Top Stories
US Government Secures Pre-Release Access to AI Models From Google, Microsoft, and xAI
The Center for AI Standards and Innovation (CAISI) at NIST announced on May 5 that it has signed new agreements with Google DeepMind, Microsoft, and xAI to evaluate frontier AI models before they reach the public. Under the deals, the three companies will hand over models — including versions with reduced or removed safeguards — for national security testing in classified environments.
CAISI, which serves as the government’s primary hub for AI model evaluation, said it has already completed more than 40 evaluations of state-of-the-art models, including some not yet publicly available. The agreements cover both pre-deployment and post-deployment assessments, and were drafted with flexibility to adapt as capabilities advance.
The move represents the Trump administration’s most concrete step toward AI oversight to date. OpenAI and Anthropic had already signed similar agreements back in 2024. CAISI said those earlier partnerships “have been renegotiated” to align with newer directives from Commerce Secretary Howard Lutnick and the administration’s AI Action Plan. The net result: every major frontier AI developer now has some form of agreement to let the federal government test its models before release.
Cerebras Files for $3.5 Billion IPO at $26.6 Billion Valuation
AI chipmaker Cerebras Systems filed an amended registration statement on May 4, targeting an IPO that would raise $3.5 billion and value the company at $26.6 billion. The Sunnyvale startup plans to sell 28 million shares at $115 to $125 each on the Nasdaq.
If the deal goes through at the high end, it would be the largest tech IPO of 2026 and the most direct public-market challenge yet to Nvidia’s dominance in AI hardware. Cerebras reported revenue of $510 million in 2025, up 76% year over year, and swung to a net income of $237.8 million after posting a $481.6 million loss the prior year.
The company’s close partnership with OpenAI has drawn attention — and scrutiny. OpenAI accounts for a significant portion of Cerebras’ revenue, and the two companies signed a $10 billion inference deal in 2024. Whether investors see that concentration as a strength or a risk will likely shape how the offering is received.
Sources: CNBC, TechCrunch
After China Blocks Meta-Manus Deal, AI Startups Reassess Cross-Border Ambitions
Bloomberg reported on May 5 that Chinese AI startups are rethinking their international strategies in the wake of Beijing’s decision to block Meta’s $2 billion acquisition of Manus, the Singapore-based AI agent startup with Chinese roots. China’s state planner described the deal as a “conspiratorial” attempt to hollow out the country’s technology base.
The block has sent a clear signal: AI talent and capabilities are now treated as national security assets by both Washington and Beijing. Manus’ founders had originally established their parent company, Butterfly Effect, in Beijing in 2022 before relocating to Singapore — a common playbook for Chinese AI companies seeking global reach. That playbook just got significantly harder to execute.
The ripple effects extend beyond Manus. Several Chinese-founded AI startups with international operations are reportedly reassessing their corporate structures and acquisition strategies. The deal’s collapse also leaves Meta without the agent technology it had planned to fold into Meta AI, at a time when competitors are racing to ship autonomous agent products.
Quick Hits
-
AI infrastructure: IREN announced it is acquiring Mirantis to strengthen its AI cloud delivery capabilities, adding another player to the rapidly consolidating AI compute infrastructure market. (GlobeNewsWire)
-
AI-powered attacks: The Hacker News reports that 2026 is shaping up as “the year of AI-assisted attacks,” with exploits arriving before patches in 28.3% of cases and AI-generated phishing campaigns becoming increasingly difficult to distinguish from legitimate communications. (The Hacker News)
-
Google Pentagon fallout: Fortune reports that while 600+ Google employees protested the company’s classified Pentagon AI deal, the internal backlash lacks the teeth of the 2018 Project Maven revolt. Employees today have less leverage — the labor market is tighter, and Google’s leadership has made clear that military contracts are part of the business. (Fortune)
-
Foxconn AI demand: Foxconn reported a 29.7% revenue increase in April, driven primarily by surging demand for AI servers, underscoring how AI hardware demand continues to reshape the supply chain. (Taiwan News)
Worth Watching
The CAISI testing framework is worth following closely. With all major frontier AI developers now under some form of pre-release government evaluation, the question shifts from “will there be oversight?” to “how meaningful is it?” CAISI’s 40+ completed evaluations suggest real work is happening, but the voluntary nature of these agreements — and the fact that companies control which models they submit — leaves significant gaps. Whether this evolves into something with real enforcement teeth, or remains a gentlemen’s agreement, will shape the regulatory environment for years.
Cerebras’ IPO could set the tone for AI chip competition. If the market rewards a Nvidia challenger at this valuation, expect a wave of AI hardware startups to follow. If it stumbles, Nvidia’s dominance gets another year of breathing room.