AI News: June 29, 2026

1. Google Limits Meta’s Access to Gemini Models Amid Compute Shortage

Google. Google has restricted Meta’s purchase of access to its Gemini models because it could not supply the full computing capacity Meta requested, according to a Financial Times report. Google told Meta around March that it could not meet the demand, disrupting several Meta AI projects and pushing Meta toward an internal model for some workloads. The episode underscores how tight compute has become, with Google itself reportedly paying SpaceX roughly $920 million a month for bridge GPU capacity. Source

2. GLM-5.2 Matches Leading US Models at Finding Security Bugs

Z.ai. Independent security evaluations from Semgrep and Graphistry reported that Z.ai’s open-weight GLM-5.2 matched or beat US frontier models at finding cybersecurity bugs, with one benchmark putting it ahead of Claude Code at 39 percent versus 32 percent and tying closed models on agentic cyber investigations. The results, surfaced in a Wall Street Journal writeup, sharpen the debate over US export restrictions on high-capability models, since GLM-5.2 is freely downloadable under a permissive license and sits outside those controls. For security teams, it points to a capable open-weight option for automated vulnerability discovery. Source

3. Coinbase Halves AI Spending by Routing to Chinese Models

Coinbase. Coinbase adopted Chinese open-weight models including GLM 5.2 and Kimi 2.7 through an automated routing system, roughly halving its AI costs even as token usage rose. Improved caching lifted its cache hit rate from 5 percent to 60 percent, compounding the savings. The move adds Coinbase to a growing list of Western firms testing lower-cost Chinese models as frontier-API pricing comes under pressure. Source

4. Princeton’s CEO-Bench Finds Only Three Models Survive a 500-Day Startup Sim

Princeton. Princeton released CEO-Bench, a benchmark that tasks AI agents with running a simulated software company over 500 days, and most models finished below their starting capital. A simple rule-based system outperformed nearly all of the LLM agents tested. The result highlights persistent weaknesses in long-horizon agentic decision-making, where a model has to plan and act consistently over an extended run rather than a single session. Source

5. Tencent Paper Argues AI Must Finish Tasks, Not Just Answer

Tencent. A Tencent research paper contends that AI systems need to move from generating responses to completing whole tasks inside persistent work environments. The authors frame durable workspaces paired with reusable skills as prerequisites for AI that can act as a reliable digital colleague rather than a single-turn assistant. The framing reflects a broader shift toward agents measured by completed work instead of answer quality. Source

6. GM-Backed Momenta Launches Hong Kong IPO Seeking Up to $751M

Momenta. The GM-backed Chinese autonomous-driving software firm Momenta opened its Hong Kong IPO, offering 19.9 million shares at HK$295.60 each to raise up to HK$5.89 billion, about $751 million, at a roughly $9 billion valuation. Around 60 percent of the proceeds are earmarked for research and development, including AI compute and data storage, with about 20 percent allocated to its robotaxi rollout. Trading is set to begin July 8. Source

7. Wall Street Eyes Micron as the Next Nvidia on AI Memory Demand

Micron. Investors are positioning memory maker Micron as a potential major beneficiary of the AI buildout, citing surging demand for the high-bandwidth memory that AI accelerators depend on. Analysts frame its role in HBM and broader memory supply as a structural tailwind comparable to Nvidia’s earlier run. The thesis reflects how AI infrastructure spending is rippling beyond GPUs into the wider hardware supply chain. Source

8. Ford Rehires Veteran Engineers After AI Fell Short on Quality

Ford. Ford said that relying on AI alone proved insufficient for quality product development and has brought back experienced senior engineers it refers to as “gray beards.” The company concluded that deploying AI without deep human expertise could not guarantee high-quality engineering outcomes. The reversal is a counter-signal to expectations that AI can wholesale replace specialized technical staff. Source

9. Payroll Data Shows Employment Falling for Young Workers in AI-Exposed Jobs

Researchers. An analysis of US payroll data spanning more than 730 occupations found that employment among workers ages 22 to 25 in highly AI-exposed roles is now contracting at roughly 3.8 percent per year. The finding adds concrete labor-market evidence to the debate over how AI is affecting entry-level knowledge work. It suggests displacement pressure is showing up first among the youngest workers in the most exposed fields. Source