Runtime: Google's new cloud firm needs leaders; China's GLM 5.2 draws a buzz

Welcome to The Stack's Runtime! Today: Google reveals more details about its new AI cloud company, Databricks says demand is rampant for a new open source Chinese LLM, and a European neobank lifts the lid on its agentic architecture - in a case study sponsored by AWS's data team.

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.

Banking on agents

You may not have heard of European bank bunq, but it's grown to serve 20 million customers and plans to expand to the US and beyond.

The company has created a multi-agent architecture for customer service called "Finn" that now handles 87% of customer interactions autonomously. Machine learning engineer Benjamin Kleppe wrote the first lines of code for Finn, building on AWS Bedrock, Aurora and OpenSearch.

He went through "a whole bunch of optimisations" earlier this year to drive down cost to serve and improve performance. Its current framework uses Opus as a central orchestrator to manage conversation state and high-level logic, but then delegates narrow operational tasks to sub-agents powered by lightweight models. bunq is particularly proud to have ensured full action traceability, with compliance and legal brought in from day one.

TPUs for me and you: Despite assembling one of the best distributed-computing networks of its time in the early 2000s, it would take Google at least a decade to start seriously using that infrastructure to go after the enterprise cloud infrastructure market. This time around it's taking no chances, and revealed more details Monday about the joint venture it has set up with Blackstone to run AI data centers powered by its own silicon.

Google is hiring for two key roles at "TPU Co.," the chief operating officer and the chief technology officer. Both roles will report to CEO Benjamin Treynor Sloss and aid in the creation of a new neocloud (of sorts) based around Google's TPU AI chips. Throw your hat in the ring?

All you need to do to get a shot at the CTO role is be proficient at "SRE [site-reliability engineering], Software Development, Networking, *and* Hardware operations," Sloss wrote on LinkedIn, which is quite a range of disciplines. The COO will be in charge of most of the traditional business roles like finance and internal IT, but will be laying the groundwork for a shift "to an agentic AI-first corporate architecture over time," whatever that is.

TPU Co is hoping to have 500 megawatts of computing power up and running by next year, which will be a tall order no matter which enterprise infrastructure wizard winds up getting the job. If enterprises show interest in a Google-Cloud-without-the-Google business model, expect AWS to think real hard about finding a dance partner for a similar venture based around its Trainium chips.

Claude costs create cheap cure: The revelation kicked off by the Claude Code craze — that large-language models have gotten much better at executing tasks in sequential fashion — led to a surge in demand for Anthropic's powerful but expensive models. After freaking out about skyrocketing AI costs, companies that want Claude-like performance on a budget are flocking to the recently released GLM-5.2 model, which is available under the MIT license and another sign that Chinese companies are taking open source very seriously.

Join peers following The Stack on LinkedIn

Mmm, synthetic data: Now that AI agents are shifting from experiments to production software inside enterprises, measuring ROI will become even more important. Prosus is now running 70,000 AI agents across its portfolio of companies, based around a "large commerce model" that is lowering costs and generating synthetic customers to test new offers.

Fun to stay at the BYOC: Traditional software vendors will need to adapt to the bring-your-own-cloud movement, which allows companies in regulated industries to get the benefits of running a third-party vendor tool while keeping control of their data. In this week's STACKUP, Noah Bovenizer discussed that trend with Tsuga CEO Gabriel-James Safar, who just raised $35 million in Series A funding for the observability startup.

Meet the new boss: Enterprise infrastructure designed for the pre-agent era is going to fall down when hit with the soaring levels of internet activity generated by agents; just ask GitHub, Patrick McFadin wrote in a guest piece for The Stack. "The explosion is real, the anxiety is earned, and the systems we built for the last wave are already starting to creak under the next one. But this is not a reason to panic, and it is definitely not a reason to wait for AI to magically fix everything," he said.

We're also reading:

Domo's data analytics business fell off a cliff last year amid shifts in the AI market and personal troubles for CEO Josh James, who was arrested for DUI last year, Business Insider reported.

The days of paying high-priced management consultants by the hour to assure C-suite executives that their plans totally make a lot of sense are numbered, according to The Wall Street Journal, as customers increasingly demand outcome-based pricing.

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.