Social Feed: AI infra deals & research
A social AI‑intelligence post bundles recent infrastructure and research signals: CoreWeave is set to power Anthropic, AfterQuery raised $30M for AI training data, and tools like SUPERNOVA and agent annotators are drawing attention. (x.com) The single feed groups compute partnerships, funding and reasoning‑tool updates that reflect active investment across the AI stack. (x.com)
CoreWeave said on April 10 that Anthropic signed a multi-year cloud deal to run and deploy Claude models on CoreWeave’s infrastructure, with compute coming online later in 2026. (investors.coreweave.com) CoreWeave said the rollout will start in phases later this year, and the company said nine of the 10 leading artificial intelligence model providers now use its platform. CNBC reported CoreWeave shares rose 11% on April 10 after the Anthropic announcement. (investors.coreweave.com) (cnbc.com) The deal landed one day after CoreWeave announced a $21 billion expansion of its agreement with Meta, according to CNBC. CoreWeave said it listed on Nasdaq in March 2025, tying the Anthropic contract to a fast buildout of public-market artificial intelligence infrastructure. (cnbc.com) (investors.coreweave.com) The other side of the stack is data: models need labeled examples and test environments before they can answer reliably. AfterQuery said on April 9 that it raised a $30 million Series A round at a $300 million valuation to build those datasets and reinforcement learning environments. (businesswire.com) AfterQuery said Altos Ventures led the round, with The Raine Group, Y Combinator and BoxGroup participating. The company said it has passed a $100 million annual revenue run rate, launched 14 months ago, and works with nearly 100,000 verified professionals in fields including finance, software engineering, medicine and law. (businesswire.com) Research updates in the same feed point to a third bottleneck: teaching models how to reason outside math and code. A paper posted to arXiv on April 9 said SUPERNOVA uses curated natural-language instruction data for reinforcement learning with verifiable rewards, reporting more than 100 controlled experiments and relative gains of up to 52.8% on the BBEH benchmark. (arxiv.org) That approach tries to turn expert-written examples into a reward signal a model can check against, instead of relying only on open-ended human preference judgments. The SUPERNOVA authors said their models beat strong baselines on BBEH, Zebralogic and MMLU-Pro, three benchmarks used to test broader reasoning. (arxiv.org) Agent annotators aim at a related problem: replacing slow manual review with systems that propose labels and explain them. In a December 2025 Nature Communications paper, researchers said their multi-agent system CASSIA improved cell annotation across 970 cell types while adding reasoning and confidence checks meant to reduce hallucinations. (nature.com) Those pieces fit together as one supply chain. Cloud contracts buy graphics processing unit capacity, startups like AfterQuery sell expert data, and papers like SUPERNOVA and CASSIA show how labs are trying to turn that compute and data into models that reason more consistently. (investors.coreweave.com) (businesswire.com) (arxiv.org) (nature.com) The social post grouped those signals into one snapshot, but the common thread is concrete: more money for compute, more money for training data, and more experiments aimed at making model outputs easier to verify. (investors.coreweave.com) (businesswire.com) (arxiv.org)
