Gavin Baker - AI Market Jitters - [Invest Like the Best, EP.485]

1. Key Themes

The AI Demand Signal is Unambiguously Positive Despite Market Panic

The market sold off 40-60% in AI names over July, yet every on-the-ground quantitative metric points to acceleration. Gavin could not find a single negative data point after actively pressure-testing.

"However you cut it, whether you cut GPU availability, whether you cut GPU rental pricing, whether you cut the spot price of DRAM this month, token growth, everything is actually accelerated." [00:03:19]

"I literally spoke to a company this morning who rented a cluster of several thousand Blackwells... somewhere in the mid $2 per GPU hour. They're renting the exact same size cluster... and they're hoping seven months later to pay just under $4 today. Like that's pretty crazy because again, you would expect a really gentle decline in prices would be bullish. Instead, we're up depending on the starting point 50 to 60% in six or seven months." [00:15:46]

The Contract-vs-Spot Pricing Gap Is the Most Underappreciated Dynamic in Tech

Everyone who signed long-term compute contracts in 2024-2025 locked in prices that are now a fraction of current spot. As those contracts roll off, hyperscaler operating cash flows will materially accelerate — meaning analysts are modeling the wrong earnings power.

"You have everyone in 24 and 25, even if you were really bullish, you thought that GPU prices would decline slowly... I don't think anyone in 24 or 25 thought that the prices of old GPUs would be going vertical. Everybody thought, hey, we're going to be smart, we're going to sign these long-term contracts... you have the contracted base of installed compute trading at a massive discount to the current spot market." [00:04:33]

"There are $1.3 to $1.4 trillion in hyperscale operating cash flow. If you just assume they monetize at a discount to current Blackwells, then it's more like $2 trillion of operating cash flow. And that kind of takes $700 billion of credit demand out." [00:12:21]

Open Source Tokens Are Bullish, Not Bearish, for AI Infrastructure

The market interpreted the rise of open-source models (GLM 5.2, Kimi K3) as negative for AI infrastructure — a fundamental analytical error. Tokens require the same compute regardless of source, so open source only transfers margin away from frontier model operators into infrastructure.

"A token is a token. And you need the exact same amount of compute to make a token all else equal. It takes the same amount of flops, the same amount of memory, the same amount of watts... All open source taking share does is take margin dollars out of the frontier model layer... you're literally just taking margin from frontier models and essentially driving more margin dollars into the AI infrastructure layer." [00:08:05]

"Jensen is the world's largest supporter of open source... Does it really stand to reason that Jensen would be the world's biggest supporter of open source if it was bad for his business?" [00:09:25]

The Credit Risk Is Real But Likely Self-Resolving

The rise in real yields and widening of credit spreads (including NVIDIA CDS blowing out) is the one legitimate concern. But Gavin argues the problem dissolves if operating cash flows accelerate as contracts reprice — meaning the self-funded buildout thesis becomes self-reinforcing.

"If we need credit to fund this buildout, this is a significant negative... debt-fueled buildouts, they demand immediate repayment. So if supply and demand get a little bit out of whack, things can unwind very, very quickly. That's what happened in the internet." [00:11:58]

"As long as we're in a compute shortage, which I'm just like desperately trying to find a single sign that we're not in one and that it's not actually getting worse almost by the day, it's almost like the problem becomes the solution." [00:20:46]

NVIDIA's New "Credit Wrapper with Revenue Share" Business Model Is Massively Underappreciated

NVIDIA has invented a novel financing structure — wrapping GPU purchases in credit with a revenue share above a floor price — that simultaneously strengthens its competitive moat, increases revenue per gigawatt, and solves the cash flow mismatch for buyers.

"They've rolled out this really clever new business model, which I would describe as kind of like a credit wrapper with a revenue share if GPU prices are above the floor. And this could lead to them having a really giant cloud business effectively through royalties really quickly." [00:37:43]

"It significantly increases their revenue per gigawatt. And then it also strengthens their competitive position." [00:40:42]

LTA Game Theory Makes Oversupply Scenarios Far Less Likely

The long-term supply agreements now signed between hyperscalers and memory/GPU makers have created a game-theoretic equilibrium where breaking a contract could destroy a company's entire AI competitive position for years.

"Let's just say Google breaks an LTA. There's an oversupply... they break their LTAs. Well, if they're breaking their LTAs, it probably means oversupply, prices are coming down. And then, you know, capacity naturally contracts. Well, what do you think is going to happen to Google's allocations? And then, you know, this is a cyclical industry and oversupply is followed by undersupply. What do you think they think is going to happen to their allocations next time?" [00:35:45]

"You might blow up your entire business and your franchise by breaking an LTA. And that was never the case before." [00:36:44]

AI Regulation Is the Largest Unpriced Risk — and the Industry Is Blowing Its PR

Regulatory backlash, driven by widely-circulated misinformation about data centers, is the one risk that could actually derail the buildout. The AI industry has done a terrible job telling its own story.

"An author made a mistake in a book. It overestimated the amount of water usage in data centers by 10,000x. Not a little bit. Like not one order of magnitude. Not two orders of magnitude. Not three. She's admitted that mistake many times. I was completely wrong. It's been super debunked." [00:53:54]

"New York, it just feels like, is the first of many. And even in some of these deep red states that are super pro-growth, they're just like, hey, you guys are not doing a good job telling your story." [00:56:23]

Tokens as % of Total Comp Is the New Leading Indicator for AI-Native Companies

Token spend relative to labor spend is emerging as the defining metric separating AI-native companies from the rest — with leading companies hitting 20-30%+ ratios.

"In the really pilled companies, it gets really high. 20%, 25%." — Patrick O'Shaughnessy [00:31:48]

"Our friend Dylan Patel at his company, he's an ASI maxi, but he's at 30%. That's probably the highest one I've heard. I've actually heard of 50. And there's $25 trillion in knowledge work. Let's take your 20% number. That's $5 trillion." [00:32:00]

SRAM-Based Disaggregated Inference Is the Overlooked Next Wave of Compute ROI

Disaggregating inference into its three components — pre-fill, attention, and feed-forward network — across specialized chips (including SRAM accelerators) will dramatically improve economics of the installed compute base.

"The ultimate holy grail is if you could do pre-fill on one chip that probably doesn't have HBMDRAM, do the attention on a super high-powered chip with HBMDRAM, and then do the feed-forward network on one of these SRAM chips... The ROI on adding these SRAM accelerators to the existing installed base of compute and new compute... you just can't beat SRAM in particular for that feed-forward network." [00:57:21]

---

2. Contrarian Perspectives

Claude (the AI Model) Is Homogenizing Stock Market Interpretation — and Creating Artificial Volatility

Most investors are now feeding every piece of news through Claude, producing correlated, probabilistic interpretations that move markets in lockstep — recreating a "Walter Cronkite" effect but for financial markets, and producing full cycles in days rather than years.

"It's almost like we're back to... Walter Cronkite, only voice of truth... it's kind of Walter Cronkite for the stock market. And everybody just believes whatever it says. By the way, it's really smart. But it's not always right." [00:21:41]

"This guy TBU... he posted this amazing chart of Japanese capacitor stocks. And he said, we've had an entire capacitor cycle in six weeks... you've already had what probably would have normally been a three-year cycle in like six weeks." [00:22:41]

Open Source Tokens Massively Inflating the Value of Frontier Tokens, Not Destroying It

The consensus view is that open source models commoditize AI. Gavin argues the opposite: cheap open-source tokens acting as task executors actually make a high-IQ orchestrator model more valuable, not less.

"It may be that these cheaper tokens massively inflate the value of the most cutting-edge frontier tokens. Because if today you have 120 IQ open source models and they're really cheap to run, doesn't that make a 160 IQ model that can orchestrate them more valuable?" [00:50:04]

NVIDIA's Valuation Is Paradoxically Cheapest at the Moment of Maximum Competitive Advantage

NVIDIA is at its lowest forward PE in 10 years, at the exact moment when its credit wrapper model, matchmaking role in land/power, and ecosystem lock-in make its competitive position arguably the strongest ever.

"NVIDIA is actually, as we record this, at its lowest forward PE of the last 10 years. Crazy. The only time the semis have been cheaper were Liberation Day and DeepSeek. And those were kind of bottoms." [00:18:17]

"NVIDIA's dominance, the current environment, the extent to which it favors NVIDIA, it is a little hard for me to understand why it's trading at such a low multiple." [00:37:14]

SpaceX's Compute Business Is Almost Entirely Absent from the Stock's Consensus Estimates

Public markets are pricing SpaceX as a satellite company with some AI adjacency. Gavin argues the compute buildout alone — if even partially executed — makes current estimates look wildly conservative.

"Grok 4.5 and Cursor, I think the sum of that probably hits a $10 billion ARR pretty quickly. Forget all of that... consensus estimate is $73 billion. And that's 8 gigs at $50 billion a gig... I think very little is built in from my perspective to that stock for the amount of compute that they might be able to bring on." [01:00:28]

AI Natives Prove Labor Substitution Through Not Hiring, Not Mass Layoffs — Which Means It's Invisible to the Market

The narrative of mass AI-driven layoffs is wrong. The actual mechanism is simply that AI-native companies don't hire humans they otherwise would — making the substitution invisible to traditional labor statistics while being highly real.

"For sure, I think in a lot of these AI natives, you're seeing labor substitution, but not because they're firing people. They're just not hiring nearly as many humans. The gross profit dollars per FTE at A16Z, Iconic, a bunch of companies that have done this work, they're vertical. Particularly relative to past generations of startups." [00:31:07]

---

3. Companies Identified

NVIDIA

Designer of GPUs dominating AI training and inference compute. Mentioned as the central company in AI infrastructure with a novel credit-wrapper-plus-revenue-share business model, lowest forward PE in 10 years despite strongest-ever competitive moat, and equity stakes in essentially every major AI lab.

"Nothing's more financeable than an NVIDIA GPU. Nothing... they've rolled out this really clever new business model, which I would describe as kind of like a credit wrapper with a revenue share if GPU prices are above the floor. And this could lead to them having a really giant cloud business effectively through royalties really quickly." [00:37:14]

Anthropic

Private frontier AI lab, operator of Claude models. Described as having been in the "pole position" of AI labs, with shareholders chomping at the bit to correct any perception of deceleration.

"Anthropic is clearly in the pole position. And oh, by the way, Grok and Cursor have also, you can see from third-party data, like July was a pretty transformational month with Grok 4.5, Grok builds coming out." [00:19:41]

OpenAI

Frontier AI lab behind GPT models. Described as having "gotten back into the game" with meaningful acceleration in July.

"Open AI has accelerated. And it's rapid. And it's rapid continues to grow really strongly and is almost certainly pumping out significant amounts of free cash flow." [00:03:45]

xAI / Grok (SpaceX AI division)

AI lab associated with Elon Musk, developing Grok models and acquiring Cursor. Cited as newly on the "Pareto frontier" of AI labs after Grok 4.5 launch.

"Grok 4.5, the cursor acquisition. Cursor has clearly accelerated meaningfully. They've shown over the last three years they can bring on more compute faster than anyone at lower prices." [00:59:09]

Cursor

AI coding assistant. Cited as having demonstrated a 15x efficiency gain by using frontier models to plan and delegating to smaller models for execution, and as having accelerated following the Grok acquisition.

"AI is speed running what we've learned amongst humans, which is you could use the frontier model to plan and then farm out tasks to the dumber models. And it's 15 times more efficient or whatever the metric was." [00:48:33]

Fireworks AI

Open-source inference cloud. Cited for its Nexus product, which in three lines of code ingests proprietary data, RL-trains a custom model, and routes queries between the custom model and frontier models — described as "the solution for every AI native."

"Fireworks did come out with a really cool product called Nexus. And if you're using Claude Code, OpenAI Codex, Grok Build, it is literally three lines of code, like 20 words. And Fireworks ingests your data. They can RL a model. And then there's a router that sends the query. And they've had amazing results." [00:47:13]

"Lit at Fireworks. She is an absolute killer." [00:58:42]

Together AI

Open-source inference cloud. Mentioned alongside Fireworks and Modal as growing nearly as fast as frontier labs while burning very little cash.

"These inference clouds have gotten really good at supervised fine tuning and reinforcement learning... Together, Modal, Base 10, they're all working in a very cash-efficient way. What's shocking about those business models is they're growing almost as fast as the frontier labs in the early days, but burning very little cash." [00:27:52]

Modal

Open-source inference cloud. Same context as Together and Fireworks — extraordinary unit economics relative to growth.

"Together, Modal, Base 10, they're all working in a very cash-efficient way." [00:50:28]

Base 10

Open-source inference cloud with similar growth and cash efficiency profile to Together and Modal.

"Together, Modal, Base 10, they're all working in a very cash-efficient way." [00:50:28]

Safe Superintelligence (SSI)

Stealth AI lab co-founded by Ilya Sutskever. Cited as focused on continual learning and sample-efficient learning breakthroughs, expecting a model release in August, and actively working with NVIDIA.

"SSI says that they're going to come out with their model in August. There's this whole generation of new labs that are focused on this." [00:25:29]

Cognition

AI lab founded by Scott Wu. Cited as a dark horse Game of Thrones-scale player, and creator of the Cognition Index showing that companies spending the most on AI are growing meaningfully faster.

"The bull case, you've seen charts from Cognition, Ramp, and Stripe, that the companies that are spending the most on AI are growing meaningfully faster. Yeah, I love that Cognition Index. The Cognition Index is wild." [00:32:55]

Harvey

AI legal platform. Cited as a model for how AI natives use frontier models plus custom RL-trained open-source models to achieve defensibility.

"Harvey, Legora, all of them. Because if you can go from just using one, two, or three frontier models to using those frontier models for whatever it is, 30% to 60% of your token consumption, and then use your own RL model, all of a sudden you're not a wrapper. You're way more defensible." [00:48:11]

Legora

AI legal platform. Mentioned alongside Harvey as an example of an AI native using router architecture for defensibility.

"Harvey, Legora, all of them." [00:48:33]

CoreWeave

Neo-cloud GPU rental company. Cited as one of only a handful of companies globally that have successfully brought on more than 500 megawatts of power in a single year.

"The only companies that have brought on more than 500 megawatts of power in a year are the hyperscalers, CoreWeave, Crusoe, and SpaceX." [01:00:55] (quote begins at [01:01:05])

Crusoe

Neo-cloud compute company. Mentioned alongside CoreWeave as having executed a 500+ megawatt power buildout.

"The only companies that have brought on more than 500 megawatts of power in a year are the hyperscalers, CoreWeave, Crusoe, and SpaceX." [01:00:55]

StarCloud

Orbital compute startup funded by Benchmark. Cited as evidence that orbital compute is becoming real, with SpaceX providing Starlink laser technology as a partner.

"Our friends at Benchmark, they funded StarCloud. And I don't know, last time StarCloud is an orbital compute company that SpaceX is kind of partnering with. They're going to, I think, let them use the Starlink laser technology, which is really important for orbital compute." [01:02:45]

SK Hynix

DRAM / HBM memory manufacturer. Discussed as the paradigmatic example of a memory company that should replicate NVIDIA's credit-wrapper business model.

"I'd do the exact same thing NVIDIA is doing right now... I would be going to the buyers of GPUs, Tradiums, and whoever and say, I'll participate in the NVIDIA credit wrapper." [00:38:53]

Micron

US memory manufacturer. Cited in the context of LTA game theory and Apple's historical ability to strong-arm memory suppliers.

"Their volume is so big that even if they super screw Hynix, Micron will of course take them." [00:36:44]

Etched

AI chip startup (ASIC-based). Cited as an example of an alternative compute provider that gives hyperscalers a credible threat if they break LTAs.

"You're an investor in Etched. If you break an LTA, they just say, OK, fine, great. You broke the price agreement. We're going to break the volume agreement. And screw you. We're going to give the volume to your competitor." [00:37:14]

Meta

Hyperscaler. Cited as a company that briefly spooked markets by appearing to rent out compute, but whose CapEx plans remained unchanged and whose model Muse 1.1 was a meaningful quality leap.

"They just reported they didn't cut CapEx... they released their best model in a long time, Muse 1.1, which is actually a very good model." [00:06:21]

Ramp

Corporate spend management platform. Cited alongside Cognition and Stripe for publishing data showing AI-heavy companies grow faster.

"You've seen charts from Cognition, Ramp, and Stripe, that the companies that are spending the most on AI are growing meaningfully faster." [00:32:55]

Stripe

Payments company. Same context as Ramp and Cognition — cited for research showing correlation between AI spend and growth.

"You've seen charts from Cognition, Ramp, and Stripe, that the companies that are spending the most on AI are growing meaningfully faster." [00:32:55]

ASML

Dutch semiconductor lithography equipment maker. Referenced in the China DUV context — their EUV machines represent the 25-year gap China is trying to close.

"The market massively overreacts. And then if this ever hits ASML's orders, maybe it hits it in five years." [00:44:48]

Black Forest Labs

AI image/video generation lab. Briefly named in the context of an essay about market homogenization driven by AI-interpreted news.

"This company, Black Forest Labs, I think that's their name. Because there is an interesting essay that got sent to me." [00:20:46]

Blackstone

Alternative asset manager. Cited as likely already approaching memory companies about adopting credit-wrapper-style financing structures.

"I'm sure our friends at Blackstone and Apollo are suggesting some variant of this to the memory companies." [00:39:17]

Apollo Global Management

Alternative asset manager. Same context as Blackstone — expected to be structuring creative financing solutions for the memory supply chain.

"I'm sure our friends at Blackstone and Apollo are suggesting some variant of this to the memory companies." [00:39:17]

A16Z (Andreessen Horowitz)

Venture capital firm. Cited for publishing data on gross profit dollars per FTE going vertical at AI-native companies.

"The gross profit dollars per FTE at A16Z, Iconic, a bunch of companies that have done this work, they're vertical. Particularly relative to past generations of startups." [00:31:31]

Benchmark

Venture capital firm. The conversation was recorded at Benchmark's offices; Benchmark also funded StarCloud, the orbital compute startup.

"We're sitting in the Benchmark offices. Yes, this is their famous table for their famous dinners." [01:03:41]

---

4. People Identified

Gavin Baker

CIO/Managing Partner at Atreides Management. Hedge fund investor with deep focus on technology and semiconductors, regular guest on the podcast. Main speaker throughout, providing all major analytical frameworks.

"I literally spoke to a company this morning who rented a cluster of several thousand Blackwells... They're renting the exact same size cluster... and they're hoping seven months later to pay just under $4 today." [00:15:46]

Jensen Huang

CEO of NVIDIA. Cited for his idealistic, patriotic pro-open-source stance — which Gavin argues is not charity but reflects Jensen's certainty that open source is good for NVIDIA's business.

"Jensen is the world's largest supporter of open source. He is like a super idealistic guy. He's a patriotic American. I think he always does what's right. But does it really stand to reason that Jensen would be the world's biggest supporter of open source if it was bad for his business?" [00:09:25]

Dylan Patel

Founder of SemiAnalysis. Cited as having tokens represent 30% of total company spend — the highest ratio Gavin had heard at the time.

"Our friend Dylan Patel at his company, he's an ASI maxi, but he's at 30%. That's probably the highest one I've heard." [00:32:00]

Dorkesh (Dwarkesh Patel)

Podcast host and commentator. Cited for an extremely bullish prediction that H100 GPU rental could reach $250,000 per year — roughly 15x spot — a scenario not even in Gavin's "considered but dismissed" outcome space.

"Dorkesh, he's a very smart guy. He's very plugged in. Then he pointed out that margins on compute are going up. The amount of compute is going up. And inference margins going up. And if you multiply those three, that's how you're getting this crazy acceleration." [00:43:03]

Scott Wu

CEO of Cognition. Named as a dark horse candidate to become a Game of Thrones-scale player in AI.

"I think our friend Scott Wu. Cognition is kind of... You're here to that one. Yes." [00:58:42]

Lit (at Fireworks AI)

Leader at Fireworks AI (first name only given). Described as "an absolute killer" — one of two individuals named when asked who could become a major unexpected player.

"Lit at Fireworks. She is an absolute killer." [00:58:42]

Dario Amodei

CEO of Anthropic. Cited for a prescient four-months-prior commentary on the existential compute dilemma — too much spend risks bankruptcy, too little risks losing the AI race.

"I think four months ago that Dario was talking about how it was a really thoughtful commentary. But he's like, it's really, really hard. Because if you buy too much compute, you could go bankrupt at the scale of these things. But if you don't buy enough, you could lose." [00:42:05]

Satya Nadella

CEO of Microsoft. Referenced for his framing that AI either drives 10% economic growth or drives labor substitution — the two ways AI ROI ultimately flows back to customers.

"Kind of Satya's comments like either we're going to start growing 10% or we're not." [00:31:07]

TBU (anonymous, "semiconductor mafia" on X)

Anonymous semiconductor analyst on X/Twitter. Cited for producing a viral chart of Japanese capacitor stocks demonstrating a full multi-year industry cycle compressed into six weeks.

"There's this guy TBU. He's like part of the anonymous semiconductor mafia on X. Actually a very smart guy. I know him in real life. But he posted this amazing chart of Japanese capacitor stocks. And he said, we've had an entire capacitor cycle in six weeks." [00:22:11]

Anish Acharya / Vishria (Anish Acharya, a16z)

Referenced as "our friend Vishria" — likely Anish Acharya at A16Z. Cited for early observation, two years ago, that companies were reaching $50M ARR and cash flow in nine months.

"It's like our friend Vishria. I think he said two years ago. I've never seen more companies go from being founded to like $50 million a year in revenue and generating cash flow in like whatever it is, nine months." [00:46:45]

Elon Musk

CEO of SpaceX and xAI. Cited extensively for SpaceX's compute buildout track record and the aphorism about making the impossible late.

"I think one of Elon's phrases is, we specialize in making the impossible late." [01:01:52] (timestamp [01:01:59])

Eric (at Benchmark)

Partner or associate at Benchmark VC who coordinated the podcast recording at Benchmark's offices.

"Eric coordinated for me, so he gets a special shout out. Thank you, Eric." [01:03:41]

---

5. Operating Insights

Use a Router Architecture to Escape the "Wrapper" Trap and Build Defensibility

AI-native companies that rely entirely on one or two frontier models are vulnerable to commoditization and margin compression. The winning operational playbook is to generate domain-specific data, use it to RL-train a custom open-source model, and deploy a router that sends 30-60% of queries to your own model and the remainder to frontier models. This simultaneously cuts token costs, improves outcomes, and creates a proprietary moat.

"If you can go from just using one, two, or three frontier models to using those frontier models for whatever it is, 30% to 60% of your token consumption, and then use your own RL model, all of a sudden you're not a wrapper. You're way more defensible." [00:48:11]

"Fireworks did come out with a really cool product called Nexus. And if you're using Claude Code, OpenAI Codex, Grok Build, it is literally three lines of code, like 20 words. And Fireworks ingests your data. They can RL a model. And then there's a router that sends the query." [00:47:13]

Track Tokens as a Percent of Total Comp as a Leading Indicator of Organizational AI-Pilling

The ratio of token spend to total compensation is the emerging signal of how deeply AI is embedded in an organization's operations. Leading companies are at 20-30%, with outliers reaching 50%. This metric predicts growth outperformance better than traditional inputs.

"In the really pilled companies, it gets really high. 20%, 25%." — Patrick O'Shaughnessy [00:31:48]

"Our friend Dylan Patel at his company, he's at 30%. That's probably the highest one I've heard. I've actually heard of 50. And there's $25 trillion in knowledge work. Let's take your 20% number. That's $5 trillion." [00:32:00]

The Cognition Index Is the Cleanest Available Proof Point for AI's Business Impact

Companies spending the most on AI are growing materially faster, according to data from Cognition, Ramp, and Stripe. Even skeptics who note the lack of industry controls find that drilling down into specific verticals (e.g., HVAC contractors, plumbers) confirms the pattern across blue-collar and white-collar alike.

"The bull case, you've seen charts from Cognition, Ramp, and Stripe, that the companies that are spending the most on AI are growing meaningfully faster. Yeah, I love that Cognition Index. The Cognition Index is wild." [00:32:55]

---

6. Overlooked Insights

Continual Learning and Sample-Efficient Learning Could Temporarily Suppress Training Demand — and Multiple Labs Are Close

Gavin briefly mentioned that "a lot of people seem to feel like they are very close to solving continual learning and sample-efficient learning." This was passed over quickly, but it is potentially the single biggest technical risk to the AI infrastructure buildout thesis. Current frontier models are trained on 300 trillion tokens; if models could be trained on 10 trillion tokens and then learn sample-efficiently in deployment, training compute demand could collapse temporarily. SSI expects a model demonstrating this in August. NVIDIA is already embedded with all these labs — suggesting they are hedging this scenario internally.

"A lot of people seem to feel like they are very close to solving continual learning and sample efficient learning... if instead of a model trained on 300 trillion tokens, you can train something on 10 trillion tokens that you let out into the world and learn sample efficiently, that doesn't sound good for training demand... SSI says that they're going to come out with their model in August. There's this whole generation of new labs that are focused on this." [00:25:00]

Orbital Compute Is Closer to Reality Than Anyone in Public Markets Is Pricing

Gavin spent time at Starbase and said "orbital compute feels more real every day." Benchmark — not an Elon-adjacent fund — funded StarCloud, an orbital compute company that SpaceX is partnering with on Starlink laser inter-satellite link technology. The fact that rational, independent capital from Benchmark is funding this without inside access is a meaningful signal that a completely new class of AI infrastructure — outside terrestrial power and land constraints — may be emerging on a shorter timeline than anyone expects.

"I did spend a lot of time at Starbase. And orbital compute feels more real every day." [01:02:33]

"Our friends at Benchmark, they funded StarCloud... an orbital compute company that SpaceX is kind of partnering with. They're going to, I think, let them use the Starlink laser technology, which is really important for orbital compute. But I do think that's like kind of a good sanity check... maybe I'm crazy. And maybe Elon's crazy. And maybe Benchmark is also crazy. And maybe the SpaceX engineers are also crazy. But man, that just doesn't seem that probable to me." [01:02:45]