Google has made a commitment to deploy $200 billion in capex this year in AI infrastructure data center buildout. Because of the capex and accelerated depreciation, making an investment in AI compute in the US right now is hugely tax advantaged. And because of the extreme demand for compute, it's a pretty obvious kind of ROIC model, return on invested capital. So if you make this sort of an investment, you have significant demand for that compute infrastructure. You're very good at running the compute infrastructure. That capital can deliver massive profit returns for you with very high confidence in some forecasted period. Building the most advanced frontier lab driven model also takes tens of billions of dollars of capital. And the question really is can you deliver the profits from the model?

Friedberg articulates the core capital allocation dilemma every big tech company faces: is the return on building frontier models competitive with simply renting out compute infrastructure?

As time has gone on and as everyone has competed on models, as we've talked about many times on the show, I think it's pretty obvious that it is very hard to get the same sort of return on capital invested in model development as it is in capital invested on compute infrastructure and being model agnostic. What Google has is probably one of the greatest install enterprise bases in the world for compute. So they have the most enterprise customers. They have the most consumers. And in both cases, they don't necessarily need to have the best model to make an incredible business. They can be model agnostic.

Friedberg lays out why Google's competitive moat may lie in being a neutral infrastructure provider rather than winning the frontier model race, reframing what 'winning' in AI looks like for incumbents.

The way I would frame it is capex is high alpha low beta in data center infrastructure. That capital and model development theoretically could be high alpha but it's very high beta. It's a very risky way to deploy capital. So if I'm the board, I'm the management, I'm deploying more capital in computing infrastructure, less capital into model development. That's what I think's going on.

Friedberg's alpha/beta framework gives investors and executives a precise vocabulary for evaluating AI capital deployment risk, distinguishing reliable returns from speculative bets.

The scientists who want to be involved in super intelligence, who want to cure cancer, who want to be on the frontier of these models, right? They're sitting there dealing with this channel conflict at Google because Google Cloud wants all of the compute in order to rent it out to Anthropic and those building the frontier models internally want that compute in order to compete with Anthropic. So you have this inherent channel conflict between those wanting to build the models. It looks like it's being resolved in favor of being more of an infrastructure company. So where does that leave — you know, telescope out for a second — SpaceX also reported this week they also have channel conflict.

Gerstner identifies 'channel conflict' as a structural force reshaping every major AI player simultaneously, explaining talent departures, strategic pivots, and stock moves across the sector.

When I saw this Google news, my reaction was and then there were two because like Brad was saying, we used to have five major companies in the hunt to be the leading frontier lab, the leading frontier model just a year ago. Now we're really down to just Anthropic and Open AI. So the market for frontier intelligence has become a duopoly. Now Elon is still on the hunt. I'm sure Google would say they're still on the hunt, but like Brad is saying, they may have contradictory incentives there because they can actually do quite well just with their compute. So I think that the market for frontier intelligence has become a duopoly. I think it's a very powerful duopoly. I don't think it's being commoditized.

Sacks makes a bold structural claim — that frontier AI has already consolidated into a duopoly — which, if correct, has major implications for regulation, investment, and the future of open-source AI.

What you're evolving to is a two-tier market structure where there's a market for frontier intelligence and there's a market for let's call it kind of commodity or lagging intelligence — whatever you want to call it — that's 6 to 12 months behind. There is a market for those tokens, those models, but the reality is you can't charge anything for the weights. You can charge for the compute, you can charge for the inference that you're providing. You can charge for essentially consulting services to help put the whole thing together. But if you're not at the frontier, you can't charge for the model layer itself. If you are at the frontier, you can charge a premium. And that's where Anthropic and Open AI are.

Sacks articulates a clear monetization theory for AI: only frontier models can extract value from the model layer itself, while everyone else competes on commoditized compute and services.

The latest we heard is Anthropic is now over 80 billion of ARR. Started the year at 10. It had forecast 100 billion as exit ARR for the year. And most people said that that would be impossible to achieve. Now it looks like they're going to do it with a couple of months to spare. So their estimates are going up. I mean 110, 120 or higher for end of year ARR. Open AI seeing acceleration. So I think what you're seeing now is a very clear bifurcation in the market.

Anthropic's 8-10x ARR growth within a single year is one of the fastest revenue ramps in software history, validating the duopoly thesis with hard numbers.

Elon comes out and says, 'Not so fast. We're entering the singularity and the frontier models are way further ahead than people think.' I believe that to be true. I think for your use case they're very similar but I don't think that's the most sophisticated use case that people are trying to train on and trying to experience. And then Jensen came out this week and said closed models are actually cheaper — you know, if you don't have to build it for yourself, if you don't have to, you know, the training costs and a lot of expertise to fine-tune and maintain and guard rail and keep it safe. So he's basically making the argument that not only are the frontier models further ahead, but that the cost differential between the two is not what everybody's making it out to be.

Gerstner synthesizes a counter-narrative from Musk and Huang: the total cost of open-source AI (training, fine-tuning, safety, maintenance) may make frontier closed models more economical than they appear.

On the connectivity side, the Starlink side, they generated $2.6 billion in adjusted EBIT. Space was kind of negative 200 million. So call it break even and AI was plus 1.1 billion. But AI, to Brad's point, it's unclear whether the pricing they're getting on compute rental today is temporary and at a premium because of the lack of compute available in the market today. And people that need compute are paying Elon a premium for that compute. So I think there's a question mark where that goes. But the connectivity piece on Starlink, 4.3 billion in the quarter and 2.6 billion in adjusted EBIT. He's got 12 million subscribers. That's doubled year-over-year.

Friedberg's segment breakdown reveals that Starlink is already a massive, profitable cash machine funding Elon's riskier ventures — a structural dynamic most investors underappreciate.

I think there was a really interesting data point that I saw in the commentary on this, which is only 30% of Air Table's sales team was making quota. They had a 30% sales attainment number and that told me a lot about this business. What it told me is — and I'm reading between the lines here, but this was a company that had a successful PLG motion, in other words, organic growth, product-led growth, and they were growing about 20% a year. But that was not good enough for its board. So, what happens? The board pressures the founders to do something that frankly is unnatural for them, which is they say, 'Look, you should bolt on a traditional sales-led motion here to get the growth up faster.' Does that work? No.

Sacks uses Airtable's 30% quota attainment as a diagnostic for a recurring VC-era failure mode: forcing a sales-led growth motion onto a PLG company that was never built for it.

Why isn't that something that the company could do on its own? And I think that's the structural problem is I think it's very hard for both VCs who are on the board and the founders to shift into private equity mode. Why? Because they're going to have to demolition what they've built, right? They've got all this loyalty to the team. They don't want to think about how do I eliminate 80-90% of the cost structure. It's just not what they do. I mean, what founders want to do and the outcome that the board members are going for is a venture-backed outcome. And I think they could have done this. They could do what Bending Spoons does, but they're not built for it.

Sacks identifies a structural mismatch between VC incentives and the private-equity playbook that could rescue many slow-growth SaaS companies — a systemic insight beyond the Airtable case.

I think AI in a way makes their job easier because in the past the reason why you couldn't eliminate like all of the talent, the infrastructure, is because you needed the institutional memory. You needed people who knew the codebase. Now AI can learn the codebase instantly. And so yeah, maintaining is easier with AI. I think maintenance mode becomes way easier with AI because you don't need the historical knowledge anymore. The AI can go in and sort of reconstitute that historical knowledge.

Friedberg surfaces a under-discussed consequence of AI coding tools: they dramatically lower the human capital required to maintain legacy software, fundamentally changing the economics of acqui-PE rollups.

My team just built something that is so mind-blowing that to buy it with off-the-shelf software would have been a quarter million dollars in software and like a million dollars in integration over two or three years. and we built it in a month and and now we have complete insight into the whole portfolio, the competitive set, the founders, everything going on.

A concrete real-world example illustrating how vibe-coding is collapsing the cost of custom software by orders of magnitude.

Keep in mind that one of the reasons why Leopold got blown out, okay? I mean, is because he bet on the SAS apocalypse. Remember, it wasn't just that he was super long these chip stocks that had a correction. He was super he was short Adobe and a whole bunch of other SAS companies. And those trades also moved the wrong way on him. So again, I just think that it's painting with too broad a brush to say that all of SAS is going to get obliterated here.

Uses a high-profile hedge fund blow-up as a cautionary tale against oversimplified macro theses about AI destroying all software categories equally.

Nobody buys Microsoft because Microsoft writes the best code. They buy Microsoft because Microsoft is the rail that everything else runs on. Active Directory is where your employee identities live. Excel is where your boardex numbers come from. Teams is where the compliance recorded conversation happens. Azure holds a Fed ramp high authorization and Department of Defense impact level 5 clearance, which means a defense contractor cannot casually swap it out for something cheaper and so on down the line.

Articulates why infrastructure lock-in, compliance requirements, and regulatory certifications — not product quality — protect incumbent enterprise software from AI disruption.

IGV is up 20% in the last six months. It's up 20% in the last five years. Snowflake's up 88% in the last 6 months. IGV is an ETF of growth software companies. So, to David's point, there was a panic about software companies. There was a big trade out. you know, honestly, they performed pretty well and as he mentioned in the month of July, they were up when a lot of the semiconductor AI stocks were down. And some of these companies, data bricks, snowflake, click house, etc. are doing extraordinarily well.

Market performance data showing data-infrastructure software companies are outperforming despite broader SaaS fears, suggesting a bifurcation between data-layer winners and application-layer losers.

I do think that for these no code a lot of these application software companies they're realizing like that that you know the game is up sell the company get what you can get you know importantly here in the air table story all the latestage investors right we passed on this in the last three funding rounds right which I think we're at 2 billion 5 billion and 11 billion but all those latestage investors which were the most venerable of growth firms they all got their money back and the early stage investors ended up making a lot so if this is a failure This is a pretty good failure for Silicon Valley.

Reframes Airtable's 90% valuation collapse as a capital structure success story, revealing how liquidation preferences protected investors across the stack.

Participating preferred is the double dip, right? Yeah. And look, we've never done that. You know, we believe in clean terms. No one's trying to be punitive towards founders. It's just it doesn't make sense for some people on the cap table to be making money while other people are losing money. It just doesn't make sense, right? Well, that that that's just a transfer of value from some people on the cap table to other people on the cap table. So, the standard thing you do is you make sure that the investors get paid back and then everybody is participating in the upside.

A practical explanation of clean vs. participating preferred terms that reveals the ethical and economic logic behind founder-friendly venture deal structures.

Forbes published an investigation called these American startups are making China's AI smarter. US data labeling startups are selling valuable training data to Chinese labs which in turn is helping them catch up with the US frontier ones. Two startups, Sergei and Meror are both valued at over $20 billion. They sell training data sets to people like OpenAI Anthropic federal agencies. Um they all sell the same data sets to top Chinese AI companies.

Surfaces a major national security blind spot: the same high-quality synthetic training data fueling American frontier AI models is being sold commercially to Chinese competitors.

Look, I think we got to decide what our objective is here. Are we trying to just get in like a full-blown economic war with China? Are we just trying to prevent all of our companies from doing business over there? Historically, the rules have been that you want to be careful about technology transfer of technology that has a dual use, right? That it has a military application. My sense of data is that it's largely a commodity. I mean, data labeling certainly is. If you basically tell them that they can't use data labeling, I guarantee you there's no shortage of labor in China that they can use to do the data labeling. In fact, they probably are.

Applies the established dual-use technology export control framework to AI training data and argues commodity data restrictions would trigger costly retaliation without strategic benefit.

I'm really glad that the first Trump administration limited the export of EUV lithography machines to China. You know, that was all the way back, I think, in 2019. So, that was a really important decision. And so, look, I think targeted strategic controls make sense. I would just make sure that this one actually meets that bar.

Uses the EUV export ban as the gold standard for effective tech containment, establishing a high bar for what constitutes a genuinely impactful strategic restriction versus performative one.

I got to say, using Kimmy and Quen and, you know, GLM 52 for the last 60 days, my lord, these things are good. And I don't think it's very patriotic to be giving them an advantage. Any of these data sets are created by experts here in America who are given like the queries that have errors in them. So when you give um you know a thumbs down to a query that's highly technical, it could be code, it could be biology and science, these are you know PhDs going in there and putting in the latest and greatest content and then verifying it, double verifying it. And that's why we're getting better and better results out of the LLMs. So essentially, you're just helping them catch up. And this could be a big advantage for America if we weren't sending it there.

A firsthand practitioner's assessment that Chinese frontier models have already reached alarming capability parity, combined with a specific mechanistic argument for why Western expert-labeled data is a non-replicable advantage worth protecting.