AI News: May 30, 2026

1. Groq Raises $650M to Pivot Toward an Inference Cloud

Groq. AI chip startup Groq is reportedly raising $650M from existing investors to expand an inference-hosting business, following its roughly $20B arrangement with Nvidia in December 2025 in which senior staff left for Nvidia and Groq licensed out its hardware technology rather than being acquired outright. The raise is essentially backstopped by committed investors, which tells you the market wants Groq alive as a neutral inference provider even after Nvidia absorbed much of its talent. The strategic read is that Groq has stopped trying to beat Nvidia on silicon and is instead selling fast inference as a service, a more defensible position than competing on chips it no longer fully controls. Source

2. XCENA Raises $135M Betting Memory, Not Compute, Is the Real Bottleneck

XCENA. South Korean chip startup XCENA closed a $135M Series B at a $570M valuation, arguing that memory bandwidth rather than raw compute is what actually limits AI workloads. Its hardware processes data near DRAM to cut expensive CPU-to-GPU round trips, a near-memory approach the company says can lower infrastructure costs if it scales to production. The bet is worth watching because the industry has spent two years throwing more FLOPs at the problem, and a credible memory-bound counterargument backed by real silicon and a $570M valuation is rarer than another accelerator startup. Source

3. CNN Sues Perplexity Over Alleged Copyright Infringement and Paywall Scraping

Perplexity. CNN filed suit against Perplexity in New York federal court, alleging the AI search engine reproduced substantial portions of its articles, including paywalled content, and generated near-verbatim excerpts when users entered story headlines. The complaint says Perplexity ignored crawler blocks and kept using CNN journalism after licensing talks collapsed the prior year, and it seeks damages plus a permanent injunction; Perplexity counters that facts cannot be copyrighted. CNN joins The New York Times, Reddit, and Dow Jones in pressing scraping and copyright claims, and the accumulating litigation is steadily defining where retrieval-augmented answers cross from fair use into infringement. Source

4. A Company Reportedly Ran Up a $500M Claude Bill in One Month

Industry. The Decoder reports that an unnamed company accumulated roughly $500M in Claude charges in a single month after failing to configure usage limits on its licenses. Whether or not the exact figure holds, the episode is a clean illustration of how agentic workloads make token consumption nonlinear and hard to forecast, and how flat pricing without hard caps can produce runaway bills. The practical takeaway for anyone deploying agents at scale is that FinOps controls and per-team usage limits are now load-bearing infrastructure, not optional hygiene. Source

5. Amazon Scraps Its Internal AI Usage Leaderboard After Employees Gamed It

Amazon. Amazon discontinued its internal “Kirorank” AI usage leaderboard after employees inflated their scores by running meaningless tasks, which drove up cloud costs; an SVP conceded the system was well intentioned but backfired. The company is shifting to tracking normalized, meaningful AI-generated code rather than raw usage counts. It is a useful real-world data point on a mistake many organizations are about to make, since rewarding raw AI usage as a vanity metric reliably produces exactly the gaming behavior it incentivizes. Source

6. Open-Source Claude Code Harness Pushes a Contract-First Agent Workflow

Open Source. A solo developer released Claude Code Harness, an MIT-licensed plugin that wraps Claude Code in a structured plan-work-review-release loop, using two source-of-truth files, spec.md and plans.md, to enforce planning, testing, and release discipline. The idea is to convert freeform agent chat into evidence-backed, reviewed repository changes rather than ad hoc edits applied straight to the tree. It is part of a broader and welcome shift toward treating agent coding as an engineering process with a harness around it, instead of trusting a model to behave well from a single prompt. Source