SpaceX's $2T Case, Nvidia's Shock Selloff, America Turns on AI, Trump Pulls AI Order, Bond Crisis?

2026-05-22 - 102 min - source - Read full transcript

Jason Calacanis (host)Chamath Palihapitiya (host)David Friedberg (host)Gavin Baker

Key insights

Anthropic hiring Andrej Karpathy to lead a new pre-training team focused on recursive self-improvement could compound gains on top of raw compute scaling.

Chamath and Gavin Baker frame it as a potential 'new Moore's law' where models improve themselves during training, on top of an already roughly 10x/year improvement rate. Both flag recursive self-improvement and continual learning (models learning from experience the way humans do) as the two remaining unsolved frontiers in AI progress.

ai-recursive-self-improvement

Cost per token could fall sharply by re-architecting toward networks of smaller, specialized models instead of one giant model.

Friedberg points to early MIT research and existing verticalized small-model deployments (e.g. the company Abacus) as evidence that a single architectural breakthrough could halve inference cost - an efficiency unlock the hosts see as separate from, and possibly as impactful as, raw scale increases.

ai-recursive-self-improvement

The public AI backlash (booed commencement speeches, AI becoming a 'four-letter word') is driven by a perceived power imbalance, not just job-loss fear.

Friedberg argues AI concentrates leverage and profit among a small group faster than the technology diffuses to the public, comparing the psychological shock to the Copernican revolution's dethroning of human centrality - 'it kind of shifts and f***s with the ego of the human, it's almost anti-humanist.' Gavin and Chamath add that some anti-AI, anti-datacenter sentiment may be amplified by state-actor propaganda, drawing an explicit parallel to Cold War-era KGB disinformation.

ai-backlash-and-regulation

Tech CEOs are actively worsening the AI PR crisis through careless layoff messaging.

Chamath singles out Cloudflare CEO Matthew Prince's memo labeling laid-off staff 'measurers' as a self-inflicted disaster, and criticizes Zuckerberg for describing internal engineers' work being used to train coding models in the same breath as cutting roughly 8,000 jobs. His prescription, echoing Palantir CTO Shyam Sankar, is to stop quoting AI-lab CEOs and instead surface frontline-worker stories (factory workers, ICU nurses).

ai-backlash-and-regulation

A Trump executive order requiring federal review of frontier AI models was drafted, leaked, and pulled hours before its announcement; Chamath counters with a US-China 'KYC' proposal instead of unilateral US regulation.

The hosts report the EO would have applied government supervision/insight-review to frontier models (not just LLMs) before release, and that Trump personally objected to the oversight provisions. Chamath argues neither pure self-regulation/courts nor a US-only regime is sufficient once China's frontier capability is under nine months behind the US; he proposes a shared bioweapon/terrorism screening standard, akin to FDA contaminant testing, that both countries already effectively do internally. Gavin is open to a China-inclusive framework but skeptical of the US regulating alone.

ai-backlash-and-regulation

SpaceX's S-1 reveals 'Elon Web Services' (EWS) - Anthropic paying SpaceX $1.25B/month, $45B over three years, to rent Colossus compute - as a major new profit engine alongside Starlink.

The AI segment did $3.2B revenue (2x YoY) but a $6.4B operating loss before this deal; the Anthropic contract alone roughly quadruples that segment's run-rate revenue, though either party can cancel with 90 days' notice. Combined with the pending Cursor acquisition (another $2-3B), the hosts argue SpaceX is becoming as much an AI-compute company as a rocket/satellite company ahead of its roughly $1.75-2T IPO (Polymarket gave 71% odds of a first-day close above $2T).

spacex-ipo

Cursor's Composer 2.5 model, trained on Colossus 2 with Cursor's proprietary coding data, became Pareto-dominant within three to four weeks.

Gavin Baker calls this the most important non-S-1 data point of the week: Composer 2.5 uses the same base model (Kimi K2.5) as its predecessor, but a short burst of reinforcement learning on Cursor's coding-token dataset (reportedly larger than all public-internet coding data) pushed it clearly outside the prior Pareto frontier - evidence of how fast proprietary data plus compute access can move a lagging lab (xAI/Grok) toward the frontier.

spacex-ipo

Nvidia's Q1 blowout ($81.6B revenue, +85% YoY, 75% gross margin) came with an explicit buyback/dividend signal aimed at countering a 'losing share to TPUs' narrative.

Nvidia raised its dividend 25x (1 cent to 25 cents/share) and added $80B in buybacks on top of a prior $100B. Gavin argues the share-loss story is misleading: Broadcom's comparable AI-semiconductor segment (stripping out China, which Nvidia has and Broadcom doesn't) grows slower than Nvidia's, and competing accelerators (TPU, Trainium) are conspicuously absent from public benchmarks like MLPerf and SemiAnalysis InferenceMAX because, in his view, they would lose.

ai-infrastructure-economics

AI-sector valuations are 'cross-sectionally inefficient' - chip, memory, power, and cooling stocks are pricing incompatible futures.

Gavin Baker's framework: memory makers trade at 3-5x P/E, Nvidia at a low-to-mid-teens multiple, while power, cooling, and optical names are priced for a much larger buildout. These valuations can't all be correct simultaneously - whichever segment's multiple turns out right implies a large repricing, up or down, in the others.

ai-infrastructure-economics

The GPU-depreciation bear case popularized by Michael Burry was undercut by CoreWeave's results and Nvidia's demonstrated multi-year chip lifespan.

Burry argued neoclouds were overstating profits by amortizing GPUs over 4-6 years against a 'true' 2-year useful life. Gavin counters that lenders are now financing GPUs on 6-7 year asset-backed loans at falling rates, and CoreWeave's CEO reports customers signing six-year contracts and expecting years 7-9 of useful life - which the hosts credit with 'saving' the neocloud financing model this quarter.

ai-infrastructure-economics

Inflation and bond-market signals are flashing red even as AI-driven earnings stay strong, and the Strait of Hormuz closure is relatively good for the US despite being bad globally.

Prediction markets put May CPI at 99% odds of 4.2%+, with professional forecasters flagging a possible 6% print; the 10-year yield hit 4.6%, Japan's 30-year hit a record 5.1%, and global debt-to-GDP sits at 310%. Separately, the 12-week-old Strait of Hormuz closure is pushing LNG prices up 100-200% outside the US while US natural gas costs fall, which Chamath and Gavin frame as a forcing function that widens America's relative advantage given its energy and food self-sufficiency - even as they concede rising rates alongside rising inflation is 'never good.'

macro-bond-and-geopolitics

The Trump-Xi summit produced no 'grand deal,' only incremental trade gestures, while China and Russia visibly deepened ties immediately afterward.

Friedberg reports the outcomes were limited to some aircraft and agricultural purchases plus H100/H200 chip sales to Chinese firms, not the broader tension-reducing deal some had hoped for. Chamath speculates the real substance was an unannounced geopolitical negotiation covering Taiwan, Venezuela, and Iran, guessing Trump may be signaling a long-term (20-100 year) handoff on Taiwan in exchange for stability during his own administration, leveraging US control over Venezuelan, American, and Brazilian oil supply as points of pressure.

macro-bond-and-geopolitics

Media referenced

Invest Like the Best (Patrick O'Shaughnessy interview with Gavin Baker) - podcast - Referenced for Gavin's point that chip-company and DRAM-company revenue multiples cannot both be correct at once - one segment's pricing is wrong
60 Minutes special on UK self-driving bans - show - Cited as evidence that the UK, Boston, and New York are leaning toward banning or restricting self-driving to protect driver jobs
a16z essay on Flock Safety - article - Cited approvingly for the argument that AI-camera crime detection makes crime effectively a policy choice, not an unsolvable problem
Wall Street Journal report on Anthropic's profitability - article - Cited as the source for Anthropic reportedly turning EBIT-positive across its language-model business

Companies

SpaceX - Filed its S-1 this week targeting a $75B raise at a $1.75T valuation, the largest IPO ever attempted; listing expected mid-June under ticker SPCX
Starlink - SpaceX's existing cash engine: $11.4B revenue last year, 50% growth, $4.4B operating income, over 10 million subscribers
Anthropic - Hiring Andrej Karpathy for recursive self-improvement research; also paying SpaceX $1.25B/month ($45B over three years) to rent Colossus compute - the 'Elon Web Services' deal
Nvidia - Reported blowout Q1 earnings ($81.6B revenue, +85% YoY), raised its dividend 25x, and announced an additional $80B buyback
OpenAI - Referenced throughout as a frontier lab and GPU customer; also named in the wrongful-death lawsuit discussion over a user's suicide
Tesla - Discussed as a likely future merger candidate with SpaceX, which the hosts argue would create the world's fourth-largest company by market cap
xAI - Was trailing Anthropic/OpenAI/Gemini until Elon granted Cursor access to Colossus compute and shipped the Grok Build agent harness
Cursor - Its Composer 2.5 coding model, trained on Colossus 2 using Cursor's proprietary coding-token dataset, became Pareto-dominant within three to four weeks; SpaceX/xAI reportedly acquiring Cursor for an additional $2-3B in revenue
Broadcom - Central to the 'Nvidia is losing share to TPUs' narrative Gavin Baker disputes, arguing Nvidia's comparable AI-semiconductor segment is growing faster once China revenue is stripped out
CoreWeave - Its CEO's six-year customer financing contracts are cited as the rebuttal to Michael Burry's GPU-depreciation bear case
Meta - Zuckerberg laid off roughly 8,000 employees around the same time he described using internal engineers' work to train coding models, which the hosts say worsened public AI anxiety
Cloudflare - CEO Matthew Prince's memo labeling laid-off staff 'measurers' is singled out by Chamath as a PR disaster that damaged the industry's messaging on AI-driven layoffs
Palantir - CTO Shyam Sankar's clip (via Fox News) is replayed for his argument that the industry should stop quoting AI-lab CEOs and start telling frontline-worker stories instead
Flock Safety - AI camera/gunshot-detection company discussed as a model for bottom-up, town-by-town crime-technology deployment with built-in privacy audit trails
Atreides Management - Gavin Baker's hedge fund, which manages 100+ positions with a 30-person team and is an investor in a space-semiconductor company he calls 'Exite Labs' (name unclear in source audio)

Techniques and frameworks

Pareto Frontier analysis - Used throughout to compare frontier AI models (Grok 4.3, Gemini 3.1 Pro, Cursor Composer 2.5) on a capability-vs-cost curve; only four companies currently have a model on the frontier
KYC for frontier AI models - Chamath's proposed US-China framework: both sides already vet training runs domestically, so a shared verification standard could prevent bioweapon/terrorism misuse without unilateral US regulation
DC-to-DC power delivery - Proposed re-architecture of data-center power (skipping DC-to-AC-to-DC conversion) that Chamath says needs a Jensen Huang/Elon Musk co-design partnership to cut power loss and cooling cost
Rapid reusability (Starship) - Distinguished from ordinary rocket reusability (SpaceX/Falcon, potentially Blue Origin/China): flying and landing the same vehicle multiple times per day, a prerequisite for Elon's Moon/Mars colony plans and orbital data centers

Summary

Gavin Baker of Atreides Management sits in for an absent David Sacks on an episode built around three big stories: Andrej Karpathy's move to Anthropic, SpaceX's blockbuster S-1 filing, and a Nvidia earnings beat set against a darkening macro backdrop. The hosts open on Karpathy joining Anthropic's new pre-training team to pursue recursive self-improvement - the idea that a model can meaningfully participate in improving its own training - which Chamath and Gavin frame as a potential second growth curve stacked on top of ordinary compute scaling. Friedberg adds a complementary efficiency angle: smaller, networked, verticalized models could cut cost-per-token sharply, an unlock he sees as at least as consequential as raw scale.

A long middle section wrestles with why the public has turned on AI, prompted by several booed commencement speeches. Friedberg's core argument is that AI concentrates economic leverage in a small group faster than the technology diffuses to everyone else, producing a psychological shock he compares to the Copernican revolution's blow to human centrality. Chamath and Gavin push a parallel, more cynical read: some of the anti-AI, anti-datacenter sentiment may be amplified by foreign (implicitly Chinese) information operations, echoing Cold War-style disinformation. The group agrees the industry's own CEOs aren't helping - Chamath delivers an extended, profane takedown of Cloudflare's Matthew Prince for a layoff memo that branded departing staff "measurers," and criticizes Zuckerberg for tying internal engineers' contributions to model training in the same breath as cutting thousands of jobs. This backlash discussion folds into news that a Trump executive order requiring federal review of frontier AI models was drafted, leaked, and then abruptly pulled; Chamath uses the moment to pitch a US-China "KYC" framework for frontier models as an alternative to either no regulation or unilateral US oversight, while Gavin is more skeptical of the US acting alone.

The back half pivots to SpaceX's S-1, filed at a $1.75 trillion target valuation ahead of a planned June listing. Beyond Starlink's now-familiar growth, the standout revelation is "Elon Web Services" - Anthropic committing $1.25 billion per month, $45 billion over three years, to rent SpaceX's Colossus compute clusters. Gavin treats this, plus Cursor's newly Pareto-dominant Composer 2.5 model (trained on Colossus 2 using Cursor's proprietary coding data), as evidence that access to compute and proprietary data can move a lagging lab to the frontier in a matter of weeks. Chamath builds out a full underwriting case for SpaceX at $2 trillion: not just a satellite-internet business but foundational AI and communications infrastructure, further boosted by speculative future moves like a Tesla merger and orbital data centers.

Nvidia's Q1 results get a shorter but pointed treatment: $81.6 billion in revenue, an 85% YoY beat, a 25x dividend hike, and $80 billion in new buybacks. Gavin uses the moment to argue two things: first, that the "Nvidia is losing share to TPUs" narrative doesn't hold up against a fair segment comparison with Broadcom, and second, that AI infrastructure valuations across chips, memory, power, and cooling are "cross-sectionally inefficient" - they can't all be pricing the correct future simultaneously. He also credits CoreWeave's results and financing terms with quietly rebutting Michael Burry's GPU-depreciation bear thesis.

The episode closes on macro and geopolitics. Inflation and bond signals are genuinely troubling - a possible 6% CPI print, a 4.6% 10-year yield, Japan's 30-year at a record 5.1%, and 310% global debt-to-GDP - prompting Friedberg's self-described "Dr. Doom" riff about an inevitable, gravity-driven unwind. Chamath and Gavin counter that the US remains relatively best positioned given its energy self-sufficiency, and note that the ongoing Strait of Hormuz closure, while bad for everyone, disproportionately hurts Europe and Asia. They close by dissecting the Trump-Xi summit, which produced modest trade gestures rather than a grand bargain, with Chamath speculating that the real content was an unannounced negotiation over Taiwan, Venezuela, and Iran conducted behind closed doors.

Notable Quotes

"What is the point of being concerned when you have ridden the roller coaster to the top and it is beginning its descent? I don't know what there is to be concerned about. The force of gravity is inevitable. The roller coaster will roll down. We will throw our hands in the air and we will scream wee as we go for the ride." - David Friedberg

"Well, I think what's important about Elon Web Services does make me laugh. But $15 billion, that means the AI business right there is going to quadruple." - Gavin Baker

"I still don't like the fact that Tesla's over here. And as I've told you, that will get merged in... That thing will look very cheap, I think, in a few years... I would say we're at the beginning of the beginning." - Chamath Palihapitiya

"There's something about AI that's very like not human-centric. And it kind of shifts and f***s with the ego of the human, it's almost anti-humanist. And I think that that's like a deep psychological current for a lot of people and their disdain for this technology." - David Friedberg

"Now when they try to get a different job, they're like, oh, you're one of the Cloudflare measurers. How does that help anybody? It didn't need to be done this way." - Chamath Palihapitiya