AI News: June 18, 2026

1. World Model Startup Odyssey Raises $310 Million at a $1.45 Billion Valuation

Odyssey. The world model startup raised a $310 million Series B at a $1.45 billion valuation, led by Natural Capital with participation from Amazon, AMD Ventures, and Google Ventures, bringing its total raised to $337 million. Founded by former Cruise and Wayve engineers, Odyssey collects physical-world data using backpack-mounted cameras and trains models that generate interactive video with accurate physics for games and robotics. As part of the deal, Amazon becomes the preferred cloud provider, with Odyssey optimizing its models for AWS Trainium chips. Source

2. Zhipu AI’s GLM-5.2 Narrows the Gap to Closed-Source Coding Leaders

Zhipu AI. The open-weight GLM-5.2 scores 74.4 percent on the FrontierSWE long-horizon coding benchmark, landing roughly one point behind Anthropic’s Claude Opus 4.8, and jumps to 81 on Terminal-Bench 2.1 from 63.5 in its predecessor. It ranks second to Opus 4.8 on PostTrainBench while beating GPT-5.5 and Opus 4.7, positioning it as the strongest open-source model for autonomous coding. The model trails frontier systems by about ten points on reasoning tests such as Humanity’s Last Exam and consumes more tokens, making it less cost-efficient outside coding. Source

3. Analysis Warns Hyperscalers May Soon Outspend Their Cash Flow on AI

Epoch AI. An Epoch AI analysis of SEC filings finds the five largest hyperscalers, Microsoft, Amazon, Alphabet, Meta, and Oracle, are growing AI infrastructure spending about 70 percent a year while operating cash flow rises only around 23 percent. If the trends hold, capital expenditures would exceed operating cash flow around the third quarter of 2026, pushing free cash flow toward zero or negative. The firms are already turning to outside financing, with Alphabet raising $85 billion in equity and Amazon and Nvidia selling bonds, though the analysis cautions it is a simple extrapolation. Source

4. Pew Study Finds Only 16 Percent of Americans Expect AI to Benefit Society

Pew Research. A new Pew Research study finds just 16 percent of US adults believe AI will have a positive impact on society over the next 20 years, while about 40 percent expect negative effects. Nearly two-thirds think AI is advancing too quickly, 67 percent doubt the government will regulate it meaningfully, and 59 percent distrust companies to develop it safely. Adults under 30 were the most pessimistic, with only 14 percent optimistic, signaling persistent public skepticism even as AI adoption spreads. Source

5. Pramaana Labs Raises $27 Million to Bring Formal Verification to AI

Pramaana Labs. The startup raised a $27 million seed round led by Khosla Ventures, with Accel, BoldCap, Nexus Venture Partners, Premji Invest, and Unbound participating. Pramaana layers a deterministic verification system over conventional language models, using the LEAN proof language to check that outputs conform to domain-specific rules. The approach targets high-stakes sectors such as law, pharmaceuticals, and taxation, where reliability matters more than raw fluency. Source

6. Nvidia Research Shows Robots That Train Themselves Using AI Coding Agents

Nvidia. Nvidia researchers detailed ENPIRE, a system in which AI coding agents autonomously set up testing environments, write their own reward functions, form hypotheses, and modify training code using behavior cloning or reinforcement learning. A fleet of eight robots shared discoveries through Git version control and reached up to 99 percent success on tasks including block manipulation, pin sorting, and cable cutting. Scaling from one agent to eight cut Push-T completion time from five hours to two, suggesting a path to continuous real-world robot improvement with little human supervision. Source

7. Canada’s CPP Investments Commits Up to $741 Million to India AI Data Centers

CPP Investments. Canada Pension Plan Investment Board committed up to 70 billion rupees, about $741 million, to India’s AI data center buildout, taking an 8.2 percent stake in Hyderabad-based operator CtrlS and funding a hyperscale joint venture in which it holds 48 percent. CtrlS runs more than 15 facilities across India, where government incentives include tax exemptions for foreign cloud providers serving overseas customers from domestic data centers. The deal joins competing commitments such as AirTrunk’s $30 billion plan and Meta’s partnership with Reliance, even as the expansion strains local power and water. Source

8. DeepL Acquires Live-Event Audio Startup Mixhalo

DeepL. German translation company DeepL acquired San Francisco-based Mixhalo, a real-time audio streaming platform for concerts, conferences, and sports events, for an undisclosed sum. Mixhalo lets attendees receive live translated audio of speakers on their phones and had raised over $39 million from Founders Fund and Cowboy Ventures since its 2016 founding by Incubus guitarist Mike Einziger. The acquisition deepens DeepL’s push into real-time speech translation and marks its expansion into the US market with a new San Francisco office. Source

9. G7 Leaders Push for AI Distribution That the US Cannot Switch Off

G7. At the G7 summit, leaders including French President Macron and Indian Prime Minister Modi voiced alarm that the US could unilaterally revoke access to American AI models, pointing to the Trump administration’s recent ban on Anthropic’s Mythos 5 and Fable 5 systems. They floated a “trusted partners” scheme to build an alternative distribution network, arguing that few buyers would commit to US AI that could disappear overnight. The episode underscores rising pressure on allied nations to develop sovereign AI even as American models keep outperforming rivals. Source

10. HSBC Bets on Google Cloud AI for Projects Worth Over $100 Million Each

HSBC. HSBC said it will deploy Google Cloud AI across its global operations through a multiyear partnership, with individual projects each projected to generate more than $100 million in added revenue or savings. The bank plans to roll out more than 200 new AI use cases over the next two years. The deal is one of the largest enterprise AI commitments by a global bank and signals deepening adoption of generative AI in heavily regulated financial services. Source

11. Microsoft’s Copilot Cowork Shifts to Usage-Based Billing and Eyes DeepSeek

Microsoft. Microsoft is moving its Copilot Cowork product to usage-based billing and may add DeepSeek models to its lineup, according to reporting. The pricing shift reflects the growing cost of running agentic AI features and a broader industry move away from flat per-seat subscriptions. A potential DeepSeek integration would mark another step in Microsoft diversifying beyond OpenAI for the models powering its Copilot stack. Source

12. OpenAI Researchers Propose Predicting How Often Models Will Fail Before Launch

OpenAI. OpenAI researchers published work aimed at forecasting how frequently AI models will fail on a task before they are deployed, rather than discovering failure rates only after release. The method seeks to estimate reliability in advance, giving teams a clearer view of where a model is likely to break down. Such predictive evaluation could help practitioners set guardrails and choose deployment thresholds for production systems. Source

13. Microsoft Researcher Builds a Neural Network From Goats in Age of Empires II

Microsoft. A Microsoft researcher constructed a working neural network entirely from goat units inside the strategy game Age of Empires II as a pointed critique of how AI research is conducted and communicated. The stunt demonstrates that computation can emerge in unexpected substrates while challenging the field’s tendency toward spectacle over substance. It circulated widely as a commentary on rigor and hype in contemporary machine learning research. Source