All-In PodcastBiggest LBO Ever, SPAC 2.0, Open Source AI Models, State AI Regulation Frenzy
CHAPTERS
- 0:00 – 8:40
Opening Banter, Hosts, and Pivot to EA Mega-Deal
The episode opens with lighthearted banter about generals’ fitness tests, pushup contests, and inside jokes among the hosts before quickly pivoting to the news that Electronic Arts is being taken private in the largest LBO ever. Jason frames the basic deal terms and investor lineup, setting the stage for a deeper discussion on gaming, sovereign capital, and private equity.
- •Hosts reintroduce themselves with joking nicknames and riff on military fitness tests and personal fitness.
- •Jason announces EA’s $55B take-private at $210/share (25% premium), led by Saudi PIF, Silver Lake, and Jared Kushner’s Affinity.
- •Context: PIF’s broader gaming and tech investments (Lucid, LIV Golf, Uber, Newcastle United, Savvy Games, Scopely, Niantic, Nintendo, Take-Two).
- •EA’s history: founded in 1982 at Sequoia’s office; major IPs like Madden, The Sims, Need for Speed.
- •Transition to Chamath for a bull vs bear breakdown of the transaction.
- 8:40 – 30:00
Bull and Bear Case for EA’s Take-Private and AI’s Role in Gaming
Chamath argues EA’s privatization is a smart, long-term play to restructure the business, embrace AI, and escape console gatekeepers. Friedberg expands on why AI will disproportionately enhance video games compared to social or traditional media, and why Saudi Arabia’s gaming spree is a coherent macro bet on future leisure time.
- •Bull case: gaming as an ‘anchor pillar’ of internet usage with ~3B daily active players; EA as an 800-pound gorilla.
- •Strategic goal: use private ownership to clean up OpEx, adopt AI tooling, and develop distribution beyond Xbox/PlayStation amid subscription price hikes.
- •Bear case: AI tools massively increase the number of games and enable distribution via social platforms, potentially eroding incumbent IP advantage.
- •Friedberg’s thesis: AI yields more engagement in gaming than social/traditional media via interactive, adaptive experiences (e.g., Fortnite’s AI bots for new players).
- •Saudi PIF’s gaming strategy: Savvy Games acquisitions (Scopely, Niantic), stakes in Nintendo, Take-Two, Activision Blizzard; positioning EA as platform.
- •Macro bet: AI-driven productivity could increase free time; entertainment demand and gaming’s share of it likely grow significantly by 2030.
- 30:00 – 41:40
Private Equity’s Boom, Overcrowding, and the Limits of 60/40
The conversation zooms out to private equity’s explosive growth and why Chamath thinks the asset class is structurally challenged. They trace how zero interest rates, the shift from 60/40 portfolios, and leverage fueled returns, but how too much capital and too many mediocre managers now threaten future performance.
- •Private equity AUM has surged to ~$5T, roughly tripling since 2015.
- •Historical context: 60/40 (bonds/equities) was the standard for risk-adjusted returns until ZIRP pushed capital into ‘alternatives’—VC, PE, hedge funds.
- •With near-zero rates, PE could borrow cheaply, juice returns, and outpace VC/hedge funds, attracting a flood of fast followers and laggards.
- •Consequence: intense competition, overpaying for assets, and under-management, leading to return compression.
- •Chamath’s key metric: DPI (distributions to paid-in capital) vs IRR; many PE funds show weak distributions over last 4–5 years.
- •Predicted outcome: capital concentrates in top performers (e.g., Silver Lake), while overflow has already begun inflating a new bubble in private credit.
- 41:40 – 56:40
IPO Market Dysfunction and the Evolution to SPAC 2.0 / 3.0
Chamath dissects the failures of traditional IPOs and direct listings from his experience with Slack and Coinbase, then explains how SPACs can evolve into a cheaper, more competitive alternative. He outlines his new compensation structure, why he wants minimal retail participation, and envisions a future ‘SPAC 3.0’ with fully pre-wired common-stock capital.
- •Traditional IPO issues: banks charge 6–8% fees, allocate underpriced shares to favored clients, generate day-one pops and post-pop drift.
- •Direct listings: no underwriting fee but first trade is often the peak, followed by a straight decline (as with Slack and Coinbase).
- •SPAC 1.0 recap: proved the vehicle’s viability but had misaligned incentives (founder shares, warrants) and mixed outcomes—some big winners (SoFi, MP Materials), others speculative or premature (Desktop Metal).
- •SPAC 2.0 (American Exceptionalism): no founder shares, no warrants, sponsor only earns if stock is up 50%, then 75%, then 100%.
- •Target profile: more mature, resilient-revenue businesses vs highly speculative, early-stage tech; heavy institutional investor base (~98.7% of capital).
- •Retail warning: both Chamath and Jason urge individual investors to largely avoid SPACs or cap exposure at <1%, stressing venture-like risk profiles.
- •SPAC 3.0 vision: sponsor pre-wires $1–3B+ of flexible, common-stock capital for a ‘pre-baked IPO’ with no conversion risk and fair pricing.
- 56:40 – 1:09:10
AI-Powered Operational Turnarounds and the Limits of Traditional Private Equity
Friedberg highlights AI as a transformative lever for traditional industries, citing Josh Kushner’s roll-up of CPA firms. Chamath contrasts that promise with his on-the-ground experience trying to sell AI transformation into PE portfolios, arguing that misaligned incentives and mediocre management teams make change difficult unless ownership is highly concentrated.
- •Example: Josh Kushner’s Thrive Capital rolling up CPA firms at EBITDA multiples and then applying AI (e.g., TaxGPT-like tools) to reinvent workflows.
- •Friedberg’s view: public investors can still ‘beat the market’ by identifying legacy companies with real AI execution capability and leadership.
- •Chamath’s experience with major PE firms: GPs love the pitch, but portfolio companies are often ‘B and C companies run by C and D folks,’ resistant or unable to adopt AI.
- •Result: 80Ninety’s rapidly scaling AI business sees almost no revenue from PE portfolios despite strong ROI case studies.
- •He argues AI transformation works best in owner-operated models (deca-billionaires or sovereigns like Saudis) and in public companies whose CEOs fear disruption and job loss.
- •Implication: a new AI-native ‘private equity 2.0’—like Josh’s model—may outperform traditional PE in executing AI-driven value creation.
- 1:09:10 – 1:26:40
New AI Media Apps, Personalized Content, and the Future of Shared Culture
The besties briefly explore consumer AI video apps (OpenAI’s Sora-based ‘SLOP’ and Meta’s Vibes) as early experiments in user-generated, AI-driven media. Friedberg argues we’re at the beginning of new media forms that will mix shared cultural context with individualized experiences, even as traditional mass culture fragments.
- •OpenAI’s Sora app: opt-in persona usage (e.g., Sam Altman) to navigate IP rights while letting people create videos with notable faces; default ‘opt-out’ stance for training data likely invites lawsuits.
- •Meta’s Vibes: seen more as a data play and early experiment than a fully formed TikTok competitor.
- •Friedberg’s thesis: future media will shift from centrally produced/broadly consumed to more distributed, personalized production layered on shared cultural memes (e.g., everyone making different ‘Sam Altman’ clips).
- •Concern: erosion of big, shared cultural moments (e.g., Tarantino films, Sopranos episodes) in favor of fragmented, personalized feeds.
- •Jason notes a backlash: people organize group screenings (e.g., buying 20 seats for a Paul Thomas Anderson film) to recreate shared cultural experiences.
- 1:26:40 – 1:36:40
DeepSeek, Kimi, and the Rise of Chinese Open-Source AI Models
The discussion turns to DeepSeek 3.2 and other Chinese open-source models that are dramatically undercutting Western API prices. Chamath explains how US companies are actually using these models via domestic providers like Groq, and Sacks frames the strategic tension: open source as a check on Big Tech versus the fact that leading open models now mostly originate in China.
- •DeepSeek 3.2 EXP: new architecture with DeepSeek Sparse Attention (DSA) significantly lowers training and inference costs, with API pricing around $0.28–0.42 per million tokens vs ~$3.15 for Claude.
- •EightyNinety’s practice: routes many workloads through Kimi on Groq, finding it more performant and far cheaper than OpenAI/Anthropic for certain use cases.
- •Operational challenge: swapping models isn’t plug-and-play; fine-tuning and prompt engineering are model-specific, so each leapfrog forces painful refactors.
- •Architecture: Groq forks open-source models (DeepSeek, Kimi, LLaMA) and runs them on US soil with US hardware and staff; no data automatically ‘phones home’ to China.
- •Security concerns: the main theoretical risk is backdoors in the model weights or code, but Sacks and Chamath note that major players aggressively red-team each other’s models, and so far no such exploit has been surfaced.
- •Strategic landscape: Meta’s LLaMA 4 underwhelmed and Meta may pull back from open source; OpenAI’s open models are far from frontier; Chinese firms (DeepSeek, Moonshot/Kimi, Alibaba’s Qwen) are leading in open-source performance.
- 1:36:40 – 1:43:20
AI’s Energy Crunch, Data Center Backlash, and Possible Off-Ramps
Chamath raises alarms about AI’s looming pressure on the electrical grid and prices, citing local resistance to data centers and an energy CEO’s forecast of doubling rates. Sacks outlines a phased response—short-term grid optimization, medium-term gas, and longer-term nuclear—while warning about public backlash if AI is blamed for soaring power bills.
- •Local pushback: Indianapolis residents blocked a $1B Google data center over fears of electricity cost inflation; Virginia’s ‘data center alley’ already consumes ~40% of state power.
- •Energy CEO’s projection: without solutions, electricity prices could double in five years due to AI/data center demand.
- •PR risk: ‘Big Tech AI doubled my power bill’ could be politically toxic and accelerate regulatory and populist backlash.
- •Off-ramp 1: cross-subsidies where hyperscalers pay premium rates while residential rates stay flat/decline, funded by big tech’s large free cash flows.
- •Off-ramp 2: home battery deployments near data centers to buffer load and insulate households from volatility and price spikes.
- •Sacks’s energy timeline: within ~5 years, squeeze 80 GW extra from existing grid by shedding ~40 peak hours to backup generators; medium-term ramp gas capacity (despite turbine backlogs); long-term nuclear build-out as a core AI power source.
- 1:43:20 – 1:48:20
Open Source vs Closed Source and Why AI Is Not ‘Digital Nuclear Weapons’
Sacks contrasts early analogies of AI to nuclear weapons with the reality that AI is becoming a ubiquitous consumer and enterprise tool. They argue that because ‘everyone will have AI,’ policy needs to accept decentralization and focus on specific harms rather than try to lock AI into a few controlled entities.
- •Old paradigm: ‘GPUs are plutonium’ and AI as ‘digital nukes’ leading to calls for extreme centralization and moratoriums.
- •New reality: AI is a mass-market tool—consumers and businesses will all want personalized AI agents, often running on their own hardware (phones, on-prem, distributed systems).
- •Decentralization: multiple US closed-source providers, several large Chinese open-source models, plus dozens of startup and vertical models ensure no single-point control.
- •Analogy: unlike nuclear technology, which everyone wants to prevent from spreading, AI is inherently proliferating because its primary use is productive, not destructive.
- •Policy implication: regulators should pivot from trying to ‘stop proliferation’ to managing well-defined harms (fraud, cybercrime, discrimination) within existing legal frameworks.
- 1:48:20
State AI Regulation Frenzy: California, Colorado, and the Case for Federal Preemption
The episode closes with a detailed critique of state-level AI bills, highlighting how vague ‘safety’ standards and disparate-impact rules could force ideological outputs and crush innovation. Sacks and Friedberg argue that only a single federal standard can preserve a unified US market and prevent ‘woke AI’ mandates from blue states from effectively setting rules for the nation.
- •Regulatory explosion: all 50 states have introduced AI bills; over 1,000 bills proposed and 118 enacted; strong momentum especially in blue states.
- •California: SB-1047 (vetoed by Newsom) was extremely intrusive; SB-53, now moving, demands frontier models publish safety frameworks and incident reports around nebulous ‘catastrophic harms’ (cyber, bio, ‘model autonomy’).
- •Colorado (SB-24-205): bans ‘algorithmic discrimination’ defined as unlawful differential treatment or disparate impact across protected classes; holds both developers and deployers liable.
- •Practical effect: even race-neutral AI tools (e.g., mortgage decisioning using credit scores) could trigger liability if outputs have disparate impacts; likely pushes developers to embed DEI layers to immunize against lawsuits.
- •Friedberg’s critique: these are oversight/control statutes, not filling genuine legal gaps—existing civil/criminal law already covers harms like cyberattacks, fraud, discrimination, and physical injury.
- •Chamath’s analogy: just two conflicting emissions regimes (US vs California) already distorted the auto industry; 50 different AI regimes would ‘render the category impotent’ economically.
- •Sacks’s federalism argument: under the Commerce Clause, a seamless national market is a key US advantage; a patchwork of 50 AI regimes would make US AI deployment resemble the balkanized EU market and hand a strategic edge to China.
- •Politics: some Republicans resist preemption out of anger at Big Tech’s past censorship, but Sacks argues that failure to preempt will let blue states hard-wire woke AI and DEI outputs into national tools.
- •Trump factor: Sacks notes Trump’s July 23 AI speech explicitly calls for a single national AI standard (similar to his stance on vehicle emission preemption), suggesting eventual federal action is likely.
