Cohere merges with Aleph Alpha, Anthropic's AI agent marketplace, and the Musk vs. Altman trial begins

Two completely unrelated announcements this weekend, but they tell the same story about where AI power is actually concentrating - and who's trying to stop it.

On one side: a transatlantic merger designed specifically to give Europe a fighting chance against American AI dominance. On the other: a US Department of Justice stepping in to shield Elon Musk's AI company from state regulators. The lines are getting drawn fast.

In today's newsletter we cover Anthropic's experiment where AI agents traded real goods for real money. But here's the trivia: what is the term for a system where multiple AI agents interact, negotiate, and transact with each other autonomously?

The answer is hiding near the bottom of today's newsletter... keep scrolling. 👇

Canadian AI startup Cohere is acquiring Germany-based Aleph Alpha in a deal backed by Lidl's parent company, the Schwarz Group. The merger has the blessing of both the Canadian and German governments, and the goal is explicit: build a credible, sovereign AI alternative for European enterprises that don't want to be entirely dependent on OpenAI, Google, or Anthropic.

This isn't just a business merger - it's a geopolitical statement. Aleph Alpha has been one of Europe's most prominent homegrown AI champions, focused heavily on privacy-first, enterprise-grade models. Cohere brings strong developer tooling and a track record of deploying LLMs in production environments. Together, they're positioning to serve the segment of the market that's explicitly uncomfortable routing sensitive data through US-based APIs.

For context: European data sovereignty concerns aren't theoretical. The GDPR, ongoing EU AI Act implementation, and general distrust of US tech policy under the current administration all create real demand for this kind of alternative. The question is whether a merged Cohere-Aleph Alpha can actually close the capability gap with frontier models fast enough to matter.

This is one to watch closely if you work in enterprise AI or European markets.

Read the full story →

This one is genuinely fascinating. Anthropic ran a controlled experiment where AI agents played both sides of a classified marketplace - some acting as buyers, others as sellers - and struck real deals for real goods with real money changing hands.

This is a meaningful step beyond demos. The agents weren't just simulating transactions - they were navigating actual negotiation dynamics: pricing, offers, counteroffers, and closing deals. The experiment was designed to explore how AI agents behave when economic incentives are real rather than synthetic.

The implications here go well beyond e-commerce. If agents can reliably represent principals in economic transactions - buying compute, procuring services, negotiating contracts - you're looking at a fundamental shift in how businesses operate. We're not there yet, but Anthropic clearly is probing the edges of what's possible.

If you're following AI agents closely, this experiment is a preview of where the agent economy is actually heading - and it's moving faster than most people realize.

Read the full story →

This is the kind of number that stops you mid-scroll. The UK government just quietly revised its estimate of carbon emissions from AI datacenters - and the new figure is more than 100 times higher than the previous one. Not 10% off. Not double. One hundred times.

The discrepancy is creating real friction between two major UK policy goals: becoming an AI superpower and hitting net zero targets. The departments responsible for each vision apparently weren't coordinating on the underlying numbers - and now the gap is impossible to ignore.

The energy math on AI infrastructure has been a quiet problem for years, but this is one of the starkest official admissions yet. Maine's governor vetoed a datacenter moratorium bill last week too - so regulators on both sides of the Atlantic are wrestling with the same tension between AI growth and environmental commitments.

This won't be the last government to discover its projections were wildly off. AI compute demand is scaling faster than most policy frameworks were designed to handle.

Read the full story →

xAI released grok-voice-think-fast-1.0 over the weekend and the numbers are hard to argue with. The model scores 67.3% on the τ-voice benchmark - the industry's most rigorous voice AI evaluation - outperforming Gemini, GPT Realtime, and xAI's own previous flagship across retail, airline, and telecom workflows.

The τ-voice benchmark is notable because it evaluates real-world enterprise voice scenarios rather than clean lab conditions. Retail, airline, and telecom workflows involve interruptions, domain-specific jargon, and complex multi-turn conversations - exactly the stuff that breaks most voice models in production.

We covered xAI's entry into voice APIs back in April when they launched their speech APIs - and it looks like the follow-up model has already leapfrogged the competition. The voice AI space is getting competitive fast.

Read the full story →

The lawsuit between Elon Musk and OpenAI's Sam Altman officially heads to trial today in Oakland, California. On paper, it's a dispute about whether OpenAI violated its original nonprofit mission by pursuing a for-profit conversion. In practice, it's a proxy war between two of the most powerful figures in AI - and it's going to surface a lot of internal documents neither side probably wants public.

Musk cofounded OpenAI in 2015, left acrimoniously in 2018, and has since launched his own competing lab (xAI) while publicly attacking Altman at every opportunity. The legal question is whether OpenAI's commercial pivot constitutes a breach of the charitable commitments that defined its founding. The real drama will be in the discovery - board communications, internal debates about the nonprofit structure, and the circumstances of Musk's departure.

This trial runs parallel to OpenAI's ongoing push to complete its for-profit restructuring. The timing is not coincidental. Whatever the legal outcome, the testimony alone is likely to shape public perception of both men and the companies they run.

Read the full story →

ComfyUI just raised $30 million at a $500 million valuation, and it's worth pausing on what that number actually represents. This is a tool that gives creators granular, node-based control over AI image, video, and audio generation - the opposite of the polished, one-click interfaces from Midjourney or Adobe Firefly.

The bet here is that a significant segment of the AI image generation market wants to go deep - to control every step of the pipeline, chain custom models, and build repeatable workflows. ComfyUI built that audience organically through open-source tooling, and investors are clearly convinced there's a sustainable business on top of it.

Speaking of building fast - if you need a website for your AI project or side business, 60sec.site lets you spin one up with AI in under a minute. Worth checking out if you've been putting it off.

Read the full story →

The answer is A - Multi-agent system! That's the established term for environments where multiple AI agents interact, coordinate, and in some cases compete with each other. Anthropic's marketplace experiment is a live demonstration of what economic multi-agent systems look like when the stakes are real. The field that studies this intersection of AI agents and economics even has its own name: agent-based computational economics.

Anthropic's agent marketplace experiment raises a genuinely interesting question: would you trust an AI agent to negotiate and spend money on your behalf today? Or does that feel like a step too far right now? Hit reply and let me know where your line is - I read every response and I'm genuinely curious how people are thinking about this one.

That's it for today - see you tomorrow with more. You can always browse past issues at the Daily Inference archive, and if you want to go deeper on any topic, check out the Daily Inference podcast. See you tomorrow.