The infrastructure layer is fracturing along lines of control and cost. Amazon preemptively dismantles NDA barriers around data centers just as SoftBank's AI ambitions outgrow its balance sheet and require DigitalBridge as a third-party arm, signaling that data center economics have become the real constraint on AI deployment, not model capability. Meanwhile, the commodity layer explodes outward: Aleph Alpha ships Kolibri as an open-weight 78.1B MoE model that activates only 3.46B parameters per token on a single H200, DeepSeek releases desktop agent harness apps with MIT licensing, IBM makes Bob self-hostable for on-premises deployment, and Microsoft launches MAI-Transcribe-2-Streaming as a ranked-first speech model at $0.54 per hour. These moves collapse the moat between frontier and accessible, pushing the actual margin battle downstream into inference serving (Prime Intellect on Blackwell, NVIDIA's DGX Spark at 1 petaflop for $199 desktops) and agent behavior (the meta-pattern: agents now speak first, moving the hard problem from what to answer to when to interrupt). The social cost surfaces in parallel: OpenAI loses a safety employee over broken culture, Meta's Muse collects detailed profiles of your friends and family at millions of downloads, and the price of anything with memory spikes because the entire commodity chain has been conscripted into AI infrastructure. What emerges is not competition between models but competition over who controls the decision layer, the inference margin, and the right to interrupt your attention.
Sloane Duvall