Inside TomorrowX: Brian Armstrong, Gavin Baker, Guillermo Rauch, Andrew Feldman & More
- 01AI demand is accelerating while public markets price in a demand cliff
- 02Speed of inference is becoming the buying criterion
- 03Hardware and physical infrastructure is the new venture frontier, but timelines are underestimated
- 04Full-stack vertical integration is more innovative and more secure
- 05Cybersecurity: AI arms race where defenders must use AI, and tokens are the new loot
- 06Agentic finance and decentralized AI
1. Key Themes
AI demand is accelerating while public markets price in a demand cliff
Gavin Baker argues the market is running a flawed "capex vs. revenue" calculation. He says the market's math undercounts open-source and enterprise on-prem demand. "Stocks are trading like, you know, the world is not, stocks are not trading like the world is ending. But stocks are trading like there is an imminent issue with demand. And I just don't think that's what we've heard here today." 00:57:54. He cites Fireworks' Lin Qiao: "Exeter top customer. It is 13x in the first nine or ten months of this year" 00:59:17. He also cites Michael Dell: "in the first two years of offering AI factories, which is 24 and 25, they sold 3,200 of them. And they've sold, I believe, 3,300 in the last nine months." 01:00:15. On the second-derivative fear he says "it may also be that this is like the railroad build out and it goes for 60 years" 01:01:33, and he frames the present as "a voting machine phase... ultimately, the market is a weighing machine" 01:01:04.
Speed of inference is becoming the buying criterion
Andrew Feldman of Cerebras says the market has flipped from asking why speed matters to demanding it. "I think a year and a half ago, some people were like, well, why do we need it to be fast? And now they're like, I use it every day. It needs to be fast. Faster." 00:08:42. He claims "we build the fastest AI compute on Earth. Not by a little bit, but 10 or 15x" 00:08:52, and says "in the last six months, we won AWS and Microsoft and Oracle" 00:09:20, on top of OpenAI as a customer.
Hardware and physical infrastructure is the new venture frontier, but timelines are underestimated
Keller Rinaudo Cliffton of Zipline describes the whiplash in capital allocation: "CapEx was like a radioactive zone for venture capital. And it's weird how fast... now actually the only thing that is economically defensible is hardware and infrastructure." 00:36:48. His warning is that investors underestimate timelines: "Tesla took 15 years to really achieve meaningful scale. SpaceX took 20 years... Zipline, we've been at it for more than a decade" 00:37:14. The real friction is mundane: "how do you make the actuator reliable? Oh, how do you handle like supplier industrialization engineering for the motor that you're producing with the supplier? Like, do we have good enough end of line?" 00:38:54. Matteo Franceschetti of Eight Sleep echoes this: "The hardware is really, really, really hard." 00:46:58.
Full-stack vertical integration is more innovative and more secure
Keller Rinaudo Cliffton: "I think SpaceX is an amazing example of like, you know, as full stack as it gets, raw materials in one side of the factory, finished product on the other, and then you're operating that product. You're not selling it to another customer. Full stack is going to be easier to secure. It's also easier to innovate because you control everything." 00:36:05. He contrasts this with "an air traffic control tower that might be buying software from like 20 different places and trying to like kludge it all together."
Cybersecurity: AI arms race where defenders must use AI, and tokens are the new loot
Brian Armstrong: "now that the attackers are using AI, we've got to be using AI on defense." 00:52:36. At Coinbase "every line of code before it goes into production... is being scanned by one of these cyber expert models" and sensitive systems are designed to be "resilient to any individual engineer or employee having malware on their laptop" 00:53:01. Guillermo Rauch of Vercel adds a non-obvious incentive shift: "the targets of this hacks are all companies that have access to AI tokens... Tokens are very precious. They're a fungible with us dollars... the reward for hacking has now increased essentially." 01:32:47. His prescription is to rebuild infrastructure "in memory safe programming languages... in ways that are almost like provably unhackable" 01:29:37.
Agentic finance and decentralized AI
Brian Armstrong frames Coinbase's AI push as a political/economic project: "We don't want to live in a world where there's one or two AI overlords... we need a decentralized marketplace of open weight models that can collaborate." 00:51:09. The mechanism: "You need prices, you need free market economies, you need financial accounts for those agents... payment rails so that agents can assemble into teams, hire each other, make payments in real time all over the world." 00:51:34.
Open-source American AI vs. closed frontier labs
Gavin Baker wants a "vibrant American open source ecosystem... I do not want to live in a world dominated by one, two or three frontier AIs. Because these AI systems are imbued with value systems." 01:07:59. The speaker identified as the poker champion and Just Games founder (Liv Boeree-style game theorist) offers a competing view: the delta between closed and open will likely widen "absent some more funding being collected for an open ecosystem. Because... over time, the model companies will stop selling the tokens and offering their models and just create useful products and services out of them" 01:25:06. Andrew Feldman of Cerebras says "I think the open source models will continue to get better and better... we need both." 00:22:45.
Export controls, Huawei, and the Jensen Huang removal
Gavin Baker argues that Jensen's pro-"first, best, and most" stance kept Huawei weak: "to make the hardware better it needs to be used. Now that Huawei hardware is going to be used by every Chinese AI lab. I think it might get better really fast." 01:06:08. The game-theorist speaker pushes back on the lock-in thesis: "China is not dumb... They are bottlenecked by chips so they can use marginal chips where we are in the opposite position." 01:19:41. Both agree the chip-sales debate is the real stakes behind the Nvidia leadership drama.
Hardware-enabled product businesses win on retention and operating leverage
Matteo Franceschetti on Eight Sleep: "We disclose publicly that we are crossing a billion dollars in cumulative sales... We sell in 37 countries" 00:46:11, still entirely direct-to-consumer. Return rates are "record low for a consumer product" and "Our retention is better" than Oura 00:48:28. Productivity per head is "probably 3 to 5x higher than any other hardware company" 00:50:35.
Employee wealth creation and radical transparency as culture
Andrew Feldman on post-IPO Cerebras: "we made more than a thousand millionaires. That's something I'm really proud of. I think that's our job as leaders, is to make our guys wealthy." 00:10:39. His transparency practice, learned from Mark Leslie, yields low attrition: "We have people who've been with me now for 25 years." 00:21:06.
2. Contrarian Perspectives
"Capex is a four-letter word" flipped to "hardware is the only defensible asset", and investors are unprepared
Keller Rinaudo Cliffton says the consensus rotation into hardware is right directionally but naive: "very few of these investors have actually any meaningful experience, like setting up factories or building infrastructure or like permitting construction projects in the real world... the next three or four years are probably going to be a lot of real world learnings." 00:39:27. He points to the most-funded companies: "even looking at companies like OpenAI and Anthropic trying to build data centers... It turns out it's hard." 00:39:55. The mental model he offers: "the performance of the project is equal to the lowest performing supplier... if there's a zero anywhere in the chain, then the overall project is a zero." 00:40:33.
Developing countries are the real AI/robotics battleground, not US data centers
Keller Rinaudo Cliffton: "We're really focused on like data centers and building data centers in the U.S. But the reality is that the vast majority of people in the world live in places that aren't the U.S. or China." 00:34:30. The geopolitical framing: these countries "want trade, not aid" and will choose between "a U.S. tech stack... [or] the Chinese tech stack" 00:34:30. Zipline began in Rwanda precisely because it was illegal to fly beyond visual line of sight in the U.S. for a decade, and that constraint forced it to build to hundreds of millions of autonomous miles abroad first.
The "lock them into our stack" case for selling chips to China is a six-dimensional-chess rationalization
The game-theorist speaker: "when you're finding yourself making a very convoluted argument for this, I think it's time to question the argument a bit more." 01:19:16. Her points: China won't actually be locked in, and she cites that "Anthropic reported that they were able to switch to AMD chips from one tech stack to another where it took them six to 12 months before to only three days" 01:20:07. She adds that arguments from shareholder-value motives get disguised as philosophy: "when someone argues with motivated reasoning and pretends it's for other benefits, that is when our conversation gets broken." 01:24:13.
Hardware as a lock-in moat is dissolving, and Nvidia is becoming "a car"
The game-theorist speaker: "Jensen said, I'm not a car... Now he is a car. Like people can just go in, use you, step out, use another one." 01:22:02. If models can port between chips in days, software-based hardware moats shrink. She still notes "He is the best car that exists."
Human enhancement drugs will mint more trillion-dollar companies
Brian Armstrong: "Eli Lilly and Novo Nordisk got a trillion dollars of market cap because of these drugs. And they started off treating disease, you know, but now they're being used by healthy people for enhancement... I think we'll see more human enhancement companies hit a trillion dollars in market cap." 00:55:28. He ties it to his biotech, NewLimit, on longevity.
3. Companies Identified
Cerebras
AI chip and inference company that IPO'd this year. Mentioned for 10-15x faster inference and customer wins. Andrew Feldman: "we have open AI as a customer. But in the last six months, we won AWS and Microsoft and Oracle." 00:08:52. Gavin Baker called the "tone from Cerebras on... the state of AI demand" encouraging 00:57:53.
Zipline
Autonomous drone delivery company, valued at $21.5B after a $1.5B round. Keller Rinaudo Cliffton: Uber Eats partnership "to grow with them to a million deliveries a day, which is about a quarter of their overall deliveries in the U.S." 00:24:50; Walmart expanding "to 300 Walmart super centers in the first phase" with "90,000 of the 100,000 SKUs" 00:25:35; "growing 10 to 15% week over week" 00:27:12; "three plus strategic partners who are each trying to buy more than a million deliveries a day" 00:26:44. Delivery time "between like two and five minutes" 00:30:20. Operates a fleet command center.
Eight Sleep
Sleep-fitness temperature-regulating mattress pods; direct-to-consumer. Matteo Franceschetti: over "a billion dollars in cumulative sales," 37 countries going to 40 00:46:40; hotel (five-star) and airplane partnerships explored, modeled on Starlink with airlines 00:43:43. Customers say "that we save their marriage" 00:43:01. Offices in New York, San Francisco, Boston, Shenzhen, Milan.
Coinbase
Crypto platform building agentic finance. Brian Armstrong: building "an AI enabled financial account, the payment rails so that agents can assemble into teams, hire each other" 00:51:57; stores "99% of all the crypto entirely offline" 00:53:27; serves "seven of the largest banks in the world, majority of the crypto ETFs, 140 government agencies, 50 of the top 100 largest hedge funds" 00:53:27.
NewLimit
Longevity biotech co-founded by Brian Armstrong. "the biotech company I co-founded, New Limit, is doing a lot of cool stuff in this area with longevity." 00:55:57.
Vercel
Frontend cloud platform (Guillermo Rauch is its CEO, though the transcript mislabels his segment). Said in the cyber segment: "when an enterprise buys our platform, we help them deploy new software that they can now produce infinitely faster because they have coding agents... roll out for sale is behind the firewall." 01:30:31.
SpaceX / Starlink / Grok
Gavin Baker: "SpaceX is the most important national defense contractor for America" 01:13:25; "85% of mass from orbit" 01:13:54; Starlink V3 and orbital compute; "I think the size of SpaceX's first-party AI business... is getting pretty close to $10 billion" 01:16:39; Starship catch, "Mechazilla." He calls Grok "by far the most valuable AI tool to me" 01:16:11. Starlink for "every school in Honduras" 01:14:52.
Fireworks AI
Open-source model serving and customization. Gavin Baker: customer growth of "13x in the first nine or ten months of this year" 00:59:17; "they're the Atreides portfolio company. So I think they're the best one doing this" 00:59:48.
Dell Technologies
Sells "AI factories" to enterprises. Per Michael Dell as relayed by Gavin Baker: 3,200 sold in 2024-25 and ~3,300 in the last nine months, "over 6,000 discrete enterprises" 01:00:36.
Astera Labs
Connectivity semiconductor. Gavin Baker: "since they reported earnings two weeks ago, business has accelerated. Things have gotten better." 00:57:22.
Nvidia
Most valuable company in the world. Gavin Baker: leadership turmoil; Jensen has "relationships with the Asian supply chain... unparalleled" 01:10:13; Colette Kress is "a superb CFO" 01:03:14.
Anthropic
Mentioned for ~$80B revenue by Gavin Baker 00:58:48, an upcoming IPO per the host 00:09:43, and a chip-switching report by the game-theorist speaker 01:20:37.
OpenAI
Customer of Cerebras; ~$40B revenue per Gavin Baker 00:58:48; possible IPO next year per host.
Uber Eats
Zipline's distribution partner: "1.2 million merchants who today deliver with Uber Eats" 00:24:50.
Walmart
Zipline partner expanding to 300 super centers.
Huawei
Mentioned by Gavin Baker as the Chinese AI chip ecosystem that benefits from US policy: "Now that Huawei hardware is going to be used by every Chinese AI lab. I think it might get better really fast." 01:06:08.
AMD
Mentioned by the game-theorist speaker: Anthropic "were able to switch to AMD chips from one tech stack to another where it took them six to 12 months before to only three days" 01:20:07.
Taiwan Semiconductor (TSMC)
Gavin Baker: Jensen has "never had a contract with Taiwan Semi... Morris Chang actually long ago asked him to be the CEO of Taiwan Semi." 01:10:13.
Eli Lilly and Novo Nordisk
Cited by Brian Armstrong as the template for trillion-dollar enhancement drugs 00:55:57.
GlaxoSmithKline
Andrew Feldman: "their use of AI is extremely interesting in helping them develop drugs in the future." 00:07:34.
Veritas
Mark Leslie's former company; background on a mentor 00:15:58.
Oura (Aura in transcript)
Wearable benchmark. Matteo Franceschetti: "Our retention is better" 00:48:28.
Microsoft, AWS, Oracle
Cerebras customer wins 00:09:20.
Starlink (as a model for Eight Sleep)
Matteo Franceschetti: hotel strategy "Similar to what kind of Starlink is doing with airlines." 00:43:43.
Just Games
Founded by the game-theory/poker speaker; named by the host 00:19:12, no details given.
Gemini, Grok, and "Muse"
Named by Gavin Baker among smaller AI revenue contributors 00:58:48.
4. People Identified
Gavin Baker (Atreides Management)
Longtime technology investor, 27 years in the business. Insightful on the demand-vs-market disconnect and a close friend of Jensen Huang and Elon Musk. "I've been doing this 27 years and I met one Jensen and I met one Elon." 01:03:14. He co-wrote a letter with Brad Gerstner to Nvidia's board asking for a slow, deliberate approach 01:03:14. (Note: the transcript labels parts of his speech under other names; the investor commentary is his.)
Andrew Feldman (Cerebras CEO)
Founder-CEO; took Cerebras public; "made more than a thousand millionaires" 00:10:39. His leadership philosophy: transparency, "most of the stuff people fear is so dead boring that if they actually learned it, right. They would never think about it again." 00:19:59.
Keller Rinaudo Cliffton (Zipline CEO)
Founder of Zipline since 2013, ten years in "obscurity." Quote: "we are scrappy. We are frugal. We are paranoid... special forces." 00:41:30.
Matteo Franceschetti (Eight Sleep CEO)
Travels "like 250 days a year" 00:45:01 and runs a team with 3-5x the revenue per employee of hardware peers.
Brian Armstrong (Coinbase CEO)
Pioneer in trusted frontier financial tech and co-founder of NewLimit.
Guillermo Rauch (Vercel CEO)
Cyber and AI deployment perspective. "the models have gotten really capable of cybersecurity. And so the bad actors have found the... new methodology for hacking." 01:32:21.
Jensen Huang (Nvidia CEO)
Gavin Baker's friend of 27 years; "I do not know a single CEO who is more patriotic about America than Jensen." 01:06:37. Origin story: arrived at nine, "went to a reform school... He had a 17 year old roommate with knife scars covered in tattoos. He worked as a janitor" 01:06:37. In Taiwan: "they've lit up all of these buildings and these highways with signs saying, bring Jensen back." 01:11:06.
Elon Musk
Cited by Gavin Baker as unique; the game-theorist speaker calls him inspiring for "the pure force of will and nature... how he pursues audacious goals" 01:26:23.
Demis Hassabis
Named by the game-theorist speaker as a top inspiration: "just the most extraordinary thinkers" 01:26:23.
Brad Gerstner
Co-signed the Nvidia board letter with Gavin Baker 01:03:14; said "they were going to put up the first orbital compute satellites in 2027" 01:13:54.
Mark Leslie
Former CEO of Veritas. Andrew Feldman's mentor on radical transparency 00:15:58.
Pierre Lamond
Legendary venture capitalist. "now at 95 or 96, still calls me and asks me if I'm driving hard enough." 00:15:58.
Michael Dell
Delivered the AI-factory data: 3,200 then 3,300 01:00:15.
Lin Qiao (Fireworks AI CEO; "Lynn"/"Lin Chow" in the transcript)
Her business grew 13x in nine to ten months 00:59:17.
Jitendra Mohan (Astera Labs CEO)
Reported accelerating business 00:57:53.
Colette Kress (Nvidia CFO)
"Colette is a superb CFO." 01:03:14.
Morris Chang
TSMC founder; reportedly asked Jensen to be TSMC's CEO 01:10:39.
Walter Isaacson
Elon's biographer, cited for the Mechazilla engineer story 01:13:25.
Rick Pitino
Basketball coach used as an analogy for succession after a legend leaves 01:03:42.
Alice Huang (Alice Wang/"Alice" per the speaker)
Referenced as an alleged spy in the Nvidia episode; no verified details 01:23:20.
Molly O'Shea (host)
Interviewer; noted as "the Molly O'Shea effect" on company valuations 00:23:27.
Becky (Gavin Baker's wife)
Mentioned as the co-creator of a planned family AI 01:08:50.
Jack Ma
Cited as an example of Chinese ideological/short-term-cost behavior 01:21:05.
Secretary Lutnick
Visited Zipline's fleet command center 00:28:30.
5. Operating Insights
Build the product to withstand the unit-of-sale problem
Keller Rinaudo Cliffton's operating plan became one customer's order: "our entire operating plan for the company became like the unit of sale, right? And so we had to take the operating plan, tear it up, throw it into the garbage, remake our plan for the next three years" 00:26:44. Lesson: plan for the scenario in which a single partner consumes your whole 3-year capacity, and pre-negotiate supplier scaling in lockstep.
Make observed behavior the product demo
Zipline's fleet command room converts skeptics: "when people stand in front of fleet command for 20 or 30 minutes... they're seeing a new layer of infrastructure" 00:29:00. Putting live operations on display beats a pitch deck for credibility in a "sci-fi vaporware" category.
Design systems assuming a compromised insider
Brian Armstrong: "We also have designed all of our sensitive systems internally to be resilient to any individual engineer or employee having malware on their laptop. That's a common attack vector." 00:53:01. Also multi-signer withdrawals (3-of-5) and liveness checks like "say the number 37" that an AI agent would struggle to generate quickly 00:55:00.
Give liquidity guidance to newly rich employees
Andrew Feldman: "there's no reason for you to have 99% of your wealth in one stock" 00:12:31; "once you have some cash, it's okay to do nothing with it" 00:14:27; decide "over a 90-day period or over a six-month period" 00:15:13. Pair with a message: "What we should watch is our sales and our growth" 00:11:10.
Radical transparency as retention tool
Andrew Feldman answered any question at all-hands including his salary and stock holdings 00:19:37, reasoning that disclosed information is "painfully boring," which reduces rumor and builds inclusion. Result: low attrition and 25-year tenures.
Offer trade-ins and generous trials to defuse high-ticket hesitancy
Matteo Franceschetti: 30-day trial with free return shipping, and "now we have a trading program... you can trade in your old gen and we will ship you a new one at a major discount" 00:47:46. Low returns then validate the product.
6. Overlooked Insights
AI tokens are functioning as a black-market currency, changing who gets hacked
In a throwaway line in the cyber segment, Guillermo Rauch says "the targets of this hacks are all companies that have access to AI tokens... Tokens are very precious. They're a fungible with us dollars. Right. And so there is, there's black markets, not only for data... but also for tokens." 01:32:47. This implies that compute-rich companies (AI labs, inference providers, and any enterprise with large API budgets) face a structurally higher threat level, and that token-governance and spend controls are becoming a security category. It also hints at distillation attacks as a state-level vector.
The market's AI-revenue math excludes the on-prem and open-source economy entirely
Gavin Baker's aside is easy to skim past: Dell's 6,000-plus enterprise "AI factory" buyers "does not show up anywhere in that calculation that the market is doing" 01:00:36, and because "the margins are lower on these open source tokens, it drags more compute demand with it" 00:59:48. The implication is that revenue-per-GPU is lower but GPU-hours-per-dollar of revenue is higher for open source, so the ratio of capex to visible revenue will always look worse than the underlying utilization suggests. This reframes the "AI bubble" debate as a measurement problem, and points to Dell, Fireworks, and similar open-source infrastructure enablers as undercovered beneficiaries.