Hi, it’s Marc. ✌️
“Nobody really wants to talk about the hedge that all the buildup of data centers is.”
Stephen Messer thinks the trillion-dollar data center buildout makes sense even if the frontier labs never earn it back, because a stockpile of general-purpose GPUs is a strategic asset the day China moves on Taiwan. He puts it plainly: even at a 3% probability, those chips are “worth gold.” Nobody in the market is pricing the buildout that way.
Stephen is worth listening to on calls like this. Last time he sat in that chair, he told us large language models would end up like web browsers: free, everywhere, and worthless as a moat. That’s roughly what happened. He invented affiliate marketing, sold LinkShare for $425 million, and has spent 18 years building Collective[i], an AI company that predicts economic outcomes instead of next words. He wrote up the pricing side of his argument as “Peak Token” on his blog, and jokes he’s been in witness protection from frontier-lab investors ever since.
This one is a playbook, not a podcast: why token pricing power is collapsing, what the forward-deployed engineer push is really about, and the one question a CEO should ask before spending an AI budget.
About Stephen Messer: Stephen is co-founder and vice chairman of Collective[i], one of the earliest AI companies (founded 2008), which trains models on pooled, proprietary transaction data to predict economic outcomes: which deals close, which buyers are real, what demand does next. Before that he co-founded LinkShare, effectively inventing affiliate marketing, and sold it to Rakuten for $425 million. He publishes two posts a week at reloadnyc.com.
“If your gap and the premium you’re asking people to pay for a language model disappears in two weeks, but it took you a year to make, that is not the sign of a moat that is going to be defensible.”
Why this matters: The enterprise revolt against token pricing went public last week. Palantir’s Alex Karp told CNBC that CEOs are “livid,” paying for “tokens that create no value” while handing their proprietary data to the labs charging them. Uber reportedly burned its annual AI coding budget in four months. Meanwhile AI agents are getting credit cards and Stripe is building payment rails for them, which means token costs are about to become a line item in every P&L, not just engineering budgets. Stephen’s framework explains what’s underneath: the models are commodities, the data is the moat, and the market has put 90% of its money on the commodity.
🎯 Jump to the best parts
00:00 Why Everyone Is Investing In The Wrong AI
00:44 Introduction
01:47 Why LLMs Are Becoming Commodities
03:16 Why Bigger Models Won't Win
05:48 Compute vs Software
07:42 The Real AI Moat
09:04 OpenAI's Business Challenge
12:09 Economic Foundation Models Explained
15:50 Building Collective[i]
17:24 Why Proprietary Data Wins
21:16 Predicting Business Outcomes With AI
24:44 The Biggest Enterprise AI Mistake
26:24 Why Most AI Wrappers Will Die
28:13 Is AI In A Bubble?
30:00 The Taiwan Chip Risk
32:25 AI, Geopolitics & National Security
34:38 Should Frontier Models Be Restricted?
36:10 Where CEOs Should Actually Invest In AI
37:46 The Most Mispriced AI Opportunity
38:40 Advice For AI Founders
40:58 Lightning Round
42:07 Where To Learn More
Important Links
Collective[i]: https://www.collectivei.com
Stephen’s blog: https://www.reloadnyc.com
Last episode with Stephen: Beyond LLMs: How AI is set to reshape global business
Watch or listen now: YouTube • Apple Podcasts
See our last episode with Stephen:
Our biggest takeaways from this conversation
1. Peak Token is here.
Stephen’s core thesis: a language model’s training data is public, so the model is a commodity, and open-source distillation now closes any frontier lead in weeks. The question that kills the business model isn’t whether frontier models are better. It’s how much you can charge for the difference.
“If your gap and the premium you’re asking people to pay for a language model disappears in two weeks, but it took you a year to make, that is not the sign of a moat that is going to be defensible.”
“How much can you charge for that little incremental improvement? ... My question is, can that cover the cost of everything else?”
The pattern: a frontier release looks untouchable for two or three weeks, then open-source models from China and Japan distill their way to an indistinguishable gap at a fraction of the cost. DeepSeek was the preview, a year ago.
New laptop-class hardware can run high-parameter models locally. Stephen’s estimate: 90% of workloads run local, and you pay frontier prices only for the rare job your machine can’t handle. That ratio breaks the labs’ economics.
The labs’ escape route is narrow premium niches, security and biology, where one result is worth real money. Hence the talent fights over those teams.
What to do with this: Price every AI vendor as if their underlying model becomes free in 18 months. If the pitch still works, it’s a real business.
Related reads:
→ Beyond LLMs, with Stephen Messer (his first appearance)
🚀 Build credibility. Drive pipeline. Win in digital assets. We produce institutional-grade research that positions you as the authority in your category, then distribute it to 100,000+ decision-makers. Let’s talk.
2. Forward-deployed engineers are a lock-in play.
The labs’ answer to commoditization is sending engineers into enterprises to build agents, positioned as a replacement for Accenture and McKinsey. Stephen’s read: those engineers have no incentive to build the one thing a token-cost-literate buyer would demand first, an abstraction layer:
