The AI Moat Split Three Ways

The AI industry spent two years betting that intelligence itself was the moat — whoever built the smartest model won, and everyone else paid for access. This summer the bet stopped paying. DeepSeek's V4-Flash now costs roughly a hundredth of what Anthropic charges for Claude Fable 5 on the same benchmarks ^[1]. Alibaba and Moonshot give their frontier models away free to students and developers, charging only companies earning more than $20 to $50 million ^[2]. OpenAI cut its Luna fees 80%, Anthropic halved prices, and the rest followed ^[1]. When intelligence itself is nearly free, the industry had to stop betting on intelligence and start arguing about where the value actually lives. That argument has split into three camps, and they are not converging.

where the moat lives

Infrastructure: The biggest money is here. Chase Bank puts total build-out costs at $5.5 trillion by 2030, more than available cash flows ^[3]. Amazon has sold over $92 billion in bonds across four non-dollar markets ^[4]; Alphabet and Amazon both posted negative free cash flow in the June quarter, and S&P downgraded Oracle to BBB- ^[5]. Washington is in too, funding gigawatt-scale projects at Savannah River and Paducah ^[6], and a new ETF now lets investors buy the power assets and grid companies directly ^[7].

Services: The labs that won the model race are not following the money into servers. OpenAI and Anthropic are spinning up implementation companies — OpenAI Deployment Company at $4 billion, Ode at $1.5 billion — to embed engineers inside mid-sized enterprises ^[8]. The bet: proprietary data and operational context create stickiness that renting a model never will.

Data: Enterprise software firms argue neither model nor server is the moat. Snowflake routes traffic across OpenAI, Anthropic, and open-source models to avoid lock-in ^[9]. Tealium's founder says computing power is now widely accessible, so situational awareness is what separates generic tools from high-value ones ^[10]. OTM's CEO says access to leading models no longer confers an edge ^[11].

The infrastructure camp deserves the closest look, because it is carrying the most capital and the most doubt. Morgan Stanley estimates a fully optimized data center running the latest Nvidia chips could cost $25 billion a year to rent while generating $23 billion in output ^[12]. MIT's NANDA initiative reports that 95% of enterprise generative AI pilots show no measurable P&L impact on $30 to $40 billion of investment ^[13]. AI infrastructure stocks sold off on August 4 as investors reassessed how fast the returns would arrive ^[14]. Texas power futures now signal a slower build-out than expected ^[15]. And Jim Cramer is telling investors to look elsewhere ^[16].

The endless focus on the data center, and the anti-data center backlash, is obscuring opportunity after opportunity away from it, and I am no longer willing to tolerate you missing these. — Jim Cramer

The services camp is the most revealing of the three, because it is where the winners of the model race chose to go. OpenAI and Anthropic built the best models and then, instead of doubling down on servers, started selling the work of putting those models to use. Boris Berat, the Carna Health founder, describes the trap they are escaping: a startup building on top of a frontier model is in a "parasitoid relationship," paying its supplier to become its competitor ^[17]. Embedding engineers inside enterprises is the way out — the lab stops renting intelligence and starts owning the operational context around it. The data camp is the thinnest position, but its logic is the cleanest. Snowflake's Sridhar Ramaswamy makes the point in one line.

the smartest model in the world cannot make sense of truly bad data. — Sridhar Ramaswamy

If the model is a commodity and the server is a commodity, the only thing a competitor cannot copy is the data you already hold and the context you already understand. The most expensive bet in corporate history is the one most actively questioned by its own returns data — Morgan Stanley's arithmetic, MIT's pilot results, the August sell-off. And the two thinner camps are not independent of it: the services and data arguments only work if the infrastructure camp has overbuilt, because if compute stays scarce and expensive, the model labs and the data holders are just renters again. So the doubt hanging over the $5.5 trillion is not a side note; it is the load-bearing assumption of the other two theories. The moat did not migrate to a single layer — it fragmented, and the fragment carrying the most capital is the one whose failure the other two are quietly counting on.