AI Agent Autonomy Crunch | AI Field Notes #14

AnalysisMistral, the Paris AI lab behind the Le Chat assistant, announced on April 29 that its Vibe coding agent (a tool that turns plain English instructions into code changes) can now run as a remote session in the cloud. Several can run in parallel and notify you when a branch or pull request is ready, without anyone babysitting a terminal. The new default model, Mistral Medium 3.5, is a 128 billion parameter open-weight model with a 256k token context window, scoring 77.6% on SWE-bench Verified (a coding test that fixes real GitHub bugs). Pricing lands at $1.50 input and $7.50 output per million tokens. The shift is that shipping AI-built software now depends on parallel review capacity more than raw model intelligence.

AnalysisHuawei expects roughly $12 billion in AI chip revenue in 2026, up 60% from $7.5 billion last year, the Financial Times reported on May 1. Most of that comes from the Ascend 950PR, an inference chip that entered mass production in March and ships with a software layer that mimics CUDA, Nvidia's programming framework that has kept Chinese labs tied to American silicon. ByteDance has committed about $5.6 billion in orders, with Alibaba and Tencent following. The 950PR delivers about 2.8 times the FP4 performance of Nvidia's H20 (the export-control-compliant chip Nvidia could legally sell into China) at roughly $16,000 per card. With Nvidia mostly shut out of the leading edge in China, Huawei could capture 60% of domestic AI chip share by year end.

AnalysisApple reported $111.2 billion in fiscal Q2 revenue on April 30, up 17%, with Services hitting nearly $31 billion and another $100 billion buyback authorized. Capex actually fell 36% year over year to about $1.9 billion, a striking gap to the more than $650 billion that Microsoft, Meta, Alphabet, and Amazon plan to spend together on AI infrastructure this year. Apple's strategy still leans on partnerships, including the Gemini deal that will power the personalized Siri later this year, and on running more inference (the cost of actually using a model, separate from training it) on-device through its own silicon. Tim Cook flagged Mac demand outpacing supply, citing what he called a "DIY agentic AI boom" in China. The result is Apple looking healthier than peers without joining the spending spree.

AnalysisPrivate equity firm KKR secured more than $10 billion in commitments this week to launch Helix Digital Infrastructure, a new vehicle that will design, build, own, and operate AI data centers along with the power generation, transmission, and connectivity to feed them. The bundling matters: the constraint is no longer GPUs alone, but the megawatts to run them. PJM, the largest US grid operator, is forecast to be six gigawatts short of reliability requirements by 2027, and roughly 30% of new data center capacity is now projected to come from on-site generation, up from near zero a year ago. The IEA reported that data center electricity use surged 17% in 2025, while AI-focused facilities climbed even faster.

AnalysisLegal AI startup Legora raised a $50 million Series D extension on April 30 led in part by Nvidia's NVentures, reportedly its first legal AI investment, lifting Legora's valuation to $5.6 billion barely a month after a $550 million round. The Stockholm-founded company crossed $100 million in annual recurring revenue and now serves about 100,000 lawyers across 1,300 organizations including Latham and Watkins, Hengeler Mueller, and the in-house team at T-Mobile. Rival Harvey, based in San Francisco, hit an $11 billion valuation last month with backing from Sequoia and Andreessen Horowitz. Both are spending hard on advertising, Harvey via the actor from "Suits" and Legora via Jude Law, which tells you the differentiation has moved from product to mindshare.

AnalysisA run of late-April research points to AI agents emerging as the most consequential unsecured asset in the enterprise. An AGAT survey published this week found 88% of organizations reported confirmed or suspected AI agent security incidents in the past year, with healthcare at 92.7%. Stanford's Trustworthy AI Lab found that fine-tuning attacks bypassed Anthropic's Claude Haiku in 72% of cases and OpenAI's GPT-4o in 57%, meaning model-level guardrails alone are not enough. The execution layer (where an agent calls a tool, opens an API, or moves money) is where the actual breaches happen. Cisco's State of AI Security 2026 and recent NIST guidance both reframe prompt injection from a chatbot trick into a structural attack on autonomous agents that read untrusted text and then act on it.

AnalysisTwo enterprise AI rounds closed on May 1 that show where money is moving inside the software stack. Hightouch raised $150 million at a $2.75 billion valuation led by Goldman Sachs Alternatives and Bain Capital Ventures to build agents that act on customer data inside marketing workflows. Netomi pulled in $110 million led by Accenture Ventures, with Adobe Ventures participating, to build agentic customer experience tools for large enterprises. Featherless.ai raised a smaller $20 million Series A from AMD Ventures and Airbus Ventures to provide serverless inference (running open-source models on demand without dedicated GPUs) for teams that want an alternative to closed APIs. The pattern: capital is funding agents that take action inside operational software, not chat interfaces bolted on top.