# Anthropic’s $65B Round, Production-Grade Agents, and Mistral’s Infrastructure Push

*By AI News Digest • May 29, 2026*

Anthropic’s massive new round led the day, but the broader story was operational: OpenAI pushed agents closer to production deployment, Onyx argued enterprises now need dedicated control planes for autonomous systems, and Mistral tied AI competitiveness to domestic infrastructure.

## What stood out

Today’s strongest signal was that AI is being treated more like infrastructure: huge capital at the top, more production-ready agent tooling in the middle, and tighter control requirements at the edge [^1][^2][^3][^4].

### Anthropic set a new financing marker — even as AI spend gets harder to justify

Anthropic said it raised **$65 billion** in Series H at a **$965 billion** post-money valuation, and said the capital will fund research and expand capacity for Claude. The company also said run-rate revenue has crossed **$47 billion**, driven by Claude moving into core operations across industries and broader everyday use [^1][^5].

At the same time, enterprise cost discipline is getting sharper. Axios reported that companies are starting to question whether soaring AI spending is delivering meaningful returns, and cited an AI consultant describing a client that spent **half a billion dollars in one month** on Claude licenses after failing to set usage limits [^6].

*Why it matters:* Capital markets are still rewarding AI scale, but enterprise buyers are paying much closer attention to cost controls and return on spend [^1][^5][^6].

### The agent stack is starting to look like production infrastructure

OpenAI rolled out major Agents SDK updates: a Codex-style harness for long-running tasks, first-class sandbox support across multiple providers, a hosted shell and containers endpoint in the Responses API, agent memory, snapshotting and rehydration, handoffs, approvals, and TypeScript support alongside Python [^2].

OpenAI also emphasized separating the harness from compute so sandboxes can stay ephemeral while state is restored when work resumes [^2]. That lands as background coding agents are becoming more practical: Latent Space described a December 2025 inflection when models such as Opus 4.5 and GPT 5.2 made spec-to-PR workflows viable with much less handholding, while Cognition said Devin’s merged PRs grew **7x** and its share of commits rose from **16%** in January to **80%** in March [^7].

*Why it matters:* The center of gravity is shifting from assistant-style chat in the IDE toward cloud-run agents that can keep state, use tools, and work over real files and environments for longer stretches [^2][^7].

### Enterprise adoption is pulling agent security forward

Onyx Security described a **secure control plane** that uses small specialized models to monitor autonomous agents and decide when a smarter review agent should step in, aiming to reduce illegitimate or incorrect actions without adding human review to every step [^3].

Its CEO said that, in a typical enterprise, **more than half** of deployed agents are now autonomous coding agents, with another roughly **45%** coming from low-code automations, and said the fastest-growing category often launches without controls [^3]. He argued that traditional identity, endpoint, and API security tools lack the context to judge *why* an agent is taking a particular action, even as incidents have included downtime and accidental publication of code or tokens [^3].

*Why it matters:* If autonomous agents keep spreading, oversight tools will increasingly need to understand intent and context, not just permissions and endpoints [^3].

### Mistral is pairing enterprise applications with sovereign infrastructure

Mistral framed high-end manufacturing as a major target market, arguing that useful models in this sector must understand physics, object dynamics, and factory optimization. The company also said it is building stack-wide sovereign infrastructure — from models and deployment engineering to controlled hosting capacity — and announced a new French data-center site on a roadmap to **200 MW** by the end of 2027 and **1 GW** by 2029 [^4][^8][^4][^8].

Related interviews said Mistral has already invested about **€4 billion** in data centers across France and Sweden, with CapEx plans in the multi-billion-euro range, and CEO Arthur Mensch said the company remains on track for **$1 billion ARR** in 2026, driven by software growth and infrastructure deployment [^9][^4]. Mensch also argued Europe has a short window to build independent AI infrastructure, calling AI a strategic and macroeconomic asset and pointing to limited chips, memory, and electricity as constraints [^9][^10].

*Why it matters:* This is one of the clearest European cases for pairing industry-specific AI applications with domestic compute capacity and customizable models [^4][^9][^10].

### Also worth tracking

- **Open-model momentum keeps building.** NVIDIA said it is adopting the Linux Foundation’s OpenMDW framework across its open model families to simplify licensing, while Factory reported open-model use more than **3x** higher relative to closed models over the last month and Fireworks said it is processing **30 trillion tokens per day** with open-model share still climbing [^11][^12][^13].
- **RL infrastructure is getting cheaper and more distributed.** Hugging Face said its science team cut async RL weight-sync bandwidth by about **100x** by sending only changed weights, reducing one example payload from **1.2 GB** to **20–35 MB** and demonstrating fully disaggregated training over HTTPS and a bucket rather than a shared cluster [^14].
- **NVIDIA highlighted sim-to-real robotics progress at ICRA.** Among the results it cited: **3x** faster multi-arm planning on Jetson with ScheduleStream, **4.5x** better navigation success with COMPASS and roughly **80%** real-world success with zero real training data, and around **75%** real-robot grasp success for Grasp-MPC versus **41%** baseline [^15].

---

### Sources

[^1]: [𝕏 post by @AnthropicAI](https://x.com/AnthropicAI/status/2060061347522433422)
[^2]: [Build Hour: Agents SDK](https://www.youtube.com/watch?v=tK32trvj_b4)
[^3]: [Building an AI Guardian for Enterprise with Onyx Security CEO Maxim Bar Kogan](https://www.youtube.com/watch?v=QDsbFLEt9ro)
[^4]: [Mistral AI Signs New Manufacturing Deals With Airbus, BMW](https://www.youtube.com/watch?v=0b58a0-C8t8)
[^5]: [𝕏 post by @AnthropicAI](https://x.com/AnthropicAI/status/2060061348818518493)
[^6]: [𝕏 post by @axios](https://x.com/axios/status/2060047174268785130)
[^7]: [The Age of Async Agents — Cognition's Walden Yan & OpenInspect's Cole Murray](https://www.latent.space/p/cognition)
[^8]: [US Strikes Iran Targets; Snowflake Jumps on Results | Bloomberg Brief 5/28/2026](https://www.youtube.com/watch?v=wsoffld61yI)
[^9]: [Mistral AI CEO: We invest 'where we think we have an edge'](https://www.youtube.com/watch?v=61fZkfsj2eg)
[^10]: [Mistral CEO: EU policymakers realize that something needs to be done on AI](https://www.youtube.com/watch?v=AvwXPQNYsrQ)
[^11]: [𝕏 post by @NVIDIAAI](https://x.com/NVIDIAAI/status/2060035668655677804)
[^12]: [𝕏 post by @matanSF](https://x.com/matanSF/status/2060005777348112734)
[^13]: [𝕏 post by @FireworksAI_HQ](https://x.com/FireworksAI_HQ/status/2060103886028046739)
[^14]: [𝕏 post by @ClementDelangue](https://x.com/ClementDelangue/status/2059989047947260203)
[^15]: [NVIDIA Research Advances Robotics From Simulation to the Real World](https://blogs.nvidia.com/blog/icra-research-robotics-simulation-to-real-world)