SpaceX’s $2T Case, Nvidia’s Shock Selloff, America Turns on AI, Trump Pulls AI Order, Bond Crisis?
Summary
- SpaceX’s filing supplied a plausible $2 trillion case, but the near-term engine is terrestrial compute rather than orbital data centers. The proposed $75 billion raise at a $1.75 trillion valuation rests on Starlink’s $11.4 billion revenue and $4.4 billion operating income, plus Anthropic’s cancellable $1.25 billion-a-month Colossus contract. Friedberg described the basic internet-infrastructure case; Chamath’s underwriting emphasized a “capital moat” accelerating technology, execution, and learning, with terrestrial data centers alone potentially reaching $100-$200 billion of revenue by 2030-32.
- Nvidia’s $81.6 billion quarter strengthened the fundamentals even as its valuation lagged richer AI-adjacent trades. Revenue rose 85% year-over-year and 20% sequentially, with $58 billion of net income, $48 billion of free cash flow, and 75% gross margins; Gavin argued Nvidia may also be growing faster than Broadcom in comparable Western AI infrastructure. The unanswered bear case is benchmarking: rival ASICs are not entering MLPerf or comparable tests, leaving the share-loss argument “fighting shadows.”
- Anthropic’s profitability is beginning to answer the AI-ROI objection, while Cursor’s proprietary coding data strengthens the model-development case. Gavin estimated OpenAI and Anthropic together at roughly $100 billion of ARR with approximately 80% inference gross margins, and saw $200-$400 billion of aggregate LLM ARR as plausible by year-end. Adding Andrej Karpathy to pursue recursive self-improvement and continual learning could put model development on Chamath’s combination of “overdrive and autopilot.”
- The AI backlash is being driven as much by concentrated gains and careless corporate messaging as by the technology itself. Friedberg argued that people see a small group collecting outsized returns before benefits diffuse, while Jason highlighted Cloudflare cutting more than 20% of staff and Meta laying off 8,000 people while recording work on employees’ computers to train models. Chamath’s verdict on branding former workers “measurers”: tech CEOs “don’t understand the moment.”
- The panel rejected a unilateral AI slowdown while leaving room for narrow, internationally coordinated guardrails. Friedberg emphasized state competition and the risk of foreign efforts to slow technological progress; Chamath used postwar nuclear proliferation as an analogy for why the U.S. and China may need to preserve strategic balance. Kevin Weil gave the cautious non-answer on testing frontier models, while Gavin argued that courts, liability, and self-regulation already create incentives for responsible conduct and that government power is a “one-way ratchet.”
- Autonomy and public-safety AI will likely advance through safety economics and competition between jurisdictions. Gavin said human-driving mandates will invite wrongful-death suits and predicted that cities without Waymo or cybercabs will eventually feel “barbaric and unsafe,” while Chamath said policymakers should first ask warehouse workers whether jobs with 35%-40% churn are actually worth preserving. Friedberg framed Flock Safety and Las Vegas’s drone-enabled policing as evidence that “crime is now a choice”; he estimated $30-$40 million annually could make Las Vegas much safer.
- The macro backdrop is unequivocally harder, but the panel split between systemic alarm and selective ownership. Oil-driven inflation forecasts reached 4.2%-6%, the U.S. 10-year hit 4.6%, and Japan’s 30-year was cited inconsistently in the transcript as both 3.1% and 5.2%. Friedberg warned that 310% global debt-to-GDP could turn rising yields into a credit crisis. Chamath’s response was to own five or fewer businesses for a decade, while Gavin held that higher rates, extraordinary AI growth, and America’s relative energy advantage can all be true.
- U.S.-China talks produced commerce rather than a grand bargain, yet continued contact may itself reduce geopolitical tail risk. Friedberg saw no durable de-escalation, whereas Chamath suggested there may have been useful private alignment on the geopolitical “tic-tac-toe” and said that America and China talking is “only good.” Gavin supported selling depreciated Nvidia GPUs to China because it could discourage a rival, more power-hungry ecosystem and keep America’s platform embedded.
Deep dive
1. Karpathy could put recursive AI on “overdrive and autopilot”
Chamath placed Andrej Karpathy in the tradition of Google Fellows such as Jeff Dean: singular technical talents positioned “at the foot” of successive computing waves. Karpathy commercialized the “bitter lesson” at Tesla—including spending what Chamath thought might have been a quarter of his time labeling driving video—then helped found OpenAI and continued experimenting independently.
AutoResearch, his open-source system for running five-minute model-training experiments, drew more than 82,000 GitHub stars. At Anthropic, Chamath viewed that curiosity as a powerful force for recursive self-learning and potentially “an order-of-magnitude improvement on a yearly basis,” making model quality “absolutely parabolic.”
Gavin called recursive self-improvement and continual learning AI’s “two final frontiers.” The first lets a model influence training during a forward pass; the second lets it learn from experience as humans do. Combined, they “might pull the future forward in a very real way”—and could make 10X annual improvement look conservative.
2. Profitability is beginning to answer the AI-ROI objection
Gavin’s key data point was Anthropic becoming positive in its latest quarter, as reported in the episode’s Wall Street Journal reference. If Anthropic and OpenAI together are near $100 billion of ARR with roughly 80% gross margins on inference, “the returns are there,” even before counting hyperscaler gains from advertising and recommender systems.
Adding Gemini, Cursor, xAI, and open source, Gavin considered $200-$400 billion of LLM ARR by year-end plausible. That excludes some of AI infrastructure’s most profitable existing applications—better ad targeting, measurement, and recommendations at Google, Meta, and Amazon.
Friedberg saw another efficiency frontier in networks of smaller, specialized models. Re-architecting how models cooperate could produce answers with less energy than a monolith; one modest breakthrough that halves cost per token would be “a tremendous efficiency gain” and appears within reach.
3. AI’s PR crisis reflects concentrated power, fear, and human ego
Chamath rejected breathless coverage of every benchmark improvement and wanted attention shifted to end users: OpenAI, with human involvement, helping solve a decades-old math problem, and drug candidates and INDs that had been sitting on the shelf becoming viable enough to approach clinical trials. A later clip from Palantir CTO Shyam Sankar added examples of nurses gaining patient time and factories adding shifts. “Nobody wins if we become Luddites and go back in time.”
Gavin’s cynical reading of model-maker alarmism was incentive-driven: Dario Amodei benefits from boundary conditions that create a regulatory moat, leaving Anthropic “inside of the tent pissing out.” Gavin said fundraising scale and the volume of public warnings could be read alongside successive financing rounds because “everybody is trading their own book.”
Friedberg offered three deeper currents behind the backlash: AI initially concentrates leverage and wealth; foreign or state actors may amplify anti-technology sentiment; and machine intelligence dislodges humans from the psychological center, as heliocentrism once did. AI feels “almost anti-humanist,” even if that is fuel rather than the root cause.
Gavin’s counter-story was a father using LLM research to identify an existing safe drug for his daughter’s rare mutation. Her neuronal activity reportedly rose from roughly 30%-40% to 80%-90%, enabling a normal life; he was then using AI to tailor a possible complete cure and was reasonably confident he could have a cure in months. “We should tell those stories.”
4. Layoff messaging is training workers to fear AI
Jason argued that displacement can no longer be dismissed entirely as post-pandemic bloat. Cloudflare’s Matthew Prince said he cut more than 20% of staff despite record growth and cash flow, characterizing people who manage people and measure data as “measurers” while promising hiring elsewhere.
Meta’s message looked even more destabilizing: alongside 8,000 layoffs, Zuckerberg described using thousands of strong internal engineers to create tasks and tools that teach models to code. Jason also said Meta was putting recording software on every employee’s computer to study work and train models. Employees who had built AI tools to make themselves more efficient consequently perceived that they had helped “train your way out” of employment.
Chamath thought Prince’s memo “could not have written a worse memo”: reducing people to a label also puts “a scarlet letter on their backs” during the next job search. If layoffs are followed by larger buybacks, dividends, and cash piles, careless CEO missives will intensify resentment regardless of the underlying economics.
5. U.S.-China parity makes unilateral AI slowdown untenable
Friedberg framed advanced AI as part of a broader state-competition problem. He argued that foreign actors have historically used disinformation and other interventions to slow progress in rival countries, and that a world where China achieved materially better models and deployment than America would be unhealthy.
Chamath supplied the postwar nuclear-proliferation analogy: once strategic capability spreads, stopping development risks intolerable asymmetric power. He considered China being less than nine months behind potentially stabilizing because comparable capability can create détente between differently organized societies: the destination is both sides looking each other in the eye and saying, “All right, weapons down.”
The White House had reportedly planned, then canceled, an executive order involving oversight of advanced AI. Friedberg said he had read the proposal as applying specifically to frontier models; Chamath said it was broader than language models. Chamath favored narrow KYC-style controls—potentially coordinated with China—to keep models away from uncontrolled actors seeking biological weapons, rather than broad supervision of all AI development.
Kevin Weil gave the cautious non-answer on pre-release frontier testing: unilateral U.S. rules felt premature, while coordinated U.S.-China verification would be more palatable. Gavin then argued that courts, liability, and self-regulation already discipline model makers, while authority granted to government is “almost never taken back” and tends to become a “one-way ratchet.”
6. Local competition will decide autonomy, policing, and privacy
Jason proposed paced autonomous-vehicle rollouts or taxes on humanoid robots to fund retraining, noting that more than 10 million Americans drive professionally. Chamath’s pushback was to ask workers directly: Amazon warehouse churn of 35%-40%, rather than 3%-4%, suggests outsiders may be defending jobs the employees themselves do not want.
Safety could overwhelm job-preservation rules. Gavin said autonomous-vehicle companies will face wrongful-death suits when human driving remains mandatory despite a solution that could have a zero-death rate. He also predicted that cities without Waymo or cybercabs will eventually feel “barbaric and unsafe,” much as cities without Uber once felt intolerably inconvenient.
Flock Safety illustrated the jurisdictional model: towns choose AI-enabled license-plate cameras, while privacy can be constrained through rolling retention periods, no facial recognition, and audit trails. Friedberg’s stark formulation was “crime is now a choice,” with residents eventually voting both at the ballot box and with their feet.
The contrast was Cambridge reportedly switching off gunshot detection while Las Vegas deploys rooftop drones and rapid visual tracking. Friedberg estimated another $30-$40 million annually could make Las Vegas much safer—a de minimis investment beside the cost of crimes it prevents.
7. SpaceX’s filing combines a cash generator with large options
The proposed IPO would raise $75 billion at a $1.75 trillion valuation, with a likely June 12 listing under SPCX; Polymarket assigned a 71% probability to a first-day market capitalization above $2 trillion. It would be more than twice the size of Saudi Aramco’s cited $29 billion offering.
Starlink supplied the operating foundation: $11.4 billion of revenue, 50% growth, $4.4 billion of operating income, and more than 10 million subscribers. The broader space business generated about $4 billion, grew 17%, and lost $650 million operationally.
xAI produced $3.2 billion of revenue—more than double the previous year—but lost $6.4 billion. SpaceX spent $20 billion on capex, with more than 60% directed toward AI compute as xAI attempted to close the gap with Anthropic, OpenAI, and Gemini.
“Elon Web Services” changed the profile: Anthropic is paying $1.25 billion monthly for Colossus 1 and part of Colossus 2, a $45 billion three-year agreement. Either party can cancel with 90 days’ notice, but the current run rate effectively adds another Starlink-sized revenue stream.
8. Colossus gives SpaceX an execution advantage and Cursor a second life
Gavin emphasized data-center construction speed: the first facility reportedly took 122 days, the second 91, and the third 66. With an anchor off-taker—and likely more partners later—SpaceX can stamp out capacity, while Nvidia has reason to allocate scarce GPUs to whoever can quickly convert “electrons into tokens.”
Cursor’s Composer 2.5 was the evidence for what proprietary data plus concentrated compute can do. After roughly three or four weeks of reinforcement learning on Colossus 2, using the same base model as Composer 2—Kimi K2.5—it moved well beyond the existing Pareto frontier.
Gavin said Cursor allegedly possesses more coding tokens than exist on the public internet; a new base model incorporating Cursor’s data could produce extraordinary results. David Sacks added that RL’ing the Cursor model on the world’s biggest coherent compute cluster was another important step. Jason’s framing was that Cursor had been compute-starved, then improved almost immediately once Elon opened Colossus.
Chamath asserted that Cursor had already been acquired, though Jason noted it was absent from the filing. The cited $2-$3 billion of fast-growing revenue and proprietary coding data would strengthen both the commercial and model-development cases.
9. xAI now has both a frontier model and the missing harness
Gavin placed one build of Grok 4.3—a 500-billion-parameter model—on the frontier beside Google’s Gemini 3.1 Pro, OpenAI, and Anthropic. Being on the frontier matters, but xAI previously lacked the runtime layer that Claude Code and Codex provided competitors.
David Sacks said Grok Build fills that gap with more than a downloadable interface: it manages state, memory, and the working environment. His claim was that in an agentic system the harness is “essentially as important as the model,” and both increasingly need to be co-developed.
Composer 2.5 becoming Cursor’s most-selected model also mattered because benchmarks miss user preference. Its “vibes” were reportedly strong as well as its measured performance, suggesting that Colossus, Cursor data, Grok 4.3, and Grok Build are beginning to operate as a coherent stack.
10. SpaceX’s $2 trillion case is a compounding moat, not a single multiple
Friedberg’s basic underwriting started with roughly $18-$19 billion of prior-year revenue and an estimate of $25-$30 billion this year. He described Starlink as potentially the most important internet-infrastructure project since the internet itself, scaling toward hundreds of millions of users because it is useful and should become cheaper.
Chamath described Starlink as a grower at roughly GDP plus 10 to GDP plus 15 and as the underlying platform that allows other businesses to develop. He then added an AI business spanning both applications and compute capability.
Chamath’s central mechanism was circular: revenue creates operating leverage, which funds a “capital moat”; that accelerates the technology moat, then the execution and learning moat. He believes the flywheel is still at “the beginning of the beginning,” despite already moving quickly.
Chamath projected roughly $40-$45 billion of revenue the following year and potentially another doubling after that, making a 20-times-revenue valuation easier to underwrite if the proceeds fund businesses that consolidate the moat. Terrestrial data centers alone could generate $100-$200 billion by 2030-32, before giving any credit to orbit.
Chamath also reiterated his belief that Tesla will eventually be merged in, creating one corpus of physical movement, intelligence, infrastructure, and connectivity—with Elon’s “one more thing” capacity deserving a premium.
11. Rapidly reusable Starship is the gate to orbital compute
Gavin distinguished ordinary reusability from rapid reusability. Conventional reusability may require extensive refurbishment of the rocket, engines, and fairing, leaving 30-60 days between flights; Starship is designed to fly, land, and fly again multiple times per day, which is necessary for lunar bases, Mars settlement, and truly massive payload deployment.
That harder design goal explains the engineering risk. Gavin declined a precise prediction but guessed rapid reuse might arrive within one or two years, possibly sooner. Even a “fireball” would generate instrumentation and learning: “If you don’t fail, you’re not learning.”
Orbital compute already has a proof point: an Nvidia H100 is operating in space, and Karpathy reportedly used it both to train a model and run inference. Gavin noted that SpaceX can qualify inexpensive, non-radiation-hardened commercial chips through payload and rocket engineering rather than demanding bespoke space silicon; Excite Labs’ chips were cited as an example.
Gavin’s point estimate for meaningful orbital compute was the second half of 2028 through the first half of 2030. Friedberg’s broader case was strategic rather than financial: space-based communications and data centers could provide a “backup for civilization” if governments restrict speech, commerce, or information flow.
12. Nvidia’s quarter leaves the share-loss narrative unproven
Nvidia reported $81.6 billion of quarterly revenue, up 85% year-over-year and 20% sequentially, alongside $58 billion of net income, $48 billion of free cash flow, and 75% gross margins. At a $5.3 trillion market capitalization, the shares were up only 16% for the year.
Capital returns are becoming material: another $80 billion of buybacks followed $100 billion of buybacks completed at the start of 2023, together representing roughly 4% of the company. The dividend rose 25-fold from one cent to 25 cents per share, and management plans to return 50% of free cash flow.
Against Broadcom’s forecast of 143% growth in comparable AI-semiconductor revenue, Gavin believed Nvidia’s hyperscaler-plus-AI-cloud business—excluding China—may still be growing faster. Nvidia is also outgrowing hyperscaler capex, making a simple ASIC share-loss narrative difficult to reconcile with observed revenue.
Chamath framed Jensen Huang’s frustration as putting up record numbers without receiving credit. Gavin’s response was that TPU v7, Trainium, Inferentia, and other ASICs are not being submitted against GB300s in clean public benchmarks such as MLPerf. “You can’t fight shadows”; without comparable tests, the claim that custom silicon is winning remains unproven.
13. Longer GPU lives strengthen neocloud economics
The AI trade looks cross-sectionally inconsistent: memory makers at roughly three-to-five-times earnings and Nvidia at a low multiple coexist with much richer power, cooling, and optical names. If the latter valuations are right, Nvidia and memory have substantial upside; if the chip multiples are right, the surrounding infrastructure should underperform.
Nvidia also expects a $20 billion CPU business this year, making it one of the world’s largest CPU suppliers almost overnight. Because Nvidia works with every major lab, it can co-design domain-specific architectures internally—challenging the assumption that specialized workloads must migrate away from its platform.
Disaggregated inference extends old hardware’s usefulness. Groq or Cerebras accelerators can handle decode in front of older GPUs, potentially giving those GPUs 10-15 years of productive life rather than two and undercutting the bear case that four-to-six-year depreciation schedules overstate neocloud profits.
CoreWeave reportedly finances hardware at about 6% and signs six-year prepaid contracts. Its CEO, Michael Intrator, said customers would not sign those contracts if they believed the hardware lacked a six-year life and expected additional useful life in years 7, 8, and 9. Chamath called that financing advantage “profound”; Gavin said Nvidia had effectively “saved” the neocloud model by validating longer asset lives.
14. Rising yields reward selectivity while America retains relative advantages
The warning lights were oil, inflation, and duration: the Iran war had reached week 12 against initial expectations of four to six weeks; Polymarket put May inflation above 4.2% at 99%; professional forecasts reached 6% CPI; and the U.S. 10-year touched 4.6%.
Jason initially said Japan’s 30-year yield was 3.1%, the highest ever recorded. Later, Friedberg referred to “the 30-year at 5.2%” as a possible credit-crisis catalyst. The transcript does not resolve the discrepancy. Friedberg connected the broader moves to global debt at 310% of GDP: currency weakness, money printing, asset inflation, and leveraged carry trades could cascade into a credit crisis. “The force of gravity is inevitable.”
Chamath’s response was portfolio discipline, not macro timing: own five or fewer companies that represent the future, understand them deeply, and hold for ten years. “There are about four things that I stay on top of”—everything else adds information and emotional volatility without enough reward.
Gavin held three truths simultaneously: rising rates are concerning; AI fundamentals are unprecedented; and the closed Strait of Hormuz may relatively favor America. He pointed to U.S. food and energy self-sufficiency and said the dollar crisis was not around the corner. Chamath added that U.S. natural gas supports electricity while LNG inputs elsewhere were up 100%-200%, making the disruption worse for Europe and Asia than for America.
15. China talks yielded no grand bargain but may have reduced tail risk
Friedberg saw the 48-hour presidential and CEO visit as incremental rather than transformative: aircraft, agricultural commitments, and some chip commerce, but no durable partnership that decisively de-escalated tension. Xi’s subsequent meeting with Putin underscored that “the story continues.”
Chamath thought the surface announcements understated possible private progress. He suggested there may have been alignment on the geopolitical “tic-tac-toe”—a division of moves across contested regions that could help both countries—even though such understandings would not appear in a detailed press release.
Gavin favored selling China depreciated Nvidia GPUs. Keeping Chinese developers inside the American ecosystem could reduce incentives to construct an alternative that consumes more power and introduces optical networking earlier, making chip sales a potentially stabilizing path for preserving U.S. leadership.
Jason argued that the energy consequences of conflict could alter the Taiwan calculus, while Chamath concluded that America and China talking is “only good” and that a more stable world could emerge after the Iran situation resolves.