David George & Jack Altman on AI, Autonomy, and the Next $25 Trillion

1. Key Themes

The answer to every "either/or" in AI is "and"

David George's core framework is that nearly every binary debate (frontier vs. open source, labs vs. applications) resolves to "both." His reasoning is that demand is so large relative to current diffusion that there is room for every layer. As he put it: "if the premise of your question is, you know, is this going to be successful or that going to be successful? The answer in AI is probably and." 00:00:00. He grounds this in the current state of supply and demand: "Right now, demand is totally insatiable" 00:03:56 and "we are massively, massively constrained. Like basically in everything in the supply chain. So basically, you know, you can't get data center capacity until 2028" 00:04:54.

AI revenue is enormous but sits on a tiny base of users, so diffusion is the opportunity

George points out that the "fastest success we've ever seen in business" rests on a very narrow base. "OpenAI-Anthropic together. And then if you add in SpaceX AI, now with Cursor, they're, call it like 120 billion of revenue. And they add more revenue per month than all the hyperscalers, except for Amazon" 00:02:45. Yet "30 million coders in the world who are the vast majority of the revenue. There's 1.5 billion knowledge workers" 00:02:45. Within coders it is a further power law: "The top million of the 30 are spending just a massive amount" 00:03:30. His summary: "It's like less than 5% diffused into the B2B economy" 00:04:24.

The infrastructure buildout is already bigger than the railroads, and he expects it to 10x

"Right now, from an infrastructure build out standpoint, we just surpassed the railroads... As a percentage of GDP" 00:01:54. On trajectory: "I think that's going to 10x... this is going to take many years... we will spend many trillions" 00:02:28. Supply is the binding constraint, not demand: "if you get like more OpenAI and Anthropic or SpaceX AI capacity live, it generally gets used" 00:05:31.

Frontier models win today on revealed preference, but N-1 and open source will be massive at 50x usage

George says frontier labs capture "95% plus of dollars" 00:11:41 and have raised "$350 billion" between OpenAI and Anthropic 00:11:41. Users pay up: "people are willing to pay, you know, 10x more for frontier tokens" 00:12:01. But as volume scales, cost matters: "The appetite for N-1 or open source technology is also going to be huge... because if you 50x the usage that happens, like people will care about cost" 00:18:57. Jack Altman adds "N-1 is going to be really good soon. You know, like N-1 will be better than 5-6" 00:19:31. Evidence at the application layer: "how long has it been since most customer support tickets required the highest end model? Yeah. Like three years or something" 00:19:40.

Vertical applications survive when they need last-mile product detail and on-the-ground go-to-market

George's framework for where labs will and won't compete: labs focus on "first-party products that are obviously in coding and within the blast radius of coding" and on products "that appeal horizontally to all 1.5 billion knowledge workers. So think like Microsoft Office and Google Apps" 00:14:36. Everything else is open. On legal: "the last details of the product really matter to the users. And then two, it requires like real go-to-market and on-the-ground efforts" 00:17:34. He estimates legal is "somewhere between number six through 15 on the priority list" for the labs 00:15:36.

Consumer AI is the most underrated opportunity

"It's so funny to me that no one talks about consumer right now... if like historical technology markets are any guide, like the big prize is going to be the killer consumer products because you can address billions of people with them" 00:23:24. Today's usage is "skeuomorphic" (search replacement) 00:22:01. The shift ahead is "from being reactive to proactive," multimodal, and action-taking: "Open claw was the first moment. I think Grokbot is a moment now" 00:22:56. Business model: subscriptions plus a new ad format: "We have yet to figure out like what the native consumption model is going to be and then what the native ad format is going to be that follows that" 00:23:54.

Autonomous driving will 10x the market the way Uber did, and the car-ownership market is larger still

George frames the history: taxis were "less than 0.1% of the overall miles traveled," and when Uber and Lyft delivered a better product, "taxis were a hundred million bucks in San Francisco and then they became a billion dollars, like three years into Uber and Lyft" 00:27:56. Autonomy offers "between 10 and 14 times safer than a human driver" 00:29:00. Economics: "owning a consumer car equates to like 80 cents per mile... riding in an Uber or Lyft is probably like two bucks and change per mile. This market is like massively elastic" 00:29:26. And "17 million cars sold a year in the U S new... minimum $10,000 a car" for autonomy features 00:30:37.

Robotics will be bigger than language, starting with narrow, high-ROI industrial use cases

"It'll be a bigger market than language... because it will have B2B applications, it'll have consumer applications" 00:31:55. The near-term path is "relatively defined use cases with like redundant tasks with safe environments" 00:32:07. He expects "we will have a ChatGPT moment for robotics... I think it's within five years, hopefully" 00:33:04.

A rare convergence of product cycles: the next wave is larger than the $25 trillion mobile/cloud era

George notes the last generation (mobile plus social, e-commerce, SaaS, cloud) "produced like 25 trillion of market cap" 00:26:43. Now: "AI, autonomy, robotics, AI's application to health... bio... American dynamism... Way more than 25" 00:27:27. He scores the cycle: product cycle "a nine or 10 out of 10," capital cycle "probably at a six" 00:38:56.

Growth is where venture returns now happen, and the power law is more extreme

"Half of private market returns get generated between the seed and the B and then half of returns get generated from the C plus" and in five years it may be "70, 30... in favor of the late stuff" 00:36:21. Why the power law is steeper: companies are early-cycle, and "in AI, if you throw more money at it, it can just scale up and get better" 00:36:49.

Narrative and "vibes" are a real economic asset

George argues vibes drive valuation, fundraising, hiring and retention: "the vibes matter because of fundraising, of your valuation, for retaining your employees, and for hiring new employees" 00:44:59. Example: Palantir and Alex Karp. Prescription for founders: "You got to go direct. You got to own the narrative. You're the face of the company" 00:49:33.

2. Contrarian Perspectives

Don't point AI at cost savings, because efficiency is a latent, capped opportunity

Where most companies and boards ask AI to cut cost, George argues the best companies are "singularly focused on front end" product expansion: "you can produce, you know, 10x, 100x more revenue potentially if you nail the next product... Whereas like on the cost side, like you can only get so much more efficient." His conclusion: "It's like shorting your own future if you are just focused on the cost side" 00:10:32.

Coding revenue is not an anomaly, and the "sellers of tokens vs. buyers of tokens" bear case is fundamentally wrong

Jack raised the best bear case he'd heard: revenue is flowing to sellers of tokens, not yet proven in buyers' ROI 00:08:16. George rejected it: "that whole argument is predicated on the fact that there won't actually be tangible, productive use of like writing more code... Writing more code far more efficiently is better... I just think that's like fundamentally wrong" 00:08:47. He cited direct observation: "spending time with Stripe, like they feel like there's very high returns to them being able to be a lot more efficient in writing code" 00:09:19.

Highly publicized vertical-AI skepticism is misplaced: legal was the surprise early adopter

Against the prevailing view that AI wouldn't penetrate legal because of hallucinations, George says: "Legal is a market that's like absolutely in takeoff... end clients are demanding that their law firms use Harvey. They're like, we care about it for product and we care about it for cost" 00:15:06. Even Kirkland's announcement that it would "spend 500 million bucks to build their own technology stack" was a "major validation event" rather than a threat 00:17:01. Jack noted the people on X doubting this are citing "18 months ago" 00:16:11.

Late-stage investing is where "venture" outcomes now happen, because the founder is the asset class

George's view is that the best late-stage founders keep reinventing their businesses beyond any model: "the outcomes of the companies can be much larger than you would ever put in a financial model" 00:42:18. Evidence: Databricks' Ali Ghodsi "has figured out seven years later, his revenue has accelerated," and SpaceX when invested "Starlink wasn't even GA yet... the economics were terrible and like it didn't work. And, you know, like here we are, that is like one of the best business segments ever created" 00:41:51. His valuation philosophy: "we like pay fair prices for great companies... the alpha you get is always on the revenue side" 00:40:31.

Capital in AI is the product, unlike prior software cycles

"If you threw endless amounts of money at ServiceNow or Workday or Salesforce, like during the SaaS era, they would get all messed up. Like we saw this experiment with like the vision fund." In AI, "if you put $40 billion towards training these models, they get much better" 00:38:00. This inverts decades of thinking that too much capital destroys companies.

3. Companies Identified

OpenAI

Frontier AI lab behind ChatGPT, Codex, and consumer AI products. Mentioned as part of the fastest-growing revenue base in business history and for its leadership in narrative. George: "the person who's doing the best job of this in the market is Tebow from OpenAI and like, he's very direct and, you know, not only is it feeding good vibes among the investor community and good vibes among employees, it's driving a lot of behavior on the customer side too" 00:49:07. Also noted: "just OpenAI and Anthropic have raised like $350 billion" 00:11:41.

Anthropic

Frontier AI lab behind Claude and Claude Code. Cited with OpenAI as the core of the ~$120B revenue base: "opening I-anthropic together. And then if you add in SpaceX AI, now with Cursor, they're, call it like 120 billion of revenue" 00:02:45.

SpaceX (including SpaceX AI)

Elon Musk's space/satellite company with Starlink and now an AI business. Mentioned as an example of founders reinventing a business beyond the original model: "When we did SpaceX, you know, like Starlink wasn't even GA yet... that is like one of the best business segments ever created. Yeah. And he's got an AI business on top of it. And he's going to have, you know, orbital compute, which is also going to work" 00:41:51.

Cursor

AI coding product, now part of the SpaceX AI revenue discussion. Described as "dead in the center of the blast radius" of the labs' coding products yet working: "And you know, like cursor also, and that is like dead in the center of the blast radius. Totally. And, you know, there's a huge market of people who use it" 00:18:57.

Replit

AI app-building platform. a16z is "the biggest investors in Replit. Like it's working very, very well" 00:18:37, despite sitting in the blast radius of Codex and Claude Code.

Lovable

AI app-building company. Grouped with Replit as evidence that blast-radius products still win: "Replit and Lovable. And you wouldn't think that these things... aren't Codex and Cloud Code going to be able to make a website? And the answer is, yeah, they can. And also these things are unbelievable" 00:18:37.

Harvey

Legal AI company; a16z is a large investor. "We're large investors in Harvey... it's a market that's like absolutely in takeoff... probably like 12 months behind coding, just in terms of like diffusion into the workforce" 00:15:06.

Legora

Legal AI company, referenced by Jack as a peer to Harvey ("we know this too, obviously. Yeah, for Ligora" 00:15:50). George said the reasoning-models breakthrough applies equally: "lawyering is reasoning" 00:16:25.

Databricks

Data and AI platform. George cites it as a "model buster": "the first one that we did out of the growth fund... right when we started the growth fund and we did a round at like $6 billion... seven years later, his revenue has accelerated" 00:41:05. Also named as one of the companies that "are going to buy for" the abstraction layer 00:20:06.

Palantir

Data/AI software company. Used as the prime example of narrative driving valuation: "the companies that have done a great job of this, like there's a direct line to their valuation... Palantir and Karp. He's got like a cult following" 00:45:35. Business is now "massively accelerated in commercial" and "seen among CEOs as like a trusted place to go implement your AI" 00:46:30.

Tesla

Self-driving and robotics player. Jack: "I just got a Tesla for the first time... it is insane. It is like the best product" 00:25:37. George lists Tesla alongside Waymo as the leading autonomy providers and notes Tesla is "very good" at robotics 00:31:55.

Waymo

Autonomous ride-hail. "Millions of miles traveled in Waymos and they're like between 10 and 14 times safer than a human driver" 00:29:00. Scale is tiny today: "There's like less than 10,000. It's crazy. In the U S." 00:31:11.

Mind Robotics

Robotics company founded by the Rivian founder (RJ Scaringe); a16z partner Sarah led the investment. "He's going to put robots on the factory floor. And so they'll do manufacturing assembly work on Rivians. It's relatively defined use cases with pretty high ROI today... you can also have the models learn from the work that you're doing" 00:32:35.

Rivian

EV maker whose factory serves as the embedded first customer for Mind Robotics. 00:32:35

Stripe

Payments company; named as a customer-side example of AI coding productivity: "spending time with Stripe, like they feel like there's very high returns to them being able to be a lot more efficient in writing code" 00:09:19.

OpenRouter

Model-routing/aggregation platform. Jack noted "the open router idea was very good there too," and George agreed with their thesis that "the economy is going to like oscillate between using dollars and tokens. And like it's going to be a new form of currency... we're big fans of those guys" 00:10:44.

Kirkland & Ellis

Law firm referenced for announcing it would "spend 500 million bucks to build their own technology stack," a moment people feared for Harvey but which George treated as validation 00:17:01.

Microsoft

Cited for the platform principle: "a platform is only a platform if all of the things built on top of it generate more revenue than it" 00:17:01, and for Office as a first-party battleground for the labs 00:15:06.

Google

Referenced for Google Apps as a target for lab first-party products 00:14:36, and as the search paradigm consumers are replacing with chat 00:22:01.

Uber and Lyft

Used as the historical analogy for autonomy's market-expansion: "I was an investor in Uber... it 10x the market size because you delivered a much better product" 00:27:56.

NVIDIA

Referenced via the "is it NVIDIA or new chip companies" "and" framing 00:01:33.

Snowflake

Named alongside Databricks and Palantir as an abstraction-layer competitor 00:20:06.

Salesforce, ServiceNow, Workday

Used as SaaS-era examples where more capital would not have produced a better company 00:37:18.

Amazon

The only hyperscaler that adds more revenue per month than the frontier labs 00:03:13.

Andreessen Horowitz (a16z)

George's firm. Notable stats: "650 employees that we have that help our companies" 00:51:23, ~20% market share in growth 00:52:51, and a deliberate marketing strategy, with Ben Horowitz telling George a firm without marketing was "the dumbest fucking idea I've ever heard" 00:44:35.

Benchmark

Jack Altman's firm, which "just raised our first growth fund in history" 00:34:17.

Anduril

Defense tech company led by Palmer Luckey, used as an example of a founder as the face of the company 00:49:07.

Vision Fund (SoftBank)

Used as a cautionary example: "Well, first of all, like, I don't think it was a bad idea. I think it was just like wrong time of the cycle" 00:37:33.

xAI/Grok ("Grokbot")

Mentioned as a moment for proactive consumer agents: "Open claw was the first moment. I think Grokbot is a moment now" 00:22:56.

OpenClaw ("Open claw")

Agent product cited as the first moment of proactive, action-taking consumer AI 00:23:24.

4. People Identified

David George

General partner at a16z leading the growth fund. The guest. Offers the "and" framework, the 1.5 billion knowledge workers vs. 30 million coders data point, and the product-cycle/capital-cycle scorecard. "Product cycle is like a nine... capital cycle... we're probably at a six" 00:38:29.

Jack Altman

Host of Uncapped and partner at Benchmark. Raised the sharpest bear case ("sellers of tokens, not buyers of tokens" 00:08:16) and noted that for customer support "the technology is fully there. It's just going to take many years to get there" 00:20:18.

Erik Torenberg

a16z partner whose framing "it's all going to work" kicked off the conversation. George: "I am very, very closely aligned with Eric on this point that it's all going to work" 00:01:54.

Ben Horowitz

a16z co-founder. Pushed George to adopt a marketing strategy in growth: "he just like looked at me and said, like, that's the dumbest fucking idea I've ever heard" 00:44:35. Also praised for the scale and ambition he and Marc Andreessen bring to the firm 00:50:26.

Marc Andreessen

a16z co-founder, praised alongside Ben for "ambition and the scale and sort of the mindset" 00:49:57.

Ali Ghodsi

CEO of Databricks. "Amazing CEO" 00:41:05; his revenue accelerated seven years after the $6 billion round because he keeps finding the next product: "it's often the founder that figures out the next thing" 00:41:35.

Alex Karp

Palantir CEO. The model of founder narrative converting into valuation: "he's got like a cult following" 00:45:35.

Elon Musk

Founder of SpaceX, Tesla and related companies. Cited for how a big valuation and a loyal following give "a bigger war chest, to buy companies, to raise more capital" 00:46:53, and for reinventing SpaceX via Starlink and orbital compute 00:42:18.

Palmer Luckey

Founder of Anduril; "ideally, if you have a dynamic founder who can be the face of the company, it's like Palmer with Anderl" 00:48:40.

Sam Altman / OpenAI leadership ("Tebow")

The transcript attributes to "Tebow from OpenAI" the best narrative-building job in the market: "he's very direct... it's driving a lot of behavior on the customer side too" 00:49:33.

RJ Scaringe (founder of Rivian, Mind Robotics)

Described only as "the founder of Rivian" who started Mind Robotics: "he's an exceptional founder. Yeah. Um, and has produced a great company with great products" 00:32:07.

Sarah (a16z partner)

Led the Mind Robotics investment: "one of my partners led investment in Mind Robotics. And so, you know, Sarah did this deal" 00:32:07.

5. Operating Insights

Build your voice early: the founder is the distribution channel

George says narrative needs to be owned directly by the founder, not outsourced: "It's like, you got to go direct. You got to own the narrative. You're the face of the company. You need to tell a compelling story. And if you don't, yeah, some people care, but a lot of people aren't going to care." 00:49:33. He ties this to concrete outcomes: higher valuation means "less dilution, more ability to raise capital. And, you know, you can use that as a weapon" 00:44:59.

Use a valuation lens that treats your own multiple as a recruiting and M&A weapon

Because retail and employee attention follow narrative, a premium multiple compounds: "The valuation is a huge benefit for them. And then it's massive for them retaining and hiring employees" 00:46:53. Operators should treat narrative investment as a hiring and retention expense, not a vanity expense.

Treat the implementation motion as your moat: field engineers and hands-on integration

George and Jack both emphasize that in applications, the defensible parts are the "messy integrations you have to do and rules that you have to follow, brand guidelines" 00:20:44, and "FDs" (forward-deployed engineers) on the ground: "those things like in the ground, like making sure they work, is a big thing" 00:17:34. The labs "aren't going to build that motion."

Don't trust one-off studies of AI productivity; re-test as models cross thresholds

Jack flagged the old finding that AI made coders 20% less effective, and George observed that it predated "the big four or five" model releases and that "the best coders in the world have like fully flipped over on this" 00:09:19. Operators should re-run internal evaluations every model generation rather than anchoring on old conclusions.

Pricing and routing abstraction: plan for the token-vs-dollar tradeoff

George echoed the OpenRouter thesis that "the economy is going to like oscillate between using dollars and tokens" 00:10:44. Operationally, that means application companies should build model-routing logic to use cheaper N-1 models where quality allows, because "that's their gross margin" 00:19:40.

6. Overlooked Insights

Taxi miles were under 0.1% of US miles traveled, so the autonomy TAM is measured against all driving

George casually noted the baseline: "taxi driving, right. Um, that was like less than 0.1% of the overall miles traveled in the U S." 00:27:56. Ride-hail already made that market 10x; the implication is that autonomy's addressable market is not "rides" but the substitution of the entire consumer-car cost structure (80 cents per mile owned vs. roughly two dollars per mile in a ride-share), where he sees an elasticity-driven explosion once autonomous costs fall below both. Combined with "17 million cars sold a year" and "250 million cars in the U S" 00:30:37, the real prize is a software attach on the installed fleet, not a taxi business.

Early-stage models already require two to three orders of magnitude less data than humans' knowledge base, which hints at a training breakthrough

George mentioned in passing "the stats on what it took you and I to like build our knowledge base... two or three orders of magnitude less than what the models actually train on. So there seems to be an opportunity there" 00:06:40. This is among the few identified paths that could slow compute demand, and a16z is already backing "teams that we've backed that are working on algorithmic breakthroughs" 00:06:40. It is a quiet signal that the infrastructure thesis has one genuine, named counter-scenario, and that the firm is hedging it.