AI is the most extraordinary wealth-creation event in capitalism's history — and the supply constraints, bubble risks, and power dynamics are just getting started
The speaker opens by framing the current AI moment as unprecedented in capitalist history, citing Anthropic reportedly adding $11B in ARR in a single month — matching the combined ARR of Palantir, Snowflake, and Databricks, each built over a decade. He argues Anthropic has done this while burning ~80% less capital than OpenAI, but contends the company is actively throttling Claude's intelligence due to compute constraints, generating 70% fewer tokens on capped plans. The shift from all-you-can-eat to usage-based pricing is, in his view, what will push OpenAI and Anthropic past $200B ARR.
On infrastructure, the speaker identifies zoning and permitting — not energy or chips — as the binding constraint on data center expansion, per a senior PE infrastructure executive. He introduces orbital compute as the long-run solution: SpaceX racks in sun-synchronous orbit, cooled via radiators, connected by laser links through vacuum, forming a virtual data center. He notes SpaceX already operates the world's largest satellite fleet (~98-99% of all satellites), claims engineers have solved the cooling problem, and argues inference workloads are naturally suited to orbital deployment. The watts shortage on Earth, he projects, begins alleviating in 2027-2028.
Taiwan Semiconductor's capacity discipline is framed as the single most important variable determining whether an AI bubble forms. The speaker argues Jensen Huang could sell $2-3T of GPUs if TSMC obliged, but TSMC's deliberate supply constraints are preventing the kind of debt-fueled overbuilding that characterized the 2000 fiber bubble. Key differences today: GPUs run at 100% utilization, and capex is funded from operating cash flows, not debt. He recommends watching TSMC capacity decisions as the leading bubble indicator.
On the competitive landscape, the speaker sees the Pareto frontier of intelligence-per-dollar now dominated by Anthropic, OpenAI, and xAI's Grok — with Google having lost its TPU cost-per-token advantage through conservative TPU v8 design choices. He raises three structural questions for AI investors: (1) will frontier tokens continue capturing the majority of model-layer economics; (2) will a violation of Sutton's Bitter Lesson emerge as ASI optimizes itself into lower compute demand; (3) will continual learning arrive, and when. At the application layer, he makes the provocative claim that AI has net destroyed value even counting native winners like Cursor.
The session closes on personal safety concerns — noting Molotov cocktails thrown at Sam Altman's home — and a warning that deepfake social engineering (e.g., a simulated child FaceTiming to request a wire transfer) is an imminent threat requiring family safe words and device-free protocols. His bottom line: the machine gun is here, and those who don't master it will be mastered by it.
A wide-ranging investor monologue covers the breathtaking speed of AI revenue growth at Anthropic and OpenAI, the infrastructure bottlenecks constraining it, and the emerging thesis around orbital compute via SpaceX. The speaker also dissects AI chip competition, the Taiwan Semi chokehold on bubble prevention, and mounting risks at the application layer.
The three biggest questions for AI investors are: (1) will frontier tokens continue commanding the overwhelming majority of economic value at the model layer; (2) will there be a bitter lesson violation driven by ASI efficiency gains; (3) will continual learning arrive and if so when.
Elon Musk's superpower of being able to raise unlimited capital whenever he wants stems from 20 years of consistently making investors money, achieved partly by systematically underpricing SpaceX equity rather than maximizing valuation
The machine gun is here. If we do not all become masters of the machine gun, we're going to get mastered.
we need to approach this with humility, recognize there's a lot of uncertainty, and be thoughtful
The Deep Seek paper was published 7 days before Deep Seek Monday, on a Monday that was a holiday in America
AWS availability zone GPU prices in Asia had already doubled by Deep Seek Monday, and GPU availability was going down
Deep Seek revealed that reasoning models are dramatically more compute-hungry during inference than non-reasoning models
Claude on Opus is generating 70% fewer tokens for the exact same question compared to before, indicating Anthropic has deprecated Claude's intelligence due to compute constraints
Closing the Strait of Hormuz caused natural gas prices in Bloomberg to fall 20% in the US while doubling or tripling in Asia and Europe
Tech essentially got as cheap relative to the rest of the market in early April 2025 as at any point over the last 10 years
SpaceX operates the world's largest satellite fleet, comprising approximately 98-99% of all satellites in orbit
No company other than SpaceX is consistently capable of landing and fully reusing an orbital rocket, even 10 years after SpaceX first demonstrated it
The watts shortage for data centers will begin to alleviate in 2027-2028, with orbital compute ultimately solving it
SpaceX's SpaceX engineers are confident they have solved the cooling problem for orbital compute racks
SpaceX now also operates the largest data center on Earth
SpaceX plans to keep orbital compute racks in sun-synchronous orbit so solar panels always face the sun and the radiator always faces away
SpaceX compounded at low 30% per year for roughly a decade
Ramp automates 85% of expense reviews with 99% accuracy
Ukraine is starting to really win the war against Russia
Jensen has never had a contract with Taiwan Semi. They do business on what seems fair in handshakes.
Every GPU is running at 100% utilization, whereas 99% of fiber was unutilized during the year 2000 bubble.
If Taiwan Semi did what Jensen wanted, Nvidia could sell $2 trillion of GPUs in 2026 or 2027, maybe 2.5 to 3 trillion.
Orbital compute is better suited for inference workloads; training will continue to be done on Earth for a long time
Elon Musk built a data center in 122 days while everybody else is taking 3 years.
Nine months ago Google dominated the Pareto frontier of AI models on intelligence-first cost; now the Pareto frontier is dominated by Anthropic, OpenAI, and Grok 4.3.
Google lost their per-cost token leadership as a result of making very conservative design decisions with TPU v8.
AI models have shifted to usage-based pricing, and subscribers on fixed $250-$300/month plans are getting severely rate-limited and a lobotomized version of the AI.
Claude now produces 70% less tokens on rate-limited plans versus usage-based enterprise plans.
The shift from all-you-can-eat to usage-based pricing is why OpenAI and Anthropic will exceed well over $200 billion in ARR this year.
Continual learning — a model that dynamically adjusts its weights in real time — is distinct from current reinforcement learning during mid-training, which only works for verifiable tasks.
Trainium is doing the best among TPU, Trainium, and AMD as GPU alternatives today
Trainium 3 has a switch scale-up network, which you really need to economically inference MoE models
If somebody does something different and it gets to 1 or 2 or 3% share, Nvidia will make that chip
Jensen saw every Taiwan Semi process when it was a twinkle in Taiwan Semi's eyes and knows more about it than a little company with 200 people can imagine
Taiwan Semi is showing Jensen everything, the same way they're showing Amazon everything, AMD everything, TPU everything
Zuckerberg is the only one of those true internet giants to have made Meta an AI first company internally
someone threw Molotov cocktails at Sam Altman's house
A Blackwell rack weighs 3,000 lbs, is 8 ft high, 4 ft deep, and 3 ft wide, and draws 100 kilowatts
one reason Copilot is so bad, or has been so bad, is just not enough compute available
CoreWeave's lowest financing was like low sevens percent
The current AI build out is still overwhelmingly funded out of operating cash flows, unlike the year 2000 internet bubble which was debt-fueled.
18 months ago, the companies focused on coding were Cursor, Cognition, and Anthropic, while OpenAI was doing everything under the sun
Microsoft probably would be an $800 stock today if they were using their GPUs to serve OpenAI and Anthropic's capacity instead of using them for their own products
Anthropic added $11 billion of ARR in 1 month, matching the combined ARR of Palantir, Snowflake, and Databricks which each took 10 years to build
If Anthropic had unconstrained compute, they would probably be doing well north of $100 billion in ARR today, possibly $150 billion
Anthropic could raise capital at probably at least a 100% premium to its latest rumored valuation mark
Starlink V3 satellites will operate at 20 kilowatts
Samsung had to give Elon Musk an office in their fab in Texas because he was so unhappy about the pace at which they were expanding and building.
YouTube data is actually really genuinely valuable in a world of robotics
Grok 4.3 is the best lowest cost 500 billion parameter model and is on the Pareto frontier.
the two companies most deeply engaged with startups are Amazon and Nvidia by a mile, then Google is next most intense
Cerebras has an overwhelming ratio of on-chip compute memory relative to shoreline IO
A Cerebras machine can theoretically run any size model
It took Cerebras three generations of chips to get it right
Disaggregation of prefill and inference means GPUs are going to have 10 or 15-year useful lives
CPUs are way more important than they were in an agentic world due to orchestration and tool calls
Jensen can probably get pretty close to the frontier with his own model whenever he wants
Deep Seek's latest or original model was only 150,000 reasoning traces
The single most useful agent is a really good summary of the points that would be interesting to me from podcasts — there's like 6 hours a day of stuff that feels like it's in my job description to watch
Nvidia in early April was essentially as cheap as it gets relative to the market in the last 10 or 12 years and very cheap on an absolute basis
lower-quality stocks get bid to the moon whereas some of the higher quality expressions have actually really underperformed
in '24 and '25 there was a nuclear bubble and a quantum bubble visible right in front of observers asking about an AI bubble
Astera is a stock I've been close to a long time and I first invested in the series C
Astera was in a lot of copper loser baskets
Astera their biggest product is going to be a switch
definitionally if you're a switch company or an accelerator company, you cannot be a copper loser because you're going to be on the other side of that connection
in '24 and '25 the AI trade traded together so you could be long GPU compute, scale-up networking, and optical scale across and short power and that trade worked
those cross-sectional correlations within AI really fell apart in January of this year
Google had that TPU advantage last year which is now gone
Google has the most compute of everyone and the biggest installed base of compute
Amazon is in a really strong position because of Trainium and you're going to see real P&L efficiencies from robotics over the next 18 months in their retail business
Microsoft flinched for a moment in early '25 and if you flinch, you lose position, you lose all these allocations, and it's difficult to get it back
there is essentially zero engagement with startups from AMD, Microsoft, and Meta
the reason Ukraine is really winning is they have the best battlefield AI outside of probably America and Israel
slavery was really ended by the British Empire
America's most reliable allies today are Israel, South Korea, and Japan — countries America rebuilt after WWII
SpaceX has already showed an illustration of a rack-sized satellite with solar wings approximately 500 ft long on each side
ASML, KLA Tencor, Lam Research, and Applied Materials wanted Taiwan Semi to catch up to Intel because they didn't want a monopsony, and their A teams worked in Taiwan.
The overwhelming majority of economic returns at the AI model layer have come from frontier tokens, which has been surprising.
someone whose daughter was diagnosed with a very rare mutation was able to use AI to identify a drug on the market that can actually impact his daughter's disease and has spun up a company to cure it
Turbo quad is a Google memory optimization written up in a paper a year ago that went viral as a supposed DRAM demand killer, but no AI engineer on planet Earth believed it would have any impact on DRAM demand.
scale up networking would go crazy while scale out was going down or DRAMs massively underperforming NAND and HDDs which had not happened
AI has net destroyed value at the application layer — even if you count Cursor and Cognition, the most successful AI natives, trillions of dollars of value has been destroyed by AI at the application layer
Anthropic has burned approximately 80% less capital than OpenAI to reach a roughly similar revenue scale
Taiwan Semi's lead over Intel and Samsung is whatever it is — 9, 12, 15 months — at the leading node edge.
America after 1945 had the nuclear bomb and no one else had it, but instead of controlling the world they rebuilt Germany and Japan
Elon Musk has done more than any living human to decarbonize the world.
The Terra Fab is a SpaceX and Tesla joint venture to build the world's largest fab in America, with a partnership with Intel.
Coding might be the shortest path to ASI and useful AI because if you really go to coding, you can write yourself code to do anything
If we don't get an AI bubble, Taiwan Semi will have single-handedly prevented it by maintaining supply constraints on wafers.
speculative, lower-quality, smaller-cap names are easy to move if you have a big presence on X or Reddit
Semi-cap equipment companies trading at 40 times next quarter's annualized earnings and DRAM companies trading at mid single digit — at the peak of the last cycle it was like five versus 12
Meta's new model from MSL is not on the Pareto of frontier with XAI, Google's one entrant, OpenAI and Claude but it's pretty close
Chinese open source models involve significant distillation from American frontier models, with many ways to launder this
it is a little dystopian that now the best AI is only available to people with a lot of money
What looks like your son or daughter FaceTimes you — it's an utterly accurate simulation of them knowing everything — and says wire me a million bucks
Because we are in shortages, the lowest quality companies are doing the best — lowest quality players with high costs and high part failure rates are sold out and raising prices
people who were very skeptical a year ago are no longer skeptical
I am more and more worried about personal safety and as AI increasingly becomes political, I worry that's going to get directed at more and more AI political leaders
Zoning and approval have surpassed energy and chips as the biggest gating factors for data center infrastructure investment, according to a senior PE infrastructure executive
A violation of Richard Sutton's bitter lesson — that more compute will always outperform human algorithmic ingenuity — is the biggest risk to the AI trade, potentially triggered by ASI making itself more efficient.
Based on Carlotta Perez's framework, every foundational new technology has always had a bubble: markets correctly identify the technology, diversity of opinion breaks down as everyone turns bullish, supply gets ahead of demand, and a crash follows — especially severe if debt-fueled.
Chip design has an iron triangle analogous to tank design: designers must trade off between competing dimensions (e.g., attack, defense, mobility in tanks), and all chip designers face equivalent fundamental trade-offs.
1% market share in AI chips is going to be worth 100 billion dollars
Prefill is taking in the context and is fundamentally a memory capacity bound problem; decode is generating new tokens and is memory bandwidth constrained
AI got to a quality where it was all of a sudden really easy for a bunch of people to get really smart on these different subsectors, start trading them, and then they get put into baskets and those baskets create price efficiency
There are two kinds of drawdowns: ones where your hypothesis is invalidated and you must crystallize losses, and ones where you profoundly disagree with the price action in companies you know well and can build pent-up alpha by leaning in
In Jensen's five-layer cake of AI, profits are accruing to energy, data centers, chips, and models — not really to applications
If you're a software or AI company of any kind, you have to be in the token path
In a real bull market for a commodity, the commodity suppliers with the highest cost go up the most because they go from on the verge of bankruptcy to just gushing cash
Orbital compute should be reframed as racks in space connected by lasers through vacuum forming a virtual data center, not Pentagon-sized floating buildings
A new prisoner's dilemma: if you're at the frontier, do you release that model via API or not? If everyone at the frontier agrees not to, Chinese open source quickly gains an advantage when one person defects
the companies doing the best today and creating the most economic value are those with the highest effective ratio of utilized GPUs per human
Watch Taiwan Semi's capacity decisions as the single most important indicator of whether an AI bubble will form.
Everybody needs to have a safe word — leave your digital devices behind, go to the ocean, and establish a family or company safe word that can't be socially engineered
“the most extraordinary moment in the history of capitalism, the history of American business”
“This is such a beautiful dream, but it's a dream for our grandchildren.”
“Picture, you know, a British naval ship from the 18th century. Prefill is loading the cannon, decode is firing it.”
“May you find something that you love as much as Gavin loves markets and companies and capitalism and history.”
Shared with Earmark · earmark-ai.com