Hi everyone!
Welcome to the latest issue of your guide to AI, an editorialized newsletter covering the key developments in AI policy, research, industry, and start-ups over the last month. First up, a few updates:
-
Our 9th annual Research and Applied AI Summit is next Friday in London! We’ll dive into best practices for building AI-first products and translating applied research with leaders from ElevenLabs, Black Forest Labs, Isomorphic Labs, DeepMind and poolside, as well as AI geopolitics with Meta, Tony Blair Institute, Delian Alliance Industries, Bloomberg and the New York Times. After the event, we’ll share write ups on Air Street Press and recorded talks on our YouTube.
-
State of AI Report 2025: I’m hiring a Research Assistant to help us build the industry’s most loved report 🙂 This is a perfect role if you love distilling complex topics in AI research, industry and politics into key messages to shape public conversation around this topic. If you know the perfect candidate, please reply/share! Here is a job description.
-
Air Street Press featured a number of pieces: biology’s scaling laws, Delian’s new autonomous strike drones, Hedera’s €15M Series A to scale liquid biopsies for cancer patients, what Wall Street didn’t see coming in AI back in 2016, and whether the EU AI Act is actually useful.
I love hearing what you’re up to, so just hit reply or forward to your friends 🙂
The AI Factory narrative: infrastructure as industrial strategy
The term “AI factory” is no longer just rhetorical flourish. It has become the defining metaphor for the geopolitical and economic infrastructure race now unfolding in the open, as state and commercial agendas become visibly entangled. At its core is a push by Silicon Valley and U.S. political leadership to recast hyperscale datacenters as nation-building infrastructure, not just technical backends. Jensen Huang’s refrain that “these are not data centers, they are AI factories” is central to this rebrand. He invokes a powerful political argument: if the U.S. reshored sneaker factories, why not AI production? What sneakers were to 1990s globalization, AI is to the new era of strategic decoupling.
This rebranding has proven politically effective. It aligns the interests of capital-heavy AI players like OpenAI, Oracle, and NVIDIA with nationalist industrial policy. It plays directly into Trump’s $600B AI Acceleration Partnership (at which every important CEO in corporate America was present), where AI infrastructure deals with Gulf states are packaged with the same gravity as oil or arms deals. The Stargate data center in Texas, co-developed by OpenAI, Oracle, SoftBank and Crusoe, is the pilot site for what is pitched as a 10-site, $500B network of U.S.-controlled AI superclusters.
Meanwhile, a 5GW Stargate-like project in Abu Dhabi would leapfrog U.S. deployments in scale. These are national projects built with foreign dependencies, an irony at the heart of techno-sovereignty. This is the heart of the “AI Factory Illusion” that I wrote about in Air Street Press: a globally distributed, highly automated infrastructure wrapped in the language of 20th-century manufacturing. As NVIDIA courts sovereign clients and neoclouds to diversify beyond Big Tech, the term serves to unlock subsidies, fast-track permitting, and deepen state-industry alignment. If anything, it helps NVIDIA print tens of billions of dollars of revenue a quarter, or $44.1B in the last quarter, to be precise.
Meta stalls, Anthropic surges, OpenAI eats business lunch
The major labs’ spring updates reveal a fragmented landscape. Meta’s internal turmoil around its “Behemoth” model perhaps underscores the collapsing marginal utility of trillion-parameter scaling absent a killer product experience. Engineers reportedly question whether the latest version represents real progress.
By contrast, Anthropic is cooking, securing integration deals with Apple (to power a new Claude-based version of Xcode) and Amazon (for Alexa+), and launching Claude 4, even though deceptive reasoning capabilities triggered ethical scrutiny.
Claude 4, particularly the Claude Opus 4 variant, has proven to be a substantial leap forward. It outperforms competitors like GPT-4.1 and Gemini 2.5 Pro on key benchmarks like SWE-bench (72.5%) and exhibits strong long-context performance. It also supports extended autonomous operation for up to seven hours without degradation, making it well-suited for persistent agents. Its hybrid reasoning, long-term memory, and tool use integration make it arguably the most capable model in production today. Anthropic has also deployed it under its new AI Safety Level 3 (ASL-3) standard for the first time.
This comes as coding startup Windsurf, which has entered into an agreement to be acquired by OpenAI, was cut off from Claude access by Anthropic directly. In a somewhat timely fashion, Windsurf announced its own foundational model, SWE-1. These turf wars are worth following to inform how companies should think of building or tuning their own models vs. procuring them from OEMs or via third parties. Is the model the product or is the product enhanced by the model?
Meanwhile, Ramp data shows just how far ahead OpenAI’s share of US business subscriptions is from the competition. In a sample of more than 30,000 American businesses and billions of dollars in corporate spend using data from Ramp’s corporate card and bill pay platform, 81% of businesses with paid AI subscriptions were paying OpenAI.
Meta’s military realignment
Meta is undergoing a dramatic shift, albeit outside the lab. In a move that would’ve been unthinkable just a few years ago, Meta has partnered with Anduril to co-develop extended reality hardware and software for the U.S. military. The flagship product is EagleEye, a battlefield AR headset that fuses Meta’s Reality Labs and Llama models with Anduril’s Lattice defense platform. Silicon Valley’s most consumer-facing, metaverse-obsessed company is now explicitly building military gear.
What was once taboo (Big Tech’s open embrace of defense) is rapidly becoming normalized. The partnership also marks a détente between Mark Zuckerberg and Palmer Luckey, signaling a realignment of AI-native leadership around national security imperatives.
More racing across modalities
The generative video race is accelerating across labs and platforms. Google’s remarkable Veo 3 now adds soundtracks to generated video, marking a step closer to multimodal narrative synthesis. Their demos showed coherence and visual fidelity approaching short-form film. This sparked viral excitement all over Twitter, as observers noted a leap in cinematic fidelity.
Startups are pushing forward too. Lightricks’ new model is capable of realistic motion rendering from text, while Odyssey is betting on interactive generative video experiences as a new frontier. Odyssey’s research preview (try it!) lets users walk through photorealistic, real-time AI-generated 3D scenes such as forests, shopping malls, and more by streaming frames every 40ms, without a traditional game engine. The model doesn’t just generate a clip: it creates an explorable, real-time world. That alone sets it apart from static or pre-rendered outputs.
On the research frontier, DeepSeek’s new R1 checkpoint blends Chinese pretraining corpora with Qwen3 backbones, continuing China’s push for frontier-scale open models. In parallel, Google released Gemma 3B and MedGemma, its health-tuned LLM family, while the open-source Gemma Diffusion shows growing ecosystem depth.
Gemma Diffusion is particularly exciting because it offers a new paradigm for text generation: using diffusion rather than autoregressive methods. It allows for full-sentence or paragraph generation in parallel, with faster outputs and the ability to self-correct. As an open-source initiative, it broadens experimentation and hints at a post-token-by-token future for language models.
The export control shuffle and sovereign stack realignment
U.S. policy is shifting to reflect this AI infrastructure race. The rescinding of the AI diffusion rule signals a recalibration of export controls around advanced semiconductors and models. Rather than blunt bans, future policy may lean on traceability, licensing regimes, and platform-level coordination.
Meanwhile, NVIDIA’s NVLink Fusion program is expanding the modularization of its stack, allowing nation-states, corporates, and “neoclouds” to build bespoke AI systems. This plays into NVIDIA’s explicit push to court sovereign and enterprise clients outside of the Big Tech oligopoly.
From Saudi’s Humain deal to Sweden’s consortium AI factory, the geopolitical stack is re-aligning around modular, NVIDIA-led platforms. Sovereignty now means not only owning the model weights, but the infrastructure, training corpus, and policy posture that governs their use. This is a new form of sovereignty: not territorial, but computational.
The factory metaphor, however ill-fitting in technical terms, is proving useful as political architecture. The future of AI will be built in these factories, whether or not they produce anything as tangible as steel.
Research papers
Practical Efficiency of Muon for Pretraining, Essential AI
In this paper, the authors investigate Muon, a simple second-order optimizer, as a replacement for AdamW in large-scale language model pretraining. They show that Muon expands the compute-time Pareto frontier, enabling faster training or reduced compute at large batch sizes without sacrificing data efficiency. Experiments across models up to 4B parameters and batch sizes up to 16M tokens demonstrate that Muon consistently requires 10–15% fewer tokens than AdamW to reach the same loss, with this advantage persisting or growing as batch size increases.
The study also addresses hyperparameter tuning by combining Muon with maximal update parameterization (muP), enabling efficient transfer of hyperparameters from small to large models. The authors introduce a “telescoping” algorithm that controls tuning overhead, keeping it modest even at scale. This work is relevant for practitioners seeking to optimize training efficiency and resource allocation in large language model development, especially in distributed or compute-constrained environments.
AlphaEvolve: A coding agent for scientific and algorithmic discovery, Google DeepMind
In this paper, the authors introduce AlphaEvolve, an evolutionary coding agent that leverages state-of-the-art large language models (LLMs) to autonomously improve algorithms through iterative code modifications and automated evaluation.
AlphaEvolve orchestrates a pipeline where LLMs generate, critique, and evolve code, guided by machine-gradeable evaluation functions. The system was tested on challenging tasks, including discovering faster matrix multiplication algorithms, most notably, finding a new algorithm for multiplying $4 times 4$ complex matrices using 48 multiplications, improving on Strassen’s 49-multiplication result after 56 years.
Beyond mathematics, AlphaEvolve was applied to optimize Google’s data center scheduling, kernel engineering for Gemini LLM training, and hardware circuit design, yielding measurable improvements such as a 0.7% recovery in compute resources and a 23% kernel speedup.
The research highlights that combining LLMs with evolutionary search and automated feedback can yield novel, verifiable solutions in both theoretical and practical domains, especially where automated evaluation is feasible.
Enigmata: Scaling Logical Reasoning in Large Language Models with Synthetic Verifiable Puzzles, ByteDance Seed, Fudan University, Tsinghua University
In this paper, the authors introduce Enigmata, a suite designed to improve and evaluate logical reasoning in large language models (LLMs) using synthetic, verifiable puzzles. The suite includes 36 tasks across seven categories, each with an automatic generator and verifier, enabling scalable data creation and precise difficulty control. The authors propose a two-stage training approach: rejection fine-tuning with high-quality solutions, followed by multi-task reinforcement learning with verifiable rewards (RLVR).
Experiments show that models trained with Enigmata, such as Qwen2.5-32B-Enigmata, outperform strong baselines like o1 and o3-mini-high on puzzle reasoning benchmarks (Enigmata-Eval, ARC-AGI, and others), and generalize well to out-of-domain tasks, including advanced math and STEM problems. Notably, adding Enigmata data to larger models (e.g., Seed1.5-Thinking) further boosts performance on challenging benchmarks.
The work demonstrates that synthetic, diverse puzzle data can enhance LLM reasoning, supporting applications in education, automated problem solving, and robust AI evaluation.
The Open Molecules 2025 (OMol25) Dataset, Evaluations, and Models, Meta, Carnegie Mellon University, University of Cambridge
In this paper, the authors introduce the Open Molecules 2025 (OMol25) dataset, a large-scale resource containing over 100 million density functional theory (DFT) calculations at a high level of theory, spanning 83 elements and a wide range of molecular systems, including biomolecules, metal complexes, and electrolytes.
The dataset is designed to address the lack of comprehensive, diverse, and high-quality data for training machine learning interatomic potentials (MLIPs) that can act as DFT surrogates. The authors provide detailed descriptions of the data generation process, including sampling strategies for chemical and structural diversity, and rigorous quality control.
Baseline models such as eSEN, GemNet-OC, and MACE are evaluated on OMol25, with eSEN-md achieving energy MAEs as low as 1.2 meV/atom and force MAEs of 12.3 meV/Å. However, challenges remain in accurately modeling ionization energies, spin gaps, and long-range interactions.
ProRL: Prolonged Reinforcement Learning Expands Reasoning Boundaries in Large Language Models, NVIDIA
In this paper, the authors introduce ProRL, a prolonged reinforcement learning methodology designed to expand the reasoning capabilities of large language models beyond what is accessible through standard RL or extensive sampling of base models. They address challenges like entropy collapse and training instability by incorporating KL divergence penalties, reference policy resets, and dynamic sampling.
Their experiments train a 1.5B parameter model, Nemotron-Research-Reasoning-Qwen-1.5B, on a diverse set of 136K problems spanning math, code, STEM, logic puzzles, and instruction following. The model outperforms its base model (DeepSeek-R1-1.5B) with average pass@1 improvements of 14.7% in math, 13.9% in coding, 54.8% in logic puzzles, 25.1% in STEM, and 18.1% in instruction following. Notably, ProRL enables the model to solve tasks where the base model fails entirely, and shows strong generalization to out-of-distribution tasks.
The work demonstrates that with sufficient RL training, models can develop novel reasoning strategies, suggesting practical benefits for deploying smaller, more capable models in real-world applications where compute and data are limited.
Learning to Reason without External Rewards, UC Berkeley, Yale University
In this paper, the authors introduce INTUITOR, a method for training LLMs using only internal feedback, specifically, the model’s own confidence (self-certainty), as a reward signal, rather than relying on external supervision or labeled data. The approach, called Reinforcement Learning from Internal Feedback (RLIF), replaces traditional reward signals in policy optimization algorithms with self-certainty scores, measured as the average KL divergence between the model’s output distribution and a uniform distribution.
Experiments show that INTUITOR matches the performance of supervised RL methods like Group Relative Policy Optimization (GRPO) on mathematical reasoning benchmarks (GSM8K, MATH500), and outperforms them on out-of-domain tasks such as code generation (LiveCodeBench, CRUXEval). The method also improves instruction-following and fosters more structured, interpretable reasoning.
A key caveat is that INTUITOR’s effectiveness depends on careful regularization (KL penalty) and online reward computation to avoid reward exploitation. This research suggests that intrinsic model signals can drive scalable, domain-agnostic learning, which could be valuable for autonomous AI systems where external supervision is impractical.
HealthBench: Evaluating Large Language Models Towards Improved Human Health, OpenAI
In this paper, the authors introduce HealthBench, an open-source benchmark designed to evaluate large language models (LLMs) in healthcare contexts. HealthBench consists of 5,000 multi-turn conversations, each graded against detailed, physician-written rubrics covering 48,562 unique criteria across seven health themes and five behavioral axes (accuracy, completeness, communication, context awareness, instruction following).
The benchmark measures both overall and theme-specific model performance, using a model-based grader validated against physician judgment. Results show rapid improvement in recent models: OpenAI’s o3 model scores 60% overall, compared to 16% for GPT-3.5 Turbo and 32% for GPT-4o. Smaller, cost-efficient models like GPT-4.1 nano now outperform older, larger models.
The paper also introduces HealthBench Consensus (focusing on 34 critical, physician-validated criteria) and HealthBench Hard (a challenging subset where top models score only 32%). The research highlights persistent gaps in context-seeking and reliability, emphasizing the need for robust, real-world evaluation frameworks as LLMs are increasingly deployed in healthcare.
Robin: A Multi-Agent System for Automating Scientific Discovery, FutureHouse, University of Oxford
In this paper, the authors introduce Robin, a multi-agent AI system designed to automate the scientific discovery process, integrating hypothesis generation, experimental planning, and data analysis. Robin was applied to identify novel treatments for dry age-related macular degeneration (dAMD), focusing on enhancing retinal pigment epithelium (RPE) phagocytosis. The system proposed ripasudil, a ROCK inhibitor, as a therapeutic candidate, which was experimentally validated to significantly enhance RPE phagocytosis.
Robin employs agents like Crow and Falcon for literature review and Finch for experimental data analysis. Experiments included flow cytometry to measure phagocytosis and RNA sequencing to explore transcriptional changes, revealing upregulation of ABCA1, a lipid efflux pump linked to RPE function.
While Robin automates key steps, it relies on human input for experimental execution and prompt engineering. This research demonstrates AI’s potential to accelerate drug discovery, particularly in repurposing existing drugs for unmet medical needs.
A Cross-Species Generative Cell Atlas Across 1.5 Billion Years of Evolution: The TranscriptFormer Single-cell Model, Chan Zuckerberg Initiative, Stanford University
In this paper, the authors introduce TranscriptFormer, a generative foundation model designed to create a cross-species single-cell transcriptomic atlas spanning 1.53 billion years of evolution across 12 species. The model integrates gene and transcript data using a transformer-based architecture, enabling tasks like cell type classification, disease state prediction, and gene-gene interaction simulation.
The experiments demonstrate that TranscriptFormer outperforms existing models, such as UCE and ESM2-CE, in generalizing across species, including those separated by 685 million years of evolution. Notably, the TF-Metazoa variant achieved the highest macro F1 score (0.778) in out-of-distribution species classification. It also excelled in human-specific tasks, matching or slightly surpassing state-of-the-art models on the Tabula Sapiens 2.0 dataset.
This research highlights the potential of generative models in biology, offering tools for cross-species analysis, disease modeling, and virtual experimentation, with applications in evolutionary biology, drug discovery, and personalized medicine.
Absolute Zero: Reinforced Self-play Reasoning with Zero Data, Tsinghua University, Beijing Institute for General Artificial Intelligence, Pennsylvania State University
In this paper, the authors introduce the Absolute Zero Reasoner (AZR), a reinforcement learning framework that trains LLMs without relying on human-curated datasets. AZR operates under the Absolute Zero paradigm, where the model proposes and solves its own tasks, guided by verifiable rewards from a code execution environment.
The research demonstrates that AZR achieves SOTA performance on coding and mathematical reasoning benchmarks, surpassing models trained on tens of thousands of curated examples. Notably, AZR-trained models show strong cross-domain generalization, with coding-trained models improving math performance by up to 15.2 percentage points.
The experiments highlight the importance of task diversity, with AZR leveraging abduction, deduction, and induction tasks. However, the authors note challenges like safety concerns and occasional undesirable outputs, emphasizing the need for oversight.
This work matters as it reduces dependency on human data, enabling scalable, autonomous learning. Potential applications include adaptive AI systems in education, coding, and problem-solving domains.
Model Merging in Pre-training of Large Language Models, ByteDance Seed
In this paper, the authors introduce Pre-trained Model Average (PMA), a novel framework for model merging during LLM pre-training. They conducted extensive experiments across model scales (from millions to over 100B parameters) with both Dense and MoE architectures.
Their results demonstrate that merging checkpoints from the stable training phase produces significant performance improvements across downstream tasks. Remarkably, applying PMA at early stages of the cosine-decay phase achieves comparable results to final-stage annealed models.
Merging with constant learning rates can effectively simulate annealed performance without the computational expense of full annealing. They also introduce PMA-init, which stabilizes training when loss spikes occur. Through mathematical analysis and experimentation, they determine optimal merging intervals scale with model size, while including more checkpoints improves performance.
Darwin Gödel Machine: Open-Ended Evolution of Self-Improving Agents, University of British Columbia, Vector Institute, Sakana AI
In this paper, the authors introduce the Darwin Gödel Machine (DGM), a self-improving AI system that modifies its own codebase to enhance its coding capabilities. The DGM combines self-improvement with open-ended exploration, maintaining an archive of diverse coding agents rather than evolving just one solution. This approach allows exploration of multiple paths through the search space.
Experiments on coding benchmarks show impressive results: performance increased from 20.0% to 50.0% on SWE-bench and from 14.2% to 30.7% on Polyglot. The authors demonstrated that both self-improvement and open-ended exploration components are essential, as removing either one significantly reduced performance gains. The improvements generalized across different foundation models and programming languages, showing the robustness of the approach.
Knowledge Insulating Vision-Language-Action Models: Train Fast, Run Fast, Generalize Better, Physical Intelligence
In this paper, the authors tackle a key challenge in robot control: maintaining knowledge from pretrained vision-language models while adding fast, continuous action capabilities. They identify a critical problem: when adding continuous action modules to vision-language models, the pretrained knowledge often degrades, resulting in poor language following and slow training.
Their solution, “knowledge insulation,” introduces three key innovations: 1) Joint training with both discrete and continuous action representations, 2) Stopping gradient flow from the action expert to the vision-language backbone, 3) Co-training with general vision-language data alongside robotics data.
The approach achieves state-of-the-art results on LIBERO benchmarks (96.0% on LIBERO-90) and outperforms alternatives on real-world tasks like table bussing and drawer manipulation. This approach could enable robots to both follow human instructions accurately and execute complex physical actions efficiently.
LaViDa: A Large Diffusion Language Model for Multimodal Understanding, UCLA, Panasonic AI Research, Adobe Research
In this paper, the authors introduce LaViDa, the first family of vision-language models based on diffusion rather than traditional autoregressive generation. While current VLMs like LLaVA generate text sequentially, LaViDa leverages diffusion models to enable parallel decoding and bidirectional context. The authors tackle key challenges with novel techniques: complementary masking for efficient training, prefix-DLM for faster inference, and timestep shifting for quality sampling.
LaViDa achieves competitive performance against autoregressive baselines on benchmarks like MMMU (43.3% vs 35.1%), Math Vista, and ScienceQA. On COCO captioning, it surpasses Open-LLaVA-Next by +4.1 CIDEr with 1.92x speedup. Most impressively, LaViDa excels at constrained generation tasks, with 100% satisfaction on poem constraints compared to <50% for autoregressive models.
This research matters because it offers flexible speed-quality tradeoffs and superior text-infilling capabilities, enabling applications requiring structured outputs or format constraints that autoregressive models struggle with.
Investments
Quantum Systems, the drone company for defense and commercial applications, raised a €160M Series C financing round from Balderton Capital, Hensoldt, and Airbus Defense and Space.
NewLimit, a biotech company developing medicines to extend human healthspan through epigenetic reprogramming, raised a $130M Series B from Kleiner Perkins, Khosla Ventures, and Human Capital.
Arondite, a UK-based AI startup focused on enhancing human-machine collaboration in defence, raised a $10M Seed financing round led by Index Ventures, with participation from Concept Ventures and Creator Fund.
Relevance AI, the San Francisco- and Sydney-based startup developing an AI agent operating system, raised a $24M Series B financing round led by Bessemer Venture Partners with participation from King River Capital and Insight Partners.
Toloka, the AI data annotation company that’s part of Nebius Group, raised a $75M funding round led by Bezos Expeditions and Mikhail Parakhin (CTO of Shopify).
Inductive Bio, a AI company working on ADMET for small molecule drug discovery, raised a $25M Series A from Obvious Ventures and a16z Bio + Health.
Parloa, the German AI agents startup for customer service automation, raised a $120M Series C financing round at a $1B valuation from Durable Capital Partners LP, Altimeter Capital, and General Catalyst.
True Anomaly, a space defense technology company developing military-class orbital systems, raised $260M in a financing round led by Accel with participation from Meritech Capital and Eclipse.
Recraft, the AI startup specializing in image generation for branding and marketing, raised a $30M Series B financing round led by Accel, with participation from Khosla Ventures and Madrona.
Samaya AI, the company building expert AI agents for financial services, raised $43.5M in a financing round led by New Enterprise Associates.
Sensmore, the robotics startup pioneering Physical AI for heavy mobile machinery, raised a $7.3M financing round led by Point Nine.
LMArena, a platform for testing and voting on AI models, raised $100M in a financing round at a $600M valuation from Andreessen Horowitz and UC Investments.
Reflect Orbital, a company developing satellite constellations to deliver sunlight on demand, raised a $20M Series A from Lux Capital, Sequoia Capital, and Starship Ventures.
HederaDx, liquid biopsies for cancer care, raised a €15M Series A led by Vsquared Ventures.
OpenAI, the AI research and deployment company, secured more than $7 billion in financing from JPMorgan to build a data center.
SpAItial, the AI startup focused on spatial foundation models, raised a $13M seed financing round led by Earlybird Venture Capital with participation from Speedinvest and several high-profile angels.
Grammarly, the AI-powered writing assistant and productivity platform, raised $1 billion in a non-dilutive revenue-based financing round from General Catalyst, with no valuation disclosed.
Wordsmith, the legal intelligence platform for in-house teams, raised a $25M Series A led by Index Ventures.
LawZero, a nonprofit research group focused on safe AI development, raised $30M in a financing round from Eric Schmidt’s philanthropic organization and Skype co-founder Jaan Tallinn.
Cast AI, the application performance automation platform, raised a $108M Series C at an $850M valuation from G2 Venture Partners and SoftBank Vision Fund 2.
Optimal Dynamics, the AI-driven decision intelligence platform for trucking and logistics, raised a $40M Series C financing round led by Koch Disruptive Technologies.
Stack AI, the enterprise AI platform for creating custom AI agents, raised a $16M Series A financing round from Lobby VC, LifeX Ventures, and Gradient.
Greenlite AI, the AI compliance automation company for financial institutions, raised a $15M Series A financing round from Greylock, Thomson Reuters, and Canvas Prime.
Octave, the AI marketing-tech startup focused on enhancing customer profiling and campaign strategy, raised a $5.5M seed financing round from Bonfire Ventures, Unusual Ventures, and Bee Partners.
Nous Research, the decentralized AI startup leveraging Solana for open-source AI models, raised a $50M Series A at a $1B valuation from Paradigm, with previous investors including Distributed Global and North Island Ventures.
Hedra, the AI-powered video generation and editing platform, raised $32M in a Series A financing round led by a16z, with participation from Index Ventures and Abstract Ventures.
Vercept, the AI startup developing a “computer interface of the future,” raised a $16M seed financing round from Fifty Years, Point Nine, and the AI2 Incubator.
Rumored investments
Decagon, the AI startup building customer service agents, is raising $100M in a financing round at a $1.5B valuation from Andreessen Horowitz and Accel. The company is at ~$17M ARR, up from ~$1.5M ARR LTM.
Abridge, the AI health startup focused on medical transcription and documentation, is in talks to raise a financing round at a $5 billion valuation; the list of investors in the new round was not disclosed.
Anysphere, the maker of the AI-powered coding tool Cursor, is rumored to have raised $900M in a financing round at a $9B valuation from Thrive Capital, a16z, and Accel. Confirmed yesterday
Aris Machina, the AI-enhanced industrial software platform targeting manufacturing optimization, is rumored to be raising a financing round from Earlybird, Village Global, and AENU.
Acquisitions
Ezra, the company pioneering affordable full-body MRIs, was acquired by Function Health; the acquisition price was not disclosed.
Windsurf, an AI-assisted coding tool formerly known as Codeium, entered into an agreement to be acquired by OpenAI for $3 billion. Windsurf had previously raised $150M in a Series C funding round led by General Catalyst at a $1.25B valuation.
Databricks, the Data and AI company, agreed to acquire Neon, a serverless Postgres platform for developers and AI agents, for a rumored $1B.
Convergence.ai, an AI agent company specializing in adaptive, intelligent systems for digital workflows, was acquired by Salesforce; the acquisition price was not disclosed.
Together AI, the AI Acceleration Cloud platform for developers and enterprises, acquired Refuel.ai to integrate its models and platform capabilities; the acquisition price was not disclosed.
23andMe, a genetics-led consumer healthcare and biotechnology company, was acquired by Regeneron Pharmaceuticals for $256 million. I include this here because of the presumed value of 23andMe’s data business for drug discovery.
io, a company focused on developing inspiring and empowering products (vague yes) that was formed by Jonny Ive and Sam Altman, was acquired by OpenAI for a whopping $6.5B.
Twirl, a data orchestrator for developing, testing, and deploying data pipelines, was acquired by Modal. The acquisition price was not disclosed.
Moonhub, the AI recruitment company that developed the world’s first AI Recruiter, was acquired by Salesforce. The acquisition price was not disclosed.
Informatica, a leader in AI-powered enterprise cloud data management, entered into a definitive agreement to be acquired by Salesforce for approximately $8 billion.
Mapmygenome, an AI-driven genomics and personalized health company, acquired Microbiome Insights to expand its North American footprint; the acquisition price was not disclosed.
Sphinx Bio, a software for AI workflows in biology, was acquired by Benchling for an undisclosed amount.