Google just cut agent token costs by up to 65% — and the buyers of AI compute are fronting the cash for the buildout. Same economy, both sides moving at once. If you're just joining: IREN's AI-cloud expansion is one of the cleanest signals we've got of GPU capacity turning into contracted revenue. The company says it signed 2.8 billion dollars in multi-year AI cloud contracts, raised its 2026 AI cloud ARR target above 4 billion, and laid out capacity growth from roughly 3 megawatts of self-built AI cloud to 480 megawatts scheduled for 2026 — with 1.2 gigawatts targeted for 2027. This is AI Daily Briefing — and today, one number finally landed where I've been pointing all week. Let's start with Gemini 3.6 Flash. Sixty-five percent off token cost on long-horizon engineering tasks. That's a per-task price change on the exact workload agents chew through, not just another leaderboard headline. IREN AI cloud expansion isn't over. Follow us wherever you're listening, and the next chapter comes to you. VentureBeat, with Carl Franzen:
Google DeepMind today released three new proprietary AI models it says are among its most token-efficient yet: Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and Gemini 3.5 Flash Cyber. The models aim to make AI agents faster, smarter, and cheaper at scale.
Okay, the number I actually care about: Gemini 3.6 Flash, $1.50 per million input tokens, $7.50 output. Flash-Lite drops to thirty cents in, two-fifty out. That's the delta that changes build-versus-buy math for anyone running agents at volume. And notice what VentureBeat's piece does not have attached: a technical report. Sixty-five percent cost cut on long-horizon engineering tasks, quoted straight from the blog post. Until someone runs it, that's still a vendor claim. Right, and I'm not spiking the ball on it. The 65% is a headline figure on their benchmark, with their task definition. What I want is that curve on my workload — the ten-hour agent run where costs compound. And there's still a 3.5 Pro coming, so the whole cost-performance ladder is moving under our feet. You can't price a build decision against a lineup that isn't done shipping. That's the trap, though. Somebody's gonna sign a two-year inference commitment this week on a number from a model that'll be superseded by August. Here's Shane Snider at Data Center Knowledge:
The company said recent contracts include customer prepayments covering roughly 45% of the GPU capital expenditures associated with those deployments, reducing IREN's net funding requirements. An IREN spokesperson told Data Center Knowledge that such prepayments are “an increasingly common feature of AI cloud contracts,” suggesting that customers are committing capital earlier to secure access to scarce AI compute.
So the IREN number came in: $2.8 billion in new multiyear contracts, and the year-end AI cloud ARR target moved from $3.7 billion to more than $4 billion. This is the follow-on to the demand question we've been chewing on all week. And the detail that actually matters is buried three grafs down: customers prepaying about 45% of the GPU capex up front. The AI labs are co-investing to lock in capacity. Right, that's the financing model Data Center Knowledge is treating as the template. The tenant fronts nearly half the build, so IREN raises less on its own balance sheet. Pay the capex early, get the compute. Which changes the whole speculative-capex story. IREN isn't building on a hope and a lease — they've got signed contracts and prepayments before the racks go in. Different risk profile from the pre-revenue infra rounds I keep side-eyeing. From SynBioBeta:
Chai Discovery, a company focused on engineering AI models for the discovery of new molecules, has announced a $400M Series C funding round aimed at accelerating its advancements in the field. This funding round, which values the company at $3.8 billion, was led by Index Ventures, with participation from Kleiner Perkins, Sequoia Capital, Dimension, and several existing investors.
Chai Discovery: $400 million Series C, $3.8 billion valuation, Index leading. And here's what I actually want to know: what's that $400 million priced on? Because a drug-discovery model doesn't bill per token like an enterprise agent. The economics are completely different — you're betting on discovery outcomes, on a molecule that clears pre-clinical, not on inference run rate. And look at that cap table: Sequoia, Kleiner, Bain, Baillie Gifford, and OpenAI's still in from earlier rounds. The customers are Eli Lilly and Pfizer. That's a very specific set of hands on the fine-tuning. Second one of these in two days, too. Same question I had on the materials-discovery raise: science-compute funding keeps getting valued like SaaS, and it just isn't the same bet. Nina Achadjian's quote is pure founder-praise boilerplate — technical brilliance, commercial clarity. What I wanted from the release was who owns the model weights when Pfizer's data goes in and a candidate comes out. Cisco Blogs, with Amin Karbasi:
Today, Cisco is introducing Antares, a family of security small language models (SLMs) purpose-built for one of the hardest, most time-consuming and expensive problems in security: pinpointing where known vulnerabilities exist within a codebase. We are releasing two of these models—Antares-350M and Antares-1B—as open-weight models now available to the broader community on Hugging Face.
Cisco dropped Antares today — two open-weight security models, 350M and 1B, both on Hugging Face, with a named team and a technical write-up attached. After the week we've had, an actual model release with receipts feels almost quaint. And the task is narrow on purpose: pinpoint where a known vulnerability actually lives in a codebase. That's the expensive, tedious part Cisco says nobody wants to do by hand. What I like: they say these beat bigger closed and open models on that one task at a fraction of the cost, and they're small enough to run locally. So you're not shipping your source code to somebody's cloud to find your own bugs. That local part is the tell. A 1B model on your own hardware, scoped to one job — that's a different agent-security bet from the off-script-reasoning pitches we heard yesterday. This one's small enough that you can actually reason about its failure modes. Now compare that with the demo drops this week, where the technical report is nowhere to be found. Cisco gave us weights, a benchmark, named authors. Hold that up as the bar. I'll believe the benchmark when it survives a messy real repo, but being able to download it and check is the whole point. It's cheaper to distrust, and I mean that as a compliment. Inside Privacy is tracking this one, with Laura Kim, Alexandra Remick, and Munseong Park on the byline. Connecticut signed a comprehensive AI bill back in May, and now they're extending it to subscription services. So the same law that touches safety is reaching into cancel flows and auto-renewals. Here's the confusion I keep hammering: one bill bundles model safety with consumer-protection subscription rules. Those are two totally different regulatory problems under one signature. And for anyone building on top of these APIs, that's the part that actually lands. Your compliance surface just grew a subscription-disclosure obligation that has nothing to do with the model. Covington's Inside Privacy has the breakdown — Kim, Remick, and Park walked through the May 27 signing. If you ship in Connecticut, read that this week, not just the launch headlines. If you follow AI Daily Briefing for the policy stakes, try Anthropic Pentagon Watch — a daily briefing on Anthropic's fight with the DoD over Claude, military AI use, autonomous weapons, and procurement blacklisting. Find it wherever you listen to podcasts.
What we're watching next: whether IREN hits its more-than-four-billion-dollar AI cloud ARR target by the end of 2026.
You'll find links to every story we covered today in the show notes. If one caught your attention, that's where to dig in. That's AI Daily Briefing for today. This is a Lantern Podcast.