1 October 2026

137 new episodes (129 hours listening) published by the 50 podcasts on the list. 30 highlight clips below (87 mins total).

Investment 1 October 2026 · 30 clips

The Knowledge Project Podcast

Bill AckmanPershing Square Capital Management

Episode

Bill Ackman: People are Going to Lose a Lot of Money

Bill Ackman founded Pershing Square Capital Management and has built its team over 22 years; Pershing Square owns 47% of Howard Hughes, where he is executive chair.

Hosted by

Shane Parrish

29 September 2026 · 1h 17m · 4 clips below

1 of 4 framework

Bill Ackman argues AI has sharply raised the disruption risk that sets a business's value, a risk Warren Buffett failed to see with the internet, and says every investor will look foolish on some holding because of it.

His test for the winners is whether the saving stays with the company: Cognition rewriting banks' COBOL in days should cut the cost of running big financial institutions, but money is a commodity, and pricing power decides whether the margin reaches shareholders or passes to customers.

3m 41s · from 30:25
My observations
- Ackman separates two things AI optimists usually run together: a company getting cheaper to run, and a company keeping the saving. Banks are his case of the first without the second, because every competitor will buy the same software to sell the same commodity.
- Pershing Square asks whether it would be happy owning a business if the stock market shut for 10 years, and on that horizon a threat several years out moves the valuation today.
- Microsoft once shipped a release of consequence every couple of years; Ackman says AI tools now change in days, which cuts how long any judgement about a moat stays current.
2 of 4 framework

Pershing Square sold Netflix within weeks of buying it, when management announced the advertising tier it had just told Bill Ackman it would never launch, and bought back later once that model had worked and the stock had halved.

Ackman's rule is that new information contradicting the thesis forces one of two moves, buy much more or exit, and he exited because the range of outcomes had widened past what a portfolio of high-certainty companies can hold. The proceeds went into Alphabet, on the view that a loss need not be recovered in the stock that caused it, and the tax loss had value of its own.

3m 24s · from 51:06
My observations
- Ackman sold on the width of the outcomes and had no quarrel with the expected value: he thought management would probably work the problem through. That separates a thesis being wrong from a thesis becoming uncertain, and a concentrated high-certainty portfolio has to act on the second.
- Pershing Square put the proceeds into Alphabet and banked the tax loss, declining to win the money back in the name that lost it, where averaging down after a guidance miss is the more common reflex.
- Pershing Square bought back only once both conditions held: the advertising model had proved itself, and the stock had halved again from a multiple Ackman found uninteresting.
3 of 4 framework

Bill Ackman puts Pershing Square's biggest investment mistake down in part to time on the road raising capital, and has moved the firm into permanent-capital vehicles where investors can sell their shares but the money stays.

Open-ended funds, in his account, lose capital both ways: institutions redeem after strong years because the position has grown too large in their portfolio and after weak years because returns are weak, while a fund that closes drops out of the diligence cycle. Charlie Munger's answer on Berkshire, that it was almost never forced to decide by circumstances, is the design Pershing Square has copied.

2m 16s · from 1:05:04
My observations
- Ackman's redemption pattern runs through the allocator's own rebalancing: trimming a manager who has outgrown its weight and cutting one in a bad patch both pull capital when a closed vehicle could be buying, and Pershing Square cites the COVID crash and the financial crisis as the moments permanent capital let it buy.
- Employees own 28% of Pershing Square Holdings and put a little over $500 million into Pershing Square USA, making the people running each vehicle its anchor holders.
- Pershing Square is arguing for the structure it runs. The cost falls on the holders: a listed vehicle gives them liquidity at whatever discount to assets the market sets, a cost the clip leaves out.
4 of 4 contrarian

Pershing Square uses AI for research and not to build its models, and Bill Ackman argues its best trades, credit default swaps before the pandemic and shorting AAA credit before the financial crisis, are ones no model would have suggested.

His reasoning is that a tool every manager has cannot set one apart, and that AI reviews what has already happened. He credits Pershing Square's record since 2010, when many value investors fell behind, to steadily higher standards for business quality and to occasional macro views expressed through instruments whose payoff is large against the capital committed.

1m 59s · from 37:42
My observations
- Ackman's best trades share a shape: a view the market did not hold, expressed through an instrument with a small cost and a large payoff, credit default swaps and a bankrupt company's stock among them. A model trained on what has already happened is built to find the consensus those positions bet against.
- Ackman keeps AI at the research stage on competitive grounds: faster reading for every manager lifts the baseline and moves nobody ahead of it.

Animal Spirits

Will Higher Rates Kill the Stock Market? | Animal Spirits 484

Episode

Will Higher Rates Kill the Stock Market? | Animal Spirits 484

Hosted by

Michael Batnick and Ben Carlson

30 September 2026 · 59m 34s · 2 clips below

1 of 2 current issue

Animal Spirits counts more than 70% of S&P 500 stocks at least 10% below their highs while the index sits within 1% of a record, a combination seen once in 30 years, just before the dot-com bust.

Duality Research finds 21 of the 28 names on the 52-week-low list in utilities, staples and real estate, the sectors most exposed to higher rates. A 22V basket of 81 AI capex names trades at 15 times EV to EBITDA, about 1.14 times its pre-2026 average, and the VIX sits at 16.

3m 33s · from 2:22
My observations
- Staples and utilities are at multi-year lows against the S&P 500 while most stocks fall, so the selling carries no flight to defence; splitting doubt from fear gives a way to read a market that is de-rating without panicking.
- Consumer staples is the only sector whose multiple has expanded this year, per Duality Research, which means the index's gains have come from earnings while valuations compressed in tech and everywhere else.
- One precedent in 30 years is too few to time anything by, and the comparison with 2000 describes how narrow the market has become and stops there.
2 of 2 current issue

Animal Spirits puts the US Aggregate bond index at its worst 10-year real return since the series begins in 1976, with its five-year return negative for the first time.

Nominal returns are about minus 0.3% a year over five years and 1.4% over ten. Inflation and rate rises were both worse in the 1980s, but yields then started far higher; this cycle combined low starting yields, a rapid rise and high inflation, and it leaves yields above 5% as the offset.

1m 15s · from 17:48
My observations
- Bondholders in the 1980s faced worse inflation and faster hikes but held coupons large enough to absorb them; this cycle began from yields too low to cushion anything, which is how the real-return record broke with milder inflation.
- Lehman built the index in the early 1980s and backfilled it to 1976; the record in question covers about 50 years, the first few reconstructed after the fact.
- The hosts present their own bond buying as a yield decision with no rate forecast attached, and disclose buying a long-duration Treasury ETF on the morning of recording.

Monetary Matters

James ElbaorMarlton LLC

Episode

Investors Are Trapped in Private Credit & Private Equity (Here’s the Opportunity)

Founder and portfolio manager of Marlton LLC, which works in closed-end funds, listed private assets, secondaries and asset managers and trades them on events such as mergers.

Hosted by

Max Wiethe

27 September 2026 · 1h 5m · 3 clips below

1 of 3 current issue

James Elbaor says Blackstone's $82 billion BCRED received redemption requests for about 10% of its shares this quarter against a 5% cap, and argues the gate is doing exactly what it was designed to do.

He calls it the classic duration mismatch of a bank: illiquid loans in a wrapper that promises periodic liquidity it cannot always deliver, with the gate written into the fund documents from the start. On his reading the roughly $8 billion requested is a queue of investors wanting out, and says nothing yet about whether the loans are impaired.

2m 14s · from 1:39
My observations
- Two questions usually run together here, whether the structure is failing and whether the assets are, and the BCRED figures answer only the first: requests at twice the cap show the gate binding and say nothing about the marks.
- At 5% a quarter against 10% requested, this cohort alone waits at least two quarters, and longer if fresh requests keep arriving at the same pace; Elbaor puts the fault on distribution, a product sold as if the gate did not exist.
2 of 3 framework

James Elbaor of Marlton argues that public BDCs at 60 to 65 cents on the dollar of NAV are pricing the cost of liquidity, and that the private marks behind them are broadly right.

If listed and gated vehicles hold the same loans at the same stated value, the gap is either the price of leaving now or a markdown yet to be taken, and he takes the first. His firm buys on events: a deep-discount fund absorbed NAV for NAV by a narrower-discount acquirer inherits the acquirer's discount, and he expects such mergers across BDCs within the next year.

4m 22s · from 29:22
My observations
- Elbaor turns one discount into two readable claims: with identical loans and identical stated values, 60 cents on the dollar means either a price of 35 to 40 cents for an immediate exit or a markdown of that size still to come, and the two lead to opposite trades.
- A fund at 38% below NAV that is merged NAV for NAV into one at 25% below takes on the narrower discount, and that is the only bet Marlton makes here: the loans carry the same risk on either side of the deal.
- Marlton is itself buying yield-producing public BDCs, Elbaor says, and does not name them.
3 of 3 explainer

James Elbaor of Marlton says the listing route that gave interval fund investors an exit from 2020 to 2024 has shut: BPRE listed and went straight to a 38% discount to NAV.

In the boom an investor could buy at NAV, see the vehicle list at a premium, margin the shares and recycle the loan into the next one. With that loop closed, managers are winding down, as Blue Owl is doing with OBDC II, or merging, usually NAV for NAV, and he expects five years of mergers, tender offers and consolidation.

3m 13s · from 5:17
My observations
- Buyers came in at NAV expecting a listing premium, then borrowed against the listed shares to buy again, so the listing window was feeding the inflows that the gates are now holding back.
- Blue Owl gave up the 5% quarterly cap on OBDC II and is returning cash as the loans repay, more than 5% in some quarters and less in others, which ends the queue by ending the fund.

Capital Allocators

Nancy ZimmermanBracebridge Capital

Episode

Nancy Zimmerman: Building a $13 Billion Fixed-Income Arbitrage Firm from a Yale Seed

Co-founder and managing partner of Bracebridge Capital, a $13 billion fixed-income arbitrage manager she founded in 1994 with Gabe Sunshine and seed capital from David Swensen at Yale. She learned market making at O'Connor and Associates, traded Treasury options at Goldman Sachs, and sits on Brown University's investment committee.

Hosted by

Ted Seides

30 September 2026 · 49m 53s · 3 clips below

1 of 3 current issue

Nancy Zimmerman says hyperscalers are bringing about $250 billion of investment-grade bonds to a market that has lost the price-insensitive buyers of long duration it relied on for two decades: Japan, then China, then quantitative easing.

The borrowers see a tournament, some playing to win and some to survive, and want as much money as they can get for as long as possible in any currency, through leases, guarantees and bonds such as Google's 100-year sterling issue. Their indifference to basis points leaves price-sensitive holders such as insurers facing a move from 4.25% to 4.61%.

3m 03s · from 34:50
My observations
- Zimmerman frames the AI build-out as a demand for duration from borrowers who do not care about price, landing in a market whose price-insensitive buyers have gone, and that pairing is where relative-value mispricing comes from.
- Data-centre leases, tenant guarantees and bonds in several currencies all trace back to a handful of issuers, so one credit can trade at different prices in different wrappers, the complex she says Bracebridge is working.
- Treasury buybacks and a shift of issuance toward shorter maturities would ease the long-end squeeze, and she treats both as live.
2 of 3 current issue

Nancy Zimmerman recalls an early Brown investment committee conversation in which the answer to needing more return than bonds paid was that bonds yield what they yield, whatever a portfolio needs.

Portfolios now hold far fewer bonds because they are no longer reliably anti-correlated with equities, and with demand to borrow this heavy she doubts anyone is confident they would hedge even a mild downturn. Watching committee members work through that is, in her account, what the seat has taught her.

1m 17s · from 41:33
My observations
- The exchange pulls apart a return target and a market yield, two numbers committees often run together: the first comes from the institution's spending needs, the second is a price, and wanting the first does not move the second.
- Heavy demand to borrow is her reason for doubting the hedge: with that much supply, bond prices may not rise when growth slips, and the stock-bond correlation that justified the old allocation may not come back.
3 of 3 explainer

Nancy Zimmerman says a central bank buying for QE against a fitted yield curve picks off exactly the cheap bonds, which leaves rates arbitrage with nothing to do.

Bracebridge lets the opportunity set decide how much capital it runs, a discipline she credits to her long dialogue with David Swensen, and the firm was once closed for almost a decade. In the opposite regime, rates are inefficient and opportunities compete for capital every day.

1m 38s · from 33:09
My observations
- A central bank that selects bonds trading cheap to a fitted curve does the arbitrageur's job for free and at unlimited size, which is why QE hurt relative-value managers in particular.
- A manager that stayed closed for close to ten years has shown capacity discipline in the one form an allocator can verify: money turned away.

Hidden Forces

Logan WrightRhodium Group

Episode

Why the China Threat Is Overblown | Logan Wright

Director of China macro and financial sector research at Rhodium Group and author of Broken China, the book the episode is built around. He moved to China in 2001 and wrote his dissertation on China's exchange rate reform.

Hosted by

Demetri Kofinas

28 September 2026 · 55m 06s · 2 clips below

1 of 2 contrarian

Logan Wright of Rhodium Group estimates China's economy probably shrank in the second and third quarters of 2026, while Beijing reports real growth of about 5%.

His gap opens in 2022, when new housing starts fell 40% in an industry worth 20% to 25% of output, Komatsu's machines logged 14% fewer hours and retail sales fell, yet the published figure was 3%. Today fixed asset investment is falling 7% in nominal terms and consumption barely rises, and the official series has never registered the property collapse.

4m 13s · from 14:37
My observations
- Komatsu's excavators report their own hours worked back to the manufacturer, a series no Chinese statistics office touches, and it fell 14% in the year the published growth rate was 3%.
- Wright separates two debates that usually get run together: whether China smoothed its data between 2014 and 2019, where he thinks the errors probably ran both ways, and the years since 2022, when every error runs in one direction because property fell out of the economy and nothing replaced it.
- An economy reporting 5% real growth with consumer inflation under 1% since 2018 and falling producer prices has almost no precedent, and Wright has to reach back to the 1890s to find one.
- macro
- geopolitics
2 of 2 explainer

Logan Wright puts China's credit expansion at $24 trillion between 2008 and 2016, against about $6 trillion of GDP, the largest in any single country in at least a century.

Until about 2011 trade surpluses and currency intervention poured new deposits into the banks; once that inflow reversed, banks bid for funding with short-dated wealth management products and trust companies did the marginal lending to developers, margin accounts and local government vehicles. Lending kept growing because everyone assumed nothing could default while new loans rolled the old.

4m 50s · from 38:11
My observations
- State and local banks in China lent the equivalent of Germany's or Japan's economy every year for eight years, on Wright's comparison, with less and less control over where it went, and he calls the idea that all of it was productive impossible to sustain.
- Between 2012 and 2016, on Wright's account, the marginal yuan of funding moved from a deposit to a wealth management product and the marginal loan from a bank to a trust company, so credit kept growing at pace on a base that now had to be rolled over.
- Measured to 2017, the total is $27 trillion; Wright offers both cut-offs and about 80 percentage points added to credit to GDP, and nothing in the argument turns on which one is used.

a16z

Sarah Wang, Alex Immerman, Santiago Rodrigueza16z

Episode

AI, Infrastructure, and the Next Investment Cycle

Partners on a16z's growth team, presenting the firm's annual State of Markets synthesis of trends in technology, AI, infrastructure and markets.

Hosted by

David George

30 September 2026 · 52m 20s · 2 clips below

1 of 2 current issue

a16z's growth team notes that US stocks are up about 20% while market multiples are down about 20%, with the S&P 500 below 20 times earnings.

The rally has come from profits, the reverse of 2000, when the largest companies ran up on valuations near 100 times. Since ChatGPT launched almost four years ago the index has returned 90%, 17% a year, which the partners set against 15% profit growth; memory makers, a cyclical group, sit at six to seven times forward.

1m 30s · from 5:01
My observations
- Splitting a return into earnings growth and re-rating is the first cut on any bubble question, and a16z's figures put this rally on the earnings side, which is exactly where 2000 was not.
- Memory makers at six to seven times forward earnings show the market still pricing parts of the AI supply chain for their cycle, a distinction that tends to disappear in a bubble.
- A forward multiple is only as good as the forward earnings, and a large share of those belongs to the same few companies spending their operating cash flow on the build-out.
2 of 2 current issue

a16z counts live AI deployments at 69% of S&P 500 companies, quantified impact at 30%, and impact tracked over time at only 2%.

Spending is concentrated: on Yipit data the top 1% of spenders on AI vendors run at about eight times the top 10%, and heavy users in AI-native firms spend $7,500 to $9,000 a month against a median of $200 to $400. AI tools cost about 1% of headcount spend in forward-leaning Fortune 500 firms and up to 10% in the most AI-built startups.

3m 55s · from 18:30
My observations
- A pilot count would put most of the S&P 500 in the AI column; asking whether impact is measured, and then tracked over time, drops the share from 69% to 2%, so few large companies can yet show what the spend returns.
- Setting AI spend against headcount cost, about 1% in forward-leaning Fortune 500 firms and up to 10% in AI-built startups, measures diffusion against the cost the tools are meant to displace.
- Most large enterprises the partners speak to still meet AI mainly through Microsoft Copilot, which puts the incumbent's bundle at the front of enterprise adoption.
- ai-adoption
- ai-and-markets

The Credit Edge

Austin CamporinMagnetar Capital

Episode

Magnetar Scours $600 Billion Stressed Debt Pile for Bond, Loan Value

Senior portfolio manager and head of special situations at Magnetar Capital, an alternative asset manager with more than $17 billion under management. He spent 16 years at Elliott Management, latterly as a portfolio manager, after high-yield research and distressed prop trading at J.P. Morgan.

Hosted by

James Crombie and Negisa Balluku

24 September 2026 · 53m 25s · 1 clip below

explainer

Magnetar's Austin Camporin puts B3-rated loans at 20% to 25% of the leveraged loan market against a usual 5% to 10%, with CLOs, which hold about 70% of loans, forced to sell on any downgrade to CCC.

Downgrades are still running ahead of upgrades, so more forced selling is coming, and pricing already shows it: CCC trades at about four times the single-B spread against a 15-year average near three. Single-B as a whole trades in line with history, which is why the index looks calm while its bottom notch does not.

Audio streamed from the publisher.

2m 11s · from 7:20
My observations
- CLO rating limits turn a one-notch downgrade into a sale at whatever price clears, and with CLOs owning most of the market there are few natural buyers on the other side; the discount comes from the holder's rules, not the borrower's prospects.
- Averaging a normal single-B middle with a bottom notch now two to five times its usual share is how an index spread can sit near its tights while stress builds underneath it.

Monetary Matters

Ben PouladianBEP Research

Episode

Meta's Muse Is the McDonald's of AI (& That's Bullish) | Ben Pouladian

Publishes BEP Research, institutional research on the AI supply chain, after starting and running an LED lighting company; he has been an Nvidia investor since 2016.

Hosted by

Jack Farley

30 September 2026 · 1h 14m · 1 clip below

contrarian

Ben Pouladian argues private credit will lend against Nvidia GPUs and not custom chips like Google's TPU or Amazon's Trainium, because a vacated GPU cluster finds its next tenant.

His analogy is prime retail frontage: when a lab hands back a GPU cluster, a queue of renters fills it, so lenders such as Blackstone and Blue Owl can underwrite the collateral. A chip built for one owner's workloads has no open-source community improving it and no second user waiting, so its street, in his image, is a less certain let.

2m 09s · from 20:38
My observations
- Lenders to data centres underwrite the next tenant as much as the current one, and a GPU any lab can run supplies one; a chip designed for a single owner's workloads leaves the collateral with a buyer list of one.
- The market can test the claim directly: whether TPU- or Trainium-backed facilities get financed at all, and at what spread over GPU-backed debt, would show whether lenders price the difference he describes.

Invest Like the Best

Noah ShinnInstinct

Episode

Inside the Personal AI Assistant Growing 10% a Day | Instinct Founder

Noah Shinn is chief executive of Instinct, an AI personal assistant he started about a year before this conversation. He puts transaction volume through the invite-only product at over $1 billion a year, half of it travel.

Hosted by

Patrick O'Shaughnessy

28 September 2026 · 1h 25m · 1 clip below

framework

Noah Shinn splits a consumer platform's revenue into the share earned from attention on the app and the share from the service underneath, and expects agents to shrink the first and grow the second.

Shinn uses Uber Eats and DoorDash for the service side, arguing, without their numbers, that every click removed from checkout lifts order volume, and an agent that orders the car or the meal unprompted removes the friction altogether. Businesses earning nearly all their revenue from time spent in the app, social media above all, are the exposed ones.

4m 25s · from 38:48
My observations
- A platform earning 70% of revenue from attention does not simply become a 30% business in his account: an agent that books the ride before it is asked for raises trips per user, so the 30% earned on the service grows.
- Amazon's upselling, food delivery and the online travel agencies sit toward the attention end of his line, which turns the exercise into a ranking of incumbents by how much an agent would take from them.
- Instinct's business model, a take rate on the transactions it routes, depends on the service-side businesses this argument says will grow, and Shinn pitches partnerships to them in the same answer.
- ai-and-markets
- equities

In Good Company (Norges Bank)

Kai YuHorizon Robotics

Episode

Horizon Robotics CEO: Car Chips, Mind-Off Driving and China's Speed | Podcast | In Good Company

Founder and chief executive of Horizon Robotics, whose chips and software run in about one in three intelligent cars on Chinese roads, with BYD, Volkswagen and Toyota among its customers. He has worked in machine learning for more than 30 years and founded the company 11 years ago.

Hosted by

Nicolai Tangen

30 September 2026 · 47m 05s · 1 clip below

framework

Horizon Robotics' Kai Yu says China's automotive compute market has consolidated to three suppliers, Nvidia, Huawei and Horizon, in a car market that launched more than 580 new models in six months.

Yu's explanation is qualification time: an automotive chip takes three to four years to be ready, then a platform cycle of about two years at Chinese carmakers and five at global ones, so a supplier waits five to eight years for returns. He calls it a terrible business that turns out to be a great one, because venture-funded entrants cannot wait that long.

2m 12s · from 12:01
My observations
- Yu's timeline explains how the most crowded car market in the world ended up with three suppliers of the computer inside the cars: the qualification cycle filters entrants years before customers see them, and each design win locks in a platform for its life.
- Yu is describing his own moat, and Horizon's place in the top three is his claim; the structural argument rests on the cycle times, which other automotive chipmakers' disclosures can confirm or contradict.
- equities
- ai-and-markets

Macro Musings

Erik ThedéenSveriges Riksbank

Episode

Erik Thedéen on QE, Central Bank Independence, and Digital Payments | Macro Musings

Governor of Sweden's Riksbank for three and a half years, after seven years running the Swedish financial supervisory authority and earlier spells at the debt office, a hedge fund and the Stockholm Stock Exchange. His first job, in 1989, was on the Riksbank's own dealing desk.

Hosted by

David Beckworth

28 September 2026 · 55m 33s · 1 clip below

contrarian

Riksbank governor Erik Thedéen questions whether round-the-clock instant payments are worth having, noting that many financial crises have been solved over a weekend while everything was closed.

Sweden, where cash has fallen to 5% of in-store purchases, now requires grocery stores and pharmacies to accept it and asks adults to keep 1,000 kronor at home for a crisis. Thedéen's other objection is anti-money-laundering checks, which are hard to run on payments that settle in seconds at any hour.

3m 11s · from 44:00
My observations
- Weekend closure works as a resolution tool: authorities have used the gap between Friday's close and Monday's open to arrange rescues, and Thedéen asks whether a 24/7 system would need a legal power to stop the banks instead.
- Swish, the mobile instant payment app, spread from paying friends to e-commerce and tipped the whole system, yet Thedéen says little has changed in the five years since: no krona stablecoin, and no tokenisation to speak of.
- macro
- risk-management

The Cognitive Revolution

Lewis Hammond, Max Nadeau, Wayne Nelms, Nick Gillian, Andrei GeorgescuCooperative AI Foundation (Hammond); Coefficient Giving (Nadeau); Ornn (Nelms); Archetype AI (Gillian); Vivodyne (Georgescu)

Episode

AI:AM: What If It Works Too Well? Colluding Agents, $200M Safety Orgs, Virtual Cells Saturate at 2%

Wayne Nelms is co-founder and chief technology officer of Ornn, which publishes a GPU compute price index built from more than 1,000 cleared rental transactions a day per index, and on which ICE has announced plans to list futures, pending regulatory approval. The other guests work on AI safety research and funding, sensor foundation models and robotic tissue testing, at the Cooperative AI Foundation, Coefficient Giving, Archetype AI and Vivodyne.

Hosted by

Nathan Labenz and Prakash Narayanan

27 September 2026 · 1h 46m · 1 clip below

explainer

Prakash Narayanan, relaying what he has heard, says the biggest compute deals are priced by whose credit the lender relies on, and that none of them reach Ornn's GPU price index.

A seller that built clusters on its own balance sheet, as he says Elon Musk did, can rent month to month and charge a premium. One that needs a five-year contract to borrow ends up on build cost plus a margin of around 20%, the arrangement he attributes to CoreWeave.

3m 16s · from 1:06:00
My observations
- Narayanan traces the price back to repayment: a bank lending against a build-to-suit data centre relies on the tenant's contract for its money, so the tenant holds the credit and sets the terms, and the landlord's margin looks more like a construction and operating fee.
- Only sellers willing to trade capacity available now feed Ornn's index, so on this account it records the cheapest and most liquid end of the market, and the bilateral deals carrying most of the financing never print on it.
- About $50 million a megawatt to Musk, against the $80 to $90 million a megawatt he says the buyer earns from the capacity, is hearsay on his own telling; a disclosed lease would settle whether the spread is that wide.
- leverage-and-financing
- ai-and-markets

Top Traders Unplugged

Nick Baltas

Episode

AI Can Build Your Portfolio—But Can You Trust It? | Systematic Investor | Ep.419

Designs systematic strategies and QIS index products, on his own description in the episode.

Hosted by

Alan Dunne

29 September 2026 · 1h 5m · 1 clip below

explainer

Nick Baltas argues an agentic portfolio cannot be backtested honestly, because a model trained today carries knowledge of the period it is asked to forecast even when its inputs are cut off at the date.

Agents that vote after seeing backtests will drift toward the best-performing scheme, putting the old data-mining bias inside the process, and because the models are not deterministic, the same prompt run twice produces two different histories. Next year's model would rewrite this year's record, and agents drawn from one provider add a monoculture risk the paper itself names.

3m 45s · from 50:23
My observations
- Baltas ranks non-determinism above look-ahead bias: two runs of the same prompt give two backtests, so a track record for the method cannot exist, only a record for one sequence of outputs.
- Restricting agents to models that existed at each date is the fix some researchers have tried, and the discussion in the clip notes those earlier models may be too weak to run this kind of analysis at all.

AI 1 October 2026 · 30 clips

a16z

Diogo AlmeidaTypeSafe AI

Episode

How Jev Turns AI Into Software That Gets Things Done

Diogo Almeida founded TypeSafe AI, maker of the Jev model, after Google Brain and OpenAI, where he worked on RLHF from late 2021 and helped release the resulting model.

Hosted by

Ben Horowitz and Martin Casado

28 September 2026 · 42m 24s · 2 clips below

1 of 2 contrarian

Diogo Almeida, formerly of OpenAI, argues that since RLHF the labs have tuned models to impress the human who grades them, and automating rote work has been left behind.

His evidence is that the intelligence needed for simple, outsourceable instructions has sat in the models for years, yet benchmarks like GPQA fall while a drive-through order still cannot be handed over. He reads OpenAI's own AGI definition, most economically valuable work, as achievable, and says the industry split toward overpromising instead.

1m 49s · from 19:58
My observations
- Almeida separates a model that pleases the person grading its answer from a model trusted to run with nobody reading the output, and argues post-training has paid for the first. Almeida uses that split to explain how GPQA can fall while a drive-through order still cannot be automated.
- If the diagnosis holds, the constraint on back-office automation is engineering for reliability, and a bigger frontier model does little to move it.
2 of 2 contrarian

The a16z panel, citing studies done at Google, puts the average pull request at a large company at about 10 lines, so a coding agent that writes them faster is optimising a small part of the work.

Those lines usually encode something learned from a customer, and writing them sooner adds no capability the software lacked. The panel goes further: with less human oversight, software written quickly may be getting worse, and new capability has to come from a new primitive inside the program, such as Jev.

1m 21s · from 33:38
My observations
- Ten lines per change puts the coding-agent productivity story in proportion: quicker small edits mean quicker iteration on what a product already does, and no change to what it can do.
- Earlier in the episode Almeida rates the agents strong on syntax and poor at architecture, perhaps 50th percentile, and the 10-line figure completes the picture: they are fastest at the part of the work that was already small.
- The panel asserts that less oversight makes software worse without putting a figure on it, and that is the claim that would matter most to anyone relying on agent-written code.
- ai-agents
- ai-economics

Big Technology Podcast

Christian KleinSAP

Episode

SAP CEO: AI Won't Kill Software, But It Will Change Your Job — With Christian Klein

Chief executive of SAP, Europe's largest software company, with about 100,000 staff and 80,000 of them using its Joule assistant.

Hosted by

Alex Kantrowitz

30 September 2026 · 58m 59s · 2 clips below

1 of 2 current issue

SAP now switches models agent by agent to get the best outcome per token, and Christian Klein says many of its agents are already on their fifth model.

Coding still runs on frontier models, but routine agents such as headcount and financial reports run on open models that do the job at a fraction of the cost. Klein's test is the P&L: a 20% productivity gain is no gain if the token bill rises 30% alongside it.

3m 19s · from 40:08
My observations
- Ramp's card data, which the host cites, has frontier models falling from 53% to 45% of AI spend in about a month; Klein describes the same shift from inside SAP as a release process, with every agent tested against several models before it ships and retested whenever a new one arrives.
- Coding on frontier models and reporting on open ones splits AI spend along the cost of an error, and that line will move each time a cheaper model clears the accuracy an agent needs.
- SAP sells the switching as a feature of its own platform, and the claim that customers prize model neutrality comes from the vendor that charges for providing it.
2 of 2 current issue

SAP brought in token budgets by job profile three to four months ago, exempted the top 1% of users who train its agents, and found model switching cut cost by up to ten times, more than the limits did.

Klein held off because a cap contradicted the company's own instruction to use AI to reskill; the design that followed sets thresholds most staff never reach and lets a manager approve more for a given month or project. Ramp's figures, cited by the host, show the heaviest users' monthly spend falling 9.7% in a month.

3m 28s · from 47:22
My observations
- Exempting the heaviest users follows the revenue: they are the ones improving agent accuracy, which SAP's product depends on, and the caps fall on developers, designers, finance and HR, where output per token is harder to see.
- Klein ranks switching as the bigger lever, a factor of ten against whatever a cap saves, which puts most of the saving in routing work to cheaper models and leaves rationing as the smaller tool.
- A manager lifting the cap for a month or a project turns token spend into a budget line run the way travel or software licences already are.
- ai-economics
- ai-agents

20VC

Alex MashrabovHiggsfield

Episode

The Untold Story of Higgsfield | Burning $4M a Month on AI Models | CEO, Alex Mashrabov

Founder and chief executive of Higgsfield, an AI video company he says went from $1 million to $1 billion in annualised revenue in 18 months, with close to 400 staff. He previously sold AI Factory to Snap for $166 million.

Hosted by

Harry Stebbings

28 September 2026 · 1h 15m · 2 clips below

1 of 2 current issue

Alex Mashrabov puts Higgsfield's spend on AI models for its own close to 400 staff at over $4 million a month, more than $10,000 a head, and expects its best engineers and creatives to reach $50,000 to $100,000 a month each.

He expects those people to ask for pay rises to match. Headcount has not fallen either: Higgsfield assumed legal and customer support would mostly be replaced, and he now counts not building those teams sooner as a mistake, with AI handling over 60% of first-line requests and failing in B2B support.

4m 53s · from 33:24
My observations
- Inference arrives on top of payroll at Higgsfield, an AI company staffed by heavy AI users: $10,000 a head a month now, heading toward $100,000 for the strongest staff, with their pay expected to rise alongside, and no cut in legal or support staff to offset it.
- In Mashrabov's account, B2B support defeats the agents because an agent is only as good as its context and rules, and Higgsfield launches products about weekly and changes those rules about twice a week.
- Some of the $4 million is experiment: Mashrabov describes a creative spending over $30,000 in one week vibe-coding an asset-organisation tool that did not reach production, and counts it as net positive for what the team learned. A per-head average folds weeks like that in with steady production use.
2 of 2 explainer

Higgsfield keeps over 80% margin on jobs run on its own or post-trained open-weight models and 20% to 30% on closed ones, and Alex Mashrabov says it chooses the model in over 40% of jobs.

His case for open weights in this market is that companies printing hundreds of ad creatives a week want the cheapest stable model, since a viral social video does not need frontier reasoning. He still expects OpenAI and Anthropic to hold over half the market by dollars, because coders move to each new release.

3m 48s · from 29:12
My observations
- Higgsfield sets much of its own margin by routing: each job it moves off a closed model onto a post-trained open one goes from 20% to 30% margin to over 80%, and it makes that call in over 40% of jobs. Two application companies with the same revenue can sit a factor of three apart on margin for that reason alone.
- Mashrabov cites OpenRouter's open-model share rising from below 30% to over 60% this year and, in the same answer, gives OpenAI and Anthropic more than half the dollars. One figure is usage and the other is spend, and he accounts for the gap: coding pays for the newest model, and ad creative buys the cheapest one that holds up.
- When a company says it builds its own model, Mashrabov reads it as post-training open weights, as Cursor did, and names the valuable kind: reinforcement learning on a customer's sequence of decisions, teaching the model to compress ten steps into one.
- ai-economics
- open-models

ChinaTalk

Julian Gewirtz

Episode

Julian Gewirtz on Trump-Xi Summit, AI, and Mao Dead Fifty Years Ago!

Author and former China director on President Biden's National Security Council staff, who has spent much of his academic career on China's turning points: the death of Mao, Tiananmen and WTO accession.

Hosted by

Jordan Schneider

22 September 2026 · 1h 5m · 1 clip below

current issue

Julian Gewirtz says Beijing has not decided how to treat Chinese open-weight models spreading worldwide, which carry both security risk and strategic benefit to China.

Xi's upbeat speech on AI to a global audience and the state security minister's essay on regime risk point in different directions, and commentators have treated each as the answer. Gewirtz's guess is that neither is, because the trade-offs are large and the technology is moving faster than the policy.

1m 06s · from 35:13
My observations
- Open weights are how Chinese models travel, so the decision Gewirtz is waiting for would land on users well beyond China's borders, and it is the one that would settle which speech described policy.
- Two speeches pointing different ways have each produced a cycle of confident interpretation. Gewirtz declines to read either as policy, and says so.

Worth listening to in full 1 October 2026

Most clips above stand alone. These are the episodes that justify the whole hour.

How Jev Turns AI Into Software That Gets Things Done

a16z · 42m 24s

Diogo Almeida, who worked at Google Brain and OpenAI before founding TypeSafe AI, argues that the automation companies have been waiting for needs a component inside the program that turns stated intent into a decision with a confidence level, and he separates reliability into uptime, determinism and consistent judgement on every call. Ben Horowitz and Martin Casado press him on whether the real world's long tail is only a data problem, and on why SaaS companies whose valuations fell on coding agents may be the best placed to use it.

Bill Ackman: People are Going to Lose a Lot of Money

The Knowledge Project Podcast · 1h 17m

Bill Ackman spends the core of the conversation on how Pershing Square invests: which companies keep the savings AI creates, why a tool every manager has cannot set one manager apart, why he sold Netflix within weeks of buying it and bought back once the advertising model had worked, and why time spent raising capital led the firm into permanent-capital vehicles. The first third is personal and the Howard Hughes segment is his own vehicle, so the argument sits in the middle of the hour.

1 October 2026

Investment 1 October 2026 · 30 clips

Bill Ackman argues AI has sharply raised the disruption risk that sets a business's value, a risk Warren Buffett failed to see with the internet, and says every investor will look foolish on some holding because of it.

Pershing Square sold Netflix within weeks of buying it, when management announced the advertising tier it had just told Bill Ackman it would never launch, and bought back later once that model had worked and the stock had halved.

Bill Ackman puts Pershing Square's biggest investment mistake down in part to time on the road raising capital, and has moved the firm into permanent-capital vehicles where investors can sell their shares but the money stays.

Pershing Square uses AI for research and not to build its models, and Bill Ackman argues its best trades, credit default swaps before the pandemic and shorting AAA credit before the financial crisis, are ones no model would have suggested.

Animal Spirits counts more than 70% of S&P 500 stocks at least 10% below their highs while the index sits within 1% of a record, a combination seen once in 30 years, just before the dot-com bust.

Animal Spirits puts the US Aggregate bond index at its worst 10-year real return since the series begins in 1976, with its five-year return negative for the first time.

James Elbaor says Blackstone's $82 billion BCRED received redemption requests for about 10% of its shares this quarter against a 5% cap, and argues the gate is doing exactly what it was designed to do.

James Elbaor of Marlton argues that public BDCs at 60 to 65 cents on the dollar of NAV are pricing the cost of liquidity, and that the private marks behind them are broadly right.

James Elbaor of Marlton says the listing route that gave interval fund investors an exit from 2020 to 2024 has shut: BPRE listed and went straight to a 38% discount to NAV.

Nancy Zimmerman says hyperscalers are bringing about $250 billion of investment-grade bonds to a market that has lost the price-insensitive buyers of long duration it relied on for two decades: Japan, then China, then quantitative easing.

Nancy Zimmerman recalls an early Brown investment committee conversation in which the answer to needing more return than bonds paid was that bonds yield what they yield, whatever a portfolio needs.

Nancy Zimmerman says a central bank buying for QE against a fitted yield curve picks off exactly the cheap bonds, which leaves rates arbitrage with nothing to do.

Logan Wright of Rhodium Group estimates China's economy probably shrank in the second and third quarters of 2026, while Beijing reports real growth of about 5%.

Logan Wright puts China's credit expansion at $24 trillion between 2008 and 2016, against about $6 trillion of GDP, the largest in any single country in at least a century.

a16z's growth team notes that US stocks are up about 20% while market multiples are down about 20%, with the S&P 500 below 20 times earnings.

a16z counts live AI deployments at 69% of S&P 500 companies, quantified impact at 30%, and impact tracked over time at only 2%.

Magnetar's Austin Camporin puts B3-rated loans at 20% to 25% of the leveraged loan market against a usual 5% to 10%, with CLOs, which hold about 70% of loans, forced to sell on any downgrade to CCC.

Ben Pouladian argues private credit will lend against Nvidia GPUs and not custom chips like Google's TPU or Amazon's Trainium, because a vacated GPU cluster finds its next tenant.

Noah Shinn splits a consumer platform's revenue into the share earned from attention on the app and the share from the service underneath, and expects agents to shrink the first and grow the second.

Horizon Robotics' Kai Yu says China's automotive compute market has consolidated to three suppliers, Nvidia, Huawei and Horizon, in a car market that launched more than 580 new models in six months.

Riksbank governor Erik Thedéen questions whether round-the-clock instant payments are worth having, noting that many financial crises have been solved over a weekend while everything was closed.

Prakash Narayanan, relaying what he has heard, says the biggest compute deals are priced by whose credit the lender relies on, and that none of them reach Ornn's GPU price index.

Nick Baltas argues an agentic portfolio cannot be backtested honestly, because a model trained today carries knowledge of the period it is asked to forecast even when its inputs are cut off at the date.

AI 1 October 2026 · 30 clips

Diogo Almeida, formerly of OpenAI, argues that since RLHF the labs have tuned models to impress the human who grades them, and automating rote work has been left behind.

The a16z panel, citing studies done at Google, puts the average pull request at a large company at about 10 lines, so a coding agent that writes them faster is optimising a small part of the work.

SAP now switches models agent by agent to get the best outcome per token, and Christian Klein says many of its agents are already on their fifth model.

SAP brought in token budgets by job profile three to four months ago, exempted the top 1% of users who train its agents, and found model switching cut cost by up to ten times, more than the limits did.

Alex Mashrabov puts Higgsfield's spend on AI models for its own close to 400 staff at over $4 million a month, more than $10,000 a head, and expects its best engineers and creatives to reach $50,000 to $100,000 a month each.

Higgsfield keeps over 80% margin on jobs run on its own or post-trained open-weight models and 20% to 30% on closed ones, and Alex Mashrabov says it chooses the model in over 40% of jobs.

Julian Gewirtz says Beijing has not decided how to treat Chinese open-weight models spreading worldwide, which carry both security risk and strategic benefit to China.

Worth listening to in full 1 October 2026

How Jev Turns AI Into Software That Gets Things Done

Bill Ackman: People are Going to Lose a Lot of Money