When generative artificial intelligence first captured the public imagination, venture capital poured into foundation model providers under the assumption that pre-trained model weights represented the ultimate defensible monopoly. The thesis was straightforward: training trillion-parameter neural networks required vast capital reserves, specialized supercomputing infrastructure, and elite research talent that only a handful of well-funded corporate giants could command. The entities that owned the base models would capture the overwhelming majority of economic surplus across the technological landscape.

However, as analyzed in Andreessen Horowitz’s landmark investigation into platform ownership, the competitive reality has evolved in an entirely unexpected direction. Rather than establishing durable monopolies, base model providers have found themselves engaged in an intense war of attrition. The rapid proliferation of highly capable open-weights models, funded by sovereign wealth, research consortiums, and technology conglomerates seeking to commoditize their competitors’ products, has introduced relentless downward price pressure across the entire foundation layer.

Each time an open-weights model matches the reasoning capabilities of the previous frontier generation, commercial API providers are forced to slash their per-token prices by fifty to ninety percent to maintain developer volume. Consequently, raw intelligence generation has rapidly transformed into a utility commodity, much like cloud storage or compute cycles. The gross margins of generic model hosting have compressed, leaving foundation model laboratories with immense capital depreciation burdens and narrow operational moats.

For technology founders and investors, this structural shift clarifies where durable enterprise value actually resides. True defensibility is migrating upward into specialized application workflows and downward into specialized hardware acceleration. The enduring winners are not those selling undifferentiated tokens, but rather the vertical software platforms that embed generative reasoning into deeply entrenched corporate workflows, proprietary customer databases, and mission-critical systems of record. When the underlying intelligence engine is a ubiquitous commodity, the prize belongs entirely to the workflow that directs it.