Welcome to Runtime! Today: The US government is thinking about making Chinese open models harder to use, why AWS sent out eye-popping bills for nearly a day, and Andy Burnham weighs the fate of DSIT.
Please forward this email to a friend or colleague! If it was forwarded to you, sign up here to get Runtime for free every week, or__ level up here__.
Carry that weight: Enterprise AI reached an inflection point over the last six months after a new crop of frontier models from Anthropic and OpenAI finally delivered on the promise of AI agents, and demand for those models surged in response. A combination of rising prices and flaky infrastructure prompted a lot of companies to take a closer look at open weight AI models from Chinese companies like DeepSeek, Z.ai, and Moonshot to run their enterprise AI strategies, and as it turns out those models might actually be more useful than their closed counterparts thanks to US government AI policy.
After Hugging Face was hit with a cyberattack last week, it attempted to use an unnamed US frontier model to analyze their logs and understand what happened, but the model refused to comply due to the guardrails the Trump administration has insisted on putting around the latest crop of frontier models from Anthropic and OpenAI. However, Z.ai's GLM 5.2 model, while less capable overall than Claude Fable or GPT-5.6, easily handled the task.
But instead of revisiting the details of the closed-model guardrails, the administration is reportedly exploring ideas to make Chinese open weight models much more difficult to use in the US. One executive from OpenAI — which would clearly stand to gain from such a move — mused that the government could simply "direct every agency to issue soft law that creates FUD."
A generation of enterprise tech executives and investors raised on open-source software pushed back on that idea over the weekend and on Monday, but even a light "ban" on open weight models made in China could stifle the enterprise AI industry just as it was getting off the ground. Thinking Machines might be on to something with last week's release of Inkling, an open-weight model that companies can train on their own data.
GRC's reputation is shifting, and GRC Engineering is at the forefront. Join Ayoub Fandi, GRC Engineer at Lovable, and Justin Pagano, Sr. Director of GRC Engineering at Vanta, to discuss Legacy GRC and what comes next. You’ll learn practical ways to start automating, with lessons from Lovable and Vanta.
Learn more Explosions in the sky: Fireworks AI is another company working on open-weight AI models, and last week it raised a huge $1.5 billion funding round. Kiera Fields took a closer look at the company and its fine-tuning services, which helps companies bend AI models around their own needs.
Anarchy for the DSIT: The decision by new UK Prime Minister Andy Burnham to dissolve the Department for Science, Innovation, and Technology and fold it into the business department has given UK IT leaders something other than open weight models to talk about. Ahead of an official announcement, techUK said the decision "would slow momentum at a time when pace is essential for both the growth of the economy and our standing on the global stage."
The S3 bill is too damn high: Some AWS customers woke up last week to find bills for trillions of dollars in cloud infrastructure services, and while AWS can be very expensive at scale it hasn't quite come to that yet. The company blamed "a configuration change in our bill computation system" that went badly, which happens from time to time, but AWS compounded the error but not realizing it was sending out enormous, erroneous bills to customers for several hours.
Reprinting the satellites: In this week's edition of STACKUP, Noah Bovenizer talked to SWISSto12's Fredrick Gustavsson, who just helped the manufacturing startup raise a $70 million Series C funding round. The company builds RF components for its satellites using 3D printers, and has launched over 2,000 components into space.
We're also reading: #
Google Cloud is designing a new AI chip that could be "6 to 10 times more efficient" at running inference tasks than its current TPUs, according to The Information.
**Cisco launched **a family of open weight small language models called Antares that were designed to help users find vulnerabilities in their code and can run locally.
Want to sponsor a newsletter and get in front of 30,000+ subscribers, including CTOs and CIOs wielding $150 billion in annual tech budget?__ Get in touch through our Partner page.____ __