Andrew Fox | BG2 Summary

1. Key Themes

SpaceX's Overnight Emergence as the #4 AI Hyperscaler

In a matter of weeks, SpaceX went from being absent in the AI compute discussion to surpassing Oracle and CoreWeave in contracted capacity. The deals with Anthropic and Google were entirely off the radar of most analysts' prior forecasts.

"In 30 days, we went from not being an AI hyperscaler to being number four. And we passed a lot of companies, including Oracle. CoreWeave is a huge business." — Gavin Baker 00:18:36

"They just signed Anthropic at 22 to 23. They just signed Google at 50. Right. So I think you can invest behind the AI business terrestrially and still be excited about it." — Andrew Fox 00:24:15

Speed Is a Structural Competitive Moat in AI Infrastructure

The ability to stand up compute faster than anyone else is not just an operational advantage — it directly compounds capital returns and supplier relationships. Jensen confirmed XAI stood up Colossus in 19 days vs. the industry norm of 3-4 years.

"Every day you're paying electricians and plumbers, that's cost. And they're now monetizing them at arguably the highest rate... Speed is literally cost." — Gavin Baker 00:03:41

"A supercomputer that you would build would take normally three years to plan. Right. And then they deliver the equipment. And it takes one year to get it all working. We're talking about 19 days." — Jensen Huang (quoted) 00:14:46

"There's only maybe two or three players now that can actually reliably engineer behind the meter data center. And there's real engineering work that goes into all of this." — Clark Tang 00:20:28

Orbital Compute as a 5x Cost Reduction on Half the Data Center Bill

The orbital compute math is compelling even before accounting for power and cooling savings. The group walked through first-principles economics that point to $5B per gigawatt in space vs. $20-25B on the ground — making this a genuine potential discontinuity in AI infrastructure costs.

"The math that you get to is it's about $5 billion per gigawatt of CapEx to put these in space. For comparison, terrestrially... that today is about $20 to $25 billion per gigawatt. So we're talking about a 5x reduction in cost on half of your bill of materials for the data center." — Andrew Fox 00:26:59

"It costs $60 billion to put a gigawatt on the ground today... You're talking about putting a gigawatt into space for $30 billion and having lower operating costs." — Gavin Baker 00:27:47

Frontier Models Are Extending Their Lead, Not Losing It

The consensus entering 2025 was that open-source models and cheap tokens would close the gap on frontier labs. Six months in, the evidence is the exact opposite — frontier tokens are capturing 90%+ of AI revenues, and long-running agents are widening the moat further.

"I think the consensus at the time... was that open source models, cheap tokens, were catching up on the frontier. That perhaps these models were beginning to asymptote. That people wouldn't really pay for premium tokens. And it seems to me that the evidence on the field six months into the year is just the opposite." — Brad Gerstner 00:46:56

"That has been decisively wrong. Probably more than 90%. And it may continue to be decisively wrong. Frontier might be 90% of the economic value. Open source might be 80% of tokens." — Gavin Baker 00:53:32

Long-Running Agents Have Unlocked a New Compute Demand Regime

Claude 4 Fable/Mythos and GPT-5.5 demonstrate that these models haven't been evaluated anywhere close to their true intelligence ceiling. The implication is that compute demand is structurally higher than anyone modeled, and that the revenue yield per unit of intelligence may be dramatically understated.

"We do not know how smart these models are... Nobody has run Mythos for a year continuously. And we may never know how smart each generation of models actually is or was. Because we don't have time to appropriately evaluate their intelligence before the next model comes out." — Gavin Baker 00:44:14

"Noam Brown post from yesterday, polynomial, is so profound... However bullish I was on compute before then, I'm just a lot more bullish." — Gavin Baker 00:44:14

The XAI/Cursor Combination Is the Most Underappreciated Variable in the SpaceX IPO

The acquisition of Cursor brings not just $10B+ run-rate revenue but one of the most valuable proprietary training datasets on earth — more coding tokens than exist on the public internet. This is being overlooked relative to the easier-to-underwrite infrastructure story.

"Cursor and Anthropic have more tokens of proprietary coding data than anyone else. And they each have more tokens of proprietary coding data than exist on the public internet." — Gavin Baker 00:05:33

"I think the thing that's getting lost is I think they've dramatically advanced their capability when it comes to building a frontier model... This is an extraordinary team that he just downloaded right into SpaceX." — Brad Gerstner 00:31:40

NVIDIA's Dominance Is Accelerating, Not Eroding

Despite years of predictions about ASIC displacement, NVIDIA has actually maintained or gained compute share. The key insight is that in a power-constrained world, tokens-per-watt equals revenue — and NVIDIA leads on that metric.

"Their biggest one thing they emphasized is we thought the world would be consuming less NVIDIA than it is. And if anything, NVIDIA is accelerating and they just continue to out-execute their competitors." — Gavin Baker [00:01:00:47]

"As long as we are in a watt-constrained world, if you can get more tokens per watt, which is literally revenue, with NVIDIA than a lot of alternatives. Just if you build your factory with another chip, you may save some money, but you're going to have less revenue and the margins may be lower." — Gavin Baker [00:01:42:10]

The AI Revenue Ramp Is Already Breaking Prior Forecasts

Dario Amodei's prior public forecast of "hundreds of billions by 2028" already may be conservative. The group believes inference revenue will exceed $200B in 2025 alone, putting the industry on a path to $1T+ before 2030.

"I think we end this year well over $200 billion in inference revenue. Well over. And so I think the math really maths." — Gavin Baker [00:01:05:27]

"It's hard for me to see that there won't be trillions of dollars in revenue before 2030." — Dario Amodei (quoted by Brad Gerstner) [00:01:06:24]

Token Pricing Is Inflationary, Not Deflationary — Opposite of Consensus

The market entered 2025 expecting monotonic token price deflation. Instead, the monetization rate per gigawatt is increasing from ~$20B to $30-40B+ per gigawatt per year, driven by demand that is far outstripping supply.

"Everyone expected token pricing, the price of compute, it's all deflationary... But I think this year what we've seen is the opposite... The monetization rates per watt are increasing." — Andrew Fox [00:01:07:09]

"The willingness to pay for these tokens, when the monetization per gigawatt is actually increasing from, you know, call it like 20 billion in the best of cases for at the beginning of the year to now like 30 to even pushing 40." — Clark Tang [00:01:09:00]

2. Contrarian Perspectives

The SpaceX Employee/Investor Overhang Is Far Smaller Than People Assume

The conventional IPO playbook assumes massive post-lockup selling. But SpaceX employees and investors have had liquidity every six months for a decade — those who wanted to sell already have. The remaining holders are self-selected long-term believers.

"The employees and to a large degree, the investors here have had liquidity every six months. For like the last 10 years. So if you're a SpaceX employee or former employee and you wanted to sell, you've had, whatever that is, close to 20 chances. And it is a matter of historical record that large investors have been able to sell." — Gavin Baker [00:01:41:39]

Open Source Winning on Tokens Is Actually Bullish for Compute, Not Bearish for AI

Most analysts treat open source model adoption as a headwind to AI profitability. The correct frame is the opposite: if open source captures margin from frontier labs, the economics must flow somewhere — and they flow into compute hardware.

"The better open source does, the better it is for compute providers. Because if the frontier models are capturing less of the margin, then you're going to spend more on compute." — Gavin Baker 00:53:58

NVIDIA Could Become the World's Dominant Open Source AI Lab — Whenever It Chooses

NVIDIA is holding back frontier open source model releases as a deliberate business choice, not a capability limitation. If the economic calculus shifts — particularly as hyperscaler customers build ASICs that compete with NVIDIA — Jensen could drop a frontier open source model and restructure the entire landscape.

"I do think NVIDIA is highly likely to be the world's dominant provider of open source AI... If all of his customers are going to compete with him... then why not compete with his customers?... I think NVIDIA can join the frontier and become one of the world's largest cloud computing companies much faster than people think." — Gavin Baker 00:57:41

Data Centers Are Not Commodities — Elon's Are Structurally Different by Design

The market treats AI compute as fungible infrastructure. Gavin's view is that Elon re-engineered data center design from first principles the same way he did rockets and cars — and the result is differentiated in ways that even the XAI team may not fully appreciate as a competitive moat.

"He looked at data center design from first principles. And he designed something fundamentally different... I did actually ask the team. I said, hey, guys, maybe it'd be a little less public about things that are very obvious to you. Right. How to design a data center. But are revelations to other people." — Gavin Baker 00:19:27

The $1.5T CapEx Spend Actually Maths Against Inference Revenue

The headline fear is that $1.5T in CapEx against $300B in inference revenue is unsustainable. But accounting for 60-70% gross margins, the fact that ~35% of spend is non-revenue-generating training, and the trajectory toward $1T+ in revenue pre-2030, the math works — and the prisoner's dilemma means no one can opt out.

"What do you think the gross margins are on that $300 billion? Let's call it 50%... that math starts to math. And what I would just say is I think that $300 billion is low, man. I just think it's low." — Gavin Baker [00:01:05:08]

3. Companies Identified

SpaceX / XAI Vertically integrated rocket, satellite, AI compute, and AI model company. Central subject of the episode — highlighted as a must-own for institutional investors, the emergent #4 AI hyperscaler, and the company with the best-in-class data center monetization rates ($50B/GW from Google deal).

"I don't know another entrepreneur or another business that's a better bet on the future than SpaceX." — Brad Gerstner 00:36:20

Anthropic Frontier AI lab. Highlighted as the existence proof that once on the Pareto frontier, revenue can scale rapidly; released Claude 4 Fable (and Mythos), cited as the revenue event that saved the broader AI market in 2025.

"It's hard to say that Anthropic's not up. After the revenue numbers they put up, after the Fable 5 release, and Mythos is evidently even better." — Gavin Baker 00:43:56

Cursor AI coding agent company acquired by XAI/SpaceX. Cited as having more proprietary coding training tokens than exist on the public internet, and as the key underappreciated variable in the SpaceX bull case.

"Cursor and Anthropic have more tokens of proprietary coding data than anyone else. And they each have more tokens of proprietary coding data than exist on the public internet." — Gavin Baker 00:05:33

CoreWeave AI-dedicated neocloud. Cited as a shared portfolio company of Altimeter and Atreides; mentioned as a company SpaceX surpassed as a hyperscaler in under 30 days.

"CoreWeave is a huge business, right?" — Gavin Baker 00:18:36

Cerebras AI hardware company and shared portfolio company. Cited in the context of the gigawatt allocation breakdown among compute providers; noted as having one gigawatt of contracted capacity.

"Cerebras, our shared portfolio company, has a gigawatt." — Gavin Baker [00:01:01:16]

Harvey AI legal tech company. Highlighted as a best-practice example of using proprietary domain data + open source models + routing to beat frontier model performance at lower cost.

"Harvey had a great blog post that they put out on X. They used their own proprietary legal data to do reinforcement learning and supervised fine-tuning with Fireworks on an open source model. And then they used a router... and they got better outcomes than Opus either 4.7 or 4.8 at a lower cost." — Gavin Baker 00:51:28

NVIDIA Dominant GPU provider. Highlighted as a company that has maintained and potentially gained compute market share against all ASIC competition; also flagged as a latent threat to become the world's leading open source AI lab and cloud provider.

"If anything, NVIDIA is accelerating and they just continue to out-execute their competitors." — Gavin Baker [00:01:00:47]

OpenAI Frontier AI lab. Cited for its Codex model appearing on the coding Pareto curve, GPT-5.5's long-running agent capabilities, and its Jalapeno ASIC being one of the few genuinely impressive custom chips built.

"You know who made a good ASIC? Jalapeno. From OpenAI. Yeah. They made a great chip." — Gavin Baker [00:01:02:23]

Google (DeepMind) Hyperscaler and frontier AI lab. Cited for its $50B/GW compute deal with SpaceX (the highest monetization rate disclosed) and Gemini 3.1 Pro's position on the coding Pareto frontier.

"XAI's deal with Google for cloud computing generates more operating profit per gigawatt than Anthropic, than Meta, than Google, than OpenAI." — Gavin Baker 00:02:53

Replit AI coding platform. Cited via founder Amjad Massad's argument that coding may be the fastest path to AGI/ASI.

"He called it bitter lesson adjacent that coding may be the fastest path to AGI and ASI. Because if you're really good at coding, you can write code... to do anything." — Gavin Baker (attributing to Amjad Massad) 00:22:28

Broadcom Semiconductor company. Cited for its TPU ASIC business (selling to Google/Anthropic), its V8I accelerator discussed at Computex, and as having ~10 gigawatts of contracted AI compute.

"OpenAI gigawatt... NVIDIA has 10. Broadcom has 10." — Gavin Baker [00:01:01:16]

MediaTek Semiconductor company. Mentioned as a new entrant in custom AI accelerators with their V8T chip, representing a more workload-specific ASIC approach discussed at Computex.

"A new class of accelerators or ASICs, MediaTek with their new V8T versus Broadcom's V8I for TPUs, actually was a big topic of discussion." — Clark Tang 00:59:33

Fireworks AI AI inference infrastructure company. Cited as the platform Harvey used to fine-tune open source models with proprietary legal data.

"They used their own proprietary legal data to do reinforcement learning and supervised fine-tuning with Fireworks on an open source model." — Gavin Baker 00:51:28

Nebius AI infrastructure / neocloud company. Mentioned as an example of the competitive neocloud landscape that SpaceX has now stepped into.

"There are probably 50 neolabs being funded in Silicon Valley right now... the Nebiuses of the world, the Irons of the world." — Brad Gerstner 00:18:47

GE Vernova Energy infrastructure company. Cited as an example of a supplier who would rationally prioritize XAI/SpaceX for gas turbines over startup neoclouds because speed of deployment equals money for all suppliers.

"If you're Vernova and you say we only have a certain number of gas combustion engines. Now we can sell them to XAI or we can sell them to one of these startup neoclouds. Who are you going to sell them to?" — Brad Gerstner 00:20:28

Databricks Data and AI platform. Cited as a quasi-public company alongside SpaceX and Anthropic — more liquid in private markets over the past three years than some public biotechs.

"SpaceX and I put Anthropic in this category as well, Databricks in this category. These things in many ways have been more liquid over the course of the past three years than some public biotech companies." — Brad Gerstner 00:42:17

Reflection AI Frontier open source AI lab. Briefly cited by Brad Gerstner as impressive, noting the team led by Misha.

"I'm very impressed by Misha and the team and what they're doing. And I very much want a frontier open source U.S. lab to win." — Brad Gerstner 00:56:07

4. People Identified

Gavin Baker CIO of Atreides Management. Cited throughout as a leading analytical voice on SpaceX, AI compute economics, and semiconductor dynamics; long-term SpaceX shareholder.

"Freida... she calculated a 55% IRR on Colossus I. You know, if you can borrow money at 6%, 7%, 8% and invest in something with a 55% IRR, I'm not the most sophisticated thinker, but that math, maths." — Gavin Baker 00:03:12

Elon Musk CEO of SpaceX/XAI. Described as an "N of one" by Jensen Huang; highlighted for first-principles data center design, speed of standing up compute, and extraordinary business deal-making in a compressed timeframe.

"I do think being all revenue will accrue to the Pareto curve... Elon, I don't think he needs liquidity. And I think he owns, what does he own, Foxy? 50% of the company." — Gavin Baker 00:38:47

Jensen Huang CEO of NVIDIA. Cited for his 2-year-old $1T compute forecast that turned out to be conservative; provided the direct "N of one" quote about Elon's 19-day Colossus build; highlighted as a strategic chess player regarding NVIDIA's open source model strategy.

"We're talking about 19 days. N of one is right." — Jensen Huang (quoted) 00:14:46

Michael (Cursor CEO) Co-founder/CEO of Cursor (acquired by XAI). Described as leading an "extraordinary team" with the most proprietary coding training data in the world; framed as the most underappreciated addition to SpaceX/XAI.

"People outside Silicon Valley may not know, you know, Michael and the team at Cursor as well. This is an extraordinary team that he just downloaded right into SpaceX." — Brad Gerstner 00:31:40

Noam Brown AI researcher (OpenAI). His "polynomial" post on the implications of long-running models was cited as one of the most profound AI observations of the year — the idea that we fundamentally don't know the ceiling of current model intelligence.

"I just think that Noam Brown post from yesterday, polynomial, is so profound... We do not know how smart these models are." — Gavin Baker 00:44:14

Amjad Massad Founder of Replit. Cited for his "bitter lesson adjacent" thesis that coding is the fastest path to AGI/ASI.

"He called it bitter lesson adjacent that coding may be the fastest path to AGI and ASI. Because if you're really good at coding, you can write code... to do anything." — Gavin Baker 00:22:28

Dario Amodei CEO of Anthropic. Cited for his prior forecast on the Dwarkesh podcast that AI revenues would reach "the low hundreds of billions by 2028" and that trillions in revenue before 2030 is hard to rule out — a forecast that now looks conservative.

"He said revenues will go into the low hundreds of billions by 2028... He said, it's hard for me to see that there won't be trillions of dollars in revenue before 2030." — Brad Gerstner [00:01:06:00]

Gwynne Shotwell President and COO of SpaceX. Briefly cited alongside Elon as part of the exceptional SpaceX leadership team responsible for the business deal-making.

"Elon is not only a great engineer, he and Gwen and the team are great at business." — Brad Gerstner 00:39:39

Alex (Whale Rock Capital) Portfolio manager at Whale Rock Capital. Cited for a compelling framing of how nascent AI agent adoption truly is.

"Alex at Whale Rock has this great way to frame it. That less than 0.2% of people on Earth are actually using AI in an agentic way." — Andrew Fox [00:01:07:35]

Freida (Altimeter) Altimeter analyst. Cited for calculating a 55% IRR on the Colossus I data center investment.

"Your colleague at Altimeter, Freida, also, she calculated a 55% IRR on Colossus I." — Gavin Baker 00:03:12

Andrej Karpathy AI researcher. Cited for his tweet about Claude 4 Fable — noting that while it tops benchmarks, what makes it special is long-running tasks.

"There's a Karpathy tweet about this yesterday. He said, you know, it's solid on all the benchmarks, but what really makes it special is long running tasks." — Brad Gerstner 00:42:45

5. Operating Insights

The "Ballast" Framework for Sizing Positions Through Volatile IPOs

Rather than trading in and out of high-conviction names around an IPO, Gavin Baker and Brad Gerstner both independently described the same portfolio construction approach: establish a base position as a set-and-forget holding, and use a smaller "ballast" tranche to adjust exposure as price dislocates — moving ballast toward the position when price is attractive, away when stretched.

"You've talked about you have ballast, you move around and you move the ballast to one side of the ship. When you want the ship to lean into the wind to go faster and you move it to the other side, we don't want the ship to tip over. I think that's a great analogy." — Gavin Baker 00:37:54

Use Claude Fable/Mythos for Cross-Model Contradiction Analysis

Clark Tang described a specific high-value workflow that was not previously possible: feeding multiple investment models simultaneously into the frontier model and asking it to identify internal contradictions across assumptions — essentially a meta-audit of your own analytical work.

"I just threw in like seven of our models and just said, okay, like, I want to create a master view of my beliefs. Given all of these assumptions of all these companies, TSMC capacity — and then produce me a report on all this stuff. And the model is able to reason through all of our assumptions. Like, actually, if you believe this, this thing is inconsistent with this." — Clark Tang 00:49:13

Harvey's Routing Architecture as the Enterprise AI Playbook

Harvey's published approach — proprietary domain data + supervised fine-tuning/RL on open source models + a router to allocate queries — beat frontier model performance at lower cost. This is the reproducible playbook for any enterprise with unique vertical data.

"They used their own proprietary legal data to do reinforcement learning and supervised fine-tuning with Fireworks on an open source model. And then they used a router... and they got better outcomes than Opus either 4.7 or 4.8 at a lower cost." — Gavin Baker 00:51:28

6. Overlooked Insights

Colossus I Delivered a 55% IRR — The Highest Disclosed Return in AI Infrastructure History

This number was mentioned only once in passing, but it is extraordinary. A 55% IRR on a hyperscale AI data center, financeable at 6-8% debt costs, represents a spread of ~47 percentage points — one of the best risk-adjusted infrastructure returns ever documented. If replicable at Colossus II and future sites, this frames SpaceX not merely as a compute provider but as the world's best-returning infrastructure fund, with captive access to the scarcest inputs (power, chips, speed of deployment).

"Your colleague at Altimeter, Freida, also, she calculated a 55% IRR on Colossus I. You know, if you can borrow money at 6%, 7%, 8% and invest in something with a 55% IRR, I'm not the most sophisticated thinker, but that math, maths." — Gavin Baker 00:03:12

Coding May Be the Fastest Path to AGI/ASI — Making XAI's Cursor Acquisition Strategically Decisive at the AGI Level

Amjad Massad's "bitter lesson adjacent" observation — that a model truly expert at coding can write code to do anything, making coding the highest-leverage path to recursive self-improvement — was dropped into conversation in a single sentence and then moved past. But if true, it means Cursor's proprietary training data is not just a coding product asset; it is potentially the most strategically valuable dataset in the race to AGI. The combination of this dataset with Grok's 1.5T parameter base model training is not a product bet — it is a direct shot at the AGI trajectory.

"He called it bitter lesson adjacent that coding may be the fastest path to AGI and ASI. Because if you're really good at coding, you can write code, if a model's good at coding, to do anything... I think coding is going to continue to be very important." — Gavin Baker 00:22:28

"One would hypothesize, based on scaling laws, that that might be a better base model. And then the cursor data is being injected into the pre-training process, not just reinforcement learning." — Gavin Baker 00:30:32