AI News: May 28, 2026

1. Cognition Raises $1B at $25B Pre-Money, Doubling Valuation in Eight Months

Cognition. The maker of the Devin coding agent closed a $1B round at a $25B pre-money valuation, more than doubling its September 2025 mark, with annualized revenue reportedly at $492M. The pace of revaluation tells you more about the market than the company: investors are pricing coding-agent leaders against a winner-take-most narrative even as enterprises remain mixed on Devin’s day-to-day reliability versus Claude Code and Codex. Source

2. Snowflake Signs $6B Five-Year Deal for AWS AI CPUs

Snowflake / AWS. Snowflake committed to $6B over five years for AWS-built AI CPUs (Graviton-derivative silicon optimized for inference), the largest non-NVIDIA AI compute commitment to date by a major data platform. The deal is interesting less for the dollar amount than for what it signals: hyperscalers’ homegrown CPUs are now credible enough that an inference-heavy customer is willing to bet a five-year capex cycle on them rather than waiting on H200/B200 allocations. Source

3. Robinhood Lets AI Agents Trade Stocks via MCP

Robinhood. Robinhood quietly added support for connecting AI agents — including Claude via MCP — to brokerage accounts so they can analyze portfolios and execute trades against pre-loaded wallet balances. Regulators have already flagged the obvious risks (unchecked agents firing market orders during volatility), and how this plays with FINRA suitability rules is going to be an interesting test case for retail AI. Source

4. YouTube Begins Auto-Labeling AI-Generated Video

YouTube. YouTube switched on automatic detection and labeling for videos featuring significant photorealistic AI content, instead of relying on creator self-disclosure. The labels are visible but YouTube has confirmed they don’t affect monetization or recommendation reach — which is the actual policy question that matters and which YouTube is conspicuously deferring on. Source

5. Microsoft MAI-Image-2.5 Catches Google’s Nano Banana 2 on Arena

Microsoft. Microsoft’s in-house image model MAI-Image-2.5 now sits third on Arena’s image leaderboard, matching Google’s Nano Banana 2 and trailing OpenAI. The interesting bit is Microsoft running its own image stack at all — through 2024 most of Microsoft’s image work shipped through OpenAI models, and standing up MAI-Image as a parallel option is part of the broader push to reduce that dependency. Source

6. China Tightens Hold on Domestic AI Researchers

Policy. A TechCrunch piece details how Chinese authorities and labs are increasingly making it harder for top AI researchers to take roles abroad, through a mix of export-control reasoning, internal pay floors, and informal travel restrictions. The story is hard to verify in detail but matches a year of compounding signals; the practical implication for Western labs is that the recruiting pipeline that fed Google Brain and OpenAI a decade ago is materially smaller than it used to be. Source

7. NVIDIA’s Annual Taiwan Spend Jumps Tenfold to $150B

NVIDIA / TSMC. NVIDIA’s spending with Taiwanese suppliers — overwhelmingly TSMC plus advanced-packaging partners — went from roughly $15B to $150B per year as Blackwell volume ramped. The number recontextualizes the geopolitical risk discussion: TSMC isn’t just NVIDIA’s most important supplier, it’s now functionally a critical-path dependency for the entire Western AI buildout. Source

8. Remote.com Claims 50% Revenue-per-Employee Lift From AI

Remote. Payroll/EOR platform Remote crossed $300M ARR and turned cash-flow positive while holding headcount flat, attributing roughly a 50% revenue-per-employee lift to internal AI automation. As always with vendor self-reports the gain almost certainly bundles in headcount discipline that would have happened anyway, but the order of magnitude — 50% rather than 5–10% — is on the high end of what mid-sized SaaS companies have publicly claimed. Source

9. Meta Launches Paid Subscriptions With AI Plans on the Roadmap

Meta. Meta rolled out paid subscription tiers across Instagram, Facebook, and WhatsApp under the Meta One brand, with AI-specific plans flagged for later in the year. The structure looks designed to give Meta a way to commercialize the heavier inference workloads its generative features increasingly run — an open question is whether premium AI lands as part of Meta One or in a separate Meta AI Pro tier. Source

10. Box CEO Coins “AI Psychosis” Among Tech Leaders

Industry. Box CEO Aaron Levie publicly suggested some tech CEOs are exhibiting “almost religious” beliefs about AI productivity gains, flagging the gap between board-deck narratives and observed delivery. It’s an inside-baseball comment but worth flagging because it lines up with a quietly growing cohort of operators noting that the productivity wins are real but lumpy, not the universal across-the-board lift that the loudest AI-maximalist CEOs are describing. Source

11. Jack Clark Delivers Cosmos HAI Lecture on AI Choice Architecture

Import AI. Jack Clark’s Import AI 458 publishes his 2026 Cosmos HAI Lab Lecture at Oxford, framing continued AI progress as a society-level choice between exploring the future and retreating from the present. The personal accounts of working with frontier systems are the more useful read for practitioners — they’re Clark’s most direct public commentary in a while on what’s changed inside Anthropic as model capability has scaled. Source