Get tomorrow's brief in your inbox
Today: Anthropic joins a $915M carbon removal fund as AI's energy bills keep climbing. Google pushes Gemini into smart speakers with a $10/month premium tier. Two-thirds of Americans say AI is moving too fast, but half are using it anyway.
The AI industry's energy problem just got a spotlight moment. Anthropic became the first pure AI company to join Frontier, a carbon removal coalition backed by tech giants like Stripe and Google, contributing to a $915 million funding round that brings total pledges to $1.8 billion. The timing is telling. While companies like Uber blew through their annual AI budgets in months and others cut Claude licenses to control costs, the carbon bill is becoming harder to ignore.
For small businesses watching AI spending, this signals where the industry is heading. If the biggest labs are now paying for carbon offsets alongside compute costs, expect those expenses to eventually trickle down to API pricing and SaaS subscriptions. Anthropic hasn't published a sustainability report yet and still favors an "all of the above" energy approach, code for buying whatever power they can get. But joining Frontier suggests they're pricing in environmental costs as AI workloads scale. If you're building on AI infrastructure today, budget for energy costs to rise, either directly through API fees or indirectly through carbon offset requirements as regulation catches up.
Source: Anthropic becomes first AI startup to join the Frontier carbon removal coalition
Zapier cuts AI costs with smarter routing. Zapier published a guide showing how to avoid blowing your AI budget by routing tasks to the right model for the job. The key insight: most workflows don't need GPT-5.5 at $30/million output tokens when Haiku costs a fraction and handles deterministic work just fine. Use rules and if/then logic for predictable tasks like lead routing. Save the expensive models for interpretation work like analyzing open-ended form responses. If you've been running everything through one model because it's convenient, audit your workflows. The savings add up fast when you stop paying for filet mignon to do hamburger work.
Source: How Zapier can minimize your AI spend
Google Home Speaker puts Gemini in your living room for $99. Google's first standalone smart speaker since 2020 ships with Gemini built in, letting you use natural language commands without the rigid phrasing older assistants required. You can say "turn off all the lights except my bedside lamp" and it works. The catch: advanced features like Gemini Live conversations and Nest camera analysis cost $10/month after a six-month trial. If you've avoided smart speakers because they're frustrating to talk to, this fixes that problem. But ask yourself whether those premium AI features justify another monthly subscription before the free trial ends.
Source: Google bets on Gemini to reinvent the smart home speaker
BrightHire connects to Zapier for interview automation. The interview intelligence platform now integrates with 7,000+ apps through Zapier, letting recruiting teams automate prep work and follow-up tasks. Before an interview, auto-create candidate records from form submissions. After it ends, trigger Slack notifications, pull transcripts into docs, or run interview details through AI for scoring and summaries. Zapier's internal talent team built an agent that fires when candidates hit final rounds, pulling every prior transcript to generate tailored prep packets. If you're hiring and tired of copying interview notes between tools, this closes the loop.
Source: Connect BrightHire to the rest of your hiring workflow
GLM-5.2 claims top spot for open-weight LLMs. Chinese lab Z.ai released GLM-5.2 under an MIT license on June 16, a 753B parameter model with 40 active parameters and a 1 million token context window. Artificial Analysis ranks it #1 among open-weight models, and it placed 2nd on the Code Arena WebDev leaderboard behind only Claude Fable 5. It's available through OpenRouter at $1.40 per million input tokens and $4.40 output, significantly cheaper than GPT-5.5 ($5/$30) or Claude Opus ($5/$25). The catch: it's text-only and token-hungry, using 43k output tokens per task compared to 24-37k for competitors. If you need a capable open-weight model for coding or long-context work and can tolerate higher output costs, this is worth testing.
Source: GLM-5.2 is probably the most powerful text-only open weights LLM
DeepL acquires Mixhalo for live event translation. DeepL bought the real-time audio startup to push its voice translation suite into conferences and live events. Mixhalo was already using DeepL as its primary translation provider, streaming audio to attendees' phones so they can hear keynotes in their own language instead of fumbling with translation apps. If you run conferences or events with multilingual audiences, this makes live translation actually usable. DeepL is opening a San Francisco office as part of the deal.
Source: DeepL acquires Mixhalo for live-event audio streaming and translation
Pinterest tests conversational shopping with Ask Pinterest. The experimental AI app lets users get product recommendations through chat instead of scrolling boards. Early days, but if you sell visual products and already use Pinterest for discovery, watch this. Conversational commerce could change how people shop on the platform.
Source: Pinterest launches an experimental AI shopping app called 'Ask Pinterest'
The Agent-to-Agent protocol is solving multi-agent chaos. Google's A2A protocol gives AI systems a common language for collaboration, using HTTP, JSON-RPC, and server-sent events for streaming updates. The problem it solves: when you need multiple agents coordinating tasks across different systems, you end up writing custom code for every integration. A2A standardizes the handoff. Each agent has a digital ID (agent card) describing its capabilities, tasks define the work units, and messages carry the data. For technical teams building workflows where agents need to talk to each other, this removes vendor lock-in and makes swapping frameworks less painful. The protocol supports async work, so one agent can send a task and keep operating while the work runs.
Source: Agent-to-Agent (A2A) Protocol: How Cross-Agent Communication Works in Production
XDOF raises $70M to build robot training data pipelines. Building capable robots requires physical interaction data that barely exists. YouTube videos and gig worker footage are too low-fidelity. XDOF emerged from stealth with funding from Thrive, a16z, Spark, and others to build data collection tools, annotation systems, and pipelines for frontier labs and robotics companies. The startup is already working with 20 customers, including several AI labs, and partnered with UC Berkeley to release ABC, the largest collection of robot training data ever assembled (130,000 trajectories, 300 hours of simulation). If you're in manufacturing or logistics and wondering when robots will actually be useful, the bottleneck isn't models anymore. It's this: getting enough real-world data to teach machines how objects move.
The ROI reckoning has arrived. NEA partner Tiffany Luck told TechCrunch that enterprises are still figuring out their AI ROI after the tokenmaxxing trend collapsed. Uber blew through its annual AI budget in months. Some companies cut Claude licenses. Meta killed its internal leaderboard. Startups are now stepping in to help enterprises track AI spend and prove value. If you pushed employees to use AI without tracking what it's actually saving, you're not alone. But the bill is due. Before renewing seats or adding models, measure what changed. Time saved, revenue gained, or tasks eliminated. If you can't connect usage to outcomes, you're spending blind.
Source: NEA's Tiffany Luck on AI IPOs, personal agents, and the ROI reckoning
Anthropic's Mythos 5 and Fable 5 blocked from export. The Trump administration blocked Anthropic from exporting its newest models on national security grounds after Amazon flagged that safety guardrails could be bypassed. At the G7 Summit, world leaders including France's Macron and India's Modi voiced concerns that the U.S. could cut off AI access overnight, damaging both their economies and the AI firms themselves. If your business depends on U.S. AI models, you now have geopolitical risk to price in. Access can be revoked without notice, and you may never know why. Diversify your model providers.
Source: World leaders want American AI. They just don't want America to be able to turn it off.
Odyssey hits $1.45B valuation building world models. The startup, founded by self-driving pioneers from Voyage and Wayve, raised $310M in a Series B led by Natural Capital with Amazon, AMD Ventures, and GV participating. World models simulate the physical world with accurate physics, and Odyssey's approach mimics Google Earth: people with cameras strapped to their backs gathering data. The company now offers models for video game creation, robotics, and generating interactive video from text prompts. AWS is now Odyssey's preferred cloud provider, and the startup will optimize models for AWS's Trainium chips. If you're building anything that needs to understand how objects move in space, world models are where the industry is heading after LLMs.
Source: World model maker Odyssey nabs $1.45B valuation backed by Amazon and other big names
CPP Investments bets $741M on India's AI data center boom. Canada's largest pension fund is acquiring an 8.2% stake in CtrlS and committing up to $317M to a joint venture developing hyperscale data centers across India. Global investors are racing to fund AI infrastructure in India as companies like Amazon, Google, Microsoft, and OpenAI ramp up investments. India has positioned itself as a hub with tax exemptions for foreign cloud providers through 2047, provided workloads run from in-country data centers. If you're evaluating cloud providers, watch where they're building capacity. Geographic diversification matters when governments can flip switches.
Source: Canadian pension giant joins race to fund India's AI-fueled data center boom
Pramaana Labs raises $27M to bring formal verification to AI. The startup led by Khosla Ventures is tackling AI reliability in high-stakes verticals like law, drug discovery, and tax prep by combining LLMs with deterministic verification layers. The system uses LEAN programming language (used to verify mathematical proofs) to check the LLM's work. For tax law, the company is working with a former IRS commissioner. For each use case, Pramaana builds its own formal verification system overseen by domain experts. If you're deploying AI where errors cost money, health, or freedom, this is the approach that makes hallucinations a solvable problem instead of an acceptable risk.
Source: Pramaana Labs raises $27M seed round from Khosla Ventures to bring formal verification to AI
OpenAI published LifeSciBench for medical AI evaluation. The benchmark includes 750 expert-authored tasks spanning seven workflows and seven biological domains, grounded in real-world life science research. AMIE, Google's medical AI system, evolved from diagnostic conversations to long-term disease management, matching clinicians in overall management reasoning and scoring higher in plan preciseness and guideline alignment. If you're in healthcare IT, this signals where medical AI is heading: from one-off diagnoses to ongoing care management.
Source: Introducing LifeSciBench, New research shows how AMIE, our medical AI, could help manage health conditions
Route AI tasks by complexity, not convenience. Before you run every workflow through your most expensive model, split tasks into three buckets: deterministic (use rules, zero AI cost), interpretation (use mid-tier models like Haiku), and complex reasoning (use GPT-5.5 or Opus only when necessary). Audit one high-volume workflow this week and see if the first two steps can run on cheaper models. The pattern: collect data with rules, analyze with mid-tier AI, escalate to premium models only when required.
Disable Gemini pop-ups in Google Docs. Click "Gemini" in the top menu, select "bottom bar preferences," and turn off the bottom bar. To disable smart features across Google Workspace, go to Gmail settings, click "See all settings," scroll to "Google Workspace smart features," and toggle them off. If AI interruptions are killing your writing flow, this takes 30 seconds.
Source: How to turn off AI in your Google Docs
The AI industry is splitting into two tracks: the hype machine pushing faster, and the infrastructure layer pricing in costs everyone ignored during the land grab. Carbon offsets, export controls, ROI audits, and formal verification aren't buzzy, but they're the work that determines whether AI sticks or collapses under its own weight. For small businesses, the play is simple: use AI where it saves measurable time or money, route tasks to the cheapest model that works, and diversify providers so one vendor or government can't shut you down overnight. Watch enterprise AI budgets this quarter. If the big players are pulling back or demanding proof, it's time to sharpen your own use cases.