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

Source: Authorized transcript of All-In Podcast, episode 274. The opening states that Sacks is absent and introduces Gavin Baker as the guest; his firm is transcribed as “a Treaties Management,” which should be checked rather than silently corrected.

Speakers identifiable from the transcript: Jason, Chamath, Friedberg/“Freeberg,” Gavin Baker, plus inserted or recounted remarks from people including Shyam Sankar and Mark Zuckerberg. Because the transcript does not consistently attach names to speaker turns, attribution below is firm only where dialogue cues establish it. Otherwise, the speaker is marked [speaker unclear].

Transcript quality: The raw file contains 6,105 caption lines, with captions repeated in overlapping three-line waves. Those mechanical duplicates have been removed for this summary. The source also contains several likely transcription errors or phonetic renderings, including “Treaties Management,” “Kimmy K K.25,” “Kodaks,” “speed hour,” “Leo Ashan Brenner,” “Michael in Trader,” “Excite Labs,” “Kevin” when Gavin appears to be addressed, and other passages listed in Section D. None has been silently corrected.

Editorial status: This document records what the speakers said. Claims, estimates, allegations, predictions and anecdotes below are not independently verified or endorsed.

Jason introduces Andrej Karpathy’s reported move to Anthropic as one of the major technology stories of the week.

He describes Karpathy as:

• 39 years old.
• Already a major figure in AI.
• A founding member of OpenAI.
• Formerly a leader of Tesla’s self-driving work.
• The person who coined “vibe coding.”
• The creator of a recently released open-source project transcribed as “Auto Research.”
• The inspiration for another tool based on his principles for Claude Code.

Jason says “Auto Research” allows AI models to improve themselves by running five-minute experiments, claims the project accumulated more than 82,000 GitHub stars, and characterizes it as something Karpathy produced largely as a weekend project that inspired people to experiment with recursive LLM systems.

He says Karpathy is joining Anthropic to lead a new pre-training team focused on recursive self-improvement—in practical terms, using Claude or other AI systems to contribute to improving future AI systems.

Chamath argues the hire matters because certain unusually strong technical people have historically driven repeated technological waves.

He compares this with Google’s culture of elevating exceptional engineers as Google Fellows, naming Amit Singhal, Shreedhar Ramaswamy and Jeff Dean.

His reasoning is that people of this type do not merely execute a known roadmap; they repeatedly find themselves at the base of new technological waves and create things that were not obvious beforehand.

He says Karpathy has repeatedly occupied that role in AI.

Chamath argues that Karpathy was an early practitioner of the philosophy expressed in Richard Sutton’s “Bitter Lesson”: brute-force computation and scalable learning methods can eventually outperform handcrafted approaches.

He connects this to Karpathy’s Tesla work. Chamath recounts a personal anecdote—while explicitly saying he does not know whether Karpathy has discussed it publicly—that around 2016–2017 Karpathy told him he spent perhaps a quarter of his own time manually labeling Tesla video data.

Chamath presents that anecdote as evidence that Karpathy was willing to work directly on the underlying data rather than treating labeling as work beneath a senior researcher.

He describes Karpathy as unusually curious, independently inventive and impressive even while acting as a “free agent.”

Chamath therefore calls the Anthropic hire “really important.”

Chamath’s larger argument is that recursive self-learning could combine with rapidly increasing compute to create an acceleration loop.

His sequence is:

1. Models become capable of helping researchers improve models.
2. The systems therefore contribute increasingly to their own research and development.
3. More compute is applied to increasingly capable systems.
4. Improvements occur more rapidly.
5. This could create something resembling a new Moore’s Law based on model capability, not merely transistor density.

He speculates that an order-of-magnitude improvement each year may become possible.

If so, he expects model quality to become “absolutely parabolic.”

Gavin calls Anthropic’s recent success “extraordinary” and “undeniable.”

He says the Wall Street Journal had reported that Anthropic was “even positive” in the most recent quarter. The wording is unclear as to the precise profitability measure and needs verification.

He argues that this matters because a major criticism of the AI investment cycle has been that enormous infrastructure spending may be circular or economically unjustified.

His counterargument:

• If OpenAI and Anthropic together are at roughly $100 billion of ARR, as he estimates;
• and inference has gross margins around 80%;
• then there is already substantial economic return behind the infrastructure spending.

He then adds Gemini, Cursor, xAI and open-source models and says it is not hard for him to imagine $200 billion, $300 billion or $400 billion of annual recurring revenue across LLM-related activity by the end of the year.

He clarifies that this framing includes Google but intentionally excludes other major returns from AI infrastructure, such as recommendation systems, advertising targeting and advertising measurement at companies including Meta/Facebook, Google and Amazon.

His conclusion is that even a narrow definition focused on LLMs could show very strong infrastructure ROI, before counting other profitable GPU workloads.

Gavin identifies two capabilities as potentially the “final frontiers” for AI:

• Recursive self-improvement: a model, or another model, influences training while the system is being trained.
• Continual learning: a deployed model learns from new experiences in a manner more analogous to human learning.

He says continual learning remains unsolved.

If both become effective, Gavin suggests Chamath’s forecast of 10x annual capability improvement might eventually look conservative.

Together, he says, they could “pull the future forward in a very real way.”

Jason asserts that Anthropic currently has a meaningful lead.

He offers several rough estimates rather than a single figure:

• perhaps three to six months ahead generally;
• six to twelve months ahead of open source;
• potentially three, six or nine months ahead of particular competitors.

He then raises the point at which AI itself could contribute more to the construction of the next language model than the human engineers orchestrating it.

He calls this crossing the “recursive valley” or reaching “super recursiveness.” These are speculative concepts in the discussion, not established thresholds.

Friedberg says he does not know when a system will literally be able to take an entire model as context and produce its successor.

He argues that there are many possible architectural paths instead.

One possibility is to build much smaller models and arrange them in networks that cooperate on a task.

If such systems can collectively produce a token at lower energy or financial cost than a single very large model, they could materially change AI economics.

He therefore sees substantial opportunity in:

• re-architecting individual models;
• changing how models communicate;
• changing how multiple specialized models cooperate;
• reducing cost per token through architecture rather than merely adding compute.

A relatively small architectural breakthrough that halves cost per token would, in his view, be economically enormous.

He says early research papers—including one he recalls sharing from MIT—suggest significant room remains.

Jason adds that small, vertical language models are likely to become important and cites a company transcribed as Abacus as an example serving corporations.

Jason says Google had recently included Gemini Nano with Chrome.

He describes it as:

• approximately 4 GB locally;
• handling tasks such as proofreading, spelling and autocomplete;
• added without users receiving what he considered sufficient advance notice.

Jason initially calls this Google “covertly installing” the model.

Chamath immediately objects that “covert” is too strong.

Jason accepts the objection and reframes his criticism as installation “without telling people, without giving people a heads-up.”

Chamath argues that discussing each incremental model improvement breathlessly is increasingly unproductive.

He believes the important story is now what end users can accomplish that they could not previously accomplish.

He gives two classes of example:

• OpenAI, together with a human, reportedly contributing to solving a mathematical problem that had remained unsolved for decades.
• Drug candidates and INDs that had previously sat unused because they appeared unpromising but, according to him, are now approaching clinical trials with AI assistance.

His point is not simply that models benchmark better. It is that the systems are increasingly producing useful outcomes.

He argues that the way AI is framed publicly contributes to hostility. Describing it constantly as a separate, threatening force creates an “us versus them” or “boogeyman” narrative.

He says AI is becoming almost a “four-letter word” to some people because of the way it is presented.

Chamath says he does not believe Google is fundamentally a “shady” actor and assumes the Gemini Nano decision was driven primarily by user utility.

A joking interjection names Meta as a company that might behave differently; Chamath declines to name a company himself.

He argues that the hosts should present both the benefits and risks of AI rather than allow discussion to become technological Luddism.

Jason, while agreeing Google is not a bad actor, reiterates that reports of users being surprised by the size and background presence of the model had prompted legitimate privacy questions. One phrase in his conclusion is transcribed as “probably a speed hour” and is unclear.

After joking that Google is “attorney maxed,” Chamath argues that Americans involved in technology have a responsibility to explain AI’s positive possibilities.

He then raises a much more serious and unverified allegation: that it is beginning to “feel or seem” as though there may be a CCP-funded campaign opposing AI and data-center development in the United States.

He presents this explicitly as a possibility rather than established fact.

He says such an effort would make strategic sense from China’s perspective but would not serve American interests.

His conclusion is that people involved in US technology should actively communicate AI’s potential benefits rather than leave the public narrative to opponents.

Jason asks whether some AI-industry leaders themselves are contributing to public fear, mentioning Dario’s warnings about employment and CEOs attributing layoffs to AI.

Chamath argues that every CEO is responding to incentives.

He specifically theorizes that it makes strategic sense for Dario to promote conditions that could create a regulatory moat around established frontier-model companies.

His logic is that once a company is sufficiently large and compliant, costly regulation can protect it from smaller entrants.

He says the intensity of alarming rhetoric can be correlated with successive fundraising rounds and increased company scale.

Chamath calls this a rational and clever business strategy, not necessarily evidence that the underlying warnings are false.

His broader advice is that observers should stop treating individual CEOs as neutral narrators and instead understand the economic incentives behind their statements.

Chamath cites Shyam Sankar of Palantir and recounts a message he says Sankar delivered on television:

“Stop breathlessly asking these model makers what they think. Go to the end user and ask the person in the factory that's using the model and ask him what he or she thinks. Ask what the doctor thinks. Ask what the scientist think and start to tell those stories. That's what we should be talking about.”

Chamath treats this as the correct shift in AI coverage: from predictions by model-company executives to observed effects on workers, clinicians, scientists and other users.

Gavin agrees with the emphasis on real users.

He recounts a sensitive medical anecdote from an event approximately ten days earlier.

According to Gavin:

• A hedge-fund manager’s daughter had been born with a very rare genetic mutation.
• Her neurons were not firing normally.
• Her expected quality of life was extremely poor.
• Her father refused to accept the initial outlook.
• He used LLMs extensively for research.
• He identified an existing marketed drug believed to be safe that might affect the condition.
• The drug reportedly worked.
• Gavin recalls the measured neural-firing figure improving from roughly 30–40% to 80–90%.
• Gavin says this improvement meant she could live a normal life, although he added that she might not reach the cognitive potential she otherwise would have.
• The father was subsequently using AI and advances in protein design to tailor the treatment further.
• Gavin says the father was reasonably confident a complete cure could be produced within months.

Gavin presents this as an anecdote illustrating how one determined person could use AI to change the trajectory of a rare disease.

The identity, condition, drug, measures and prognosis are not supplied and require especially careful verification before publication.

Gavin also says:

• Elon’s description of a future in which work becomes optional may be inspiring to some and frightening to others.
• A four-day workweek may be a more immediately appealing framing.
• Jensen is doing a good job as a positive AI advocate.
• The industry should concentrate more on benefits than fear.

Asked why young people were reportedly booing AI-related commencement speeches—including one by Eric Schmidt—Friedberg gives a three-part explanation.

He argues that people intuitively understand that technology can create enormous leverage for a small number of owners.

AI appears especially powerful because a relatively small group can:

• control systems;
• profit from them;
• capture large returns;
• potentially reshape work for everyone else.

Friedberg accepts that technologies usually diffuse and commoditize eventually.

His concern is time to diffusion. If the gains remain concentrated for a long period, society experiences a major asymmetry before broader benefits arrive.

The public currently hears that some people are becoming extraordinarily rich without being able to answer:

• How does AI benefit me?
• When will it benefit me?
• Will I still have economic power?
• Will I still have a job?

He argues that this makes backlash unsurprising.

He compares the psychological legacy to nuclear technology, which in his view caused many people during the twentieth century to become more skeptical of science and technological elites because specialized knowledge could create enormous power while exposing everyone else to risk.

Friedberg next argues that anti-technology sentiment can be amplified by foreign state actors.

He says this has a long history going back to Soviet/KGB information operations during the Cold War and has become increasingly sophisticated.

Importantly, he explicitly says he does not know enough about the specific techniques or actors operating today.

He also says the United States probably conducts analogous influence operations against competitors.

His narrower proposition is that major countries have a strategic incentive to slow technological progress in rival countries.

Friedberg’s third explanation is psychological.

He compares AI with the Copernican revolution.

Heliocentrism did not merely alter astronomy; in his telling, it disrupted institutions and a worldview in which humanity and Earth occupied a privileged place.

He argues AI similarly challenges the human ego because intelligence may no longer appear uniquely human.

He calls this quality “almost anti-humanist.”

He does not say this is the sole cause of opposition, but believes it adds emotional energy to the backlash.

He concludes that AI opposition has no simple cause: economic power, international competition, institutional interests and human psychology all overlap.

Asked directly whether AI should be slowed, Chamath says no.

He uses nuclear proliferation as his analogy.

His historical account is that:

• scientists worked on the Manhattan Project despite moral objections because US security required it;
• nuclear knowledge later reached the Soviet Union;
• some secrets were deliberately leaked because some people feared a world in which only the United States possessed nuclear weapons;
• once proliferation began, neither side could simply stop;
• a balance of power became necessary.

He applies that logic to AI.

If the United States stops while China continues, he argues, China could acquire a dangerous unilateral advantage.

But he also says a world in which the United States alone possessed sufficiently powerful AI would not necessarily be healthy.

His preferred outcome is therefore balance rather than unilateral monopoly.

In his view, AI has already entered an international proliferation dynamic analogous in some respects to the postwar arms race.

Jason offers two examples of proposed ways to slow labor displacement:

1. Paced autonomous-driving deployment so taxi, Uber, bus and truck drivers retain employment for a transitional period. He says more than 10 million Americans drive professionally.
2. A tax on humanoid robots, such as Optimus or Figure robots used in warehouses, potentially assessed per hour and used to finance worker retraining.

Chamath responds that the first question should be whether the people occupying those jobs actually want them preserved.

He says there is a meaningful difference between:

• a stable job a worker enjoys and uses to support a family; and
• a high-churn job workers themselves may prefer to leave.

He claims Amazon warehouse turnover is around 35–40%, contrasting it with a hypothetical 3–4% turnover rate for an exceptionally desirable job.

His criticism is that commentators often decide on behalf of workers what should be protected without actually surveying those workers.

Chamath then refers to a presidential AI executive order that, according to the discussion, had been announced and then pulled or “scrubbed” at the last moment.

He says:

• a list of expected attendees had leaked;
• it included major “neo lab” and hyperscaler CEOs;
• Nikesh Arora was among them;
• the president reportedly disagreed with parts of the proposal;
• those elements appeared to involve federal supervision, insight or review.

Another speaker says the leaked proposal specifically covered frontier models; Chamath broadens it to AI because future frontier systems may not all be language models.

Chamath says he thinks it may actually be beneficial that China is, by his estimate, less than nine months behind the United States.

His reasoning is that comparable capability can create the conditions for détente: each side knows the other has meaningful capacity and therefore has an incentive to reach stable rules.

He also argues that because the US and China are differently organized societies, René Girard’s mimetic-theory framework may imply less direct rivalry than between two structurally identical societies.

He proposes a form of KYC for advanced AI access, potentially coordinated between the US and China.

His concern is preventing powerful capabilities from reaching uncontrolled actors who might, for example, attempt to create a biological weapon.

He claims China already performs KYC-like controls and reviews training runs before models are released.

His desired end-state is one in which the United States and China can effectively say to one another: “All right, weapons down.”

Jason addresses the next question to “Kevin,” which appears inconsistent with the surrounding conversation and should be treated as [transcript unclear].

The respondent says the issue is extremely complicated and that it feels too early for a sweeping answer.

He is uncomfortable with the United States imposing restrictions that no other major AI power adopts.

A jointly designed US-China framework, with alignment, trust and verification, would be much more acceptable to him.

Jason reframes the proposal as a common safety battery for known extreme risks—bioweapons, terrorism and similar hazards—analogous to testing food or drugs for contaminants.

The response emphasizes that the United States already has:

• industry self-regulation;
• courts;
• tort liability and other legal remedies for parties harmed by irresponsible conduct.

The speaker argues these mechanisms already create incentives for model companies to act responsibly.

Jason cites a current lawsuit involving OpenAI and a suicide following use of an OpenAI model, although the transcript’s wording is imprecise and must be verified.

The respondent warns that government powers are rarely surrendered once granted and calls regulation a potentially one-way path; another speaker summarizes it as a “one-way ratchet.”

Jason says political resistance to autonomous vehicles has appeared in the UK and in Boston and New York, partly to protect driving jobs.

A panelist responds that municipalities could eventually face wrongful-death litigation if they knowingly prevent deployment of technology that is materially safer than human drivers. The speaker describes the hypothetical alternative as having a “zero death rate,” which should not be treated as an established statistic.

Gavin predicts that eventually visiting a city without Waymo or cybercabs will feel “barbaric and unsafe,” analogous to traveling to a city without Uber during Uber’s early expansion.

He recalls approximately:

• 50,000 US automotive deaths annually;
• one million globally.

He explicitly qualifies those numbers with “if I recall correctly.”

He expects safety, convenience and quality-of-life improvements to make long-term bans difficult to sustain.

He also argues that America’s patchwork of state and municipal policies creates competition between jurisdictions, which can expose which policies work better.

Jason brings up Flock Safety as an example of AI deployment from the bottom up.

He describes it as an AI-assisted camera system used to detect or investigate crime and acknowledges privacy concerns.

His point is institutional: towns and states can decide how such systems are used rather than waiting for a single national AI regime.

He expects autonomous-driving regulation to develop similarly.

A panelist then says:

“Crime is now a choice.”

The speaker argues that governments increasingly have technology capable of preventing or solving crime and that jurisdictions will make different policy decisions about whether to deploy it. Attribution is not reliable enough in the raw transcript to assign this statement confidently.

A speaker says the Cambridge, Massachusetts City Council had voted two days earlier to turn off gunshot detectors.

The discussion then becomes explicitly political.

One panelist advances a theory, not an established explanation, that opponents could be concerned that gunshot detection might lead to the apprehension of an undocumented migrant firing a weapon.

That motive is an allegation/speculation in the podcast and must not be represented as the City Council’s verified rationale.

The group also references an Andreessen Horowitz/A16Z essay on Flock and argues that voters will eventually compare the crime outcomes of jurisdictions using different policies.

A panelist whose identity is not reliably labeled says he and his wife spent an afternoon in Las Vegas with Ben Horowitz and Felicia Horowitz.

He praises Felicia’s involvement with Las Vegas police and describes seeing a command center in which:

• drones could be deployed from the police building;
• incidents could be observed within minutes;
• suspects could be tracked;
• the system created a strong feeling of safety.

He argues the investment required is small compared with the economic and human cost of crime.

Another speaker estimates that an additional $30–40 million annually could make Las Vegas the safest city in America. This is an opinion/estimate, not a sourced budget analysis.

Jason says Flock also demonstrates ways to place limits around surveillance.

He recalls:

• a rolling database;
• limits on how long license-plate records can be retained;
• no facial recognition;
• audit trails recording access.

He believes the retention limit is two or three years but signals uncertainty himself.

Jason argues that the public’s fear is rationally connected to what technology companies themselves are saying and doing.

His first example is Cloudflare CEO Matthew Prince.

Jason characterizes a Prince communication as saying that Cloudflare had:

• laid off more than 20% of its workforce;
• done so despite record revenue growth, strong free cash flow and customer growth;
• targeted people described as “measurers”—managers and people measuring work;
• considered those roles increasingly unnecessary because of AI;
• expected hiring to continue in other areas.

His second example is Meta/Zuckerberg.

Jason says:

• approximately 8,000 people were being laid off;
• some were on H-1B visas and therefore faced personal immigration disruption;
• this followed earlier layoffs totaling tens of thousands;
• Meta remained highly profitable;
• employees were simultaneously being used to create higher-quality training data and workflows for AI;
• recording software was reportedly being placed on employee computers to study work for model training;
• some employees who created AI tools to make themselves more productive later found themselves laid off.

Jason argues that employees can reasonably infer that they are being asked to help train systems that may subsequently replace parts of their own work.

He summarizes the fear as keeping a job temporarily while “train[ing] your way out of it.”

An inserted Zuckerberg clip says Meta believes its own employees have substantially higher average capability than outside contractors used for model-training tasks.

For coding, Zuckerberg argues that using thousands of Meta’s strong engineers to create tools and solve training tasks could improve model coding capability faster than competitors without an equivalent internal engineering base.

Chamath says the Matthew Prince note was “horrible.”

He uses an ableist slur in criticizing its PR quality; that wording is preserved in the quote bank because the user requested preservation of controversial exact language, not because this summary endorses it.

His substantive criticism is that labeling humans as “measurers” reduces people to a category and then publicly marks everyone dismissed under that category.

He argues this creates a “scarlet letter”: future employers may view a former Cloudflare worker as one of the supposedly unnecessary “measurers.”

His point is not that companies can never reduce headcount. It is that executives should understand the downstream consequences of the language used to explain those decisions.

A clip attributed to Shyam Sankar says the public is listening too much to the inventors of AI.

He recommends listening instead to:

• a factory worker who says AI enabled a third shift and more hiring;
• an ICU nurse who says AI created more time with patients and improved safety during shift changes.

Chamath extends that argument to corporate CEOs: technical brilliance does not imply brilliance in public communication.

He bluntly advises technology executives to do their jobs rather than publish poorly considered “missives.”

The group notes that fear will intensify if companies:

1. eliminate large numbers of jobs;
2. explain those eliminations as AI-driven efficiency;
3. subsequently announce large share repurchases, dividends and growing cash balances.

Chamath’s conclusion is that there may be legitimate business reasons for restructuring, but there is a responsible and an irresponsible way to communicate it.

He says current PR and communications teams often fail to understand the historical moment.

Jason says SpaceX had filed an S-1 on Wednesday and gives the following claimed terms:

• $75 billion capital raise.
• $1.75 trillion valuation.
• Expected listing around June 12.
• Proposed ticker SPCX.
• If completed on those terms, more than twice the size of Saudi Aramco’s roughly $29 billion IPO.

These claims require verification directly against the filing.

Jason describes three primary business areas.

Starlink

• $11.4 billion revenue in the prior year.
• 50% growth.
• $4.4 billion operating income.
• More than 10 million subscribers.
• Jason believes it could eventually support hundreds of millions of paying users.

Launch/space

• Approximately $4 billion revenue.
• 17% growth.
• Approximately $650 million operating loss.

AI

• $3.2 billion revenue.
• More than doubled year over year.
• $6.4 billion operating loss.

Capital expenditure

• Roughly $20 billion in the prior year.
• More than 60% allegedly directed toward AI compute.

Jason calls SpaceX’s compute-rental activity “EWS” or “Elon Web Services,” a nickname used by the show.

He says Anthropic is paying SpaceX approximately:

• $1.25 billion per month;
• for Colossus 1 and part of Colossus 2;
• under a $45 billion, three-year arrangement;
• equivalent to roughly $15 billion per year.

He argues that this has effectively added another Starlink-sized revenue stream.

He also discusses Cursor, initially saying “if” SpaceX buys it; another panelist corrects him and says the acquisition has already occurred.

They estimate Cursor could contribute another $2–3 billion in revenue and is itself growing roughly 2x.

Jason cites Polymarket as assigning a 71% probability that SpaceX closes its first trading day above a $2 trillion market capitalization.

Gavin says the $15 billion Anthropic contract effectively quadruples the AI business described in the filing.

He cites claimed construction times for successive data centers of roughly:

• 122 days;
• 91 days;
• 66 days.

His key argument is that SpaceX/xAI appears capable of building data centers unusually quickly and cheaply.

Once there are reliable offtake customers, he expects the company to replicate these facilities rapidly.

He also says Jensen Huang cares strongly about ensuring Nvidia GPUs are actually used; therefore, scarce GPUs will tend to flow toward operators who can install them and immediately turn electricity into useful computation.

Jason adds an important contract qualification: according to him, either side can cancel the Anthropic arrangement with 90 days’ notice.

Gavin describes Cursor’s newly released Composer 2.5 as a major data point.

He says:

• it became “Pareto dominant” relative to comparable models;
• the improvement followed only three or four weeks of reinforcement learning on Colossus 2;
• Cursor’s proprietary coding data was used;
• Cursor allegedly possesses more tokens of coding data than exist on the public internet;
• Cursor and Anthropic may possess the largest proprietary corpora of coding tokens.

He emphasizes that Composer 2.5 reportedly uses the same underlying base model as Composer 2, transcribed as “Kimmy K K.25” [transcript unclear].

Therefore, he attributes the improvement largely to reinforcement learning plus proprietary data rather than a new base model.

He expects the next step to combine:

• a new base model;
• Cursor’s model/data;
• reinforcement learning;
• the enormous Colossus compute cluster.

Jason says Cursor had previously lacked sufficient compute and was falling behind “Kodaks” [transcript unclear], Google and Anthropic, but access to Colossus changed its trajectory.

Jason speculates it could become the dominant player within a year.

Asked whether Elon might eventually sell compute to Google and OpenAI, Gavin says Google is easier to imagine and believes it will happen.

He says there had already been posts suggesting such interest.

He also argues Google would want exposure to orbital compute.

Gavin claims people deeply involved in data centers or space—naming Google, Anthropic, Amazon and Nvidia—generally take orbital compute seriously.

Gavin says the latest Grok, identified in the transcript as Grok 4.3, is on the frontier with leading models.

He describes one build as approximately 500 billion parameters.

He names the frontier group as:

• xAI/Grok;
• Google;
• OpenAI;
• Anthropic.

His qualification is that Grok previously lacked an effective harness.

He contrasts:

• Claude + Claude Code;
• OpenAI + Codex;
• Grok + newly released “Grok Build.”

Jason initially describes a harness in user-facing terms as a downloadable application with integrations into tools such as Notion, Gmail and Slack.

Gavin says that understates it.

To him, a harness is also:

• a runtime;
• an execution environment;
• a state-management system;
• a memory system.

He argues that in agentic AI, the harness can be almost as important as the underlying model and should be developed alongside it.

He also says Composer 2.5 had become the most-selected model in Cursor and that user “vibes” were very positive.

Together, Cursor’s data, Composer’s improvement and Grok Build suggest to him that xAI’s ecosystem is improving rapidly.

Asked about the broader Musk ecosystem and a potential future combination of Tesla and SpaceX, Friedberg shifts away from valuation.

He argues that governments around the world increasingly seek control over:

• information;
• speech;
• commerce;
• purchases;
• physical data-center construction.

Because the internet ultimately depends on physical infrastructure located inside jurisdictions, he sees value in an alternative infrastructure layer outside ordinary terrestrial control.

A space-based network combining:

• communications;
• data centers;
• storage;
• routing;
• wireless connections back to Earth

could function, in his view, as a backup for civilization and technological progress.

He discloses that he does not own SpaceX shares and says he is not merely “talking his book.”

Another panelist recounts an early SpaceX story.

He says Elon originally explored obtaining Russian launch capability because he wanted to “back up the biosphere.”

After encountering the underlying rocket problem, Elon supposedly concluded he should simply build his own rockets.

The speaker says Elon envisioned geodesic structures in space containing plants, wildlife and other organisms.

This is presented as a personal recollection and should be verified against primary historical sources before publication.

Jason argues the scale achieved by Starlink suggests orbital data centers are conceptually plausible.

Another speaker corrects him: a compute satellite would be substantially larger, with much larger structures and different physical requirements.

The speaker nevertheless says the concept is technically possible and expects Elon to be among the first to solve it.

Chamath offers a more conventional financial framework for a roughly $2 trillion valuation.

He says SpaceX did approximately $18–19 billion of revenue in the prior year and might produce $25–30 billion in the current year.

That makes the headline multiple expensive.

His bullish case rests on several components:

• Starlink becoming foundational global internet infrastructure.
• Potential expansion toward hundreds of millions of users.
• AI applications and compute becoming a major second business.
• SpaceX’s ability to construct gigawatt-scale data centers unusually quickly.
• Revenue funding additional infrastructure, which funds technical advantages, which improves execution.

Chamath describes this as a reinforcing sequence:

capital moat → technology moat → execution/learning moat.

He forecasts roughly $40–45 billion of revenue in the following year and believes growth could remain extremely high after that.

He also expects Tesla ultimately to become part of the broader Musk ecosystem, although this is speculation.

He gives Elon a premium analogous to Steve Jobs’s “one more thing”: investors are buying not only identifiable businesses but also the possibility of an unforeseen new business created at civilizational scale.

For underwriting purposes, he suggests terrestrial data centers alone might eventually represent $100–200 billion of revenue by 2030–2032.

He says achieving even a nameplate gigawatt data center is extremely difficult.

Chamath describes another potential advantage: Jensen Huang’s push toward more direct-current architecture inside data centers.

Today, he says, electricity goes through repeated AC/DC transformations before reaching racks.

Those transformations introduce:

• losses;
• cost;
• extra power requirements;
• cooling requirements;
• complexity.

A more direct DC-to-DC architecture could reduce those inefficiencies but would require substantial redesign.

Chamath sees Elon as a natural engineering and design partner for Nvidia on such a system.

Jason says being inside Starship feels comparable in volume to a large commercial aircraft, estimating room for roughly 300 people.

He also cites a chart of cumulative payload launched since 1957 and argues that SpaceX’s payload since around 2012 is approaching the historical total launched by the rest of the world.

Gavin notes an important qualification: most of that SpaceX performance occurred before operational Starship and was achieved largely with Falcon.

Gavin distinguishes ordinary reusability from rapid reusability.

A reusable rocket may require extensive refurbishment of:

• engines;
• fairings;
• other hardware,

and may not fly again for 30–60 days.

Rapid reusability means landing and flying the same vehicle repeatedly, potentially multiple times per day.

He says this is far harder technically but is necessary for ambitions such as:

• a lunar base;
• settlements on the Moon or Mars;
• very large-scale lunar infrastructure.

That requirement explains why Starship is such a difficult engineering project.

Asked when rapid reusability will be achieved, Gavin refuses a confident prediction.

He says even a test ending in a “fireball” can generate valuable information because a new rocket and booster carry extensive instrumentation.

He eventually guesses one or two years, perhaps sooner.

His central argument is that SpaceX deliberately learns through repeated test failures rather than trying to eliminate all failure before testing.

Gavin says a working Nvidia H100 is already in space and says Andrej Karpathy has both trained a model and performed inference with it.

Jason adds that Nvidia is working on hardware designed specifically for space.

He says space hardware has different requirements:

• different heat-management systems;
• lower need for some terrestrial hardware;
• much greater mechanical reinforcement to survive launch acceleration and vibration.

Gavin adds that SpaceX has historically been unusually good at using cheaper commercial semiconductors that were not originally radiation-hardened for space.

He references his firm, Seraphim Space, as an investor in a company transcribed as “Excite Labs”, whose chips he claims will appear in essentially every Starlink.

He predicts meaningful orbital compute sometime between the second half of 2028 and the first half of 2030.

Jason gives the following Nvidia figures:

• Q1 revenue: $81.6 billion.
• Year-over-year growth: 85%.
• Quarter-over-quarter growth: 20%.
• Net income: $58 billion.
• Free cash flow: $48 billion.
• Gross margin: roughly 75%.
• Market capitalization: roughly $5.3 trillion.
• Stock gain for the year: roughly 16%.
• New buyback authorization: $80 billion.
• Earlier buybacks: approximately $100 billion.
• Combined effect described as roughly 4% of the company.
• Quarterly dividend increase is garbled in the transcript as from one cent to “25% cents per share.”
• CFO reportedly said the company intends to return 50% of free cash flow to shareholders.

Every figure should be checked against the relevant earnings release and filing.

Chamath asks Gavin about remarks from an earlier Patrick O’Shaughnessy interview.

A person’s name is rendered as “Leo Ashan Brenner” [transcript unclear].

Chamath says this person’s fund rapidly grew from approximately zero to $5 billion and appears to have had large put positions against semiconductor stocks.

Gavin says the investor is brilliant and appears to have produced excellent returns, but cautions against over-reading a 13F.

He says the filing reflected positions at the end of Q1 during intense geopolitical fear and that many investors had hedges then.

Those puts may no longer exist.

Therefore, he would not automatically interpret the filing as a current bearish thesis on semiconductors.

Gavin identifies a cross-sectional inconsistency.

He says:

• memory manufacturers trade at roughly three to five times earnings;
• Nvidia trades at a comparatively low multiple;
• some accelerator companies are reasonably valued;
• power, cooling and some optical businesses imply much more optimistic assumptions.

His conclusion is binary:

• If the high valuations in power/cooling/optics correctly reflect future AI demand, Nvidia and memory may be substantially undervalued.
• If Nvidia and memory’s restrained multiples are correct, many surrounding infrastructure stocks may underperform.

He calls the AI market “cross-sectionally inefficient.”

Gavin says Nvidia changed its reporting structure, making simple comparisons difficult.

He notes Broadcom had guided to roughly 143% year-over-year growth in AI semiconductor revenue.

However, he argues that an apples-to-apples comparison should examine Nvidia’s hyperscaler and AI-cloud business while stripping out China because Broadcom did not have the equivalent Chinese exposure.

On that basis, he believes Nvidia’s Western data-center AI business is growing faster than Broadcom and faster than many firms presented as beneficiaries of ASIC share gains.

Gavin says Jensen appears increasingly frustrated that Nvidia is described as losing share while its revenue grows faster than hyperscaler capital expenditure.

He also argues that competing ASICs are not being submitted to common benchmarks such as MLPerf because they would lose to Nvidia.

That is his inference, not an independently demonstrated reason for their absence.

He says Nvidia cannot effectively rebut competitors that are not benchmarked:

“You can’t fight shadows.”

He wants transparent comparisons such as GB300 versus TPU v7 versus Inferentia/Trainium.

He also observes that very large companies often trade in stair-step patterns: their multiple compresses as investors question whether growth can continue at scale, then the stock rerates if results establish a higher sustainable floor.

Gavin highlights Nvidia’s statement that its CPU business could reach roughly $20 billion in the year.

He says that would immediately make Nvidia one of the world’s largest CPU manufacturers.

He attributes the opportunity partly to Nvidia’s unique position working with essentially every major AI laboratory, allowing it to co-design hardware around where model architecture is going.

Chamath adds that the domain-specific architecture transition he expected to occur through independent specialist chip companies is, surprisingly, happening inside Nvidia itself through these co-design programs.

The discussion attributes a major bear thesis to Michael Burry, although his surname is inconsistently transcribed.

The argument:

• hyperscalers and neo-clouds depreciate GPUs/CPUs over four, five or six years;
• the true economically useful life may be only around two years;
• therefore reported profits could be overstated.

Gavin says disaggregated inference changes the useful-life calculation.

With specialized accelerators, an operator can combine:

• newer Groq/Cerebras-style accelerators for particular stages such as decode;
• older GPUs for other parts of the workload.

This allows older GPUs to remain economically useful for much longer.

He suggests some could have useful lives of 10–15 years.

He also emphasizes financing.

CoreWeave is said to have obtained GPU-backed financing at approximately 6%.

If GPUs can be financed as durable, valuable assets at favorable rates, that is a major structural advantage for the neo-cloud model.

Jason recounts an interview with a CoreWeave executive whose name is garbled as “Michael in Trader.”

According to Jason, the executive said:

• customers willingly sign six-year contracts;
• some pay in advance;
• demand exists even for older hardware;
• useful work may remain in years seven, eight and nine.

Jason argues that customers would not sign six-year commitments if they believed the hardware became worthless after two years.

Jason describes a deteriorating macro backdrop.

His claims include:

• the Iran conflict had reached approximately week 12 despite an earlier expectation of four to six weeks;
• oil remained elevated;
• Polymarket assigned a 99% probability to May inflation at 4.2% or higher;
• professional forecasters were discussing CPI reaching 6%;
• discussion had shifted from further Federal Reserve cuts toward possible rate increases;
• the US 10-year yield had reached approximately 4.6%;
• Japan’s 30-year yield was approximately 5.1% and described as its highest recorded;
• UK yields were the highest since the global financial crisis;
• German yields were the highest since 2011;
• Korean retail investors were borrowing record amounts to buy AI semiconductor stocks.

He compares Korean enthusiasm with earlier leveraged enthusiasm for crypto and treats it as a possible speculative warning sign.

Friedberg says the time for being “concerned” may already have passed if the system is effectively beginning its descent.

He cites global debt-to-GDP of 310%.

His macro sequence is:

1. Governments continually spend to sustain economic growth and service existing leverage.
2. Eventually the structure becomes unstable.
3. Currency value falls.
4. Inflation accelerates.
5. Authorities print more money.
6. Nominal asset prices rise.
7. Higher nominal asset values allow debt servicing to continue.
8. The spiral reinforces itself.

He believes Japanese rates could potentially become a catalyst for a credit crisis because of large carry trades.

He explicitly cautions that more active market participants and economists may have better views.

His metaphor is that this is “water leaking out of the bucket.”

Chamath agrees that warning signals are flashing.

His response is not to speculate more aggressively.

He says investors should identify a small number of businesses that genuinely represent the future, underwrite them carefully and, if the case survives scrutiny, be prepared to own them for ten years.

Everything else should generally be avoided.

He says he learned through experience that market volatility no longer provides enough upside psychologically to justify the downside.

His own approach is:

• a small number of companies;
• very concentrated holdings;
• little interest in tracking dozens of public stocks;
• “five or less” public equities that he can follow meaningfully;
• approximately four things he actively stays on top of.

He declines to disclose the percentage of his net worth represented by his largest positions.

Gavin contrasts his professional setup.

He manages more than 100 positions, but with a team of more than 30 people.

He says three facts can simultaneously be true:

1. Rising interest rates are concerning.
2. AI fundamentals are extraordinarily strong.
3. Global geopolitical and energy disruption can hurt the world while relatively advantaging the United States.

He describes Anthropic as growing faster at massive scale than any company in history; the panel additionally says it is now larger than more than 100 countries and profitable. These are source claims requiring definition and verification.

Gavin notes that the dot-com bubble occurred with much higher long-term Treasury yields.

He compares Cisco, which he says traded at roughly 100 times forward earnings, with Nvidia today.

He estimates Nvidia at a low-to-mid-teens multiple of “real earnings,” which he describes as buy-side consensus rather than a quoted market statistic.

His implication is that spectacular AI growth does not automatically imply dot-com-style valuation excess.

Gavin argues that a global crisis can be bad in absolute terms while still improving America’s relative position.

He cites US advantages in:

• food self-sufficiency;
• energy self-sufficiency;
• oil and gas production;
• energy exports;
• currency strength;
• public companies;
• private technology companies.

He describes America as the “best house” in a global neighborhood burdened by high debt.

He argues that closure of the Strait of Hormuz could function as a forcing mechanism for US reindustrialization because American industry is less dependent on imported LNG than major European and Asian economies.

He says US natural-gas benchmark “NG1” was down for the year while LNG input costs elsewhere were up 100–200%.

He therefore calls Hormuz closure:

• bad for everyone;
• relatively better for the United States;
• relatively aligned with Trump’s reindustrialization goals.

He speculates that this is one reason Trump may not be in a hurry to reopen the Strait.

He stresses again that rising rates and inflation remain negative even if the United States is relatively advantaged.

Gavin says AI fundamentals have previously shown some seasonality.

He suggests:

• university students use ChatGPT and Claude less during summer;
• people may work less when the weather is good;
• therefore usage may soften seasonally.

He says it remains unclear whether the new generative-AI economy will preserve that pattern.

Jason compares this with familiar seasonal shifts in consumer technology: some apps and subscriptions weaken when people travel, while companies such as Uber or DoorDash may benefit.

Jason describes a roughly 48-hour trip involving the president and technology CEOs to meet Xi.

He says the public optics were positive—handshakes, friendly interactions and good atmosphere—but argues that the visible policy outcome appeared limited.

He mentions:

• soybeans;
• aircraft;
• Nvidia H100s;
• something transcribed as “A200s” [transcript unclear];
• sales to Baidu and other Chinese firms;
• possible implications for tariffs;
• possible implications for the Strait of Hormuz.

He asks Friedberg whether the meeting was largely performative.

Friedberg says he does not believe a sweeping long-term US-China agreement emerged.

He acknowledges:

• stated intent to continue cooperating;
• additional trade discussions;
• aircraft purchases;
• agricultural commitments.

But he says the hoped-for “grand deal” that fundamentally reduced strategic tension did not materialize publicly.

He notes Putin was meeting Xi immediately afterward and sees that meeting as another symbolic relationship-building exercise.

His broader conclusion is that competition between China and the United States remains a long-running geopolitical story rather than one resolved by a single summit.

Chamath is more positive.

He says the trip was successful and distinguishes public announcements from private negotiation.

He speculates that the most important conversations concerned the geopolitical “tic-tac-toe” involving Venezuela, Iran, Taiwan and other strategic questions.

He imagines there may be a way to divide or sequence issues in a manner that helps both governments.

He explicitly says “I’m just guessing.”

Asked whether allowing Nvidia to sell more GPUs into China is good policy, Gavin argues that supplying depreciated Nvidia GPUs could actually help maintain US advantage.

His reasoning:

1. If China cannot buy American hardware, it has a stronger incentive to create an entirely independent hardware/software ecosystem.
2. Gavin believes such an alternative could be much more power-hungry.
3. Continued dependence on the American ecosystem gives the United States more influence.
4. Therefore, continued sales can arguably stabilize the competition while preserving US leadership.

He acknowledges reasonable people can disagree.

Chamath clarifies that he does not disagree and says, bluntly, “sell everything to them.”

His original point was that the most important reciprocal commitments are unlikely to appear in a press release.

The panel agrees that direct US-China communication is beneficial.

A term is transcribed as “the these trap”, almost certainly requiring verification before publication.

The speakers say Chinese officials themselves discuss the concept and had brought it up by name.

Their argument is that talking directly is an essential step in avoiding the conflict pattern the term describes.

A speaker praises Trump’s ability to build personal rapport with dictators, monarchs and Gulf rulers.

He calls this Trump’s “greatest superpower” and argues that personal comfort with adversarial leaders can reduce conflict.

The speaker then makes a highly speculative prediction that Trump could eventually signal that Taiwan may ultimately become China’s, provided no confrontation occurs during his administration, potentially through a 20-, 30-, 100- or even 200-year transition arrangement.

No evidence is supplied for such an agreement. It is explicitly framed as something the speaker “wouldn’t be surprised” to see.

The same speaker reasons that China’s dependence on oil from Iran, Venezuela and Russia could constrain it during a major war.

His hypothetical scenario is that the United States could deny China:

• Venezuelan oil;
• American oil;
• Brazilian oil;
• Gulf/Middle Eastern oil through closure of the Strait of Hormuz.

China would then be disproportionately dependent on Russia.

The speaker says that would leave China poorly positioned against a coalition including the United States, Japan, South Korea, Australia, the UK, France and Germany, although he becomes less certain when discussing Europe as a whole.

He is especially confident Japan, Australia and South Korea would participate.

He concludes that the world could become more stable “on the other side of Iran,” however that conflict resolves. This is geopolitical speculation, not a forecast supported with evidence in the transcript.

The hosts thank Gavin for appearing and note Sacks’s absence.

They ask for Gavin’s father-in-law’s name; Gavin says Jeff Painter.

The hosts invite Painter to future events and joke that offering him backstage/VIP access helped persuade Gavin to appear.

Gavin replies that he is always there for his father-in-law.

The remaining material is outro music and non-substantive comedic audio fragments. It contains no additional argument requiring preservation.

Speaker: Chamath
Exact quote: “this idea of recursive self-learning puts these models on a combination of overdrive and autopilot.”
Context: Explaining why Karpathy’s Anthropic role could accelerate model improvement.
Supports: Recursive self-improvement / AI acceleration.

Speaker: Chamath
Exact quote: “there's an order of magnitude improvement on a yearly basis. So like this new form of Moore's law.”
Context: Speculating about annual AI capability gains if self-improvement works.
Supports: Capability scaling / recursive learning.

Speaker: Gavin
Exact quote: “continual learning is the holy grail, where the model learns from experiences the way humans do.”
Context: Describing what he sees as one of AI’s remaining technical frontiers.
Supports: Continual learning / frontier AI.

Speaker: Chamath
Exact quote: “I think breathlessly talking about every model improvement is a waste of time. There's no ROI in it.”
Context: Arguing that AI coverage should move from benchmark improvements to end-user outcomes.
Supports: AI public relations / user utility.

Speaker: Chamath
Exact quote: “nobody wins if we become Luddites and go back in time.”
Context: Calling for discussion of both AI benefits and risks.
Supports: Technology adoption / AI narrative.

Speaker: Shyam Sankar, quoted in the episode
Exact quote: “Stop breathlessly asking these model makers what they think. Go to the end user and ask the person in the factory that's using the model and ask him what he or she thinks.”
Context: Chamath cites Sankar’s argument that real-world users should drive AI coverage.
Supports: AI utility / reporting framework.

Speaker: Friedberg
Exact quote: “It's almost anti-humanist.”
Context: Describing the psychological discomfort created by intelligence no longer appearing uniquely human.
Supports: AI backlash / psychology.

Speaker: Chamath
Exact quote: “we are in this existential race and we need to get to the place where each of us, meaning us and China, can look each other in the eye and say, ‘All right, weapons down.’”
Context: Arguing for mutual capability plus safeguards rather than unilateral US slowdown.
Supports: US-China AI policy.

Speaker: Gavin/response addressed as “Kevin” [attribution requires verification]
Exact quote: “once you give something give a power to the government it's almost never taken back and it seems to grow.”
Context: Warning about premature frontier-model regulation.
Supports: AI regulation / government authority.

Speaker: Gavin
Exact quote: “I think going to a city where you can't get in a Waymo or a cyber cab is going to feel barbaric and unsafe”
Context: Predicting autonomous transport will become a normal expectation.
Supports: Autonomous vehicles / local regulation.

Speaker: [speaker unclear]
Exact quote: “Crime is now a choice.”
Context: Discussion of Flock Safety and local decisions about surveillance/crime-prevention technology.
Supports: AI policing / municipal policy.

Speaker: Jason
Exact quote: “the most you can hope for here is you keep this job for some amount of time and train your way out of it”
Context: Describing how employees may interpret AI training and subsequent layoffs.
Supports: AI employment / worker fear.

Speaker: Chamath
Exact quote: “You could not have written a worse memo. It's like you reduce humans to a label called the measurer and then you're like I'm going to lay off all the measurers.”
Context: Criticizing the communication surrounding Cloudflare layoffs.
Supports: Corporate communications / AI layoffs.

Speaker: Chamath
Exact quote: “This was like from the PR school of retards.”
Context: His deliberately harsh description of the Matthew Prince communication. Contains an ableist slur.
Supports: Tone and intensity of Chamath’s criticism.

Speaker: Shyam Sankar, replayed clip
Exact quote: “We need to be listening to the frontline factory workers who are using AI saying, ‘Wow, I was able to add a third shift. I was able to hire more workers.’”
Context: Contrasting user outcomes with predictions from AI inventors.
Supports: Labor / practical AI benefits.

Speaker: Gavin
Exact quote: “GPUs will be allocated to who can plug them in, turn them on, and start converting electrons into tokens.”
Context: Explaining why fast data-center builders may receive Nvidia supply.
Supports: AI infrastructure / SpaceX compute.

Speaker: Friedberg
Exact quote: “it's almost like a backup for civilization, but it's a backup for progress.”
Context: Describing space-based communications and compute beyond terrestrial government control.
Supports: Orbital infrastructure / civilizational resilience.

Speaker: Gavin
Exact quote: “They learn from failure. If you don't fail, you're not learning.”
Context: Explaining SpaceX’s testing philosophy.
Supports: Starship / engineering process.

Speaker: Gavin
Exact quote: “you can't fight shadows.”
Context: Arguing that Nvidia cannot answer competitors that are not submitted to common benchmarks.
Supports: Nvidia / ASIC competition.

Speaker: Gavin
Exact quote: “The AI market is cross-sectionally inefficient right now”
Context: Explaining why valuations across memory, Nvidia, power, cooling and optics imply mutually inconsistent expectations.
Supports: AI investing / valuation.

Speaker: Friedberg
Exact quote: “The roller coaster will roll down. We will throw our hands in the air and we will scream, ‘Whee!’ as we go for the ride.”
Context: His deliberately fatalistic description of global debt and rising rates.
Supports: Macro risk / debt cycle.

Speaker: [speaker attribution unclear]
Exact quote: “America and China talking is only good.”
Context: Discussion of avoiding great-power conflict after the Trump-Xi meeting.
Supports: US-China relations.

Speaker: [speaker unclear]
Exact quote: “I do think he's probably giving them the green light on like, ‘Hey, Taiwan's yours. Just let's not have it during my administration.’”
Context: Explicit speculation about what Trump might privately negotiate with China. No evidence is supplied in the transcript.
Supports: Taiwan / geopolitical speculation.