SpaceX’s $2T Case, Nvidia’s Shock Selloff, America Turns on AI, Trump Pulls AI Order, Bond Crisis?
In a Nutshell
**AI progress is accelerating through recursive self-improvement, with Karpathy’s move to Anthropic and compute breakthroughs like Cursor’s Composer 2.5 and Colossus signaling potential order-of-magnitude yearly gains in model capability.** SpaceX’s $1.75T valuation and Starlink’s $11.4B revenue underscore its position as critical infrastructure, while Nvidia’s $81.6B quarter and domain-specific architecture shifts reinforce its dominance despite mixed market reactions. Public backlash, regulatory uncertainty, and macro pressures like rising yields create friction, but the episode frames AI and space infrastructure as existential advantages the U.S. must lead rather than restrain.
These notes were generated by AI and may contain inaccuracies.
Gavin Baker from Atreides Management joins as guest with David Sacks absent. Episode covers SpaceX and OpenAI IPOs, Andre Karpathy joining Anthropic, Nvidia performance, and multiple AI-related developments.
Andre Karpathy, age 39, is a recognized figure in AI. He previously co-founded OpenAI and led the self-driving team at Tesla. Karpathy will deliver a keynote at the Liquidity event on Tuesday. He coined the term "vibe coding" and developed Auto Research, an open-source training tool enabling AI models to improve themselves through five-minute experiments, which accumulated over 82,000 GitHub stars from a weekend project. This led to community development of recursive LLMs. He also released the Andre Karpathy Skills tool based on his principles for Claude code.
Karpathy will head a new pre-training team at Anthropic focused on recursive self-improvement, enabling Claude to improve itself. Anthropic has previously discussed AI improving AI.
Gavin Baker notes that Google historically recognized singular technical talent through the Google Fellows program, naming Ahmed Singhal, Shradar Ramaswami, and Jeff Dean. Jeff Dean remained at the forefront of multiple AI waves. Karpathy has similarly been positioned at successive AI inflection points. While leading FSD at Tesla, he was among the first to commercialize Richard Sutton's "Bitter Lesson" essay emphasizing brute-force computation. Karpathy reportedly spent approximately 25% of his time hand-labeling video data from Teslas in 2016-2017.
Karpathy later co-founded OpenAI. As a free agent, his subsequent work has been notable. Baker views the Anthropic move as significant, describing Karpathy as a highly curious individual capable of independent invention. Recursive self-learning places models on both overdrive and autopilot, potentially enabling order-of-magnitude yearly improvement analogous to a new Moore's Law, producing parabolic gains in model quality through compute scaling and rapid learning.
Gavin Baker states Anthropic achieved EBIT positivity in the most recent quarter per Wall Street Journal reporting. This development supports the broader AI narrative. OpenAI and Anthropic are projected to reach approximately $100 billion combined ARR with roughly 80% gross margins on inference. Adding Gemini, Cursor, xAI, and open-source models could yield $200-400 billion ARR by year-end at high margins across language models.
Baker excludes non-LLM GPU returns such as improved recommender systems, ad targeting, and ad measurement at Facebook, Google, and Amazon. He focuses strictly on LLM economics. The analysis suggests strong ROI this year even when excluding major economically significant GPU use cases.
Gavin Baker identifies recursive self-improvement and continual learning as the final two frontiers for AI. Recursive self-improvement involves a model contributing input to its own training during forward passes, or another model contributing to training. Chimera's statistic of 10x yearly improvement may prove conservative. Continual learning remains the "holy grail," allowing models to learn from experience analogously to humans.
Anthropic currently holds a 3-6 month lead over competitors and 6-12 months ahead of open source. With Karpathy's addition, the point at which AI contributes more to language model improvement than humans in the loop may arrive. Architectural paths include networks of smaller models aggregating to lower cost-per-token than single large models. Early MIT papers indicate substantial remaining gains from model rearchitecture and deployment strategies. Small language models and verticalized models represent the future trajectory.
Google integrated the Gemini Nano model into Chrome browser, a 4GB on-device model handling proofreading, spelling, and autocomplete, installed without explicit user notification. Discussion addresses whether this constitutes a privacy concern versus a utility-focused rollout. Google is described as already attorney-maxxed.
Emphasis shifts from model improvement announcements, which are considered low-ROI to discuss, toward end-user achievements previously impossible. Examples include OpenAI solving a long-outstanding math problem and drug candidates entering clinical trials that were previously considered non-viable. The narrative should prioritize user utility over "us versus them" framing to avoid creating AI as a four-letter word.
Gavin Baker argues Americans in technology bear responsibility to advocate for AI's positive possibilities amid what appears to be a potential CCP-funded campaign against AI and data centers in America. Dario Amodei is characterized as reading his own book by creating regulatory boundary conditions for a moat, with breathlessness scaling alongside fundraising rounds. Each CEO operates with clear incentives at current scale.
Sham Sankar, Palantir CTO, advocates asking end users (factory workers, doctors, scientists) rather than model makers about AI impact. Elon Musk's framing of work becoming optional with universal basic income is noted as potentially anxiety-inducing despite abundance messaging. Jensen Huang is viewed as an effective advocate.
A hedge fund manager's daughter born with a rare genetic mutation causing minimal neuron firing was treated after LLM-assisted research identified an existing safe drug. Neuron firing increased from 30-40% to 80-90%, enabling a normal life. The father continues using AI to develop a potential complete cure within months, alongside broader protein design advances.
Three commencement speeches faced booing, including Eric Schmidt's. Young people's vocal opposition is analyzed through multiple lenses:
- Technology creates leverage for a small group, producing power imbalances and asymmetric returns before diffusion occurs
- External state actor influence on media sentiment, with historical precedent from KGB operations during the Cold War
- AI as psychologically anti-humanist, shifting focus from human-centric worldviews similarly to the Copernican revolution's disruption of church authority
Post-Manhattan Project precedent shows research labs retained scientists to prevent knowledge leakage to adversaries. Nuclear secrets were leaked to Russia to establish counterbalance. Once proliferation began, it could not be halted. Current AI dynamics mirror this: if the United States slows, another nation advances. Analysis of China possessing sufficiently advanced models and scaled deployment indicates an unhealthy global position. Balance is required rather than unilateral dominance or unilateral restraint.
Discussion addresses whether self-driving vehicles and humanoid robots (e.g., Figure robot sorting packages at Amazon) should follow paced rollouts to provide job security for the over 10 million Americans employed in driving occupations rather than abrupt displacement.
Discussions around taxing humanoid robots per hour to fund worker retraining have emerged as potential policy approaches. Surveys specifically targeting truck drivers and package sorters about whether they want to keep their current jobs are notably absent from policy conversations. If workers genuinely love their jobs and can provide for their families, there's a reasonable claim to preserve those roles. However, Amazon's 35-40% warehouse churn rate suggests these may not be highly desirable positions, compared to an expected 3-4% churn for truly good jobs.
Critics argue that much opposition comes from people taking moral high ground without consulting the actual workers affected. The conversation should focus on asking package sorters and warehouse workers what they would prefer to do instead.
A presidential executive order on AI was announced and then scrubbed at the last minute. The leaked attendee list included major AI lab CEOs and hyperscaler CEOs. The president reportedly disagreed with aspects requiring federal government supervision, insight, and review of frontier AI models. This would have applied across different types of AI systems, not just language models.
China is less than nine months behind the US in AI capabilities. This creates an opportunity for détente where both nations reach similar capability levels, potentially enabling peace and abundance. The differing societal structures between the US and China may increase the probability of finding peaceful outcomes compared to nations with identical organizational structures.
A KYC framework for AI models makes sense to prevent dangerous applications like biological weapons. Both countries already implement internal review processes for training runs before model releases. The goal should be establishing reasonable ground rules while recognizing this as an existential race where both nations need to reach parity before considering de-escalation.
Testing frontier models before release through a battery of safety assessments covering bioweapons and terrorism risks was discussed. This would function similarly to FDA testing for poisons or contaminants in food and drugs. Self-regulation, court systems, and existing legal recourse already incentivize responsible behavior by model makers. OpenAI is currently being sued by a family whose child died by suicide after interacting with their model.
Government regulatory power tends to expand once granted and rarely gets rolled back. Existing legal frameworks provide post-harm recourse mechanisms that encourage responsible development.
Cities including those in the UK, Boston, and New York are considering banning or severely limiting self-driving vehicles to preserve human driver jobs. Autonomous vehicles present different risk profiles than package sorting automation. With approximately 50,000 automotive deaths annually in the United States and one million globally, the potential for zero-death-rate solutions creates strong liability exposure for cities that reject safer alternatives.
Early Uber adoption demonstrated how quickly people adapt to superior convenience. Cities without autonomous vehicle access may eventually feel unsafe and inconvenient by comparison. America's patchwork of state and municipal regulations allows different approaches to compete, historically leading to positive outcomes.
Flock Safety uses AI-powered cameras to monitor criminal activity at the municipal level rather than through top-down federal mandates. States can regulate deployment, creating a bottom-up approach to AI governance. Crime reduction becomes a policy choice that different jurisdictions make differently, with voters ultimately deciding through their feet and ballots.
Cambridge City Council recently voted to disable gunshot detection systems. Las Vegas implemented comprehensive crime prevention including gunshot detection and rooftop-deployed drones that provide site access within minutes. This system costs $30-40 million annually but delivers measurable safety outcomes.
Flock Safety implements privacy safeguards including rolling databases with maximum retention periods of 2-3 years, no facial recognition capability, and mandatory audit trails. These backend protections address privacy concerns while maintaining system effectiveness.
Tech layoffs now extend beyond addressing previous over-hiring to include AI-driven workforce reduction. Cloudflare CEO Matthew Prince laid off over 20% of staff despite record revenue growth, specifically targeting "measurers" - people who manage others and measure data. The company plans to add headcount in other positions.
Meta conducted additional layoffs of 8,000 people while simultaneously installing recording software on all employee computers to train AI models. Employees who built AI efficiency tools during internal hackathons were subsequently laid off. The perception has emerged that the best workers can hope for is temporary employment while training the AI systems that will eventually replace them.
Labeling laid-off employees as "measurers" creates lasting stigma that follows workers into future job searches. Tech CEOs excel at product development but often fail at communicating difficult workforce decisions. The contrast between executives stating AI will eliminate jobs while simultaneously authorizing massive share buybacks and dividend increases amplifies public anxiety.
SpaceX filed its S-1 on Wednesday targeting a $75 billion raise at a $1.75 trillion valuation, which would be the largest IPO ever. The listing is expected mid-June with ticker SPCX. Starlink generated $11.4 billion in revenue last year with 50% growth and $4.4 billion in operating income, serving over 10 million subscribers. The space business contributed $4 billion in revenue with 17% growth but recorded $650 million in operating losses.
The AI business generated $3.2 billion in revenue with more than double year-over-year growth but posted $6.4 billion in operating losses. Total capex reached $20 billion, with over 60% allocated to AI compute infrastructure.
Anthropic pays SpaceX $1.25 billion monthly to rent Colossus 1 and portions of Colossus 2, representing a $45 billion deal over three years or $15 billion annually. This single contract effectively quadruples the AI business revenue. The agreement includes 90-day cancellation provisions for either party.
Data center construction timelines improved from 122 days for the first facility to 91 days for the second and 66 days for the third. SpaceX builds data centers faster and at lower cost than competitors. Cursor's Composer 2.5 model achieved state-of-the-art performance after 3-4 weeks of reinforcement learning on Colossus 2 using Cursor's proprietary coding token dataset.
The jump from Composer 2 to Composer 2.5 demonstrates what happens when an appropriate amount of reinforcement learning is applied using Nate data, or when it is injected into the pre-training of a new base model. Composer 2.5 uses the same base model as Composer 2, which is Kimmy K.25. This progress occurred over just three or four weeks, and Composer 2.5 is now paro dominant. When plotted on a curve of the blue dots, Composer 2.5 sits well outside the paro frontier after only three weeks.
The next development will be a new base model that includes the Cursor model, followed by reinforcement learning on the Cursor model using the world's largest coherent compute cluster. This development is extremely significant for XAI and Cursor. Cursor had been struggling with access to compute and falling behind Codex, Google, and Anthropic, but after being granted access to Colossus, their models began growing faster. There is a possibility that within a year, Cursor could become the dominant player, and Elon could be selling compute to Google and OpenAI.
Google is already positioned to participate in orbital compute. Companies deeply involved in data centers and space, including Google, Anthropic, Amazon, and Nvidia, are convinced that orbital compute will become reality. SpaceX is extraordinarily well positioned for this development. The Composer 2.5 data point is considered powerful and worth monitoring.
The newest version of Grok 4.3 sits on the paro frontier for all frontier models. The companies currently on the frontier are XAI with one build of Grok 4.3, a 500 billion parameter model, Google 3.1 Pro, OpenAI, and Anthropic. These are referred to as the four horsemen. Google each have one dot on the paro frontier, though having multiple dots is preferable.
Grok previously lacked a harness, which Claude and OpenAI already had with their respective tools. With Grok Build, a harness is now available to Grok as a downloadable app that translates into English and integrates with tools like Notion, Gmail, and Slack. Without this harness, using the model is comparable to using a basic chatbot from years ago. Grok Build is more than just an app. It functions as a runtime and environment that manages state and memory, making the models dramatically more useful.
In an agentic world, the harness is considered essentially as important as the model itself, and the two need to be developed together. The release of Grok Build and the pace of iteration at XAI are encouraging. Elon had been publicly frustrated with the state of affairs at XAI but is now less frustrated and shipping faster.
A potential merger between Tesla and SpaceX would create an entity with a market cap placing it among the fourth largest company in the world. The discussion addresses whether an anti-tech, anti-AI, anti-progress societal trend could lead governments to restrict information access, freedom of speech, and commerce. The internet has long provided an open alternative to physical commerce through digital channels, but it must sit physically somewhere.
The current assault on data center buildouts in the United States highlights the importance of an alternative internet built from the ground layer up. A communication network not restricted or controlled by governments on Earth could serve as a backup for civilization and for progress. Space-based communication networks, space-based data centers, and space-based communication back down to Earth via wireless represent a valuable backup system against potential restrictions on speech, commerce, and information flow.
When Elon was starting SpaceX, the original idea was to back up the biosphere. He considered putting geodesic domes in space containing plants, wildlife, and creatures. The necessity of getting materials into space led to the decision to build rockets in-house. The concept of putting data centers in space appears technically doable, especially given that Starlink now has 10,000 satellites in orbit despite earlier skepticism. While a Starlink satellite and a data center satellite are physically different, the conceptual leap is not insurmountable with Starship, which is ten times bigger than previous rockets.
Underwriting SpaceX at a $2 trillion valuation requires examining last year's $19 billion in revenue, with expectations of $25 to $30 billion this year. The investment case centers on purchasing the most important internet infrastructure project since the internet itself, which will scale to hundreds of millions of users as it becomes cheaper and more useful. This represents a delivery infrastructure growing at GDP plus 10 to 15 percent. The underlying platform enables everything else.
The valuation also includes an AI business spanning apps at the top level and compute capability at the bottom layer. Colossus is valuable to Anthropic because few entities can deliver an actual gigawatt data center. Next year's revenue is projected at $40 to $45 billion, potentially doubling the following year, placing the company at 20 times revenue. Revenue multiples are justified because revenue provides operating leverage to invest in businesses that consolidate differentiation and competitive mode through a capital mode that accelerates technology mode, execution mode, and learning mode.
Terrestrial data centers alone could generate $100 to $200 billion in revenue by 2030 to 2032 just from building. Getting a gigawatt nameplate working is extremely difficult. Jensen Huang is pushing for DC to DC power delivery to eliminate inefficient power transformations, reduce costs, power requirements, and cooling needs. This requires a fundamental re-architecture where direct current comes in and goes directly to the rack, and Elon is positioned as a natural design and thought partner for this effort.
Starship is large enough that 300 people could fit inside, comparable to the interior space of a 747. Starship is expected to get significantly larger according to the roadmap. SpaceX is about to dwarf the rest of the world's cumulative payloads into space from 2012 to today. Most historical mass to orbit was achieved by Falcon, which is reusable. Starship is designed to be rapidly reusable, a critical difference from standard reusability.
Standard reusability involves extensive refurbishment that may allow reflights in 30 to 60 days. Rapid reusability enables the same rocket to fly and land multiple times per day. This capability is essential for achieving a moon base, a colony on the moon, a colony on Mars, and mass drivers on the moon. A year or two is the consensus timeframe for achieving rapid reusability, though SpaceX learns from every attempt regardless of outcome.
A working Nvidia H100 GPU is already in space today, and Andre Karpathy has both trained a model on it and used it for inference. Nvidia is developing a space-designed version with a different heat sink, reduced weight, and reinforcement for the journey to space to withstand G-forces that data center GPUs are not built to handle.
SpaceX has demonstrated the ability to use semiconductors not originally designed for space, as seen with chips in Starlink satellites that passed radiation testing despite not being radiation hardened. A point prediction for orbital compute becoming reality is the second half of 2028 to the first half of 2030.
Nvidia reported Q1 revenue of $81.6 billion, representing 85 percent year-over-year growth and 20 percent quarter-over-quarter growth. The company achieved $58 billion in net income and $48 billion in free cash flow at 75 percent gross margins. The stock has risen 16 percent year-to-date despite this performance. Nvidia announced an additional $80 billion in buybacks on top of the $100 billion initiated at the start of 2023, buying back approximately 4 percent of the company. The quarterly dividend was raised from 1 cent to 25 cents per share, and the CFO stated the company plans to return 50 percent of free cash flow to shareholders.
Valuations across AI companies appear inconsistent. Memory makers trade at 3 to 5 times PE, Nvidia trades at a relatively low PE, and some accelerator companies have reasonable multiples, while power, cooling, and optical names discount very different outcomes. If power, cooling, and optical multiples are correct, Nvidia and memory companies should rise substantially. If Nvidia and memory multiples are correct, the other segments will likely underperform. The AI market is cross-sectionally inefficient.
Nvidia changed its reporting structure to separate data center and AI from edge, with AI further divided into hyperscalers, AI clouds, and industrial and enterprise. On an apples-to-apples basis within the Western AI world, excluding China, Nvidia's AI business is growing faster than Broadcom's, which guided for 143 percent year-over-year growth in AI semiconductor revenue. Jensen Huang has expressed frustration with the stock's performance relative to record results.
Nvidia continues to gain market share despite narratives suggesting otherwise. The company is growing faster than hyperscalers or capex even without adjustments. A key frustration is that competing chips like ASICs are not being submitted for benchmarks—they are absent from SIM analysis in inference max and ML Perf. The reason appears to be that they would lose in direct comparison, making it impossible to fight shadows until clean benchmarks exist comparing GB300 versus TPU V7, Inferentia, or Tranium. Tranium is viewed as being in a strong position but similarly avoids submissions.
Once a company becomes the largest in the world, it tends to trade in stairstep patterns where the multiple compresses repeatedly due to skepticism about size, followed by rerating that establishes a new floor at a higher valuation. Nvidia's Q1 results included the announcement that their CPU business is projected to reach $20 billion this year, making them overnight one of the world's largest CPU manufacturers.
Nvidia's unique position stems from working with every lab, which puts them in the best position to architect chips through co-design aligned with where models are heading. This reflects a broader shift toward domain-specific architectures happening inside Nvidia rather than as an external threat.
The Grok transaction highlighted that domain-specific architecture evolution is occurring within Nvidia itself. The company engages in direct DC-to-DC design programs with partners like Elon Musk for Colossus 3, representing game-changing developments.
A significant insight addressed amortization schedules for CoreWeave and other neo-clouds. The bear case argued that GPUs and CPUs were being amortized over four, five, or six years—too short given a true two-year lifespan—meaning profits were overstated. Michael Bur popularized this view.
With disaggregated inference and domain-specific accelerators, components can be mixed and matched. The GPU remains central to the constellation, with accelerators like Grok or Cerebras handling decode in front of older GPUs. This extends older GPUs' useful life to 10 or 15 years.
CoreWeave secured asset-backed financing at 6%, representing a profound advantage when financing GPUs at lower rates than other chips. Nvidia's quarter single-handedly saved the neo-clouds. CoreWeave CEO Michael Intrator stated that customers have no problem financing GPUs over six-year periods and are signing six-year contracts with advance payments. If customers doubted the six-year lifespan, they would not commit to six-year contracts.
Oil remains elevated with potential settlement talks in the Iran conflict now in week 12, far exceeding the original four-to-six-week timeline. Oil is driving inflation higher. Polymarket shows a 99% chance May inflation reaches 4.2% or higher, while professional forecasts project CPI hitting 6% in Q2—a major revision from the recent 3% handle. The narrative has shifted from anticipated Fed rate cuts to potential rate increases.
Bond yields are rising in response: the 10-year Treasury reached 4.6%, exceeding Bessen's target of under 4%. Internationally, Japan's 30-year yield hit a record 5.1%, UK yields reached the highest level since the global financial crisis, and German yields hit the highest since 2011. Korean retail investors are borrowing record amounts to trade AI chip stocks, echoing their previous crypto peak behavior.
Global debt-to-GDP stands at 310%. The spending problem at federal, state, and local levels across countries creates a cascading effect that eventually breaks, leading to massive inflation as currency value collapses, followed by money printing that inflates asset values to service debt.
The 30-year Treasury at 5.2% and Japanese yields could catalyze a credit crisis through carry trade unwinds. However, with America being the most advantaged by current conditions due to energy self-sufficiency, food self-sufficiency, and status as the world's largest oil and gas producer and exporter, the dollar faces no immediate crisis. The U.S. maintains the best currency, best economy, best public markets, and best private market companies.
The recommended approach focuses on owning businesses representing the future for a 10-year horizon while avoiding speculation. Portfolio concentration should be limited to five or fewer stocks to allow proper oversight and sleep-well-at-night ownership. Top holdings might represent 20-40% and 15-20% of net worth respectively, with only four key areas tracked closely.
Managing over 100 positions with a team of 30-plus professionals allows for different perspectives. Rising rates are concerning, yet AI fundamentals show Anthropic growing faster than any company or country in history at massive scale and now achieving profitability. The tech bubble occurred with 10-year and 30-year yields much higher than today, and Cisco traded at 100 times forward earnings while Nvidia trades at low-to-mid teens multiples of real earnings.
America's overwhelming use of natural gas for electricity generation provides a cost advantage, as NG1 is down this year while LNG prices elsewhere rose 100-200%. The Strait of Hormuz closure, while terrible for everyone, is relatively good for America and Trump's policy goals of re-industrialization and self-reliance. Every day the strait remains closed benefits American energy independence.
AI fundamentals have historically shown seasonality due to reduced ChatGPT and Claude usage by college students during summer. With enterprise AI adoption, whether fundamentals remain seasonal remains to be seen.
The 48-hour tech CEO and presidential visit to China produced handshakes and positive vibes but no definitive policy breakthroughs on tariffs or chip sales. Some movement occurred with H100s and A200s sold to BYD and top entities, along with soybean and aircraft sales. No grand deal reducing tensions materialized.
Putin met with Xi following the U.S. visit, creating a relationship-bonding moment between China and Russia. The visit was viewed as useful and productive behind closed doors despite limited public announcements, with alignment achieved on geopolitical tic-tac-toe regarding Venezuela, Iran, and Taiwan.
Selling deprecated Nvidia GPUs to China may lower the probability of China developing alternative ecosystems that would be more power-hungry due to earlier optical adoption for scale-up fabrics. This approach could stabilize the world and represent the highest-probability path for America maintaining AI leadership.
American engagement with China remains beneficial to avoid the Thucydides trap. Trump's strength lies in his ability to bond with dictators, monarchs, and Gulf monarchies. Speculation exists that Taiwan arrangements might involve 20-, 30-, 100-, or 200-year handoff deals with the condition that conflict does not occur during the current administration.
The communicated message to China emphasized that wars consume vast amounts of oil. With Venezuela, Iran, and potentially Russia off the chessboard, China would face severe constraints if conflict arises. Russia alone cannot supply sufficient oil, and closing the Strait of Hormuz would eliminate Middle Eastern oil access, leaving only Russian supply.
A more stable world is anticipated following resolution of the Iran situation. The episode concluded with appreciation for guest Gavin Baker and acknowledgment of his father-in-law Jeff Painter.
Keep All-In Podcast in your library
Save the videos and channels worth coming back to, and find them again in one place.
