What it costs to run AI models in production, and the supply forces behind those prices.
Q3 revenue rose 45% to $13.4 billion, and Anthropic, Meta, Microsoft, OpenAI and Oracle brought their first Helios racks online.
Our takeCommerce teams pay for AI inference, and a second chip supplier shipping at scale to Anthropic, Meta, Microsoft, OpenAI and Oracle could help ease GPU and token prices over time.
The roughly one-trillion-parameter model is free to self-host and commercialize, and Mistral halved its API prices.
Our takeFor European retailers and agencies with data-residency rules or a mandate to rely less on US vendors, this is Mistral's largest self-hostable model yet. Its trillion-parameter scale needs serious multi-GPU hardware, so most teams will likely use it through Mistral's API or a cloud marketplace.
Microsoft says demand exceeds capacity, Amazon holds 2026 capex near $220 billion and Alphabet flags a significant 2027 step-up.
Our takeMicrosoft's numbers matter most: about 75% of its capex buys short-lived hardware and demand exceeds supply, so providers have little reason to cut inference prices and good reason to favor committed customers.