IPOs and SPACs are Back, Mag 7 Showdown, Zuck on Tilt, Apple's Fumble, GENIUS Act passes Senate

Summary

• Meta’s reported $100 million talent offers and $14 billion-plus Scale AI stake are rational insurance against an AI threat to perhaps half of its $1.7 trillion value. Thomas Laffont calculates that spending 4%-5% of the roughly $850 billion at risk makes sense if it marginally improves Meta’s odds. Chamath’s warning is that labeling and agent knowledge supply only two parts of “the compounding of secrets”; without tightly coupled compute and silicon, Meta remains “on their back heel.”
• The Mag 7 has fractured into distinct AI wagers: Meta +18%, Microsoft +13%, Nvidia +8%, Amazon -3%, Google -8%, Tesla -20%, and Apple -21% in the cited period. Five-year picks clustered around Google and Tesla because both can integrate models, proprietary infrastructure, distribution, and physical products. Nvidia remains protected by GPUs, but Friedberg sees a “low probability but very high severity risk” from Chinese semiconductor innovation.
• Apple drew the episode’s sharpest disagreement: its large installed device network could become an ambient AI moat, or its cash-cow culture could make reinvention impossible. Friedberg imagines one “ethereal and ubiquitous” assistant moving across AirPods, watches, phones, cars, and rooms; Jason wants a humanoid robot. Chamath sees a company optimizing cables, replacement devices, and buybacks—“a this-and-that strategy is not a strategy”—while Thomas argues Apple’s recurring-profit transition proves it has reinvented itself before.
• IPOs and M&A are reopening because investors need exposure to scarce growth after SaaS decelerated from 17% median growth in 2021 to 9% today. CoreWeave reportedly quadrupled to an $81 billion market cap, Circle rose roughly sixfold to $48 billion, and Chime initially gained 40% before retreating 20%. The old growth basket is fading: only 5% of the cited SaaS cohort still grows above 25%, versus 25% in 2021.
• AI’s economic upside comes from expanding service throughput while replacing bloated software and operating expense. OpenEvidence was said to reach one-third of US physicians, often used ten times daily, while Chamath described development gains of 50%-70% at successive workflow stages that compound into dramatically smaller teams. His trade: find businesses capable of replacing hundreds of millions in licenses with tens of millions in custom software.
• The labor outcome remains unresolved even among the panel: Microsoft’s roughly 250,000-person workforce could grow, remain flat, or shrink depending on whether AI creates revenue faster than it removes work. Chamath says today’s coding agents produce too much “crap” on long, complicated tasks for layoffs to be credited to them; Friedberg thinks that limitation may disappear within three or four years. All agreed that owning a synthetic basket of AWS, Azure, and Google Cloud could capture the infrastructure demand whichever platform leads.
• The Senate’s 68-vote GENIUS Act would bring stablecoin issuers onshore, require quarterly audits and one-to-one reserves, and give legacy offshore issuers three years to comply. It also preserves banks’ position by prohibiting issuers from passing reserve interest to token holders—a compromise Sacks hopes will eventually be revisited. His framing was a reversal from “regulation through prosecution” toward rules the crypto industry can actually price and follow.

Deep dive

1. Friction is pushing growth out of Los Angeles—and AI may fill the gap

• Laffont cited restaurant recovery per location running 50% behind the national average and Los Angeles filming down 50% from its peak. His contrast was structural: San Francisco is levered to an expanding AI economy, while LA remains tied to entertainment in “secular decline” and is losing productions to competing geographies.

• Jason added that production in California was quoted as roughly 30% more expensive, with staffing, paperwork, and setup speed compounding the gap. Friedberg’s Beast Games example carried the argument: tax credits favored Las Vegas and Toronto, while season two secured a major Saudi deal, built sets there, and planned to leave them there.

• The conference’s macro counter-thesis was that AI could improve the debt-to-GDP equation by lifting US productivity over five to ten years, potentially moderating the rate pressure implied by debt alone. Friedberg said the US could capture AI’s economic surplus first, through both knowledge-worker productivity and some onshoring of physical industry.

• OpenEvidence was presented as an early specimen: already used by roughly one-third of US physicians, often ten times daily, with particular oncology traction. Friedberg’s mechanism was conditional but important: if AI lets a doctor see ten times as many patients, lower prices can coexist with more diagnostic care and a larger GDP contribution. Jason’s viral Veo 3 dentist-ad example showed similar leverage in customer acquisition.

2. Meta is spending like half its value is at risk

• Jason treated reports of $100 million signing bonuses—and potentially more than $100 million in annual compensation—as extraordinary but unverified. Sam Altman said “none of our best people” had accepted so far, arguing employees still saw OpenAI as likelier to deliver superintelligence and perhaps become the more valuable company.

• Meta’s $14 billion-plus purchase of a 49% Scale AI stake was framed by Jason as a “shadow acquihire”: Alexandr Wang could join the new superintelligence team immediately while Scale nominally remained independent. Jason’s chessboard read was that OpenAI and Google cancelling Scale contracts could leave Meta with capabilities and data that rivals had relied upon, though the episode did not establish that Meta would receive exclusive access.

• Laffont called the expenditure “highly rational.” If perhaps 50% of Meta’s roughly $1.7 trillion capitalization—about $850 billion—is endangered by AI, spending 4%-5% of that amount is sensible if it slightly improves the odds. His Onavo analogy was exact: Facebook once bought a uniquely valuable mobile-engagement data service, internalized it, and removed an investor tool from everyone else.

3. The winning AI stack compounds secrets from labels to silicon

• Chamath traced today’s problem to Facebook’s HTML5 error. Facebook Zero made sense in developing markets where a browser could evade carrier tolls, but making HTML5 the main strategy sacrificed native integration. His preferred full phone, full-stack, native-app plan lost politically; Mark Zuckerberg later called the alternative the company’s “single biggest mistake.”

• The current analogue is “the compounding of secrets” across training, models, infrastructure, and compute. OpenAI gained Azure’s tightly integrated training environment through o3; Google couples Gemini to TPUs; Anthropic has deliberately used TPUs; and Chamath inferred similar hardware coupling at DeepSeek and xAI. Meta, by contrast, had been training generically on Nvidia and releasing Llama as open source.

• Scale supplies training secrets, including expert reasoning data—not merely labeling a photograph “dog,” but constructing outcome sets where even “2 plus 2 equals 4” becomes reasoning material. Nat Friedman and Daniel Gross could contribute application and agent-building knowledge from their investments. Chamath’s diagnosis: Meta would then have labeling and app secrets, but still lack infrastructure and hardware secrets.

• Friedberg’s failed transpiler bet, made eight or nine years earlier and later unwound, changed his view. Redirecting CUDA workloads to arbitrary chips sounded attractive, but transformer attention mechanisms had to be “hand-tuned for every single target of silicon.” A new Amazon chip means little without models built for it; likewise, a generically deployed model cannot capture the gains of dedicated compute architecture.

4. The Mag 7 trade has fractured into seven arguments

• Chamath’s cited performance table showed Meta +18%, Microsoft +13%, Nvidia +8%, Amazon -3%, Google -8%, Tesla -20%, and Apple -21%. After years of correlation, he read the dispersion as the market beginning to decide “who are going to be the winners and losers” rather than buying the group indiscriminately.

• Friedberg cautioned that policy and operating conditions distort the AI signal. Tesla faced falling vehicle demand and was losing solar and EV tax credits; Apple faced tariffs, onshoring demands, and Chinese supply-chain exposure; Amazon also carried tariff sensitivity. Those “influenced market forces” help explain why price action does not map cleanly onto technical position.

• A second lens was control: Tesla and Nvidia clearly control crucial parts of their destinies, while Amazon lacks its own foundation model and Microsoft owns a large share of OpenAI without controlling it. Chamath also observed that companies increasingly say “superintelligence,” not AGI; he interpreted that shift as evidence that AGI is no longer viewed as imminent, while multiples of human intelligence sound more achievable.

5. Google and Tesla dominate the five-year winner picks

• Laffont selected Nvidia first because “all roads still lead to the GPU,” even as other architectures expand the market. His dark horse was Tesla because its potential vertical integration could span silicon, models, and physical products. Friedberg separately described humanoid robotics as a “low probability, high upside” call option embedded in Tesla.

• Chamath chose Tesla first and Google second. Tesla could combine leading vision models, xAI’s language and reasoning systems, Dojo, cars, robotaxis, and robots; Google combines Gemini, TPU, quantum work, and billions of users. Even if search declines, he thinks Google can pivot its economic north star from “price per click” to “price per token.” Veo 3, he predicted, could leave Hollywood “done” within a year.

• Friedberg ultimately favored Google, then Tesla when valuation was excluded. Google offers a portfolio of potentially enormous outcomes—Waymo, quantum computing, Isomorphic’s biologics work, weather models, and multimodel agent systems. His Nvidia hedge was China: isolation creates incentives to cross the lithography moat, and a demonstrated one-nanometer process suggested America might again be surprised, as it was by DeepSeek.

• Jason also chose Elon Musk’s ecosystem and Google, arguing Tesla and xAI should merge so Colossus, X’s real-time data, FSD, Optimus, and engineering talent point in one direction. Sacks rejected a strictly poker-like, zero-sum framing: Tesla robots, Google-generated media, and Nvidia infrastructure could each create trillion-dollar businesses in the same ecosystem.

6. Apple looks more like a cash cow than an AI contender

• Jason criticized Siri, Apple’s lack of visible AI progress, Project Titan’s reported $10 billion spend before closure, and the company’s limited acquisition activity. Sacks said Siri was still barely useful after “year 27,” while Thomas described the situation as “regime-change” territory and noted that Apple’s largest acquisition was Beats.

• Chamath classified Apple as a classic transition from growth company to cash cow. Executives with 20- or 30-year tenure provide stability but can lose the energy and lived experience needed to imagine the future; elite AI recruiting conversations feature OpenAI, Meta, and Google, “but what you don’t hear is Apple.”

• Sacks and Jason used HP, Lotus, Intel, and General Electric as precedents for creative destruction. Laffont conceded that Apple no longer controls the decisive model layer, making it resemble PC manufacturers that owned hardware but not the operating system.

• Friedberg defended Apple’s earlier reinvention: one-time iPhone hardware once supplied more than 90% of gross profit, versus roughly 40% now. Chamath floated the extreme possibility of Apple buying OpenAI for $500 billion, and Jason thought Apple’s stock might rise on the announcement.

7. Apple’s installed base is either the assistant moat or the trap

• Friedberg’s product answer was an ambient AI assistant, not necessarily a single new device. He owns “30 freaking Apple devices,” making him an easy convert if intelligence can move continuously across computers, phones, AirPods, watches, cars, and rooms—an “ethereal and ubiquitous” agent that retains identity and context without forcing users to stare at a screen.

• Jason preferred a humanoid robot, while Laffont noted that AirPods alone generate roughly three times OpenAI’s current revenue. In the cited interview, Craig Federighi argued Apple already offers environmental audio, visual capture, wearables, and a glanceable screen; other AI form factors may emerge, but existing devices are “pretty hard to beat.”

• Friedberg then said he did not think Apple had “any chance of anything great.” Chamath instead emphasized that Apple’s executives remain highly competent at making money through the existing model, while Jason argued that AirPods, cables, and replacement-device revenue create a “this-and-that strategy,” not a future strategy.

• Jason warned that the scale of the AirPods business could itself produce internal complacency: “some smart-ass MBA” might dismiss OpenAI as smaller than Apple’s AirPods business and shut down a more ambitious discussion.

8. IPOs and M&A have reopened around scarce growth

• The episode cited CoreWeave, Circle, and Chime IPOs on March 28, June 5, and June 12, but did not explicitly map each date to a company. CoreWeave had risen roughly fourfold to an $81 billion capitalization; Circle was cited as 25 times oversubscribed and about six times its opening price at $48 billion; Chime initially gained 40%, then fell 20%, leaving roughly $12 billion.

• Jason’s M&A board included Google’s $32 billion Wiz deal, SoftBank’s $6.5 billion purchase of Ampere, OpenAI’s two acquisitions—one for $3 billion and one for $6.5 billion—Databricks buying Neon for $1 billion, and an $8 billion Salesforce acquisition. Windsurf was separately cited at $3 billion, while Jony Ive’s io was described as developing an AI hardware device. DoorDash and Uber added smaller deals, reinforcing Jason’s claim that “M&A is back on the menu.”

• Jason floated the demand mechanism: institutional managers had entered 2021-22 overallocated to private assets, then spent roughly three years unable to make new crossover investments. With most public-company profit growth anemic, they may now be hungry for fresh high-growth issuance; Chime’s roughly 18-times subscription was consistent with that pent-up demand.

• Sacks added the ecosystem test: “if you put a dollar in, you need to get a dollar out.” IPOs and acquisitions are finally returning money to private-market investors. CoreWeave and Circle also offered direct exposure to AI and crypto—the open-ended themes buyers want when older sectors no longer compound fast enough.

9. AI is ending SaaS’s easy-growth era

• Sacks’s cohort data quantified the break: median SaaS growth fell from 17% in 2021 to 9% today, while the share growing above 25% dropped from one-quarter to 5%. Investors can no longer buy a broad SaaS index and assume durable compounding; they must find businesses capable of sustaining roughly 25% growth for five to ten years.

• Jason’s explanation was that customers increasingly recognize another vertical tool can add cost, people, bloat, implementation delay, and price escalators without adequate return on equity. Since 2023, buyers have increasingly expected AI to rebuild that software. “The jig is totally up for software,” because development from scratch is becoming easier than maintaining accumulated vendor complexity.

• At 8090, Chamath described an end-to-end process running from product requirements to functioning code. Gains of 50%, 60%, or 70% at successive steps compound, letting a 30-person team transact hundreds of millions of dollars of work. His broader prediction was that “the entirety of the software that runs the world” will be rebuilt “soup to nuts.”

• Chamath offered an “order-of-magnitude correct” marker: Anthropic, the level-zero supplier behind code-generation companies such as Cursor, added about 70% of the public SaaS industry’s net new ARR in Q1. Friedberg said SaaS incumbents were responding by moving from per-seat to consumption pricing; Sacks argued that variable pricing ultimately pushes customers toward Postgres, Supabase, and cheaper alternatives.

10. The S&P 493 could split between rebuilders and relics

• Chamath put average S&P 493 margins near 12% and growth in the single digits, leaving legacy businesses exposed to “a couple kids in a garage” using OpenAI or Grok. His emerging trade was to be less long the past and own a few category killers capable of applying AI to durable real-world assets. Thomas said the environment might even support going short the S&P while selecting winners.

• The operating unlock requires leadership to cross an organizational language barrier. Chamath described CEOs, CFOs, and boards speaking English while IT departments speak Mandarin, allowing incomprehensible spending to persist; one CIO cited an $18 billion annual IT budget. A private-equity owner who can mandate change might replace hundreds of millions in licenses with tens of millions in customized software.

• He remained skeptical of buying and AI-enabling rollups in accounting, law, or IT services because the terminal buyer may disappear. Andrej Karpathy asking why Google login was not simply “one click behind the scenes” generalized to the category: if agents make services automatic, who buys the rollup in seven years? Chamath would instead screen for defensible offline assets—specialty chemicals and necessary lubricants were his specimen.

11. The public window is open, but structure still matters

• After nearly 58,000 people voted on his SPAC poll, Chamath said he was “heavily leaning” toward another one, partly because respected Wall Street and crypto investors encouraged him. His retail warning was unusually categorical: any future documents would include the poll and community note, and ordinary listeners should “stay as far away as possible.” The rationale was simply, “Fate loves irony.”

• Laffont saw real-time evidence that markets are “open for business,” with Circle, CoreWeave, Chime, Caris, and filed candidates such as Figma forming a new cohort. He did not care whether a strong business used a SPAC, direct listing, or conventional IPO; the decisive question was what the asset could be worth five years later.

• Structure still affects execution. First, the offering must provide enough dollars for a large investor to build a meaningful position; second, Laffont wants a broad float. Friedberg said roughly 20% was, in his opinion, a minimum for truer pricing and less manipulation; third, lockups determine how quickly supply appears. A direct listing without a lockup may reach genuine price discovery faster.

• Scar tissue cuts both ways: excluding SPACs, the cited 2021 IPO cohort was down roughly 40% after one year and 50% after five. Chamath’s Slack experience taught him that the first-day sale could be the best direct-listing trade, informing his Coinbase sale at $335. Yet Spotify’s roughly sevenfold gain supported Laffont’s insistence that business quality ultimately outruns listing mechanics.

12. AI lifts revenue per employee—and turns Amazon into a kingmaker

• AppLovin’s revenue per employee rose from roughly $3.6 million in 2021 to $7.6 million as headcount fell from about 1,000 to 750. Amazon’s Andy Jassy similarly told employees that extensive AI use should reduce the corporate workforce even as the company builds advertising, seller, product-page, shopping, and Alexa systems.

• Laffont adopted Jensen Huang’s demographic answer to displacement: an aging population will need more doctors, nurses, and caregivers from a smaller young workforce, so society “better get a lot more productive.” He expects flexible knowledge workers to redeploy their skills and sees AI making the economy richer, though the panel did not resolve the transition cost.

• Chamath called Amazon retail a kingmaker because it can absorb “a gajillion” successful robots or delivery drones from Figure, Tesla, or others. AWS is harder: its strength as a marketplace for everything inhibits a decisive stack. Andy Jassy may eventually need to differentiate Amazon silicon, make a real model bet, or even buy Anthropic and tightly couple code generation to AWS—choices requiring hundreds of billions.

13. Microsoft’s headcount became a referendum on AI’s real productivity

• Asked whether Microsoft would employ more people in five years than its roughly 250,000 peak, Jason quickly answered “more.” Chamath predicted roughly the same level—about 225,000 to 250,000—while Friedberg predicted fewer employees alongside possible revenue decline. Thomas did not give a separate forecast in this exchange.

• Chamath challenged Microsoft’s percentage-of-code-generated metric as a “dangerous vanity metric.” Current tools can help in single-player tasks, but over long, complicated enterprise jobs their errors compound until the output is worthless—hence “app crappers.” He expects that problem to be fixed eventually, but thinks AI presently provides convenient air cover for layoffs management already wanted.

• Friedberg agreed that AI code may be poor today but rejected extending that limitation three or four years. His bearish Microsoft mechanism was customer selection: legacy enterprises using Microsoft are likelier to die, while new winners build native software and workflows rather than buying the old application stack. In that world, cloud competition and shrinking customers pull both revenue and headcount lower.

• Chamath’s conversations with large-company CIOs complicated the winner-take-all cloud thesis: enterprises deliberately run Microsoft, Google, and other clouds to diversify exposure, not merely to obtain the best price. The panel agreed that a synthetic basket of AWS, Azure, and Google Cloud could capture the infrastructure demand; if one accelerates, its upside could more than offset weakness in the other platforms’ surrounding businesses.

14. The GENIUS Act marks a bipartisan reversal on crypto

• Sacks called the Senate’s 68-vote passage—including 18 Democrats—a “huge milestone,” because ordinary legislation needs 60 votes. He expected House action within weeks and credited Bill Hagerty as principal author, alongside Tim Scott, John Thune, Cynthia Lummis, Kirsten Gillibrand, Angela Alsobrooks, and House leaders preparing the next step.

• His baseline was the prior year’s “regulation through prosecution.” Gary Gensler invited startups to meet the SEC without giving them rules, but Sacks said enforcement staff recorded those conversations and companies soon received Wells notices—effectively a “honeypot.” Trump’s campaign promise and first-week executive order then reversed the signal and began removing Biden-era restrictions.

• Sacks identified Sherrod Brown’s Ohio loss to Bernie Moreno as one reason the political calculation changed, noting that the crypto industry had backed Moreno against Brown, whom Sacks portrayed as a legislative blocker aligned with Elizabeth Warren. With roughly 50 million US wallet holders—about one in five adults as cited—Democrats had reason to ask, “Why are we dying on this hill again?” Regulatory certainty became the bipartisan answer.

15. Stablecoins get audits and onshoring, but holders get no yield

• The bill would govern US-dollar stablecoin issuers and give legacy offshore operators such as Tether three years to conform and operate onshore. Sacks argued hostility had previously left banks uninvolved while issuers moved abroad; a domestic framework now allows regulated US companies and banks to compete rather than surrendering the market.

• Every issuer would face quarterly “real audits,” not merely attestations, verifying one-to-one reserves in dollars, US Treasury bills, or money-market accounts. The investor and consumer promise is redemption certainty: whenever a holder cashes out a token, “there’s a real dollar waiting there.”

• Sacks suggested that a noncompliant offshore issuer could lose exchange support and would be in violation of US law. He avoided alleging that Tether was undercollateralized; the point was that uniform audits eliminate uncertainty rather than requiring consumers to trust competing claims.

• The compromise is that stablecoin issuers cannot pass reserve interest to token holders. Community banks feared a stablecoin paying 5% would drain deposits and put them out of business. Sacks found that concern understandable but unpersuasive and hoped lawmakers would eventually revisit the prohibition; Jason suggested that could become easier once banks themselves participate.