Debt Spiral or NEW Golden Age? Super Bowl Insider Trading, Booming Token Budgets, Ferrari's New EV
Summary
AI agents are expanding knowledge work before replacing it. An eight-month study by two UC Berkeley researchers inside a 200-person technology company found that AI users worked faster, covered more tasks, and extended their hours, while also reporting more stress and burnout. Sacks expects “task-based jobs” to become “purpose-based jobs,” with AI-native employees completing days-long assignments in two hours and driving enterprise adoption from the bottom up.
Jason’s venture firm is already reorganizing around agentic coworkers with claimed 10-20x leverage. Four of 20 employees focus on adoption; OpenClaw “replicants” receive email, Slack, Notion, and Google Docs accounts, while an “OpenClaw Ultron” agent supervises four others. Jason said agents now perform roughly 20% of an average investment-team member’s work and 30% of his assistant’s former workload; Friedberg’s surprise was that “the output is what’s recursive,” rather than the model itself.
Confidentiality and control could make on-premises AI strategically necessary again. Chamath argued that using public endpoints can send prompt-response metadata and agent traces involving confidential strategy, models, and operating data back to model providers, invoking the example of GEICO actuaries exposing proprietary risk-pricing data. His conclusion: “AI may be the reason you can’t afford not to be on-prem,” whether through local desktops, centralized private infrastructure, or enterprise-secured agent platforms.
Token expenditure is becoming a second payroll—and may exceed compensation for superstar developers. Jason’s agents rapidly reached $300 per day each through cloud APIs while doing only 10-20% of the work, implying about $100,000 annually per agent and requiring employees to become at least 2x more productive. Jason suspects hardware and model advances will eventually push token costs toward one-tenth of current levels, but Friedberg said today’s Bedrock and CoreWeave options offer no economical solution at scale.
Prediction markets are approaching a pre-Reg FD world where sharps monetize information asymmetry and squares supply the winnings. Super Bowl volume reached more than $1 billion on Kalshi and $700 million on Polymarket; suspiciously informed accounts correctly called halftime details, while an account tied to alleged military-information misuse reportedly won over $150,000. Chamath said such markets can surface truth and misconduct faster, but their ephemeral nature makes beneficial whistleblowing difficult to separate from markets deliberately created or exploited by insiders.
The CBO baseline describes a debt spiral that becomes much worse if refinancing rates or unfunded pensions break the assumptions. The episode cited a $1.9 trillion deficit, debt rising from $31 trillion to $56 trillion by 2036—about $2.5 trillion a year on average from 2026 to 2036—and Social Security exhaustion in 2032. Friedberg calculated that rates near 5% would add about $650 billion of annual interest and push total interest toward $2 trillion. His larger tail risk is federal assumption of state and local pensions, including nearly $1 trillion of unfunded California obligations, becoming “the concrete that breaks the camel’s back.”
Sacks sees an AI-led “new golden age” where growth outruns the CBO’s fiscal arithmetic, while Chamath still favors durable-asset hedges. Sacks rejected the CBO’s 2.2% real-growth assumption for 2026 and 1.8% thereafter, pointing to 4%-plus and 5%-plus recent quarters, 172,000 January private-sector jobs, and $600 billion of projected hyperscaler CapEx—roughly a 2% GDP tailwind by his estimate. Chamath countered that globally synchronized debt-to-GDP may matter less relatively, yet currencies can still “fall off a cliff,” making gold and other real assets the practical protection.
All-In’s new Liquidity event aims to open closed-door capital-allocation networks. The May 31-June 3 wine-country event is for capital allocators, LPs, and GPs. Chamath described bringing together public-market, hedge-fund, private-market, growth, and credit investors; LPs representing trillions of dollars; and leading technology CEOs for best-idea presentations, relationships, possible investments, and allocations to emerging managers. Applications are at allin.com/events.
Ferrari’s first EV may preserve the luxury experience just as autonomy makes human driving a niche activity. The car is slated for May 2026 with 1,000-plus horsepower, four motors, sub-2.5-second acceleration, 330 miles of range, and a 5,100-pound weight. Sacks liked its tactile, screen-and-button interior but disliked the provisional exterior. Chamath’s broader call was that FSD, Waymo, insurance costs, and safety economics will shrink human driving into “smaller and smaller places and less and less often,” leaving Ferrari buyers paying explicitly for the experience.
Deep dive
1. AI turns task-based jobs into purpose-based work
Jason opened with a study published in HBR on Monday, February 9. Two UC Berkeley researchers spent eight months inside a 200-person technology company and found that AI users moved faster, assumed broader responsibilities, and worked longer hours. They felt more productive, but also reported greater stress and burnout.
Sacks connected that result to his contrarian prediction that AI will increase demand for knowledge workers. Employees voluntarily did more because automation removed “menial tasks” and up-leveled their contribution, shifting employment from “task-based jobs to purpose-based jobs.”
The differentiating skill, in Sacks’s framing, will be structuring work for both oneself and one’s agents. Early adopters will “appear to have superpowers,” delivering a presentation or spreadsheet in two hours when colleagues still expect several days.
Enterprise deployment therefore may arrive as a “fait accompli”: employees bring consumerized agents into daily workflows while centrally managed transformation teams spend months studying vendors and running RFPs. Sacks likened the pattern to bottom-up adoption of consumerized SaaS.
2. Agentic coworkers are becoming an operating layer
Jason said four of his 20 employees already focus on agents and exhibit 10-20x the leverage of the other 16. He expects firm-wide adoption to take roughly six months and believes it could create a 10x advantage over competing firms.
One podcast workflow automatically finds or clips moments, deposits them in Google Drive, analyzes YouTube, Instagram, and TikTok performance, and proposes ways to improve distribution. The reporting disappears; humans receive the clip and recommended action.
His firm moves another 5-10% of its work to OpenClaw each week. Its “replicants” receive Notion, Slack, Google Docs, and email identities—access companies had previously avoided granting because they did not want to be responsible for agents mishandling passwords, Bitcoin keys, or other sensitive information.
A supervisory agent called “OpenClaw Ultron” now checks four replicants, asks what they are doing, and summarizes their work. Jason estimated that agents absorbed 30% of his assistant’s workload and 20% of an average investment professional’s. His strongest claim was: “They don’t forget to do work. They don’t make mistakes.”
3. Recursive output, not recursive training, unlocked the surprise
Friedberg said researchers had expected recursive improvement to come from continuously retraining a model on itself once context windows became large enough. Instead, “it may be the case that the output is what’s recursive.”
Chamath described agents completing a task, receiving critique from another agent, acquiring new skills or best practices, and repeating the cycle on a schedule. Friedberg saw the resulting performance gains as resembling the effect computer scientists had expected from recursive model development itself.
His posture remained explicitly open: “Let’s see how far it goes.” The mechanism is producing unexpectedly rapid gains, but he did not claim the process has removed underlying limits.
4. Confidentiality could swing enterprise compute back on-prem
Chamath’s first open question was, “Is on-prem the new cloud?” Cloud infrastructure won through shared scale, lower CapEx, and lower OpEx, but AI may make leakage of confidential strategy, models, and operating data a deep problem for some enterprises.
His example was a strategy PDF, PowerPoint, or critical model interrogated through a public ChatGPT, Gemini, or Claude instance: the prompt and response metadata go back to the provider. Agentic workflows add detailed execution traces. GEICO, he argued, would not want actuaries interrogating proprietary risk-pricing data in an open LLM instance.
Chamath also cited a judicial ruling that, in his telling, denied attorney-client privilege inside such cloud environments and treated the material as public-domain information. Combining that claim with unavoidable AI adoption produced his dilemma: enterprises need AI to survive, yet public endpoints may sacrifice “all control, all security, all confidentiality.”
Proposed architectures ranged from powerful local desktops to VAX-like centralized private computers serving “dumb terminals.” Friedberg said a local large language model may require a Mac Studio with 512 GB of memory or two daisy-chained systems. He cited 8VC’s experience as a top-20 Bedrock customer and said cloud costs were already too high; CoreWeave is oriented toward training, while spot pricing creates unusable surges.
Sacks saw an opportunity to make OpenClaw enterprise-grade and secure. Chamath’s conclusion was categorical: “AI may be the reason you can’t afford not to be on-prem.”
5. Token budgets are beginning to rival salaries
Jason put a price on the problem: his agents “hit $300 a day per agent” almost immediately through cloud APIs while doing only 10-20% of the work. Annualized, that is roughly $100,000 for each agent.
That forces explicit token budgets for top developers and a new hurdle rate: an AI-intensive employee must become at least 2x as productive, “otherwise I’ll run out of money.” Asked when tokens overtake salary, Jason said superstar developers might already be there.
Friedberg found no economical infrastructure answer today. Bedrock carries provider overhead; CoreWeave capacity is oriented toward training and requires future commitments, while spot pricing creates intolerable surges—hence, “There is no solution today that makes any sense.”
Jason suspects NVIDIA, Grok, Google, AMD, and others will increase density and drive token prices sharply lower, perhaps toward one-tenth of current output cost. That would ease the bill without removing the confidentiality incentive for private deployment.
6. Prediction markets reward the informed before the merely opinionated
Super Bowl trading reached critical mass: more than $1 billion on Kalshi and $700 million on Polymarket. One day-old account correctly predicted 17 of 20 halftime outcomes but earned only $17,000; another new account accurately traded Bad Bunny’s set list.
A more serious case involved Israeli personnel allegedly betting with classified information about military operations. The cited account reportedly earned more than $150,000, went dormant for six months, then returned to wager on the timing of an Israeli strike on Iran.
Friedberg asked whether knowledge-based advantage in a voluntary side bet should count as insider trading. His concern was platform durability: a few well-capitalized accounts with informational or arbitrage edges can steadily consume the many small accounts that “get burned through very quickly.”
Sacks supplied the market structure: “sharps” possess an edge; “squares” are “grist for the mill.” Unlike monitored sports, prediction markets can concern events insiders know or control, yet remain too fluid, dynamic, and ephemeral to regulate cleanly as securities.
7. Information asymmetry is both the product and the poison
Sacks compared prediction markets with equities before Regulation FD, when selective disclosure was “not illegal.” A CFO could tell one manager that a quarter was a blockbuster, creating networks that monetized privileged information.
He used a displayed chart of Warren Buffett’s pre- and post-Reg FD performance to argue that Buffett’s returns fell from roughly double the market to market-like performance once information became symmetric. His conclusion: “Markets thrive when there’s asymmetry.”
Unregulated prediction markets, on that framing, will resemble “the stock market pre-Reg FD,” with sharps taking money from squares. Regulation may eliminate the product; non-participation may be the only dependable defense for an uninformed trader.
Chamath nevertheless preserved the social upside: insider incentives might expose corruption, misdeeds, or current events faster than formal whistleblower systems. The unsolved problem is distinguishing those truth-producing markets from wagers created by someone who already controls the answer.
8. Liquidity aims to open closed-door capital-allocation networks
The Besties announced Liquidity, a May 31-June 3 wine-country retreat and summit for capital allocators, LPs, and GPs. It is not intended as a general-admission event; applications are available at allin.com/events.
Chamath framed the event as opening traditionally closed idea-sharing venues. The planned group includes public-market and hedge-fund investors, private-market, growth, and credit investors, LPs representing trillions of dollars, and CEOs of fast-growing technology companies.
The program will feature presentations, best ideas, relationship-building, and possible investments. Chamath also said the organizers may allocate places to emerging managers who need capital and can show good returns.
9. The debt spiral hinges on rates and hidden public liabilities
The CBO figures presented were a $1.9 trillion deficit—nearly 6% of GDP—debt rising from $31 trillion to $56 trillion by 2036, or roughly $2.5 trillion per year on average from 2026 to 2036, and debt-to-GDP moving from 120% to 135%. Social Security exhaustion moved forward to 2032.
Friedberg attacked the assumed 3.1% short-term refinancing rate. At roughly 5%, he calculated another $650 billion of annual interest, taking total interest close to $2 trillion; borrowing that interest compounds the stock of debt and creates the spiral.
His larger conditional risk arrives if Democrats win the midterms and the White House in 2028: federal support for insolvent state and local pensions. California alone was said to carry nearly $1 trillion of unfunded public-retirement obligations omitted from the CBO baseline.
Chamath suggested bankruptcy or another restructuring mechanism for state and local fiscal or pension obligations. Without it, Friedberg said federalization would be not merely “the straw” but “the concrete that breaks the camel’s back.”
10. Growth is the escape hatch, but currency debasement remains the hedge
Sacks challenged the CBO’s 2.2% real-GDP forecast for 2026 and 1.8% thereafter against growth above 4% in Q3 and a preliminary figure above 5% in Q4. If AI infrastructure earns a return, those assumptions may prove far too meager.
His fiscal prescription was to freeze federal spending until growth pulls outlays from roughly 23% of GDP back toward the historical 20%, while receipts remain near 17%. Thereafter, spending could resume growing alongside the economy.
Sacks highlighted 172,000 January private-sector jobs versus expectations near 70,000, alongside 42,000 fewer government jobs and 4.3% unemployment. Since Trump’s second term began, he cited 615,000 added private-sector jobs and more than 300,000 fewer federal employees.
Four hyperscalers alone were expected to spend $600 billion in CapEx, which Sacks estimated as a roughly 2% GDP tailwind before software and productivity returns. “We are at the beginning of an economic boom,” he argued, potentially a “new golden age.”
11. Debt may be relative even when purchasing power is not
Chamath’s 300-year framing was that major economies’ debt-to-GDP ratios generally climb together, interrupted by wars that initially raise debt and later reduce the ratio. If every country reaches 200-300%, relative positioning may change little.
He therefore said debt-to-GDP “mostly” does not matter in isolation; decoupling matters. Inflation, import costs, exports, earnings, and household wealth remain concrete consequences even when sovereign leverage rises in unison.
Political appetite offered little comfort: even a conservative Congress had not enacted much of the waste, fraud, and abuse cuts identified by the White House and DOJ. Chamath expects governments to remain “addicted to spending” absent an extraordinary external shock.
His portfolio response was to own “real, durable assets,” with gold likely benefiting as confidence in dollar-denominated wealth erodes. Sacks added that a tighter Fed balance sheet under Kevin Warsh could lift Treasury yields and increase pressure on Congress and the administration through rising interest expense.
12. Wage policy and immigration enforcement exposed the sharpest split
Jason cited roughly 7 million job openings, unemployment near lifetime lows, and labor-force participation around 62% versus a Clinton-era peak near 68%. Geographic and skill mismatches matter, but he argued many vacancies simply pay too little.
He floated a populist Trump increase from the $7 federal minimum wage by $1-2 annually for three years. Friedberg’s rebuttal was that a floor above someone’s labor value makes hiring that worker illegal, raises unemployment, and accelerates automation—otherwise, “why wouldn’t you make the minimum wage $100 an hour?”
Jason countered with Seattle, San Francisco, New York, Los Angeles, Australia, and Scandinavia: he claimed wages and prices may rise 10-20%, yet populations become happier. He accepted automation risk but viewed a higher floor as a potential antidote to socialist politics.
On immigration, Jason targeted employers in construction and leisure or hospitality, which he said employ 2.5 million illegal aliens. Sacks and Chamath pressed him on surveillance, benefits, and criminal removals; Jason maintained that employer-focused surveillance, pay-stub checks, and fines are separate from criminal enforcement, with jobs the dominant migration incentive.
13. Ferrari’s EV debuts as human driving becomes a luxury
Ferrari’s first EV was described with more than 1,000 horsepower, four motors, zero-to-60 under 2.5 seconds, 330 miles of range, and a 5,100-pound weight versus roughly 3,000 for the F40. Launch is expected in May 2026.
Jony Ive and Marc Newson were involved in an interior mixing Apple-like glass and screens with tactile, memory-based controls. Sacks liked that compromise and its physical sounds; he disliked the projected exterior’s Corvette, Trans Am, Model 3, and hatchback cues, while acknowledging it was not final.
Chamath trusted Ferrari CEO Benedetto Vigna to deliver the “Ferrari experience,” but said autonomy is “racing against time” against car culture. FSD, Waymo, safety, and insurance economics could confine human driving to “smaller and smaller places and less and less often,” leaving Ferrari as a paid experience rather than ordinary transportation.