U.S. bans Anthropic contracts & China’s OpenClaw agent craze - Tech News (Mar 16, 2026)

We’ll start with that high-stakes showdown in U.S. AI policy. The Trump administration has ordered the termination of all federal contracts with Anthropic, after CEO Dario Amodei declined to remove restrictions that limit how Claude can be used—particularly around mass surveillance and fully autonomous weapons without human oversight. Anthropic is suing, arguing this is unlawful retaliation and that the government’s “supply chain risk” framing could ripple outward—effectively pressuring other defense contractors and partners to drop the company too. The Pentagon’s counter-argument is blunt: contractors shouldn’t get to dictate battlefield rules. What makes this especially consequential is how embedded Claude reportedly is in sensitive workflows. If that’s accurate, replacing it won’t be a simple swap, and the dispute becomes a live test of how quickly government can change course once AI tools are already in the bloodstream.

Staying with AI—but shifting from Washington to everyday life—China is in the middle of a full-on open-source agent frenzy. The tool at the center is called OpenClaw, and people are reportedly lining up at major tech offices or even paying strangers online to get it installed. What’s fascinating here isn’t just adoption speed, it’s the way users are improvising new roles for an AI agent: everything from monitoring markets to acting as a wingman for blind dates—translating messages, suggesting replies, and steering conversations. Some are even letting it place trades, which has already produced the predictable result: a few people say the agent made incorrect orders or bad calculations, and they lost money. And now comes the hangover phase—security officials are warning about misconfigured setups leading to cyberattacks or data leaks, and some agencies and state-linked firms are restricting the tool on work devices. It’s a tidy snapshot of the AI agent era: easy to spread, quick to become “normal,” and equally quick to magnify financial and cybersecurity risk.

On the infrastructure side of AI, Nvidia is using its upcoming GTC conference to underline a message that might surprise people who think the whole story is GPUs: the CPU is becoming a bottleneck again. As AI shifts from chatbots to agentic systems that juggle many tasks and shuttle lots of data, Nvidia is pitching its Arm-based Grace CPU—and its next-generation Vera CPU—as the “host” processors that keep expensive accelerators busy instead of waiting around. One notable signal is Nvidia’s multiyear deal with Meta, which includes large deployments of standalone Grace CPUs and a plan to adopt Vera later. The bigger takeaway: demand is getting lumpy across the stack. Analysts are even warning about a quieter supply crunch in server CPUs, not just GPUs. In plain terms, the AI boom is no longer only a chip story—it’s a whole data-center choreography problem.

Now to Meta, which is showing up in the news for both money and responsibility. First, Reuters reports Meta is planning sweeping layoffs that could reach a significant share of the workforce, as it tries to offset the soaring cost of AI infrastructure. Meta is disputing the report, but the context is clear: the company wants to spend aggressively on AI and still hit profitability expectations, and headcount becomes the easiest lever to pull. Second, a separate thread of scrutiny keeps tightening around platform safety. Whistleblowers and former insiders told the BBC that TikTok and Meta made product and policy calls that increased exposure to harmful content during a high-pressure “algorithm arms race.” The allegation is that engagement and growth goals nudged safety concerns into the background. And in Los Angeles, a major trial against Meta and Google is putting platform design itself on the stand—features like infinite scroll, autoplay, and attention-grabbing notifications are being framed as intentionally habit-forming for children. The companies deny the core claims. Why this matters: if juries or regulators start treating design choices as foreseeable harm, it changes the legal risk profile of the entire social media business model.

Elon Musk’s corner of the tech universe had two very different signals—one about supply, the other about execution. On supply, Musk posted that his chipmaking venture—branded the “Terafab Project”—will launch in seven days, pointing to March 21st, 2026. Details are thin, and chip manufacturing is famously slow to ramp, but the intent is obvious: if AI compute is the new oil, Musk wants a refinery. On execution, reports say xAI is going through another round of job cuts and reorganizations after Musk became dissatisfied with performance and adoption of its products. Sources describe talent churn, leadership shakeups, and managers pulled in from other Musk companies to audit teams and data quality. Whether or not every detail holds, the theme is familiar: in AI right now, massive compute budgets don’t automatically produce a product people pay for—or a stable organization that can ship predictably.

In China, regulators have approved the country’s first invasive brain-computer interface product for sale—an implant system aimed at certain adults with paralysis from spinal cord injuries. The reported clinical results focus on improved hand functions when paired with assistive hardware. The clearance is limited in scope, but it’s still a major milestone: it signals Beijing is willing to move BCIs from labs into regulated medical reality. It also sharpens the global competitive landscape, with Chinese startups now moving in the same broad arena as Neuralink and other U.S. efforts. Expect more investment hype around this category—alongside louder debates about safety, long-term support, and who owns the data coming out of a human brain.

Quickly, a space update with real-world implications for planetary defense. A new study says NASA’s 2022 DART mission didn’t only change the orbit of the small asteroid Dimorphos around Didymos—it also measurably changed the pair’s trajectory around the sun. That’s the kind of subtle effect you can’t confirm quickly; researchers needed years of follow-up to detect it. The interesting twist is that debris from the impact added extra momentum, meaning the shove was more effective than the spacecraft collision alone would suggest. For future asteroid-deflection planning, that’s valuable: it improves our ability to predict outcomes instead of guessing based on models.

One more science-meets-AI story—this time from veterinary medicine. An Australian entrepreneur worked with university researchers to develop an experimental personalized mRNA cancer vaccine for his rescue dog, Rosie. The team sequenced the tumor and used AI-driven analysis to help design a vaccine targeting that specific cancer. After treatment, one tumor reportedly shrank significantly, improving comfort, though the cancer hasn’t disappeared. Researchers called the response surprising, while other scientists caution it’s easy to over-read a single case. Still, it’s a compelling preview of where medicine is heading: more personalization, faster design cycles, and AI as a practical accelerator—especially in early experimental settings.

And finally, a reminder that AI’s influence isn’t only productivity—it’s also confusion at scale. Researchers say AI-generated images and videos about a widening Middle East conflict are still spreading widely on X, including fabricated scenes that look like breaking news. X has announced penalties for creators who post synthetic war footage without disclosure, including temporary suspensions from monetization. But enforcement appears uneven, and some high-visibility accounts reportedly continue to rack up massive views. The unsettling detail: automated chat tools can sometimes validate fakes instead of debunking them, which makes the information environment even noisier. In wartime especially, the question isn’t just whether content is false—it’s whether platforms can reduce the incentive to manufacture it faster than fact-checking can keep up.