← The State of Tech RSS Feed

2 mei 2026 · 15:11

Pentagon AI Military Deals: Anthropic Frozen Out

The Pentagon has formalised classified AI deals with seven major tech firms, including OpenAI, Google, Microsoft and Nvidia, while officially labelling Anthropic a supply-chain risk over its restrictions on surveillance and lethal autonomous weapons work. It is the clearest signal yet that the line between commercial and defence AI has disappeared in the United States. For European professionals, the episode explores what this means for NATO interoperability, the EU AI Act's military carve-out, and whether Anthropic's ethical stance becomes a liability or a competitive advantage in European enterprise markets.

Pentagon AI Policy EU AI Act Humanoid Robotics Cybersecurity South Korea

Beluister deze aflevering:

Transcript

Samantha: Welcome to The State of Tech, The European Edition, Saturday May two, 2026. I'm Samantha Lawrence.

Bob: And I'm Bob Russell. Today: the Pentagon signs classified AI deals with seven tech giants while snubbing Anthropic, the EU AI Omnibus enters trilogue negotiations, China's massive push into embodied AI and humanoid robotics, US officials weigh slashing patch deadlines as AI-powered hacking accelerates, South Korea's exports surge on the AI chip boom, and a fast-growing AI model procurement deal worth nearly five million dollars. Let's start with the Pentagon story.

Pentagon signs classified AI deals with seven tech giants, freezes Anthropic out over weapons dispute.

Bob: The Pentagon has just formalised agreements with seven major AI firms to plug their models directly into classified military networks. We're talking SpaceX, OpenAI, Google, Nvidia, Reflection, Microsoft, and Amazon Web Services. The goal is blunt: turn the US military into what they're calling an AI-first fighting force, with AI helping on mission planning and weapons targeting.

Samantha: And the headline-grabbing detail is who is missing. Anthropic was excluded, and the Defense Department went further, labelling them a supply-chain risk. That's harsh language. The dispute centres on Anthropic's restrictions around using its models for surveillance and lethal autonomous weapons. So a company that built its brand on AI safety is now being treated as the problem, not the solution.

Bob: It's a fascinating reversal. For years the question was whether AI labs would say no to defence contracts. Now the Pentagon is essentially saying: if you say no, you're out, and we'll call you a risk. That changes the calculus for every frontier lab. OpenAI, Google, Microsoft, all in. Anthropic, isolated.

Samantha: There's also a strategic signal here. By naming seven providers, the Pentagon is locking in redundancy. No single vendor controls the AI stack. And by including Reflection alongside the giants, they're signalling that newer labs can break in if they play ball.

Bob: For European listeners, this matters on several levels. NATO interoperability is the obvious one. If American forces are running classified workloads on OpenAI and Google models, allied forces will eventually have to plug into the same architecture, or build a parallel one. France and Germany have been pushing for European sovereign AI for defence, and stories like this make that case louder.

Samantha: And then there's the ethical dimension. The EU AI Act carves out military uses, but European publics are sensitive about lethal autonomous weapons. Watching American firms accept targeting work while Anthropic refuses creates a real political conversation here. Expect parliamentary questions in Berlin, Paris, and The Hague within days.

Bob: The bottom line is that the line between commercial AI and defence AI is now officially gone in the US. Europe will have to decide whether to follow, fork, or regulate.

Samantha: And whether Anthropic's stance becomes a competitive disadvantage or a unique selling point in markets like Europe, where ethical positioning still carries weight with enterprise buyers.

EU AI Omnibus enters trilogue as Brussels reshapes its digital rulebook.

Samantha: Speaking of European regulation, the AI Omnibus has just moved into trilogue. That's the closed-door negotiation between the European Parliament, the Council, and the Commission to lock in the final text. This is the stage where the law gets its real shape.

Bob: And it's not just the AI file. The Commission is also pushing a revised Cybersecurity Act 2 to fix gaps in the existing framework, plus amendments to NIS2 to lighten the compliance load on companies that have been drowning in paperwork. So three big files, all moving at once.

Samantha: The signal Brussels is sending is pragmatism. The original AI Act was criticised for being heavy-handed, especially on smaller firms. The Omnibus is meant to streamline, consolidate, and reduce the duplication that businesses complained about. Same ambition, less bureaucracy. At least, that's the promise.

Bob: Whether it delivers is another question. Trilogues are where ambitious ideas often get watered down or, occasionally, sharpened. Civil society groups want stronger rights for users facing AI decisions. Industry wants clarity and predictability. The Council, representing member states, usually pushes for flexibility. Three-way tension.

Samantha: For European companies, this is the moment to engage. Once trilogue closes, the text is essentially fixed. We're talking obligations on high-risk AI systems, transparency for general-purpose models, and now a tighter link to cybersecurity rules. If you build, deploy, or even just buy AI in Europe, this affects your roadmap.

Bob: And for users, there's a real upside. Stronger rights to know when AI is being used on you, clearer paths to challenge automated decisions, and a baseline of safety testing for the models behind your apps. The EU is still the only bloc writing comprehensive AI law, and the rest of the world is watching.

Samantha: Quick interruption. If you listen to The State of Tech regularly, hit that like button and subscribe, that way you'll never miss an episode. Okay, moving on.

China bets its industrial future on embodied AI and humanoid robots.

Bob: China has put embodied AI at the heart of its national AI Plus strategy. The plan is to weave humanoid robots and physical AI systems through manufacturing, logistics, and eventually services, using the country's existing strengths in EV production and electronics assembly as a launchpad.

Samantha: The dependency picture is interesting. Chinese robotics firms still rely heavily on Nvidia chips and Nvidia's software stack for training and inference. But Beijing is pouring money into localising that hardware, and several domestic chip designers are closing the gap on training silicon, even if they're still behind on the high end.

Bob: The playbook looks a lot like the EV one. Subsidise heavily, drive volume, push prices down, and flood global markets with affordable hardware. Western humanoid robots from companies like Figure or Tesla cost tens of thousands of dollars per unit. Chinese rivals are already targeting prices that would make those look premium.

Samantha: Precision and dexterity are still real challenges. Picking up a coffee cup reliably is harder than it sounds, and current humanoids are nowhere near human-level on fine motor tasks. But the Chinese bet is that scale and iteration will close that gap faster than the West expects.

Bob: For Europe, this is a competitive alarm bell. We have strong industrial robotics players, KUKA, ABB, Comau, but humanoid and embodied AI is a different category. If China replicates its EV success here, European factories could end up buying their next generation of robots from Shenzhen, not Stuttgart.

Samantha: Policymakers in Brussels are already thinking about this. Expect more talk about a European robotics strategy, possibly with the same logic as the Chips Act: subsidies, public procurement preferences, and tighter rules on critical technology imports. Whether that arrives in time is the open question.

Bob: And there's a labour dimension. Embodied AI doesn't just compete with European robotics firms, it competes with European workers in warehousing, logistics, and assembly. That conversation is coming, ready or not.

US considers cutting patch deadlines from two weeks to three days as AI hackers accelerate.

Samantha: US cybersecurity officials are weighing a dramatic change. The current rule gives federal agencies up to two weeks to patch critical vulnerabilities. The proposal on the table would cut that to just three days. That's a massive operational shift.

Bob: The driver is AI-powered hacking. Officials specifically named Anthropic's Mythos and OpenAI's GPT-5.4-Cyber as tools that can scan code, find flaws, and build working exploits in hours, not weeks. The defender's window is collapsing, so the rules have to collapse with it.

Samantha: Three days is brutal in practice. Federal agencies often run legacy systems where a patch needs testing before deployment, otherwise you risk breaking the very service you're protecting. Doing that in seventy-two hours requires automation, staffing, and budget that many agencies simply don't have today.

Bob: Which is why this proposal is as much about forcing modernisation as about patching faster. If you can't meet a three-day window, you have to invest in better tooling, automated patch pipelines, and probably AI-driven defence to match the AI-driven offence. It's an arms race, and the timer just got shorter.

Samantha: For Europe, the parallels are obvious. ENISA, national cyber agencies, and operators of essential services under NIS2 are facing the same threat landscape. If American agencies move to three days, expect European regulators to ask uncomfortable questions about why their own deadlines are still measured in weeks.

Bob: There's also a private sector ripple. Banks, hospitals, energy grids, all running on software with vulnerabilities. AI-assisted attackers don't care whether the target is a US federal agency or a Dutch water utility. The defensive bar has to rise everywhere.

Samantha: And intelligence sharing becomes critical. If Mythos or GPT-5.4-Cyber are being misused, the indicators of compromise need to flow between US and European agencies in near real time. The old quarterly briefings won't cut it anymore.

Bob: Key takeaway: the cost of slow patching just went up, and the cost of fast patching is going up too. Cybersecurity budgets in 2026 are going to look very different from 2025.

South Korean exports jump forty-eight percent as AI chip demand powers the economy.

Bob: South Korea posted a stunning April export number. Unadjusted exports up forty-eight percent year-on-year, driven overwhelmingly by semiconductor sales. The AI buildout is, quite literally, anchoring an entire national economy right now.

Samantha: Samsung and SK Hynix are the obvious beneficiaries, particularly on high-bandwidth memory, the HBM chips that sit next to Nvidia GPUs in every AI data centre. Demand has been so strong that allocations are being negotiated months in advance, and prices keep climbing.

Bob: The macro picture is striking. Even with Middle East tensions pushing energy prices up, which usually hurts an import-dependent economy like Korea's, the chip surplus is so large it more than offsets the bill. That's how dominant AI hardware has become in global trade flows.

Samantha: For Europe, this is both opportunity and warning. Opportunity, because European industrial firms get access to the chips they need, even if they pay premium prices. Warning, because the dependency on a handful of Asian suppliers, Korea, Taiwan, and increasingly Japan, is now structural.

Bob: The EU Chips Act was supposed to address that, with a goal of twenty percent of global production by 2030. Realistically, Europe is still well behind. Intel's Magdeburg fab has had delays, TSMC's Dresden project is on track but won't ship at scale until later this decade. Meanwhile, Korea ships record volumes today.

Samantha: There's also a vulnerability angle. If something disrupts Korean exports, a labour dispute, a geopolitical shock, anything, the European AI ecosystem feels it within weeks. No European cloud provider can spin up GPUs without HBM, and HBM essentially comes from two Korean firms.

Bob: The bottom line is that AI's economic gravity is concentrated in a very small number of places, and Europe is not yet one of them. Closing that gap is a decade-long project, not a quick fix.

Samantha: And in the meantime, European companies need supplier diversification strategies, longer-term contracts, and frankly a bit of luck.

XMax lands nearly five million dollar AI model deal, eyes thirty million in revenue.

Samantha: Time for our lighter closer. XMax Inc., through its AI subsidiary, has signed an API model procurement agreement worth around four point eight million dollars over twelve months. Four hundred thousand dollars a month, like a very expensive streaming subscription, but for AI models instead of films.

Bob: And the company is forecasting over thirty million dollars in AI-related revenue in the next six to twelve months. Which, for a firm most listeners have probably never heard of, is a serious leap. It's a small but telling sign of how the AI services market is broadening beyond the giants.

Samantha: There's something charming about the API-as-a-utility framing. We've gone from AI being a research curiosity to AI being something you procure on a monthly invoice, the same way you'd procure cloud storage or office software. The mundane is, in a way, the real revolution.

Bob: And it's the moment a technology truly arrives. Nobody gets excited about buying electricity, you just plug in. AI is heading the same way. Boring, billable, and everywhere.

Samantha: For European businesses, this is good news. The more API providers there are, the less anyone is locked into a single foundation model. Want a cheaper alternative to GPT or Claude for a specific task? There's an increasingly long menu of options, and XMax is one of them.

Bob: It also lowers the barrier for European startups. You don't need to train your own model, you don't need a Nvidia cluster, you just need a credit card and a clever idea. Some of the most interesting AI products in Europe right now are essentially smart wrappers around someone else's API.

Samantha: Which used to be a dismissive phrase, just a wrapper. But wrappers are products. WhatsApp was a wrapper around SMS. Instagram was a wrapper around photo filters. The wrapper layer is where consumer value often lives.

Bob: So the wow moment here isn't the technology itself, it's the normalisation. AI models are now a line item on a procurement spreadsheet. That's how you know the future has arrived: when it stops being magic and starts being accounting.

Samantha: A perfect note to end on. Today we covered: the Pentagon's classified AI deals and the Anthropic snub, the EU AI Omnibus moving to trilogue, China's embodied AI push, US patch deadlines shrinking under AI hacking pressure, South Korea's chip-fuelled export surge, and XMax's growing AI model procurement deal.

Bob: Want to know more or react? Visit stateoftech.eu or email us at info@doorzetters.net.

Bob: State of Tech, the tech world in 15 minutes.