Latest edition
1 November 2026
133 new episodes (123 hours listening) published by the 50 podcasts on the list. 50 highlight clips below (144 mins total).
Investment 1 November 2026 · 50 clips
Masters in Business (Ritholtz)
Rich SteinmeierLPL Financial
Chief executive of LPL Financial since 2024 and with the firm since 2018, after running Merrill Lynch's remote advice channel and serving as chief digital officer at UBS. LPL supports 32,000 independent advisers with $2.6 trillion in client assets.
-
current issue
LPL Financial's research team has moved its model portfolio from 60/40 toward something closer to 40/30/30, with 30% in alternatives and other non-correlated assets, for a platform of 32,000 advisers.
Steinmeier says getting there took five years of building a product shelf that runs from ETFs to locked-up funds, plus a dedicated support desk, and the firm now speaks out in Washington for legislation to open 401(k) plans to private assets. His reason is that companies stay private longer; he concedes nobody has settled the account size below which lock-ups stop making sense.
Audio streamed from the publisher.
The Credit Edge
Ross PamphilonImpax Asset Management
Chief investment officer for fixed income at Impax Asset Management, which he joined in 2023 after Wells Fargo and Merrill Lynch; he co-founded European Credit Management. Impax was founded almost 30 years ago and manages around $30 billion.
-
1 of 4 current issue
Ross Pamphilon of Impax suggests Oracle may be a weak link in AI credit: its roughly $120 billion of BBB-minus debt would make up something like 6% to 8% of the high-yield market if cut to junk.
On his account about half of Nvidia's revenue comes from the hyperscalers, whose $2.3 trillion of contracted backlog is half OpenAI and Anthropic, and neither of those has much of a moat while token prices fall and open-weight models sell cheaper. Further down, CoreWeave's senior unsecured bonds yield 15.5% and risk being crammed down by delayed-draw term loans.
Audio streamed from the publisher.
-
2 of 4 contrarian
Impax estimates about 12% of the US investment-grade index is now AI or AI-related, with the hyperscalers alone around 4%.
Pamphilon expects that share to rise, so a manager tied to the benchmark ends up owning more tech, as one tied to the aggregate index owns more Treasuries, which he put earlier in the episode at approaching 50% against perhaps 20% to 30% ten to fifteen years ago. His alternative is an unconstrained mandate with a reference index the client can use to model risk, and wide boundaries around it.
Audio streamed from the publisher.
-
3 of 4 framework
Ross Pamphilon of Impax counts close to $700 billion of AI-related issuance across global credit markets, about $230 billion of it from the hyperscalers, whose borrowing he expects to double next year.
It is arriving as project finance that can carry construction and permitting risk, as asset-backed deals and as straight bonds, and he sees a risk of saturation. For creditors he ranks the outcomes: flat or modestly rising capex with more equity funding is best, accelerating capex funded by expensive debt is worse, and a material capex cut is arguably worse still, because capex is underpinning US growth.
Audio streamed from the publisher.
-
4 of 4 prediction
Ross Pamphilon of Impax thinks US high-yield spreads may peak around 450 basis points within six months, near their Liberation Day levels, in a shallow default cycle.
He sees a financing-led cycle with no earnings recession: sponsor loans from the 2020 and 2021 buyouts, heavy in US software, fall due over the next 12 to 18 months, and liability management exercises and drop-downs are spreading. He expects spreads to stop short of past peaks because 55% to 60% of US high yield is now BB, though a Fitch measure of private credit defaults has reached 6.3%.
Audio streamed from the publisher.
The Compound and Friends
Alex AltmannBarclays
Managing director and global head of equities tactical strategies at Barclays, leading an 11-person team that turns trading-desk ideas into research for clients. Previously a macro portfolio manager at Millennium and a decade running equity trading strategy at Citi, after starting as a single-stock analyst at JP Morgan in 2004.
-
1 of 3 current issue
Barclays estimates 401(k) contributions put $800 to $900 billion a year into US stocks regardless of price.
Altmann ties the flow to a household balance sheet with 35% of net worth in equities, the highest on record and the widest margin over home equity in data back to the 1960s, as unaffordable mortgages send younger savers' deposits into shares. The hosts call the result a relentless bid, add the move of brokers into fee-based ETF allocation, and see it in how quickly volatility spikes fade.
-
2 of 3 explainer
Barclays traces this year's calm index and violent single-stock moves to multi-manager hedge funds, whose assets it puts at about $1 trillion, roughly triple the 2019 level.
On Altmann's account, with cash yielding about 5%, a high-gross, low-net pod earns interest on the capital it does not post as collateral, so the model boomed after 2022. Every pod neutralises the same factors with the same risk models in a shrinking pool of about 3,000 tradable stocks, which pushes crowding into exposures managers may not see and makes each unwind hit harder.
-
3 of 3 prediction
Alex Altmann of Barclays puts the AI revenue needed by the end of 2028 to pay for committed capex at $1 trillion a year, against a run rate he expects to be near $400 billion.
Altmann sees spending at its terminal point, moving by perhaps $100 billion either way, so the shortfall has to close from sales, in an S&P 500 that books about $17 trillion of them and $5.5 trillion of profit.
Excess Returns
Kai WuSparkline Capital
Kai Wu is a quant at Sparkline Capital who builds portfolios on intangible value and has written more than two dozen research papers on AI, trade, innovation and human capital. The platform he demonstrates covers more than 10,000 mutual funds and ETFs.
-
1 of 3 current issue
Kai Wu counts half the S&P 500, by index weight, as tied directly to AI infrastructure.
Wu builds the figure from holdings: technology is about 35% of the index, chip makers alone 17% and the Magnificent Seven 29%, with data-centre firms such as Oracle added and Nvidia counted once. He reads July's unwind of the Situational Awareness hedge fund and October's chip sell-off as the first signs that holders of a 500-stock index have noticed they own one trade.
-
2 of 3 current issue
Kai Wu finds emerging market equities, long held as a diversifier away from developed markets, now as exposed to AI infrastructure as the S&P 500, with TSMC, Samsung and SK Hynix as the index's top holdings.
Measured on returns in excess of the S&P 500, Wu puts the rolling one-year correlation between emerging markets and AI funds near zero on average and, over the past year or two, higher than its 2021 peak of about 0.5. The usual fixes, value, developed ex-US, equal weight and small caps, cut the infrastructure exposure but lift the share in AI laggards from about 24% to between 42% and 82%.
-
3 of 3 current issue
Kai Wu puts Microsoft's and Meta's capital spending, as a share of sales, above that of many utilities today.
Wu cites two findings from decades of data: companies growing capex aggressively tend to underperform, the investment factor in the Fama-French model, and within the same sector the more capital-intensive firms have thinner moats and lower returns on invested capital. On that evidence the hyperscalers are, for better or worse, turning themselves into utilities.
All-In
Anthropic S-1 Shocker, Greatest Business Ever? Google’s $500B Wipeout, Midterm Predictions
No guest this week; the four hosts take Anthropic's S-1 filing, Alphabet's third-quarter results and the US midterms.
-
1 of 3 framework
Chamath Palihapitiya argues that a handful of AI companies now need to borrow on a scale equal to 30% to 40% of US Treasury issuance, so either they or the Treasury will have to pay more.
If buyers prefer government paper, the hyperscalers and labs pay a higher coupon, the neoclouds below them pay more again and the risk of default comes forward; if the AI borrowers clear at reasonable rates, the Treasury pays up instead, and so does every borrower priced off two- to 10-year yields. He calls it a market risk he has not seen before.
-
2 of 3 framework
Chamath Palihapitiya's team estimates that about 1,900 of Anthropic's customers spend more than $1 million, and that once AI-native start-ups are stripped out probably only a couple of hundred large companies have made the return case.
His explanation is sign-off authority: a director or vice president can approve several experiments within a delegated budget, while a multi-million contract needs a cross-functional case built on cost of ownership. On that reading, sales in the hundreds of billions over three to seven years need 25% to 30% of the possible base plus existing buyers expanding.
-
3 of 3 current issue
David Sacks reads Anthropic's S-1 as showing that its widely reported $80 billion loss was a non-cash charge on convertible notes issued to infrastructure partners such as Amazon.
On his reading the notes were revalued upward as the company's valuation rose, and issuing equity instead would have produced no charge. He puts cumulative spending at about $9 to $10 billion against roughly $110 billion raised, gross margin near 85%, and the cost of a given task falling about tenfold a year while prices fall by around half, which widens margin as tokens get cheaper.
The Real Eisman Playbook
Anthropic’s Growth is Slowing: Is AI Capex Next? | The Weekly Wrap
-
1 of 3 framework
Steve Eisman likens CoreWeave to a specialty finance lender, borrowing in the bond market to build data centres and rent out compute, so it could lose if that market does not regain confidence.
Its 9% coupon now trades at a yield near 15%, which is what new sites would cost to fund, and total debt of $36 billion passes $50 billion once leases and other obligations are added. He has seen such firms fail when fixed income confidence went, and the hyperscalers it competes with pay for the same capacity from far larger balance sheets.
-
2 of 3 current issue
Steve Eisman relays an electrical contractor's view from a data centre site that grid connection, with new transmission lines 24 months or more away, will be what slows AI construction from 2027.
The contractor says builders pay his crews $150 a day on top of a $65 hourly wage just to turn up, that substations take at least one to two years, and that 100 transformers can sit in a yard with no power poles near the site; he calls the result stranded capex. He expects Amazon and Meta to feel it more than Microsoft and Google, with CoreWeave and Oracle further back.
-
3 of 3 current issue
Steve Eisman reads Anthropic's numbers as the end of enterprises buying AI tokens regardless of cost: its sequential revenue growth fell from 147% in the second quarter to 45% in the third.
His efficiency ratio, revenue divided by compute expense, rose from 0.69 times to 2.08 times through the boom and slipped to 1.70 in the third quarter, while the S-1 lists about $800 billion of compute and equipment commitments over the next several years, which on his reading could push compute costs sharply higher from 2027. He doubts October's token price cuts of up to 50% lift volume enough to offset the lower prices.
20VC
David TischBoxGroup
Founder of BoxGroup, which invests from a $275 million seed fund and a $275 million opportunity fund; the seed fund will back around 150 companies. An early-stage investor for 17 years, with early positions in Cursor, Baseten and Clay.
-
1 of 4 contrarian
BoxGroup's David Tisch estimates concentrated seed funds win about 20% of the deals they want, not the 90% they claim, so they offer terms to 120 to 150 companies to land 30.
Tisch calls it adverse selection: a fund that needs 10% ownership ends up holding the companies that accepted its terms, which are not necessarily the 30 it most wanted. BoxGroup, with no ownership requirement, claims to get into more of the companies it is excited about.
-
2 of 4 framework
BoxGroup's investment committee does not vote: a deal needs one partner who wants it, and partners who have not met the founder may not argue against it.
Tisch argues that any early-stage company can be shown to fail, and on his rough figure 70% to 90% do, so a memo of risks is obvious and irrelevant; the job is to see why one will work. Ethical or reputational issues with a founder remain open to anyone.
-
3 of 4 framework
BoxGroup's David Tisch argues $10 to $20 million seed rounds still work for venture returns because the fastest-growing companies now give up 3% to 8% in later rounds, against 12% to 20% in the past.
Tisch adds that outcomes have grown: a $10 billion company, far-fetched ten years ago, is now an outcome an early-stage investor has to squint towards as normal. His answer to a seed at $200 million pre-money is to balance entry prices across the whole portfolio and still say yes to the individual deal.
-
4 of 4 framework
BoxGroup's David Tisch says the firm generally does not sell early in secondary markets, because taking a company from $2 billion to $8 billion is easier than taking one from zero to $1 billion.
On his account, selling early risks the relationship with the founder, and he does not accept that founders taking secondary licenses investors to do the same. Every late turn of the multiple, he argues, is a big win for the fund's LPs.
Alt Goes Mainstream
Chris Sacca and Caie KelleyLowercarbon
Chris Sacca is the largest investor in Lowercarbon, which manages more than $3 billion; at his earlier firm, Lowercase, five people ran $6 billion at the peak. Caie Kelley is the first general partner Lowercarbon promoted from within and runs its growth fund, after investment banking at JP Morgan and a $250 million fund at Sound split across OpenAI, Anthropic and Perplexity.
-
1 of 2 framework
Chris Sacca says Lowercarbon will not back a business that has to wait on a regulator's approval, a rule that has kept it in fusion and, for the most part, out of fission.
He wants every risk a company carries to sit between it and its customer, where execution can retire it. On his account fusion is regulated to roughly the standard of an MRI machine, and a Virginia community that drove out a gas plant welcomed a fusion reactor on the same site, while data centres stall in council processes.
-
2 of 2 contrarian
Chris Sacca puts Lowercarbon's power and equipment companies in the most bankable layer of the AI build-out, with years of demand left even if AI valuations were cut in half.
On his account the power is needed for air conditioning and electrification whether or not AI pays, while open-weight models leave pricing at the model layer in doubt. Antora shipping 200 units against orders for thousands is his case for skipping a demonstration phase, and he reads the Blue Owl financing of the Stargate site in Abilene, as the host describes it, as lenders treating compute as real estate.
MacroVoices
Luke GromenForest for the Trees
Founder of Forest for the Trees, which aggregates public data to identify developing economic bottlenecks for its clients; he has worked in investment research for 31 years.
-
1 of 2 contrarian
Luke Gromen calculates that every 5% of US consumer debt that defaults because AI takes white-collar jobs would consume about 40% of the US banking system's equity.
The base is $19 trillion of mortgages, card and car loans against $2.6 trillion of bank equity on the Fed's figure, and he cites staff cuts at Kroger, American Express, Barclays and Microsoft, which he ties to AI, as the job-loss case gaining ground. Banks would then have to stop lending, sell Treasuries, issue equity or let the economy slow, and he suspects the lag in bank, card-lender and private-credit shares is an early sign.
-
2 of 2 prediction
Luke Gromen tallies up to $5.2 trillion a year of Treasury supply in a recession set off by an AI bust, and finds no buyer for it except the Fed.
The sum starts from a $2 trillion deficit that the last three recessions widened by 6 to 10 points of GDP, then adds banks selling Treasuries against credit losses, foreign central banks on pace for about $600 billion of sales since February, and hedge funds, which he says own 8% of the market, cutting exposure.
Forward Guidance
The Fed Is A Sideshow While The AI Trade Implodes | Weekly Roundup
-
1 of 2 framework
Forward Guidance argues that an over-levered oil producer keeps pumping as prices fall, and expects neoclouds and chip makers to do the same with compute.
Fixed interest costs mean every extra dollar of operating margin helps service the debt, so output rises into a falling market and the glut outlasts the turn in demand. Data centre owners carrying debt-funded fixed costs face the same arithmetic, the hosts argue, and the credit market settles it at the expense of equity holders.
-
2 of 2 current issue
Forward Guidance reads this week's hyperscaler capex guidance as the start of the AI investment slowdown, with Microsoft at $38 billion against a $45 billion estimate and Google at $51 billion against $55 billion.
The hosts put the hyperscalers at the top of a pyramid: computer equipment investment has made its largest contribution to US GDP since 1959, so spending guided below estimates reaches growth through every supplier beneath them. They treat this week's GDP print, and the Fed's upgrade to a "strong pace", as rear-view data.
20VC
Jason Lemkin and Rory O'DriscollScale Venture Partners
Jason Lemkin is a SaaS investor whose portfolio includes Algolia, Talkdesk, Owner, RevenueCat and Salesloft. Rory O'Driscoll is a general partner at Scale Venture Partners, where he led investments in Bill.com, Box, DocuSign and WalkMe.
-
1 of 2 current issue
Jason Lemkin says Broadcom had to guarantee most of a debt deal that let Anthropic buy its chips, at higher interest rates than expected, and reads it as bank appetite for AI lab debt running out.
Rory O'Driscoll accepts the banks have hit their limit and argues the financing is moving to equity: Anthropic's planned issue of about $100 billion of stock, Meta's $70 billion and Google's roughly $80 billion earlier this year, and the SpaceX listing. Companies with a cheap cost of capital are choosing dilution, so on his account the spending continues.
-
2 of 2 current issue
20VC's panel says US stocks are at an all-time high on a 10-year earnings multiple yet look cheap on forward earnings, partly because profits now include one-off gains on stakes in private companies.
It describes 40 or 50 tech and AI stocks carrying the S&P 500 while the other 450 or so quietly lag, and few IPOs because the market has been waiting for the trillion-dollar listings. On its account, with listings stalled and private companies short of capital, it suits the mega venture platforms, which LPs say raised bigger funds, spent faster and are coming back to raise next year, to deploy heavily.
Excess Returns
Jay HatfieldInfrastructure Capital Advisors
Jay Hatfield is founder and chief executive of Infrastructure Capital Advisors, which runs a family of income-focused ETFs and strategies, billed at about $4 billion. He came to it from utility investment banking, power project finance and the hedge fund side.
-
1 of 2 framework
Jay Hatfield says AI is not a bubble while the market price of computing capacity stays above the rates in contracts signed a year or two ago.
Hatfield borrows the test from power and LNG, where a plant contracted below market earns more when the contracts roll, as with Cheniere's export contracts near $12 against European prices near $30. He separates contracted build-out from speculative supply: 1990s fibre was laid ahead of demand and took 25 years to fill, and office developers now need 60% to 70% pre-let before they build.
-
2 of 2 framework
Jay Hatfield says the 2017 cut in the US corporate tax rate, from 35% to 21%, breaks comparisons of today's valuation and dividend ratios with their history.
On his account the cut probably lifted after-tax returns on invested capital from about 13% to 15%, and many companies spent the saving on wages and investment, so growth now accounts for a larger share of a stock's value. He values on PEGY, price to earnings over growth plus yield, because a regulated utility yielding 5% and growing 5% is worth the same as one paying nothing and growing 10%.
Odd Lots
Alexander AltmannBarclays
Global head of equities tactical strategies at Barclays, whose team travels the world talking to investors and has surveyed them on AI financing. A returning guest, last on Odd Lots to discuss levered ETFs.
-
1 of 2 current issue
Barclays' Alexander Altmann expects about two-thirds of next year's S&P 500 earnings growth to come from semiconductors, and about 75% from the wider AI ecosystem.
He calls it the KOSPI-fication of the US market and sets it beside capex he puts, taking forward consensus figures at face value, at about 12% of US GDP, larger on his reckoning than the railroads or the Louisiana Purchase. On his account, companies with cash-rich legacy businesses can fund that build from operations.
Audio streamed from the publisher.
-
2 of 2 explainer
Barclays' Alexander Altmann traces calm index volatility above violent factor rotations to multi-manager fund assets, up from about $300 billion in 2020 to over $1 trillion.
On his account the pods post only margin, park the rest of their capital in high-yielding instruments and lever through prime brokers, so more than triple the money trades a stock universe that has not grown. Every fund neutralises the same style factors, which only moves the crowding somewhere else, and many have yet to map an AI factor whose betas keep shifting, as software went from a popular short to near 12-month highs.
Audio streamed from the publisher.
ChinaTalk
Meg RithmireHarvard Business School
James E. Robison Professor at Harvard Business School and author of Precarious Ties: Business and the State in Authoritarian Asia, which compares China with Malaysia and Indonesia. With Peter Cowhey of UC San Diego she wrote "The Cost of Keeping China Out" for Foreign Affairs.
-
explainer
Meg Rithmire says China's zombie firms have moved out of property and into the industrial-policy sectors, where they make a great deal and earn no profit.
On her account Beijing reacts as much as it plans: once real estate reached 20% of the economy with half the developers overextended, it restricted the flow of resources into property and pushed it towards industrial sectors. The loss-makers that absorbed it now crowd those markets, so even BYD and the solar makers must keep adding scale and compete on extras such as massage chairs, because price no longer separates them.
Top Traders Unplugged
Andrew McDermottMission Value Partners
Founder of Mission Value Partners, a Japan-focused investment firm he started in 2010, after working at JP Morgan and then Southeastern Asset Management.
-
1 of 2 current issue
Andrew McDermott argues Japanese bond investors have less reason to take currency risk abroad to meet their return targets, with JGB yields around 3% set against hedged US Treasuries.
GPIF's objective, which he believes is 1% over real wage growth, is the anchor: against a fixed objective the home market may now clear a bar that once sent savings overseas. Because Japan is so large a net investor elsewhere, he argues even a small reallocation moves other markets, with France the biggest relative outlier.
-
2 of 2 contrarian
Andrew McDermott argues Japan's public debt stops looking like an outlier once US and Japanese pension promises are counted the same way, citing a Federal Reserve paper.
On his account Japan restructured its pension promises so they are fixed in nominal terms while the assets behind them are real, which lets inflation work for the state's balance sheet. The US arrangement runs the other way, with real promises set against a Social Security fund held in Treasury bills.
No Priors
Jake StauchServal
Co-founder and chief executive of Serval, a two-and-a-half-year-old enterprise automation company of 200 people that competes with ServiceNow and counts Fox Corporation among its customers. He found the category selling physical security products to IT teams at Verkada.
-
current issue
Serval's Jake Stauch says the CIOs among his customers are trying to pull back from token-maxing on AI coding while still planning to raise their spend.
On his account most companies are early enough that each will run the same cycle of rapid uptake, pullback and measurement, so aggregate spend keeps rising. What they ask for is proof of return, and he doubts anyone has yet worked out how to measure what a token buys; Serval prices on workflows run, which he sees as closer to work done than token counts that also reward verbosity.
a16z
Jay MadheswaranEve
Co-founder and chief executive of Eve, which automates plaintiff law from the first client call to settlement; he says Eve has probably served a couple of hundred thousand clients, and that one Los Angeles firm on it went from about 100 to 500 clients per attorney.
-
framework
a16z argues venture investors underestimated plaintiff law software by counting what firms spent on Microsoft Office, when the real market is the legal labour those firms cannot hire.
The firms deliver outcomes worth hundreds of billions of dollars a year and are capped by staff, so software that does that work is bidding for the wage bill. Madheswaran adds that legal software money went to the defence side because that is where the budgets were, leaving plaintiff firms running on a phone and an email address.
Latent Space
Joseph KimOpenAI
Leads ChatGPT for Financial Services at OpenAI, launched with Morgan Stanley and Evercore as design partners; he came to it from the research side.
-
current issue
OpenAI says some big-name financial data connectors it tested ran error rates of 40% or more, which is why it bundled earnings transcripts and other data into its finance product.
The first-principles case against bundling, that customers may already hold licences and connectors already exist, failed on contact with the data, Kim says. Charging customers again for data they may already license was the price of controlling the experience end to end.
Latent Space
Aman SinghArca
Aman Singh leads the AI platform team at Arca, building the agent harness that runs the firm's back office; Arca has raised $64 million to build an AI-native wealth management firm.
-
explainer
Aman Singh says frontier models will confidently use last year's tax brackets, so Arca's harness nudges a model toward a base of verified current facts, or a web search, whenever it is about to make a financial recommendation.
Code the agents write in a sandbox is translated into a fixed format before it touches client data, and each finished plan passes deterministic checks for raw account numbers and small-model checks for promised returns. Singh argues the automated review catches more than human compliance staff would, because it does not tire.
AI 1 November 2026 · 50 clips
Odd Lots
Thai Duong and Bruce DangCalif
Thai Duong is chief executive of Calif, a hacking firm that tests models and infrastructure for AI labs and neoclouds, which he founded after 12 years hacking Google's products before release. Bruce Dang, a Calif engineer, spent nearly a decade in Microsoft's Security Response Center, where he helped reverse-engineer Stuxnet, then worked on iOS security at Apple.
-
1 of 4 current issue
Calif says AI-driven attackers spread through two of its banking clients' networks between a Monday and a Tuesday, where its own testers usually need about a month.
Calif cites OpenAI's responders in the Hugging Face incident, who saw about a hundred compromised points inside the network at once against the usual one or two. Detection, the advice Calif used to give clients, does not help against a hundred intruders, and shutting down is not open to a bank, so on Calif's account the defence has to be redesigned.
Audio streamed from the publisher.
-
2 of 4 current issue
AI is turning up about ten times more bugs than expected in software thought well defended, Calif says, including around a thousand vulnerabilities reported in Chrome.
PHP was never known for security, so a flaw there surprised nobody; Chrome had years of investment both in removing bugs and in mitigations that make the remainder hard to use. Calif's explanation is that either security engineers missed them or there was more code than anyone had time to read, which leaves little room for the hope that AI is only clearing legacy technical debt.
Audio streamed from the publisher.
-
3 of 4 contrarian
Calif argues that a model good enough to automate coding is necessarily good at hacking, so AI labs cannot train out one without giving up the other.
On its account the money is in selling tokens, most now spent on software work, and an agent that does that work must read code, debug and control computers, which is a hacker's job description. Calif answers Jensen Huang's call to fix the model, as the hosts relay it, the same way, noting that he does not make money unless the AI companies do.
Audio streamed from the publisher.
-
4 of 4 explainer
Calif says AI has created what it calls minus-one-day vulnerabilities: a fix posted publicly on GitHub can in principle become a working exploit within hours, before users can install it.
Attackers used to wait for a patch to ship and then spend a long time building an exploit, which gave users time to upgrade. On Calif's account groups now run what it calls n-day factories, watching Microsoft, Apple, PHP, Python and Java for any sign of a fix and pointing AI at it, and the window is closing on every major platform.
Audio streamed from the publisher.
20VC
Jason Lemkin and Rory O'DriscollScale Venture Partners
Jason Lemkin is a SaaS investor whose portfolio includes Algolia, Talkdesk, Owner, RevenueCat and Salesloft. Rory O'Driscoll is a general partner at Scale Venture Partners, where he led investments in Bill.com, Box, DocuSign and WalkMe.
-
explainer
Rory O'Driscoll reads Anthropic's S-1 as showing that 80% of its revenue comes from models less than six months old.
A footnote puts 70% to 80% of compute spending on training, so on his reading each model is very expensive to build and very profitable to serve for six to 12 months before it must be replaced, unlike software that can sell for perhaps 10 to 15 years. He argues the short shelf life favours a leader willing to keep writing $10 to $12 billion compute cheques, because few rivals can.
Eye on AI
Jeff LunsfordTealium
Chief executive of Tealium, a real-time data collection and distribution layer he helped fund in 2011 that connects to about 1,300 vendors and processes about 20 billion events a day; on his account its customers include 10 of the world's largest banks and six of the 10 largest telcos.
-
current issue
Tealium says AI agents now account for about 20% of the legitimate inbound website traffic its enterprise customers receive.
Lunsford puts revenue per session from these visitors about 38% above that from people. On his account the machines read a structured data layer beneath each page that no person sees, so the site matters more as infrastructure, and formats such as JSON-LD affect whether a delegated task can be finished there.
All-In
Anthropic S-1 Shocker, Greatest Business Ever? Google’s $500B Wipeout, Midterm Predictions
No guest this week; the four hosts take Anthropic's S-1 filing, Alphabet's third-quarter results and the US midterms.
-
framework
David Friedberg traces each dollar Nvidia spends making chips to about $18 of revenue for Anthropic, with margins of 75% or more at every layer.
His stack runs from Nvidia's 75% margin through SpaceX earning about 83% on its Colossus cluster on five-year depreciation to Anthropic's roughly 85% incremental gross margin. Customers spending $100,000 or more rose from 200 at the end of 2024 to 11,900 by September, against an addressable base he puts near one million; the customer payback of two to 10 times that closes the chain is an assumption, and the panel says so.
a16z
Jay MadheswaranEve
Co-founder and chief executive of Eve, which automates plaintiff law from the first client call to settlement; he says Eve has probably served a couple of hundred thousand clients, and that one Los Angeles firm on it went from about 100 to 500 clients per attorney.
-
current issue
Eve says US plaintiff firms turn away 96% of employment cases, because about 10 hours of screening comes before a firm knows whether a case clears its roughly $5,000 fee floor.
Firms lose money on many cases even above that floor and live on a few large payouts, Madheswaran says, so a claim below an insurance deductible gets no lawyer at all. The constraint is legal labour rather than demand, and AI that answers the intake call and does the research at once moves the floor down.
Training Data (Sequoia)
Chase LochmillerCrusoe
Chief executive of Crusoe, which builds power generation, gigawatt-scale data centres and a managed AI cloud, has passed 2,000 employees and recently closed a round at a $31 billion valuation. He was a quant portfolio manager before founding it.
-
framework
Crusoe's Chase Lochmiller models the company on Exxon, arguing that an AI supermajor spanning power, data centres and cloud carries the same built-in hedge an integrated oil major has between the wellhead and the refinery.
On Lochmiller's account, cheaper crude squeezes margins at the well and widens them in processing, so the exposure is left unhedged; the layered position also means serving rivals, as a pipeline moves competitors' gas. He grounds it in a view that this market has room for several large winners, so partnering with firms it competes against elsewhere costs little.
Latent Space
Joseph KimOpenAI
Leads ChatGPT for Financial Services at OpenAI, launched with Morgan Stanley and Evercore as design partners; he came to it from the research side.
-
contrarian
OpenAI does not think a good benchmark for financial work exists yet: the finance questions its models still missed traced to flaws in the evals themselves.
Kim says OpenAI has stopped reporting GDPval, and that when a new model's finance scores barely moved, the team found the misses came from wrong answers in the key and poorly specified prompts. He thinks researchers dig into an eval only when its numbers disappoint; his team is now building an investment banking eval meant to predict whether a junior analyst's work can be automated.