AI News Roundup: China Blocks Meta’s $2B Manus Deal, Microsoft-OpenAI End Exclusivity, DeepSeek V4 Launches

China Blocks Meta’s $2 Billion Acquisition of AI Startup Manus

In a dramatic escalation of US-China tech tensions, China’s National Development and Reform Commission on Monday ordered the cancellation of Meta’s $2 billion acquisition of Manus, the Singaporean AI agent startup with Chinese roots. The decision came after a months-long regulatory probe and marks the first time Beijing has blocked a major AI deal on national-security grounds.

Manus, which gained viral attention in early 2025 for its autonomous AI agent platform, was founded by Chinese engineers and maintains significant R&D operations in China. The block signals that Beijing is willing to extend its regulatory reach beyond its borders to prevent strategic AI talent and technology from flowing to American tech giants. Meta has not yet commented publicly on whether it will challenge the decision.

Microsoft and OpenAI Gut Their Exclusive Deal

One of the most consequential partnerships in the AI era got a sweeping rewrite on Monday. Microsoft will no longer have exclusive access to OpenAI’s technology, clearing the way for the ChatGPT maker to sell its products on rival cloud platforms including Amazon Web Services and Google Cloud.

The restructuring was triggered by Amazon’s massive $50 billion investment in OpenAI, announced in February — a deal that almost certainly conflicted with the original Microsoft exclusivity terms. Under the new arrangement, Microsoft retains a non-exclusive license to OpenAI’s intellectual property through 2032, but will no longer share revenue for OpenAI products sold on Azure. The move may also help Microsoft fend off antitrust scrutiny in the UK, US, and Europe over whether its OpenAI tie-up gave it an unfair cloud advantage.

DeepSeek Drops V4: 1.6 Trillion Parameters, Fraction of the Price

Chinese AI lab DeepSeek released its long-awaited V4 model family on Thursday, and the benchmarks are turning heads. The flagship V4 Pro packs 1.6 trillion total parameters (49 billion active) using a mixture-of-experts architecture, making it the largest open-weight model available. A leaner V4 Flash variant runs at 284 billion parameters with just 13 billion active.

DeepSeek claims V4 Pro outperforms OpenAI’s GPT-5.2 and Google’s Gemini 3.0 Pro on several reasoning benchmarks, with coding performance “comparable to GPT-5.4.” Both models support 1-million-token context windows. The real story, though, is pricing: V4 Flash costs just $0.14 per million input tokens — a fraction of what frontier competitors charge — reinforcing DeepSeek’s strategy of making top-tier AI radically affordable. The models remain text-only, lacking the multimodal capabilities of closed-source rivals.

Snap Cuts 1,000 Jobs as AI Writes 65% of Its Code

Snapchat’s parent company laid off 16% of its workforce — roughly 1,000 employees — on April 15, in what may be the starkest example yet of AI-driven workforce displacement in Big Tech. CEO Evan Spiegel said AI now generates more than 65% of Snap’s new code, allowing smaller teams to match the output of much larger ones.

The cuts will reduce Snap’s annualized cost base by more than $500 million by the second half of 2026. Investors cheered the move, sending shares up 8% on the announcement. The restructuring drew criticism from labor advocates who called it “AI washing” — using AI narratives to justify cost-cutting that may be driven by broader business pressures. Either way, Snap’s explicit framing sets a precedent other companies are likely to follow.

Alibaba Unleashes Qwen 3.6 Model Family

Alibaba’s Qwen team had a prolific April, releasing multiple models in its latest generation. Qwen3.6-Max-Preview, dropped on April 20, topped six major coding benchmarks and posted strong gains in world knowledge and instruction following. A week earlier, the open-weight Qwen3.6-27B became arguably the most capable 27-billion-parameter dense model available, outperforming 397-billion-parameter mixture-of-experts models on agentic coding benchmarks.

The releases underscore how quickly Chinese AI labs are closing the gap with Western frontier models, particularly in coding and agentic tasks. Alibaba also released Qwen3.6-Plus for enterprise agentic AI deployment, with all 35,000+ employees at partner Publicis Groupe receiving access as part of a broader Microsoft co-marketing deal.

19 New AI Laws Enacted Across US States

The wave of state-level AI regulation continues to accelerate. Over the past two weeks alone, 19 new AI bills were signed into law across multiple states, according to the AI Governance Watch tracker. Utah Governor Spencer Cox has signed nine AI bills into law just this year, making the state one of the most active in AI policymaking.

Key themes include transparency requirements, restrictions on AI in health insurance decisions (Indiana, Utah, and Washington all enacted laws prohibiting insurers from using AI as the sole basis for claim denials), and alignment with the White House’s National Policy Framework released in March. With more than 600 AI bills introduced across state legislatures in 2026 so far, companies operating nationally face an increasingly complex patchwork of compliance obligations.