AI News: August 22, 2026

1. Nvidia Acquires Poolside’s Model Factory for $6 Billion

Nvidia. Nvidia agreed to pay $6 billion for Poolside’s “Model Factory,” the software system the startup used to build AI models, along with 109 employees who worked on its Laguna coding model. Nvidia is also investing $1 billion in Poolside at a $12 billion pre-money valuation, with the deal structured so Poolside continues operating independently under its founders. The transaction expands Nvidia’s own model-building capabilities and its Nemotron line, following similar deals with Groq and Enfabrica. Source

2. DeepSeek Releases Experimental Vision Model That Rivals Claude Opus 4.8

DeepSeek. The Chinese AI company released Deepseek-V4-Flash-Vision-Exp, an experimental multimodal extension of its V4-Flash model built for agent-based applications that combine tool use with image understanding. The model supports up to 600 images per request at a flat cost of 384 tokens per image and is compatible with both OpenAI’s Chat Completions and Anthropic’s Messages API formats. On DeepSeek’s internal agent benchmarks, the model approaches or beats Anthropic’s Claude Opus 4.8. Source

3. Waymo Builds Its Own Chip to Cut Reliance on Nvidia

Waymo. Alphabet’s self-driving unit Waymo developed a proprietary processor, delivering more than 1,000 TOPS and built by TSMC on a 5-nanometer process, that is already running in its latest robotaxis. The chip handles sensor processing and onboard AI models for navigation and will also ship in a new robotaxi co-developed with China’s Zeekr. The move reduces Waymo’s dependence on Nvidia and AMD chips, a notable supply-chain shift for one of the largest autonomous vehicle operators. Source

4. Starcloud Raises $250M to Expand Orbital Data Centers

Starcloud. The satellite startup raised a $250 million extension to its March Series A, bringing total funding to $420 million at a $2.3 billion valuation, led by Manhattan West Ventures with participation from Nvidia, Cisco, and Benchmark. Starcloud runs Nvidia H100 GPUs in orbit and is co-developing a space-specific chip with Nvidia called Vera Rubin Space-1. CEO Philip Johnston said rocket launch capacity, not funding, is now the company’s biggest bottleneck as SpaceX prepares to retire the Falcon 9 in 2028. Source

5. US Pushes Allies to Pick Washington or Beijing on AI

Government & Policy. The US State Department is drafting a letter urging allied nations to commit exclusively to either the US-led “Pax Silica” coalition or China’s AI framework, warning that countries trying to straddle both risk exclusion. The push aims to secure supply chains for chips, AI systems, and critical minerals as Pax Silica, now roughly two dozen countries plus the EU, competes with China’s newly launched World Artificial Intelligence Cooperation Organization. Kazakhstan remains the only nation currently participating in both blocs. Source

6. Survey Finds Data Center Opposition Nearly Doubled in a Year

Research. A Heatmap News and Embold Research survey found opposition to new data centers being built nearby jumped from 42 percent to 75 percent over the past year, with 61 percent of opponents describing their opposition as strong. A separate Gallup poll from May found similar results, at 71 percent opposition. Respondents cited rising electricity costs, water use, and land impact as top concerns, and data centers have become a bigger issue in some local campaigns than topics like international conflict. Source

7. DOJ Investigates Andreessen Horowitz Over Portfolio Board Seats

Andreessen Horowitz. The Department of Justice is investigating whether Andreessen Horowitz violated antitrust law by holding board seats at two competing data-pipeline companies, Databricks and Fivetran, invoking the rarely used Clayton Act. The nearly year-long probe raises questions about how venture firms manage board representation as their portfolio companies increasingly compete in overlapping markets. A ruling against the firm could reshape board governance norms across AI-focused venture capital. Source

8. Investigation Finds Jailbreak Bypasses Claude Opus 4.6 Safety Filters

Anthropic. A TechCrunch investigation found that Claude Opus 4.6 generated sexually explicit content in all 10 attempts using a jailbreak technique that gradually escalates fictional roleplay scenarios, despite Anthropic’s policy against such output. The researcher behind the technique showed the model could be talked out of its refusals by framing restraint as inconsistent treatment of characters. The findings raise compliance concerns given new laws like Colorado’s age-estimation requirement and reports that roughly 3 percent of US teens use Claude despite an 18-plus age policy. Source

9. Runlayer and Rippling Drop Dueling Lawsuits Over MCP Security Product

Runlayer. AI security startup Runlayer and HR platform Rippling dropped lawsuits against each other on August 20, with no settlement or payment exchanged. Runlayer, which builds MCP gateway security for enterprises managing AI agent access, had sued after Rippling tested its product for over a year without buying it, then released a competing gateway; Rippling countersued on patent grounds. Rippling has since shipped its own MCP gateway, illustrating how quickly larger platforms can move into a startup’s niche once it proves out a market. Source