Deep Dive: The Pentagon's AI Push — Tech, Politics, and Capital in a Three-Way Battle

The Core Event

On May 1, 2026, the US Department of Defense (Pentagon) announced agreements with seven AI giants — SpaceX, OpenAI, Google, NVIDIA, Reflection, Microsoft, and Amazon AWS — to connect their advanced AI technologies to the Pentagon's most classified networks (IL6/IL7 level). This followed the Trump administration's 2025 AI Action Plan and Palantir's $10 billion Army contract.

But Anthropic was not on the list.

The company known for its "AI safety" stance was first labeled a "supply chain risk," then barred from federal government use by Trump's order, after refusing the military's demand for "unrestricted use" (banning mass surveillance and fully autonomous weapons). Meanwhile, OpenAI CEO Sam Altman stated in an internal memo that he shared "red line" principles with Anthropic — and announced the Pentagon deal hours later.

This is not just tech news. It is a three-way battle of technology, politics, and capital.

Part 1: Why Was Anthropic Blocked? — Not a Technical Issue, but One of Sovereignty

Timeline

| Date | Event | |------|-------| | 2024.06 | Anthropic becomes the first frontier AI company to deploy models on government classified networks | | 2025.07 | Anthropic signs $200M prototype development agreement with DoD | | 2026.01.09 | DoD issues AI strategy memo: "Abandon utopian idealism, embrace pragmatic realism" | | 2026.01 | US operation capturing Maduro allegedly used Anthropic models; Anthropic inquired with partner Palantir | | 2026.01 | Hegseth speech at SpaceX HQ: DoD AI "unconstrained by ideology," applicable to "any lawful use" | | 2026.02.23 | Hegseth summons Anthropic CEO Dario Amodei to the Pentagon, demands dropping "safety filters" | | 2026.02.27 | Trump orders federal government to stop using Anthropic; Pentagon labels it a "supply chain risk" | | 2026.02.28 | OpenAI announces Pentagon deal; Anthropic announces it will sue | | 2026.05.01 | Seven AI companies connect to classified DoD networks; Anthropic remains excluded |

The Core Disagreement

Anthropic's two red lines: 1. No mass domestic surveillance (of US citizens) 2. No fully autonomous weapon systems (lethal weapons beyond human control)

The DoD position: the military must be able to use AI for "any lawful purpose," including autonomous weapons development.

Hegseth's metaphor is telling: "When the Pentagon buys a Boeing plane, Boeing doesn't dictate where the plane flies."

The Deeper Logic

This is not a contract dispute — it is a question of who holds sovereignty.

• Anthropic's view: AI companies have the right to set usage boundaries; technology providers bear moral responsibility for misuse
• DoD's view: national security needs trump corporate policy; the military is buying "ownership," not a "lease"
• Political view: Trump framed Anthropic as "left-wing lunatics," politicizing its safety stance

Anthropic's dilemma: having pioneered deployment on classified government networks in 2024 (via AWS compute), it sought to build trust as a "partner." But when it tried to impose limits under that partner role, the military saw it as overstepping — you sell tools, not rules.

> Commentary: Hegseth's Boeing metaphor is logically coherent but technically absurd. A plane doesn't decide where to fly on its own, but an AI model generates its own outputs. When "use" includes "generating content," the boundary of responsibility between provider and user blurs. Anthropic's point: we're not just selling hammers; we want some say in how they're used. The DoD's reply: once sold, the hammer is mine. The essence of this debate: in a world where AI systems can autonomously make life-and-death decisions, where is the line between "tool" and "agent"?

Part 2: Seven Giants Enter Classified Networks — The Formation of an AI Military-Industrial Complex

Deal Details

• Networks: Impact Level 6 (Secret) and Level 7 (Top Secret/Sensitive Compartmented Information)
• Companies: SpaceX (Starlink/communications), OpenAI (GPT series), Google (Gemini), NVIDIA (GPUs/models), Reflection (open models), Microsoft (Azure/OpenAI cloud), Amazon AWS (cloud computing)
• Later addition: Oracle (databases/cloud)
• Platform: GenAI.mil (DoD AI platform, 1.3 million DoD personnel users within 5 months of launch)

Key Figures

• FY2027 budget: DoD requests $54 billion for the "Drone & Autonomous Warfare Group" (DAWG), a 24,000% increase over last year
• OpenAI contract: signed June 2025, ceiling of $200 million
• Palantir Army contract: ten years, valued at $10 billion
• GenAI.mil users: 1.3 million DoD personnel (within 5 months)
• 2025 retail inflows: $301 billion into US stocks (14% above the 2021 peak)

Why These Seven?

On the surface, technical capability; in reality, geopolitical narrative:

• OpenAI: America's flagship AI name, signaling "commercial AI supports national security"
• NVIDIA: controls the upstream of the AI arms race via compute infrastructure
• Google: full coverage of search/cloud/models; "Don't be evil" yields to defense contracts
• Microsoft: Azure cloud + OpenAI partnership provides deployment infrastructure
• Amazon AWS: the world's largest cloud provider; military data depends on it
• SpaceX: Starlink's combat value proven in Ukraine
• Reflection: packaged as "the American DeepSeek," despite having released no public model — evidence the Pentagon is buying narrative, not just technology

Reflection's inclusion is the most interesting. TechCrunch reported it raised $2 billion at an $8 billion valuation "without having released its first public model." But as the representative of the "open model" route — in the current open-vs-closed AI ideological war — the DoD needed an "open" company to hedge OpenAI's "closed" image.

Part 3: Retail Investors Sell Tech for Gold — Risk Aversion or a Crisis of Faith?

Gold's Surge

• 2025 gain: gold rose 65%
• January 2026: up another 30%+ in under a month
• January 26, 2026: spot gold above $5,100/oz
• Goldman Sachs target: raised from $4,900 to $5,400 (end of 2026)
• Volatility: gold's 30-day volatility indicator hit 44%, the highest since the 2008 financial crisis

Capital Flows: Tech Stocks vs Gold

| Indicator | Data | |------|------| | 2025 retail inflows into US stocks | $301B (14% above 2021 peak) | | Retail share of trading volume | 20-25%, hitting a record 35% in April | | Gold ETF outflows (March 2026) | $7B+ | | Total commodity ETF outflows (March 2026) | $11B | | Silver single-month volatility | surpassed 100% |

Why Switch to Gold?

Layer 1: Geopolitical hedging

• US ambitions toward Greenland
• Market turmoil from Trump administration pressure on the Federal Reserve
• Continued conflicts in Ukraine and the Middle East
• Central banks steadily accumulating gold (China: 12 consecutive months, reserves at 2,303.5 tonnes)

Layer 2: AI bubble anxiety

• DeepSeek (January 2025) broke the "compute = leadership" narrative, training at roughly 1/30 of OpenAI's cost
• Retail investors began diversifying away from AI-concept stocks
• Concentration risk in tech stocks (reminiscent of the 2021 meme-stock frenzy)

Layer 3: Deep unease about AI militarization

• OpenAI, Google, and others simultaneously serving civilian and military markets — the "safety" narrative collapsing
• 900+ Google employees and 100+ OpenAI employees signed letters urging their companies to side with Anthropic
• Caitlin Kalinowski, OpenAI's robotics and consumer hardware lead, resigned on March 8, 2026
• Investors realized AI militarization could trigger regulatory backlash, ethical controversy, and international retaliation

> Commentary: Retail investors moving to gold is, on the surface, "risk aversion" — in reality, a "faith transfer." From 2021-2024, the tech narrative was "AI changes everything," and retail bought the future. From 2025-2026, with AI's military uses exposed, Anthropic blacklisted for holding its safety line, and OpenAI preaching "red lines" while signing Pentagon contracts — investors saw companies that are not idealists changing the world, but pragmatists serving power. Gold has no ideals, and no betrayals. In this sense, swapping stocks for gold is not an investment choice — it is moral hedging.

Part 4: Tech Revolution or Ponzi Scheme? — Three Dimensions of Judgment

Dimension 1: Is the technology genuinely advancing?

Yes, genuinely.

• GPT-4, Gemini, Claude and others show real breakthroughs in reasoning, coding, multimodality
• DeepSeek proved high-performance models can be trained cheaply (though later questioned for "distilling" OpenAI)
• AI applications in medicine, science, and education have real value

But militarization is not a necessary component of "progress."

Dimension 2: Is the capital structure sustainable?

Bubble signs exist.

• Reflection valued at $8 billion "without a public model"
• OpenAI's shift from nonprofit to for-profit, removing "safety" from its mission statement
• The DoD's $54 billion budget (up 24,000%) reads more like political theater than an executable roadmap
• AI data center construction outpacing actual demand (e.g., "Stargate")

But unlike the 2000 dot-com bubble:

• There is real revenue (OpenAI's annual revenue in the billions)
• Government contracts (DoD, federal agencies) provide a floor
• There are clear application scenarios, not just concepts

Dimension 3: Are the ethical risks controllable?

This is the most dangerous part.

• Autonomous weapons: UN data from December 2025 showed 156 countries supporting resolutions on autonomous weapons; the US shifted to opposition
• Surveillance risk: AI militarization inevitably spills into domestic surveillance (see Palantir's history)
• Diffused responsibility: "human in the loop" ≠ "human control" — when AI marks the target red, writes the rationale, and computes the probability, an operator under a countdown of seconds will essentially just click "approve"
• Technology spillover: military AI (target recognition, situational awareness) shares underlying architecture with civilian systems; flaws affect both

Conclusion: This is not a Ponzi scheme, but it may be more dangerous than one.

A Ponzi scheme's victims are investors. AI militarization's victims could be civilians, soldiers, and surveilled citizens. It doesn't "steal money" — it "steals lives": using the rhetoric of "human in the loop" to outsource lethal-decision responsibility to human operators while crediting AI for "efficiency."

Part 5: Overall Assessment

> The core of this matter is not "whether AI can be used militarily." Technology has no morality, but the use of technology does. > > The real question: when an AI company serves both civilian markets and the military, how much is its "safety" pledge worth? > > OpenAI said "no mass surveillance and no fully autonomous weapons" — then signed the Pentagon contract. NVIDIA said it "won't provide usage restrictions beyond legal requirements" — then agreed to support autonomous weapons development. Google said "Don't be evil" — then connected to IL7 top-secret networks. > > Anthropic was the first company to try drawing contractual boundaries. It failed — not because it was wrong, but because it underestimated how the military defines "sovereignty." To the Pentagon, "purchasing" means "total control"; any conditions attached are an infringement of corporate power on state power's turf. > > But Anthropic's failure has value of its own: it proved this industry's safety rhetoric is made of paper. When profit and safety conflict, profit wins. > > Retail investors buying gold, in a sense, saw through this. They no longer believe in the "AI changes the world" fairy tale — at least not one that can accommodate both "changing the world" and "changing the battlefield." > > This is not the end of the tech revolution. It is the tech revolution facing its own shadow for the first time.

---

Sources referenced:

• Caixin: "Trump Orders Federal Government to Stop Using Anthropic" (2026-02-28)
• Caixin: "After Federal 'Blacklisting,' Anthropic Announces It Will Sue" (2026-02-28)
• Xinhua: "OpenAI Signs with the Pentagon" (2026-03-02)
• Cailianshe: "Pentagon Requests $54 Billion Budget" (2026-04-23)
• IT Home/Reuters/Guancha (2026-05-01): DoD seven-company AI agreement coverage
• Jin10 Data: "US Retail Stock Investment Funds Hit Record in 2025" (2025-12-24)
• Yicai: "Spot Gold Above $5,100" (2026-01-26)
• MIT Technology Review: "Google and 6 Other AI Companies Sign, Fully Unlocking Lethal Autonomous Weapons Restrictions"
• 21st Century Business Herald: "Gold Trading Volatility Will Persist" (2026-04-01)