Anthropic export controls on Claude & Custom AI chips heat up - Tech News (Jun 25, 2026)

A top-tier AI model was pulled offline after a White House order—and now its maker is trying to negotiate its way back into public use. Welcome to The Automated Daily, tech news edition. The podcast created by generative AI. I’m TrendTeller, and today is June 25th, 2026. Let’s get into what moved the tech world in the last 24 hours—and why it matters.

Anthropic export controls on Claude

We start with a story that sits right at the intersection of AI, national security, and who gets access to frontier models. Anthropic has reportedly been negotiating with the Trump White House for nearly two weeks to undo an export-control directive that effectively blocked broad access to its consumer-facing Claude Fable 5 model. According to reporting cited by Gizmodo, the company took Fable 5 offline on June 12 after being told it needed to prevent non‑U.S. nationals from using it—and officials were already worried about potential access by China-linked actors. What’s especially notable here is the political and operational signal: this isn’t just about one model. It’s about whether the U.S. is willing to treat advanced AI systems more like sensitive technology exports, with access gated by nationality and geography. And it’s also a reminder that “safety” arguments can quickly become “distribution” constraints—especially when jailbreakability enters the conversation. The report also claims the talks improved after Anthropic shifted the lead role from CEO Dario Amodei to co-founder Tom Brown, alongside its policy lead. If that’s accurate, it underlines a blunt reality: in 2026, AI capability is only half the battle—governance and negotiation are the other half.

Custom AI chips heat up

Staying with AI power, the silicon arms race keeps accelerating—and it’s no longer just chip companies making the running. OpenAI and Broadcom unveiled OpenAI’s first custom AI chip, called Jalapeño, aimed at inference—the work of serving models to users at scale. OpenAI says the design went from start to finish in nine months, and that its own AI tools helped speed development. The first physical sample is arriving now, with initial deployment targeted for late 2026. The big takeaway isn’t the name on the chip; it’s what it represents. When a model provider starts designing its own hardware, it’s a bet that demand will stay high and that controlling efficiency—cost, power, and supply—will be a competitive advantage. It’s also a pressure-release valve on the ongoing shortage and pricing power around top-end GPUs.

IBM’s sub-1nm chip prototype

Meanwhile, Qualcomm is making an unusually loud play for the data center. The company says Meta is the first named customer for its Dragonfly C1000 data-center CPU, slated for 2028. That’s far off, but the commitment matters because hyperscalers don’t put their name on a roadmap lightly. Qualcomm is also buying AI software firm Modular in a stock deal valued around $3.9 billion. Modular’s pitch is portability—helping developers run models across different chips without rewriting everything. If that vision holds, it challenges the idea that one vendor’s software ecosystem gets to be the default gravity well for AI. The caveat is simple: most proof points are still in the future. Between now and 2028, execution—and real-world performance—will decide whether this is a serious platform shift or just ambition with good branding.

Agents, prompts, and real software

On the far horizon of chip progress, IBM says it has a prototype approach that could push below the one-nanometre mark—claiming an effective process size of about 0.7 nanometres. IBM’s message is that traditional shrinking is getting brutally hard, so the next era may rely on stacking and more three-dimensional structures rather than just flattening transistors further. Even if this is years away from production, it speaks to a bigger constraint the whole industry feels: compute demand is rising faster than easy efficiency gains. And that’s why power, cooling, and data-center buildouts have become board-level problems.

AI coding workflows with pull requests

Let’s talk about how people are actually trying to use AI in software—because the best takes this week had a common theme: stop confusing language with logic. One essay argues the real breakthrough of systems like ChatGPT isn’t “conversational” software—it’s software that can accept natural-language instructions and turn them into predictable actions. The phrase to remember is “language-native software”: language becomes the main interface, while chat is what you use to resolve ambiguity. Crucially, the author draws a hard boundary between probabilistic understanding—figuring out what you meant—and deterministic execution—doing the thing in a way you can audit and trust. That distinction is becoming a design principle for modern apps, especially in regulated environments where ‘the model said so’ is not an acceptable explanation.

Energy buildout for AI demand

A separate critique is even more direct: many companies are effectively “programming in Markdown”—stuffing simple business rules into long prompts and then acting surprised when the result is slower, pricier, and easier to exploit. The point isn’t that LLMs are useless; it’s that they’re the wrong tool for crisp policy logic. If your process is basically “if these conditions are true, do this,” traditional code is still the safest and cheapest way to run it. Use AI where the work is inherently fuzzy—like interpreting messy language, summarizing, or classifying—not where you need strict guarantees.

Europe weighs under-16 social media limits

And if you do want agents in the loop, one practical pattern keeps winning: make them produce reviewable artifacts. Instead of a chatbot that spits out paragraphs you have to copy and verify, the argument is that agents should generate things like GitHub pull requests. That moves the output into a workflow built for scrutiny—diffs, automated tests, approvals, and the ability to reject by default. Paired with guarded permissions—bot branches, scoped tokens, and “never merge automatically”—it’s a simple way to keep agents useful without giving them the keys to the kingdom.

Global rules for autonomous driving

Now, the energy side of the AI boom: the U.S. government is putting serious weight behind new nuclear—and companies are pitching ways to squeeze more out of the grid faster. The U.S. Energy Department says it will provide $17.5 billion in loans to accelerate projects that could build ten large nuclear reactor units, using Westinghouse’s AP1000 design. Sites aren’t final, but the intent is clear: bring construction forward, reduce financing friction, and meet the growing appetite for reliable, low-emissions electricity. And in a very different approach, Sunrun, Tesla, and Renew Home announced a plan to aggregate home batteries and smart devices into a virtual power plant they claim could scale dramatically over time. The near-term focus is Virginia’s data-center corridor, where they say meaningful capacity is available quickly—if customers enroll and regulators cooperate. Put these together and you get a pragmatic picture: some solutions are decade-scale infrastructure, others are faster “capacity now” tactics. Data centers are forcing both.

Health bets: universal vaccines and antivirals

Over in Europe, leaders say the European Commission is preparing proposals that could restrict social media access for children under 16. The important angle here is the shift from national experiments to a possible EU-wide approach. If Europe moves as a bloc, platforms may face a more uniform set of rules on age verification and youth protections—harder to route around, but also harder to implement without raising privacy and enforcement questions. The details will matter, but the direction is becoming unmistakable.

Amazon seller bribery and data leaks

On transportation policy, the UN’s vehicle standards body—UNECE—approved what it calls the first global regulations for fully autonomous driving systems. This is less about letting robotaxis roam tomorrow and more about reducing regulatory chaos. A shared framework for testing, lifecycle safety management, and post-deployment monitoring could make it easier for companies to ship across markets—while also making it easier for regulators to demand evidence and accountability when things go wrong.

China’s EV export surge

Two health-related stories also stood out—both driven by the idea of getting ahead of respiratory viruses rather than reacting late. A new $500 million fund called Intercept launched with the goal of reducing common respiratory infections like colds and flu, backed by a mix of donors including some tech names. The ambition is to push a couple of candidates through early clinical trials and then hand off to pharma for the expensive final stages. Separately, researchers at Cambridge say AI-assisted vaccine design could help create broader “universal” vaccines that protect against entire virus families. They reported early human trial results for a universal Sarbeco coronavirus candidate with no significant safety concerns in a small group, with larger studies next. If this approach scales, it could shorten the time between “new outbreak” and “meaningful protection”—which is exactly where the world has been slow in the past.

Hubble spots reionization-era escape

A quick, darker note on marketplace security: an Amazon seller says a middleman offered to bribe an Amazon employee to unfreeze funds after an account suspension, and appeared to have access to internal-looking account details. Amazon says the implicated employee had already been fired for unrelated misconduct and that cases like this are rare. Still, the broader pattern is worth watching: when support is hard to reach and enforcement is opaque, underground “fixer” markets pop up—eroding trust for legitimate sellers and creating incentives for insider abuse.

Internationally, China’s electric vehicle exports reportedly rose sharply in May to a new record, with the jump being linked to higher oil prices and supply disruptions tied to the Iran conflict. It’s another reminder that geopolitics can accelerate electrification in unpredictable ways. When oil feels fragile, EVs—and broader electricity tech—look like a stability play, not just a climate play. At the same time, China’s solar export slump after policy changes shows how quickly trade flows can pivot when incentives change.

Finally, a story from the deep universe that still connects back to modern science: astronomers using Hubble detected escaping ionising ultraviolet light from an early galaxy called MXDFz4.4, seen just 1.4 billion years after the Big Bang. Why it matters is simple: it’s direct evidence for a mechanism that may have cleared the Universe’s early ‘hydrogen fog’ during the era of reionisation. In plain terms, tightly packed bursts of star formation can punch holes through surrounding gas, letting high-energy light escape and change the state of the cosmos. This result is backed up with data from Webb and other instruments, strengthening the case that small, intense galaxies played an outsized role in making the Universe transparent.

That’s our run for June 25th, 2026. If one theme ties today together, it’s control—control of who can access frontier AI, control of the hardware that runs it, and control of the systems that decide what happens when automation meets the real world. Thanks for listening to The Automated Daily, tech news edition. I’m TrendTeller. If you want, send this episode to someone who still thinks the AI story is only about chatbots. Talk tomorrow.

Anthropic export controls on Claude & Custom AI chips heat up - Tech News (Jun 25, 2026)

Our Sponsors

Today's Tech News Topics