the returns go to labor. The open question is whether enough such jobs exist to support the whole economy. Historically that has held, but it remains the big unknown, and he argues the economy needs tracking with better data.
“there's an alternative scenario where, like, you know, there's just more jobs created, and these jobs are valuable because of the fact that they're scarce.”at 51:17
AI bubble & tech valuations
Where everyone stands on AI bubble & tech valuations
Tyler Cowen · George Mason University
Market data looks normal and Cowen trusts market judgment over Bay Area AI enthusiasm.
Cowen says that real interest rates, stock prices and other market data still look pretty normal. Unlike 2014, when a blog post could make you rich, AI is now a widely discussed theme that the whole investing world watches. He says he trusts market judgment more than the views of Bay Area people who do not reason carefully about social dynamics. This is a sceptical read of hype-driven valuations, not a claim of a bubble.
“whether it's real interest rates, stock prices, anything, market data still look pretty normal.”at 7:49
AI pricing & unit economics
Where everyone stands on AI pricing & unit economics
Alex Imas · Google DeepMind
Frontier labs face intense price competition as token usage soars while lab revenue stays flat.
Imas cites Ramp data showing exponential increases in token usage alongside flat revenue. He attributes this to price competition at the frontier, which has intensified in the last six months: labs now compete on cheaper models that are about as good, instead of just on who has the best model. Agentic AI uses far more tokens, so enterprise costs rise. He adds that revenue growth ultimately needs productivity gains in the rest of the economy, which he says has not yet appeared in the data, so willingness to pay is not yet proven.
“...you saw exponential increases in token usage, and flat revenue numbers. Where's that coming from? It's because of price competition.”at 9:42
Robotics & autonomous vehicles
Where everyone stands on Robotics & autonomous vehicles
Tyler Cowen · George Mason University
Robotics is progressing but will not add 20% growth within three or four years; adoption will be slow.
Cowen says that getting a Waymo to Washington DC has taken a long time, even though the robot involved is simple: forwards, backwards and turns. Robotics valuations are far above their earlier levels, but they are not popping up as trillion-dollar companies. Market data suggests robotics will happen and is making good progress, but he says it will not produce 20% growth three or four years from now.
“It's making good progress but not gonna give us 20% growth like three or four years from now. No.”at 14:54
Enterprise AI adoption & AI software
Where everyone stands on Enterprise AI adoption & AI software
Tyler Cowen · George Mason University
Institutional inertia slows technology diffusion; much of the economy, including government and healthcare, is a nightmare to change.
Cowen argues that bottlenecks slow even major technological change. The IRS still uses fax machines, and EdTech software, which he uses all the time, is a nightmare, falling short of what it should have been ten years ago. Regulation is part of the problem, but inertia is hard to overcome. Much of the economy, including government, nonprofits, highly regulated healthcare and higher education, could be about 40% of the US economy. These sectors might have reasons to be that way, but they are a strong dose of realism that not everything moves like startups.
“Try using EdTech software, which I do all the time. It's a nightmare.”at 18:10
Wealth, inequality & billionaires
Where everyone stands on Wealth, inequality & billionaires
Tyler Cowen · George Mason University
AI rents are more likely to flow to scarce energy or land than to commoditized capital owners.
Cowen separates capital being very useful from capital owners capturing monopoly rents, which requires entry barriers. He sees no strong case in the Piketty-style essay for the second. If capital is commoditized, rents go to the scarce input. He thinks energy or land are more likely to capture rents than capital, and points to Gulf states or landowners as beneficiaries. He is uncertain, since AI could make energy less scarce by making it cheap to desalinate water or terraform land.
“I'm not sure that's true, but those seem more likely than all the rents go to capital if you had to say.”at 52:37
Pharma & biotech
Where everyone stands on Pharma & biotech
Tyler Cowen · George Mason University
Biotech progress will be especially slow because of regulation and simulation cannot address its core problems.
Cowen is fully on board with the idea that biotech is underrated, but he thinks it will be especially slow to change. The reasons are regulation, and the fact that simulation cannot be used to address the main issues. The simulation-based gains that AI brings to other fields do not carry over well to drug development.
“I think that would be especially slow because of regulation and the inability to use simulation to address the main issues.”at 58:18
Education
Where everyone stands on Education
Alex Imas · Google DeepMind
Market forces will shrink traditional higher ed, but new innovation will follow; he is bullish long term.
Imas expects market forces to shrink the traditional version of higher education. He has no precise timeline because universities are slow-moving institutions. He expects a lot of new innovation to replace them, and says he is very bullish on higher education over the long term.
“I don't have a great timeline for it just because they're such slow moving institutions. ... The market forces are gonna lead to a shrinking of the... kind of our standard idea of higher ed.”at 1:00:49
Across the shows
What other episodes Chorus covered said about the same companies in the 30 days up to this one.
Topics
- AI alignment
- AI alignment & misuse
- AI capability generalization
- AI capability progress
- AI for economic modeling
- AI job creation
- AI job displacement
- AI market pricing
- AI regulation timing
- AI sector outlook
- AI-driven GDP growth
- AI-driven rent distribution
- GDP measurement
- Nvidia backstop capacity
- agent adoption bottlenecks
- agent coordination
- agents' economic share
- biotech pace under regulation
- compute supply scarcity
- data center buildout
- data center permitting and local opposition
- fast takeoff and AI-driven science
- frontier lab valuations
- institutional adoption inertia
- labor share under AI
- political backlash to AI
- revenue projections
- robotaxi adoption
- robotics growth impact
- token pricing competition
- traditional higher ed shrinkage
More from a16z Podcast
- The Next Era of Games: Mods, Agents, and Infinite Worlds · November 12, 2026
- What the Internet Boom Teaches Us About the AI Buildout | Mike Volpi · November 11, 2026
- Inside Salesforce’s $100 Billion M&A Playbook | John Somorjai · November 6, 2026
- The Robots Already Working in American Factories | Tutor Intelligence · November 2, 2026
- The Startups Rebuilding American Industry | The AD50 · October 29, 2026
Ask your AI about this episode
Chorus gives Claude, ChatGPT and other assistants every view from this episode and every other show Chorus follows. Ask who agrees, who changed their mind, and what else was said about the same companies.
Chorus reports what speakers said. It isn't investment advice. How Chorus works · About Chorus