Pentagon vs Anthropic: $60M AI Showdown Over Military Access

So, What Actually Happened?

We scanned 190,000 articles this week, and the story that swallowed everything else wasn't about funding rounds or product launches. It was about a Cold War-era law being aimed at an AI company.

The Pentagon invoked the Defense Production Act against Anthropic, demanding unrestricted military access to Claude. Dario Amodei's response: ”cannot in good conscience accede.” While that standoff dominated headlines, ASML quietly achieved a breakthrough that could lift chip output by 50% and make would-be competitors irrelevant. Physical AI got its own data infrastructure layer when Encord raised $60M to challenge Scale AI. And Forbes declared data sovereignty an infrastructure problem, not a compliance checkbox, right as leaked diplomatic cables revealed Washington lobbying against EU data sovereignty rules.

The Bottom Line: When the Pentagon reaches for Cold War tools to compel access to AI models, the question is no longer whether AI matters for national security. The question is who gets to set the guardrails.

The Tracks That Matter

1. The Pentagon Just Invoked a Cold War Law on an AI Company. Anthropic Said No.

This is the first time the Defense Production Act has been aimed at an AI company, and the implications reach far beyond one contract dispute. Politico reports that Defense Secretary Pete Hegseth threatened to designate Anthropic a national security supply chain risk if it failed to comply by 5:01 PM Friday. The company's two conditions: no mass surveillance of Americans, no fully autonomous weapons. The Pentagon's contract language, according to Anthropic, made ”virtually no progress” on either.

What makes this more than a contract negotiation is the mechanism. The DPA was designed to force factories to make tanks, not to compel software companies to remove safety guardrails. AI policymakers on both sides called the approach ”incoherent”, questioning whether the law even applies to model access restrictions. The Financial Times reports Anthropic rejected the Pentagon's ”final offer,” while legal scholars at OpinioJuris argue this sets a precedent that every AI company should be watching.

Dario Amodei published his position directly: ”Our strong preference is to continue to serve the Department and our warfighters, with our two requested safeguards in place.” The LA Times frames it as a test of whether safety commitments survive government pressure, while the Federal News Network breaks down the legal mechanics of how the DPA could actually be enforced against a software company.

”[T]hese threats do not change our position.”
— Dario Amodei, CEO, Anthropic

Here's what works: If you're an enterprise customer of any major AI provider, this is your signal to review your contracts for government access provisions. The precedent being set here will define what ”AI safety” actually means when national security is on the other side of the table.

2. ASML's EUV Breakthrough: 50% More Chips, Zero Competition in Sight

While everyone was watching the Pentagon drama, ASML announced a breakthrough that quietly reshapes the entire semiconductor supply chain. The company's researchers have developed a way to significantly increase the power of the critical light source inside its extreme ultraviolet chipmaking machines. The result: their EUV light source can now consistently produce 1,000 watts under production conditions, potentially lifting chip output from 220 wafers per hour to 330 by the end of the decade.

The financial backdrop makes this even more significant. ASML generated $39.16 billion in net sales and $11.5 billion in net income in 2025, with $16.77 billion in net bookings during Q4 alone. More than half of those bookings were tied to EUV systems. The stock gained 101% in the past 12 months, outperforming the semiconductor index by nearly 40 percentage points.

The competitive angle is what matters most. ASML is now targeting the most technically demanding part of the EUV system: generating light with enough power and stability to support high-volume production. Any company hoping to compete in EUV lithography needs to solve this exact problem, and ASML just moved the finish line. The company expects 2026 net sales between $40.72 billion and $46.7 billion, signaling they see continued demand acceleration.

Here's what works: If your AI infrastructure strategy depends on chip availability (and whose doesn't), ASML's monopoly position just got stronger. The capacity expansion coming by decade's end means more chips, but through a single supplier. Factor that concentration risk into your procurement planning.

3. Encord Raises $60M: Physical AI Gets Its Data Infrastructure

The robots are coming, but they need data infrastructure that doesn't exist yet. Encord just raised $60M in Series C funding to build AI-native data infrastructure specifically for physical AI: robots, drones, autonomous vehicles. TechFundingNews frames it as a direct challenge to Scale AI, positioning Encord in the emerging ”physical AI data” category that barely existed two years ago.

What makes this raise notable is the timing. Physical AI is hitting an inflection point where the bottleneck has shifted from algorithms to data. Training a robot to navigate a warehouse requires different data pipelines than training a language model on text. The Morningstar press release emphasizes that Encord's platform handles the multimodal, spatially-aware data that physical AI systems need: video, 3D point clouds, sensor fusion, all with annotation workflows designed for robotics teams.

The investor roster signals where the smart money thinks the next infrastructure layer will be built. This isn't a bet on any single robot or drone manufacturer. It's a bet that every physical AI company will need a data layer purpose-built for the physical world, and that the data tools built for language models won't cut it.

Here's what works: If your organization is exploring robotics, autonomous systems, or any form of physical AI, audit your data infrastructure. The tools you use for text and image data were not designed for 3D spatial data, sensor fusion, or real-time annotation at robotics scale. The gap is real, and it's where the next infrastructure companies will be built.

4. Data Sovereignty Is No Longer a Compliance Problem. It's an Infrastructure One.

Data sovereignty surged 107% in PageRank influence this week, and the articles behind that spike tell a story that goes far beyond GDPR compliance. Forbes published a deep analysis arguing that the combination of cloud adoption, shifting geopolitics, AI workloads that need to stay local, and data residency regulations that are actually getting enforced has turned sovereignty into a strategic infrastructure question. Most organizations, they conclude, are nowhere near ready for that conversation.

The scale of the problem is becoming quantifiable. Kiteworks' 2026 Data Sovereignty Report found that one in three organizations experienced a data sovereignty incident in the past year. These aren't theoretical compliance risks; they're operational failures affecting regulated industries, defense contractors, and critical infrastructure providers across Canada, the Middle East, and Europe.

The geopolitical dimension is escalating. At the India AI Impact Summit, open source gained ground as a sovereignty strategy, but tensions persist between global AI collaboration and national data control. Meanwhile, leaked US diplomatic cables revealed American diplomats actively lobbying against EU data sovereignty provisions, framing it as a new AI ”front” in transatlantic relations.

”The practical result is that organizations are asking harder questions about what it actually means to store data 'in country' when access to that data might still route through third-party infrastructure operating under foreign jurisdiction.”
— Forbes

Here's what works: Map your data flows against the jurisdictions they touch. Not just where data is stored, but where it's processed, who can access it, and under which legal framework. If your AI workloads use cloud inference APIs, your data sovereignty exposure is larger than you think. The Forbes piece nails it: ”compliance requirements set a floor. The harder question is whether meeting those requirements actually reflects your real risk posture.”

5. NTT DATA and Ericsson: Building the 5G Backbone for Physical AI

While software AI gets all the attention, physical AI needs something the cloud can't provide: real-time connectivity at the edge. NTT DATA and Ericsson announced a partnership to deliver Private 5G as a managed global service with edge AI embedded directly into enterprise connectivity. The partnership focuses on four priority areas: global Private 5G at scale, AI at the edge, repeatable industry solutions, and a unified go-to-market.

This isn't a press release partnership. The target industries are specific: manufacturing, transportation, ports and logistics, energy and mining, smart cities. These are environments where autonomous systems need to make real-time decisions and can't afford the latency of cloud round-trips. Ericsson brings over a decade of enterprise connectivity, while NTT DATA brings global managed services capability.

”Private 5G is the backbone for scaling AI in production, where autonomous systems must operate reliably and at scale, but integration complexity often remains the final hurdle.”
— Alejandro Cadenas, Associate Vice-President of Worldwide Telco Research, IDC

The integration complexity point is the quiet connector. We saw Encord raising $60M for physical AI data infrastructure. We saw RLWRLD raise $26M for industrial robotics AI. Now NTT DATA and Ericsson are building the connectivity layer. Three separate stories, one pattern: the physical AI stack is assembling itself, piece by piece.

Here's what works: If you're running any kind of edge AI or autonomous system pilot, the connectivity layer is probably your weakest link. Private 5G managed services change the economics of edge deployment. The question isn't whether you need it; it's whether your current network architecture can support the latency requirements of real-time AI decisions.

6. Nimble Raises $47M: Agentic Web Search Gets Its Own Platform

The web search market just got an unexpected entrant. Norwest led a $47M round into Nimble, whose platform is built from the ground up for AI agents that need to search, extract, and reason over web content. This isn't another chatbot wrapper on top of search results. Nimble provides the infrastructure for AI agents to interact with the open web as a structured data source.

The timing reflects a shift in how AI systems consume information. As agents move from answering questions to executing multi-step tasks, they need web access that goes beyond keyword search. Nimble's platform handles the extraction, structuring, and rate-limiting that individual agents would otherwise need to build themselves. Norwest doesn't throw $47M at incremental improvements, which signals they see agentic web search as a new infrastructure category.

The broader pattern here: every layer of the AI stack that was previously handled by general-purpose tools is getting its own specialized infrastructure. Search had Google. Agentic search needs something purpose-built for non-human users navigating the web at scale.

Here's what works: If you're building AI agents that need to access web data, the build-vs-buy decision just shifted. Dedicated agentic web infrastructure handles rate limiting, anti-bot navigation, and structured extraction at a level that DIY approaches can't match. Evaluate whether your agent architecture treats web access as a solved problem or as the infrastructure gap it actually is.

7. Prophecy v4: When AI Agents Meet Visual Data Prep

Data preparation has been stuck in the same paradigm for a decade: write code, hope it works, debug when it doesn't. Prophecy just launched v4, which introduces AI agents for visual data preparation and analysis, running natively on Databricks, Snowflake, and BigQuery. The pitch: what took days or weeks now takes minutes, and the AI-generated logic is visual and reviewable rather than hidden in code.

The key insight isn't the speed improvement. It's the trust layer. Prophecy v4 addresses what CEO Raj Bains calls the bottleneck of validating AI-generated logic: ”Coding without AI has become unthinkable in the past two years. We are bringing the same power to millions of business data users.” Every workflow is stored as native code in Git, governance inherits from the data platform, and the visual interface means non-engineers can actually verify what the AI built before it runs in production.

This matters because data prep remains the unglamorous 80% of every analytics project. While the industry obsesses over model architecture and inference speed, the data that feeds those models still gets cleaned by hand. Prophecy's bet: AI agents plus visual verification will do for data prep what GitHub Copilot did for coding.

Here's what works: If your data team still spends most of their time on preparation rather than analysis, evaluate whether AI-assisted data prep tools have matured enough for your environment. The critical factor isn't speed; it's whether the output is auditable. Git-stored, visually reviewable workflows are the minimum bar for production data pipelines.

Signal vs. Noise

🟢 Signal: Dario Amodei and the Anthropic safety stance saw mention growth of 250% with 37.4% PageRank influence growth. This isn't hype; it reflects a genuine structural moment where AI safety commitments meet government power. When a CEO publicly defies the Pentagon and the market doesn't punish the company, it signals that safety positioning has become a defensible business strategy, not just PR.

🔴 Noise: AI Governance and Regulation saw mentions double but gained only 9.8% in PageRank influence. Everyone's talking about governance frameworks, but the actual influence of governance discourse isn't keeping pace. The gap between governance talk and governance action continues to widen, and most of the content is recycled position papers rather than new regulatory movement.

From the 190K

We scanned 190,000 articles this week. Here's what no one's talking about:

The Physical AI Stack Is Assembling Itself

Three funding rounds, one partnership announcement, and a robotics research breakthrough landed in the same 48-hour window. Individually, they're footnotes. Together, they're a blueprint. Encord raised $60M for physical AI data infrastructure. NTT DATA and Ericsson partnered on Private 5G for edge AI. RLWRLD raised $26M for industrial robotics AI. And researchers published new ”contact-anchored” learning methods that give robots a firmer grip on the physical world.

The software AI stack took years to assemble: cloud compute, model hosting, vector databases, orchestration frameworks. Physical AI is following the same pattern but compressing the timeline. Data layer (Encord), connectivity layer (NTT/Ericsson), application layer (RLWRLD), and foundational research (contact-anchored learning) are all getting funded simultaneously. If you're waiting for ”physical AI” to become a category before paying attention, you're already late.

🔍 Below the surface: Data Security appeared in 89 articles this week but made zero mainstream headlines. Here's how you spot real infrastructure: when something shows up everywhere but headlines nowhere, it means practitioners are dealing with it daily and marketing hasn't figured out how to make it exciting. Data Security's Katz centrality (foundational importance) ranks #1 across all entities this week, yet its PageRank growth is a modest 31%. The infrastructure layer everyone depends on, nobody talks about.

By The Numbers

• 50% chip output boost — ASML's EUV breakthrough could lift wafer processing from 220 to 330 per hour by end of decade
• $39.16 billion — ASML's full-year 2025 net sales, with $11.5 billion in net income
• 1 in 3 organizations — experienced a data sovereignty incident last year according to Kiteworks
• $60M Series C — Encord's raise for physical AI data infrastructure
• $47M — Nimble's raise for agentic web search, led by Norwest
• +107% Data Privacy PageRank — The largest single-day PageRank surge across all entities this week
• 89 articles on Data Security — highest Katz centrality of any entity, zero mainstream headlines
• 101% ASML stock gain — in 12 months, outperforming the semiconductor index by 37 percentage points

Deep Dive: The Defense Production Act Meets AI

Like a DJ pulling out a dusty record that nobody expected to hear again, the Pentagon just dropped the Defense Production Act into the AI mix. This 1950 law, designed to compel factories to produce military hardware during the Korean War, is now being aimed at a company whose product is software running on servers.

Why the DPA Matters Beyond Anthropic

The precedent isn't about one company's contract. If the DPA can be applied to compel access to AI models, it could theoretically be applied to any AI company. The legal question, as OpinioJuris analyzed, is whether model access restrictions qualify as the kind of ”industrial action” the DPA was designed to address. Nobody has tested this in court, and Anthropic's refusal may be the case that forces the question.

The Safety Premium

Something remarkable happened after Amodei's public refusal: nothing. No stock crash, no customer exodus, no mass partner defections. For years, AI safety has been dismissed as a PR strategy that would collapse under pressure. The Anthropic case suggests the opposite: in a market where enterprise customers care deeply about how their AI providers handle government requests, principled safety stances may actually be good for business.

The Technical Gap Nobody's Discussing

The Pentagon's own spokesman acknowledged they have ”no interest” in mass surveillance or fully autonomous weapons. If that's true, Anthropic's two conditions should be easy to meet. The fact that contract language ”made virtually no progress” on either condition suggests the real dispute isn't about surveillance or autonomous weapons. It's about who controls the guardrails, and whether a private company has the right to set limits on government use of its technology.

What Actually Works

1. Audit your government exposure: If your AI provider serves government clients, understand what access provisions exist in their contracts. Anthropic's public stance is unusual; most companies haven't disclosed their government access policies.
2. Build safety into procurement criteria: The Anthropic case demonstrates that safety commitments face real tests. Evaluate whether your AI vendors' safety policies would survive government pressure.
3. Watch the legal challenge: If the DPA is formally invoked, the resulting legal battle will define the boundaries of government power over AI companies for a generation.
4. Diversify your model providers: Regardless of how this resolves, concentration risk in AI models is now a geopolitical risk, not just a technical one.

The festival is happening whether you buy a ticket or not. But this week, someone just tried to change the setlist by law. How the crowd responds will tell us everything about what kind of show we're actually watching.

What's Coming

Amazon's $50 Billion OpenAI Bet Comes With Strings

Reuters reports Amazon's reported $50 billion investment in OpenAI may be contingent on either an IPO or an AGI milestone. If true, this is the most conditional mega-investment in AI history. It suggests even the biggest backers are hedging against the possibility that current AI capabilities plateau before reaching transformative returns.

The M&A Machine Keeps Accelerating

CNBC reports the global M&A boom is rolling into 2026, with AI sparking a deal frenzy. Capital is concentrating, smaller players are getting acquired, and the consolidation window is closing. If you're a startup wondering whether to raise or sell, the market is giving you a clear signal: the buyers are hungry.

Oaktree Capital Weighs In on AI's Trajectory

Oaktree Capital published a memo titled ”AI Hurtles Ahead,” providing one of the most measured institutional investor perspectives on where AI is heading. When a $190 billion asset manager puts their AI thesis in writing, it's worth reading closely for signals about where institutional capital will flow next.

For Your Team

Monday's meeting prompt: ”If the government invoked a Cold War-era law to access our AI tools, would our vendor contracts protect us, or would we find out the hard way that 'safety commitments' have an asterisk?”

The AI Vendor Stress Test Framework:

1. Government access provisions — Review every AI vendor contract for clauses about government data requests, compelled access, or national security exceptions. If the clause isn't there, the answer is ”they haven't decided yet.”
2. Safety commitment durability — Ask your AI vendors directly: under what circumstances would they modify their safety policies? Anthropic published their answer. Have your other vendors?
3. Data sovereignty mapping — The Forbes analysis is clear: sovereignty is infrastructure now. Map every AI workload against the jurisdictions it touches, including inference APIs you call that process your data elsewhere.
4. Physical AI readiness check — Three companies raised $133M combined this week to build physical AI infrastructure. If your roadmap includes robotics, autonomous systems, or edge AI, the infrastructure layer is being built now. Get in early or pay the integration tax later.

Share-worthy stat: ASML generated $39.16 billion in revenue and its stock gained 101% in 12 months while most people couldn't name the company. The most important infrastructure often belongs to the companies with the lowest name recognition.

Go deeper: Track data sovereignty and AI governance trends in real-time →

The Track of the Day

”The compliance requirements are the easy starting point. The harder question is whether meeting those requirements actually reflects your real risk posture, or whether you're checking boxes against regulations that were written before the current threat environment existed.”
— Forbes, on Data Sovereignty

That right there is the difference between passing an audit and actually being secure. One of them lets you sleep at night. The other one just lets you file the paperwork.

We scanned 190,000 articles this week so you don't have to. Data Pains → Business Gains.

Published: February 27, 2026 | Curated by Yves Mulkers @ Ins7ghts