NASA nuclear ship to Mars & Moon base plan and costs - Tech News (Mar 25, 2026)

Let’s start in space, where NASA approved development of the Skyfall mission. The headline: three small helicopters are headed to Mars, building on what Ingenuity proved—but this time, the goal isn’t just “can it fly?” It’s practical scouting for future exploration, like spotting landing hazards and hunting for subsurface water-ice using radar. The even bigger subplot is the ride: SR-1 Freedom, described as the first nuclear-powered interplanetary spacecraft, using a fission reactor to drive electric propulsion. If this holds up technically and politically, it’s a major step toward faster, more capable missions that aren’t so tightly constrained by fuel.

NASA also unveiled a broader strategy for a sustained moon base, with an estimated cost around twenty billion dollars. The messaging is clear: fewer flag-planting visits, more living-and-working infrastructure. The reason this matters is less about a single outpost and more about precedent—how the U.S. plans to fund, supply, and govern long-duration off-world operations with industry and international partners. The tough questions, as always, are timelines, budgets, and what gets cut when priorities shift.

Back on Earth, a noticeable theme today is AI moving from “chatting” to “doing.” Accio introduced Accio Work, a local-first desktop agent platform aimed at running end-to-end business tasks through one conversational interface. The pitch is broad—sourcing, supplier coordination, store operations, marketing, CRM-like workflows—but the interesting part is the architecture choices: actions require explicit permission, execution stays on the user’s machine, and activity is logged for auditing. That’s a sign of where this market is going: less novelty in clever replies, more pressure to safely touch real files, real browsers, and real business systems.

Sticking with agents, Anthropic added an “auto mode” for Claude Code permissions. It’s meant to reduce the constant approval prompts without going fully open-loop. A separate model judges whether a requested action looks risky or out of scope, and blocks it when needed. Critics—including Simon Willison—are still wary, because AI-based gatekeeping can be inconsistent, especially when context is messy. The broader takeaway: as coding agents get more capable, the industry is still debating the best safety strategy—probabilistic “AI says it’s fine” checks versus deterministic sandboxing that simply prevents dangerous access by default.

Apple, meanwhile, is reportedly preparing a major AI reboot centered on Siri, targeted for iOS 27 and macOS 27 and expected to show up at WWDC on June 8th. The testing reportedly includes a standalone Siri app that can remember and search past conversations, accept documents or photos for analysis, and offer a more systemwide “Ask Siri” flow that can pull in content you’re already looking at inside Apple apps. Why it’s notable: Apple has been playing catch-up in consumer AI, and the biggest lever it has is deep integration—turning the assistant into something closer to an operating-system layer, not just a voice command box.

The money-and-narratives corner of AI had two big signals today. First, OpenAI Foundation—OpenAI’s nonprofit controller—says it will distribute a billion dollars in grants over the next year, focusing on life sciences, health, and programs aimed at reducing AI downsides like job disruption and mental health impacts. It’s notable because OpenAI’s nonprofit origins have been under renewed scrutiny as the company’s commercial value exploded. This grant pledge looks like an attempt to reassert the public-benefit mission at a moment when AI’s real-world externalities are hard to ignore.

Second, Nvidia CEO Jensen Huang said he believes AGI has already been achieved—based on a very business-centric definition: the ability to create and run a billion-dollar enterprise. That framing is controversial, and it’s a reminder that “AGI” can mean whatever the speaker needs it to mean. Investors and policymakers should hear this as a signal about incentives: define AGI as “commercially transformative,” and the urgency to buy more compute and build more data centers conveniently rises.

Speaking of compute, Arm launched its first in-house processor for data centers, a major departure from its long-running role as the neutral company licensing CPU designs to everyone else. The chip is aimed at AI infrastructure, especially workloads where general-purpose CPU work can bottleneck the rest of the system. The strategic tension is obvious: when Arm sells its own silicon, it potentially competes with customers who used to be “just” partners. Whether this becomes a one-off or a long-term shift will shape how the AI server ecosystem evolves.

On the longer-horizon chip front, Norway-headquartered startup Lace raised forty million dollars to pursue a new lithography approach using a helium atom beam, pitching a path toward dramatically smaller features than light-based methods. Even if the timelines stretch—and they often do in chip manufacturing—this story matters because the industry is hunting for options beyond the current dominant tooling. AI demand is pushing semiconductor roadmaps hard, and that pressure tends to fund more moonshot manufacturing bets.

Now for the infrastructure idea that sounds like a pitch deck—until you realize serious players are circling it: orbital data centers. Ars Technica dug into whether large-scale compute in space can ever pencil out, with the argument that you’d get steadier solar power and sidestep some of the political and grid constraints that make terrestrial data centers unpopular. The catch is brutal scaling: launch costs, satellite manufacturing costs, and the environmental and astronomy impacts of massive constellations. Even optimistic models still imply staggering deployment costs, plus a much brighter, more crowded sky.

In U.S. tech policy, the FCC issued a ruling that effectively blocks import and sale of most new wireless routers if they’re considered “foreign-made,” citing national security and cybersecurity risks. The definition is broad enough that it could squeeze consumer choices and disrupt how ISPs provision equipment—because, realistically, a huge share of routers are built or designed overseas. Existing devices can still be used, but “new supply” is the pressure point, and exemptions may be hard to obtain. This is another step in the wider trend: supply chains are now treated as part of national security, not just economics.

On platform accountability, a New Mexico jury found Meta’s platforms harmful to children’s mental health and safety, issuing a three-hundred-seventy-five-million-dollar penalty after a multi-week trial. Meta plans to appeal, but the significance is the precedent: this is the first verdict in a wave of child-safety cases targeting social media design and disclosures. If more juries agree that product choices—like engagement-optimizing feeds—cross legal lines, it could force changes that regulation has struggled to deliver, and it could reshape how platforms think about liability risk.

Meta also showed up indirectly in a cultural post-mortem: Neal Stephenson responded to reports that Meta has shut down its Metaverse effort after spending tens of billions. Stephenson’s main point is that shared 3D worlds aren’t “dead”—they’re already thriving on flat screens in massive games. The failure, in his view, was the industry betting too heavily on headsets that many people find awkward, and on platforms that can vanish and take creators’ work with them. The practical takeaway: the next “metaverse-like” step is more likely to build on existing ecosystems and tools than to arrive as a brand-new walled garden.

In robotics, Amazon acquired Fauna Robotics, a startup building humanoid robots designed to be approachable and developer-friendly. Amazon has spent years scaling robotics in warehouses, and it’s been experimenting with delivery automation too—so this extends a pattern: Amazon wants a bigger footprint in embodied automation, not just software assistants. The open question is whether humanoids become a general-purpose platform, or remain expensive specialists deployed only where the economics are undeniable.

Two quick notes for builders shipping software in the age of AI. PostHog shared lessons from relaunching its in-product AI assistant, arguing many teams should start simpler—exposing capabilities through standardized tool interfaces—before jumping into complicated agents. And an open-source project called Ossature is pushing a “harness” approach: stronger specs, clearer plans, verification loops, and logs—basically, guardrails to keep LLM-generated code consistent across a real project. The common message is refreshing: reliability and clarity beat flashy demos once users depend on the output.

Zooming out even further, astronomers reconstructed the formation history of spiral galaxy NGC 1365 using what they call chemical archaeology—mapping element signatures across star-forming regions and comparing them to huge simulation libraries. It’s a clever way to read a galaxy’s past from its present chemistry, and it opens the door to testing whether the Milky Way’s story is typical or unusual across the universe.

And finally, a biotech manufacturing story with potentially big implications: researchers reported a nanoparticle-based platform to streamline production of engineered exosomes—tiny cell-released particles being explored for therapies. The headline isn’t one disease result; it’s scale and consistency. If exosome-based treatments can be manufactured more reliably, it could speed up how quickly these therapies move from promising lab results to real clinical pipelines.