September 17, 2026
TechCrunch reported that both Instinct and Meta's Muse have added the ability to make phone calls on a user's behalf -- booking a restaurant table, canceling a subscription, that kind of thing. It sounds small. It isn't. Voice is the one channel that still resists automation, because a lot of the world -- doctor's offices, landlords, legacy vendors -- has no API, no chat widget, no app. It just has a phone number and a human answering it. If an agent can now handle that call convincingly, the last holdout channel for business automation starts to fall.
For a business owner, the appeal is obvious: fewer hours spent on hold with a supplier, fewer missed callbacks. But think through what happens when the agent gets it wrong -- cancels the wrong plan, mishears a price, agrees to terms it shouldn't have. Who owns that mistake? Meta and Instinct aren't publishing much yet on error rates or dispute handling, and that's the part worth watching before anyone plugs this into a real workflow. This is exactly the kind of agent-to-agent, agent-to-human handoff we've flagged before -- the risk isn't that the technology fails outright, it's that it succeeds just often enough that nobody double-checks it. Teams thinking seriously about where agents fit into daily operations without losing accountability should look at how workflow automation is built to keep a human in the loop, not just how many tasks it can offload.
Base Labs -- the research arm Baseten spun out earlier this year -- announced a partnership with Hugging Face and Goodfire to develop and publish open methods for training and monitoring open-weight models, per TechCrunch. That's notable mostly for what it isn't: another closed lab telling us to trust its internal safety team. Open-weight models are already everywhere in production, often with far less scrutiny than the flagship models from OpenAI, Anthropic, or Google get. Publishing monitoring methods that any team can adopt is a genuinely useful contribution, not just a PR move -- assuming the output is actual tooling and not a paper nobody implements.
I'd rather see three organizations compete on making open models safer than watch them compete on making closed models bigger. The honest caveat: publishing a method is not the same as enforcing it. Nothing stops a company from ignoring Base Labs' guidance entirely, and open weights mean anyone can fine-tune the safety right back out of a model. Still, for businesses evaluating vendors that build on open-weight foundations, having an independent, published standard to check against is a real improvement over the current shrug. It's part of a broader pattern we've tracked -- from questions about whether independent AI auditors are real to earlier debates about safety-washing -- where the industry is slowly, unevenly, building the referee it should have had years ago.
Google, Nvidia, and Anthropic are backing a coalition with Emerald AI to find 100 gigawatts of grid capacity for new data centers, TechCrunch reported. That's a staggering number -- roughly the output of dozens of large power plants -- and it tells you plainly that compute is no longer the bottleneck for AI's growth. Electricity is. When three of the industry's biggest players jointly fund a group to go hunt for grid space, it means they've concluded the market alone won't solve this fast enough.
This matters to ordinary businesses in a roundabout but real way: every AI feature you rely on, from a chatbot to a forecasting model, runs on infrastructure that's about to get more contested and more expensive to build. Add Huawei's plan to launch its Ascend 960DT chip in Q1 2027 to challenge Nvidia, and you've got a compute-and-power arms race on two fronts -- chips and electricity -- happening at once. None of it changes what a business should do this quarter, but it's a good reminder that the AI tools you depend on sit on top of a supply chain that's still being built in real time, not a finished utility.
Which of these worries you more as a business owner: an AI agent making phone calls on your behalf, or the infrastructure race underneath it running out of runway before the demand does?
Sources