Saturday. The throughline today isn't a new model or a new number — it's a question that used to be procurement and is now foreign policy: whose AI are you allowed to build on? Cost is dragging developers one way, geography is dragging governments the other, and the vendors in the middle are consolidating to keep you where they want you. Here's the operator's read.
Chinese models are winning on the invoice
This is the story that should change a budget meeting. On the developer marketplace OpenRouter, Chinese models accounted for 57%–67% of tokens used in the week of September 14, up from just 6%–13% in February; on Vercel, their share rose to 55% in August from 11% in January. The pull is performance-per-dollar: companies like DeepSeek, Z.ai, and Alibaba have shipped models with major gains in tasks like coding, and their rapid, low-cost adoption has global developers switching. The trend line is steep — in a year, Chinese models' global usage climbed from under 15% to over 54% in a single week, led by DeepSeek, per OpenRouter's leaderboard data. Washington's worry isn't only market share; the fear is that adopting Chinese models pulls countries into a Chinese technology sphere of influence that hardens into geopolitical alignment.
The operator's take: the price gap is real and your engineers already know it — an open-weight model that's 4x to 100x cheaper on inference is going to show up in someone's side project before it shows up in your risk register. Get ahead of it: decide deliberately where a Chinese open-weight model is fine (self-hosted, air-gapped, non-sensitive workloads) and where it's a hard no (regulated data, government contracts, anything touching export controls). "A developer picked it because it was cheap" is not a data-governance strategy.
Microsoft folds everything into one Copilot
The US incumbents aren't answering China first — they're answering each other. Microsoft is updating its Copilot app for corporate workers as it seeks to challenge Anthropic's Claude, combining coding capabilities with a tool for building agents in a single app. The framing is one surface for asking a question, building an app, and getting work done on your behalf; Microsoft's pitch is bringing coding and productivity together in Copilot to challenge Anthropic and OpenAI in the workplace.
The operator's take: bundling is a lock-in move dressed as convenience, and convenience usually wins the pilot. That's fine if you go in clear-eyed: a single-pane Copilot lowers your integration cost and your negotiating leverage at the same time. Take the productivity gains, but keep at least one capability — agent orchestration or code gen — running on a model you can swap out, so "it's all in Copilot" never becomes "we can't leave Copilot."
A court backs calling a US lab a supply-chain risk
And the trust question isn't reserved for foreign models. A U.S. appeals court upheld the Pentagon's designation of Anthropic as a supply-chain risk. It's a striking marker: the same frontier lab that's raising money at a near-trillion-dollar valuation is being treated, by part of its own government, as a vendor to scrutinize rather than simply approve.
The operator's take: vendor risk for AI is no longer just uptime and pricing — it's regulatory and geopolitical, and it now applies to American names, not only Chinese ones. If a supplier's standing with regulators or the government can flip, that's a dependency you have to be able to route around. Qualify a backup model, keep your prompts and orchestration portable, and treat "approved today" as a status that expires.
Also on my radar
- Meta ships a consumer AI device first. Meta got to market ahead of rivals with a consumer AI device, though the strategy remains unproven — being first to the hardware doesn't mean being first to a business model.
- No AI czar for Bessent. Trump said Treasury Secretary Scott Bessent won't be the administration's AI czar — worth watching, because whoever does get the seat will shape how hard the China-model containment lever gets pulled.
- Tesla scales the Semi. Tesla aims to ramp heavy-duty Semi truck production with a Nevada factory opening — a reminder that while the headlines chase tokens, the margin fight in physical logistics is still very much on.
The throughline: capability is converging, so the fight has moved to trust. Chinese models are winning developers on cost while governments try to make that a loyalty test; Microsoft is bundling to own the workflow; and a court just reminded everyone that even a US frontier lab can be labeled a risk. The operator's job in all of this is boring and essential — stay portable, decide your red lines before your engineers decide them for you, and assume any single vendor relationship can change on you overnight. Own the exit and you can shop the market. That's the Signal for today.
Paul Sapio is the CIO of Mikhail Education and a full-stack AI engineer. Open to contract work in security, networking, AI, and SaaS development — reach out.