How US companies flipped from "tokenmaxxing" to "thrift-maxxing", mixing cheaper Chinese models with OpenAI and Anthropic, threatening the labs' IPO valuations
THE SO WHAT
Model-mixing with cheaper Chinese options is turning inference into a procurement game, not a loyalty game—CFOs are now in the loop on which model runs which workload. If you’re building on a single premium lab, assume your customers are already routing non-critical tokens elsewhere and design pricing, SLAs, and architecture for a multi-model world.
READ THE SOURCE
MORE FROM THE WIRE
Applied AIThe next AI race will be fought over trust
The next defensible moat in AI is less about raw capability and more about verifiable behavior—governance, auditability, and predictable failure modes. Operators should be instrumenting trust today: log every critical AI decision, define escalation paths, and make “can we prove what it did and why” a buying criterion.
Applied AIInstagram, Facebook Ran AI ‘Nudify’ Ads from China, Report Says
AI ‘nudify’ ads slipping through Meta’s policies via a Chinese partner show how quickly abuse can route around centralized controls—ad networks are now an AI safety surface, not just a brand-safety one. If you rely on paid social, expect sharper scrutiny on creatives and landing pages and build internal checks before regulators or platforms impose blunt restrictions.
Applied AIChina’s Moonshot to Release Breakthrough AI Model for Download
A Kimi K3 download release would put a frontier-leaning Chinese model directly into the global open stack—outside US export controls and into every serious lab’s eval matrix. If you build on open models, plan now for a world where top-tier Chinese weights are technically accessible but politically sensitive to adopt.
Applied AISources: Nvidia is in talks to provide a ~$250B backstop for OpenAI as part of a 10GW data center project that SoftBank is developing in Ohio
A 10 GW Ohio build with a ~$250B Nvidia backstop and U.S.-controlled power would turn AI compute into regulated national infrastructure, not just cloud capacity. For operators, this points to a future where access to frontier-scale compute is mediated by government-aligned hubs—plan architectures that can flex between hyperscalers, regional clouds, and your own smaller clusters.