Key Takeaways
- Chamath Palihapitiya is staying risk-off because sticky inflation, an elevated volatility index, and climbing 10-year Treasury yields crush multiples for high-growth tech equities.
- Jason Calacanis forecasts a 500x explosion in enterprise token volume as companies shift from single queries to automated agent swarms.
- Enterprise adoption trails startup experimentation by a predictable window of two to three years.
- David Friedberg argues that spending compute to generate existing office work yields weak financial returns, while spending compute on unsolved frontier science yields massive value.
The Disagreement
Tech markets are split between macro anxiety and aggressive AI optimism. The divide on the All-In Podcast centers on where equity values go next, and what actually creates enterprise value.
Palihapitiya looks at the financial environment and sees danger for tech valuations. As the All-In hosts pointed out, “The 10-year unfortunately keeps ticking up. That's bad for equities. V is elevated. That's bad for equities. And inflation is sticky. That's bad for equities.” (play from 23:35) Higher discount rates make distant cash flows worth far less today. In Palihapitiya's view, cheap capital is gone, and multiple expansion will not save struggling software companies.
Calacanis takes the counter position based on raw usage demand. He sees software engineers and knowledge workers delegating complex jobs to autonomous systems. That shift requires an order of magnitude more infrastructure. As the All-In hosts explained, “What that means is at least 500 times the token usage is what's going to happen. Now, where do those tokens come from becomes the question.” (play from 26:26) In Calacanis's view, startups are already running these workflows, and enterprise buyers will follow within 24 to 36 months.
Friedberg rejects the premise that simple token growth creates real economic value. Automating internal summaries or writing drafts is a race to zero. Friedberg stated, “I think the biggest opportunity for AI is on the frontier that is where basically businesses and people can do things that humans and businesses have never been able to do for and that means crafting solutions to really difficult engineering problems to really difficult scientific problems to previously untractable problems.” (play from 31:30) For Friedberg, burning capital on routine task replacement will destroy margins.
Who's Right (and When They're Wrong)
Palihapitiya is right about equity pricing right now. When the 10-year Treasury yield climbs, investors stop paying 20 times sales for businesses with negative operating margins. Founders who assume a macro recovery will bail out their burn rate are misreading credit markets.
Calacanis is right about the direction of consumption. Startups build agent loops that run dozens of calls to resolve one support ticket or write one test suite. As the All-In hosts noted, “What startups do is what the big companies eventually do. It takes two or three years.” (play from 25:39) The demand for inference will explode.
Yet Friedberg exposes the structural flaw in Calacanis's thesis. Just because token volume jumps 500 times does not mean the application owners capture durable profit. Friedberg explained the math plainly: “The return on dollars spent on tokens or whatever is some small number when you're doing it to produce the same stuff versus it being a much bigger number if you use it to produce things that humans have never produced before like some new cancer drug or some new material science.” (play from 32:47)
If your software uses tokens to summarize emails, your customer will treat you as a utility and squeeze your margins. If your software uses tokens to solve an unsolved engineering bottleneck, you create an entirely new asset class.
What to Do With This
Audit your product roadmap by profit potential. Rank every planned feature by whether it produces existing work faster or produces an outcome your customer could never achieve before. Cut the bottom half of routine workflow automation and redirect engineering hours toward domain problems where output quality creates a defensible monopoly.