Claude dethrones ChatGPT in US & Pentagon deals split AI vendors - Tech News (Mar 2, 2026)

Let’s start with the consumer-facing ripple effect. Anthropic’s Claude has climbed to the top spot for free apps in Apple’s U.S. App Store, pushing ChatGPT to number two. Reporting ties the surge to backlash after Sam Altman publicly discussed OpenAI working with the U.S. Department of Defense on deployments inside classified networks. Anthropic’s CEO, Dario Amodei, has been vocal about drawing hard lines—specifically against mass domestic surveillance and fully autonomous weapons. Whether you agree with Anthropic or not, the striking part is that everyday users appear to be voting with downloads. Anthropic says free users are up sharply since January, with daily signups setting records, and paid subscribers more than doubling this year.

Underneath that popularity swing is a much bigger policy and procurement story. Talks between the Pentagon and Anthropic reportedly came down to last-minute contract language, especially around what “lawful surveillance” could mean in practice. Negotiations then collapsed, and Defense Secretary Pete Hegseth publicly labeled Anthropic a security risk—an extraordinary move for a major U.S. tech company. Within hours, OpenAI said it reached a deal to supply AI to classified military networks, and Altman emphasized that OpenAI’s contract still prohibits mass surveillance and autonomous lethal weapons—calling them core safety principles that the Pentagon accepted. One detail worth watching: reports also describe internal industry blowback, with employees across AI companies urging leaders not to be played against each other by shifting government demands. If this becomes the new normal—public pressure campaigns plus contract brinkmanship—it could reshape how AI firms write policies, and how they prove compliance.

Now to the money fueling all of it. OpenAI is also raising a new funding round targeting $110 billion, valuing the company at roughly $730 billion pre-money and about $840 billion fully diluted. The headline investors include Amazon, Nvidia, and SoftBank. Amazon alone is slated to put in up to $50 billion, and OpenAI says it will use two gigawatts of compute capacity powered by Amazon’s Trainium chips. There’s also an important structural point: AWS becomes the exclusive third-party cloud provider for OpenAI Frontier—its enterprise platform for building and managing AI agents—while Microsoft remains the exclusive cloud provider for OpenAI APIs and continues hosting first-party products on Azure. In other words, OpenAI is slicing its cloud relationships by product line, not picking one winner for everything.

This all feeds into what developers are actually doing day to day—because the development workflow is changing fast. Cursor’s Michael Truell argues we’re entering a “third era” of AI-assisted software building. First came autocomplete that excelled at repetitive code. Then came synchronous agents where you steer the model step by step. The third era, he says, looks more like building a software factory: fleets of autonomous agents running in the cloud, iterating for hours, running tests, and returning artifacts you can review—logs, recordings, previews—not just a diff. Cursor claims around 35% of its internally merged pull requests are now created by agents working autonomously on separate cloud machines. If that number holds up as the tooling spreads, it’s a genuine shift: engineers spending less time typing code, and more time framing tasks, setting constraints, and reviewing outcomes.

And if you’re building systems for agents rather than just humans, the plumbing matters—especially APIs. Nate Meyvis shared an “AI-first” set of notes that boils down to something refreshingly practical: if your product needs an API, build the API, because AI tools are unusually good at accelerating that work. His recommendations include exposing documentation programmatically—think an endpoint like /api/help—so AI clients can discover capabilities without you stuffing long docs into a context window. He also argues for safer, non-destructive designs for AI-driven actions. For example, let write operations create “candidates” that require review before anything becomes official. And he flags a subtle risk: AI-generated implementations are often too eager to add fallbacks. Those can hide bugs or accidentally open security holes, so the advice is to review carefully—and even use a second AI pass specifically to hunt for dangerous fallback behavior.

On the platform side, Cloudflare is jumping into this agentic moment with “Cloudflare Agents,” an SDK and toolkit for building agentic apps on Cloudflare’s stack. The pitch is a full workflow: collect input via chat, email, or voice; reason with models either on Workers AI or through external providers via AI Gateway; manage state with Durable Objects and orchestration via Workflows; and then take actions through tools like browser rendering, vector search, or databases. Cloudflare’s cost angle is notable: Workers charges for CPU time rather than wall-clock time, which matters when agents spend a lot of time waiting on APIs, LLM calls, or humans. It’s an attempt to make long-running, tool-using agents feel less like a runaway meter.

Regulation is also tightening, and today’s date matters here. Vietnam’s new AI law took effect yesterday, March 1st, making it the first Southeast Asian country with a comprehensive AI framework. The law focuses heavily on generative AI risk, requires human oversight, and mandates labeling for AI-generated content—like deepfakes—when it’s not clearly distinguishable from real media. It also requires services to tell users when they’re interacting with an AI system rather than a human. Vietnam is also pairing governance with industrial policy: plans include a national AI computing center and more investment in Vietnamese-language models. The open question will be enforcement and the detailed decrees that turn principles into day-to-day compliance checklists.

Australia is taking a more enforcement-forward posture. Its eSafety regulator warned it could go after “gatekeepers” such as app stores and search engines to block AI products that don’t implement age assurance. A key deadline is March 9, when services must prevent under-18 users from accessing pornography, extreme violence, self-harm, and eating-disorder content, or face steep fines. A Reuters review suggested many popular AI services still haven’t publicly demonstrated compliance steps. This is one of the clearest signals yet that regulators may not only target AI makers—they may also pressure the distribution layer that makes these tools easy to find and install.

All of this—agents, regulation, enterprise adoption—funnels into the biggest bottleneck: infrastructure. A new wave of reporting frames AI as a capital-expenditure arms race spanning data centers, chips, power generation, and grid upgrades. Hyperscalers are forecasting eye-watering 2026 spend, and the tension is simple: they’re betting the demand curve will keep rising long enough to justify the buildout. Two infrastructure stories stood out. First, Google says it will build a new data center in Pine Island, Minnesota, backed by 1.9 gigawatts of wind and solar plus a long-duration battery system from Form Energy rated at 30 gigawatt-hours—designed to discharge for up to 100 hours. That’s aimed at covering multi-day renewable lulls, even if iron-air batteries trade efficiency for lower cost and longer duration.

Second, Nvidia says it will invest $4 billion across two U.S. photonics firms—Lumentum and Coherent—to strengthen optical supply chains for data center networking. The key phrase from Jensen Huang was “gigawatt-scale AI factories.” That’s not just marketing; it’s a hint that the next constraint, after GPUs, is increasingly the plumbing between them: lasers, optics, and high-speed links that keep giant clusters from choking on their own data movement.

Before we wrap, a quick tour of two other frontiers: defense tech and big bets. Israel’s Defense Ministry confirmed the first operational combat use of its Iron Beam laser air-defense system, designed to complement missile-based defenses with a potentially far cheaper per-shot intercept—though weather and atmospheric conditions still matter a lot for laser performance. And the U.S. military announced its first combat use of one-way attack drones in strikes on Iran, underscoring how quickly low-cost loitering munitions have become mainstream.

On robotics, a reality check: analysts argue general-purpose humanoid home robots are still not close in 2026. The biggest obstacles aren’t just dexterity or models—it’s deployment realities and training data. Cars get billions of human-driven miles; humanoids don’t have that kind of installed base generating behavior data, and collecting it at home raises privacy and safety issues. The near-term winners remain narrower, task-specific robots that can scale fleets and learn incrementally.

And finally, SpaceX may be approaching the public markets. Bloomberg reports the company is weighing a confidential IPO filing as soon as March, potentially aiming for a June listing. The numbers being floated are enormous—possibly the biggest IPO ever by proceeds, with valuations discussed north of a trillion dollars. It’s not final, and timelines can slip, but it’s another sign that capital markets may soon get a direct line into the space-and-infrastructure story that’s been largely private so far.

One more network note to close the loop: at Mobile World Congress, Nvidia and a coalition of telecom operators and infrastructure players announced a push for AI-native, open, secure 6G platforms. The framing is that 6G won’t just connect phones—it could underpin “physical AI,” meaning fleets of machines, sensors, vehicles, and robots. If that vision holds, security and interoperability have to be designed in from the start, not bolted on after the first major incident.