Pentagon Deploys AI on Classified Networks: 7 Companies, Not Anthropic

The Department of War approved seven AI companies to deploy their models on classified networks for operational military use on May 1, 2026. The seven: OpenAI, Google, Microsoft, Amazon Web Services, NVIDIA, SpaceX, and Reflection AI. One company was notably absent: Anthropic, which declined to sign the agreements after refusing to accept terms permitting unrestricted operational military use of its models.

This is the moment AI becomes a formal component of the US military apparatus, not an experimental program. Impact Level 5 and IL6 classification covers data up to the Secret level and sensitive unclassified defense information. Models running on these networks will process intelligence data, support command and control decisions, and assist targeting and logistics operations.

What "Deployed on Classified Networks" Actually Means

Impact Level (IL) ratings are the DoD's classification framework for cloud services handling different categories of government data:

• IL2: Non-controlled unclassified information (public government websites)
• IL4: Controlled unclassified information (acquisition-sensitive data)
• IL5: Controlled unclassified national security systems and some secret-level operations
• IL6: Secret-level classified information — the DoD's most sensitive non-Top Secret data

The seven approved companies have cleared the security assessment and authorisation process (SA&A) required to operate at IL5 and IL6. This is not a partnership announcement or a pilot programme — it is a formal authorisation to process classified data.

What military AI at IL5/IL6 does in practice:

Intelligence, Surveillance, and Reconnaissance (ISR): AI models analyse satellite imagery, signals intelligence, and sensor data faster than human analysts. A model running on classified networks can correlate multiple intelligence feeds, flag anomalies, and surface patterns that would take human teams hours or days. In an active conflict (the Iran operation is ongoing as of May 1), ISR AI means faster target identification and reduced analyst workload.

Command and Control (C2) decision support: Logistics optimisation, supply chain routing for combat operations, personnel deployment recommendations, and mission planning assistance. These are not autonomous targeting decisions — they are decision-support outputs that a human commander acts on.

Cyber operations: AI models running on classified networks can assist with offensive and defensive cyber operations, analysing network traffic for intrusions, suggesting responses to active attacks, and generating code for specific security applications.

Natural language processing of intercepted communications: Translating, summarising, and flagging relevant content from intercepted communications at a volume and speed that human analysts cannot match.

The authorisation means these AI companies are now formal defence contractors. Their models will operate inside the most sensitive computing environments the US government runs.

Why These Seven? And What Each Brings

The selection reflects both technical capability and contractual willingness:

OpenAI: GPT-5.5 and the Operator tier API. OpenAI has been building government relationships through its OpenAI for Government programme. The classified deployment likely runs a government-specific model version with enhanced logging, no training on classified data, and dedicated inference infrastructure.

Google: Gemini models via Google Public Sector Cloud, which already operates classified infrastructure for US government customers. Google has been building IL5/IL6 infrastructure for years. This extends existing government cloud to AI workloads.

Microsoft: Azure Government Secret and Top Secret clouds are the most mature classified cloud infrastructure in the commercial sector. Microsoft has operated IL6 and IL7 (Top Secret) Azure regions since 2018. Copilot for Security and Azure OpenAI Service on government cloud is the likely deployment path.

Amazon Web Services: AWS GovCloud and the classified C2S (Commercial Cloud Services) contract — the $10B defence cloud that AWS won in 2013 — provides the infrastructure foundation. AWS AI services on classified networks extend an existing customer relationship.

NVIDIA: NVIDIA is not typically thought of as a software company, but its AI Enterprise platform and the Hopper/Blackwell GPU infrastructure running classified DoW workloads are explicit here. NVIDIA's inclusion signals that the hardware layer itself (GPU drivers, NIM inference microservices, CUDA stack on classified servers) is being formally authorised alongside model deployment.

SpaceX: Starlink provides communications infrastructure in active conflict zones. AI models that run on Starlink-connected edge nodes — including potentially on battlefield devices in areas with degraded connectivity — represent a capability that the other six companies cannot provide. SpaceX's AI inclusion is likely tied to Starlink edge inference capabilities.

Reflection AI: The most interesting inclusion. Reflection AI was founded in 2024 by researchers who left Google DeepMind. NVIDIA led a $1 billion funding round with a $500 million personal commitment from Jensen Huang, which explains the relationship. Reflection AI is currently raising at a $25 billion pre-money valuation. The Pentagon inclusion at fewer than two years old is the fastest defence contractor authorisation for an AI frontier lab in recorded history.

Anthropic's Refusal: The Specific Disagreement

Anthropic declined to sign the classified deployment agreements. The reason: the agreements require companies to permit "unrestricted operational use" of their models in warfighter support contexts, with the DoW retaining authority over how the models are applied to operational decisions.

Anthropic's position: its responsible use policy does not permit deploying Claude in contexts where the model's outputs could contribute to lethal targeting decisions without additional safeguards that the DoW agreements did not accommodate. Specifically, Anthropic required that any Claude deployment in lethal autonomous or assisted-lethal contexts include human review requirements and red-teaming obligations. The DoW declined to accept those terms as part of the standard agreement structure.

Anthropic CEO Dario Amodei has previously stated publicly that Anthropic's models are developed with explicit commercial safety constraints and that national security applications requiring removal of those constraints are not compatible with how Claude is built. The DoW decision reflects this: Claude is not deployed on classified military networks.

The commercial implication is significant. Every government agency and defence contractor that builds AI-assisted systems on top of the seven approved companies gains an alignment advantage in defence procurement over those building on Anthropic models. Claude remains available for civilian government applications (non-classified). The classified operations market — worth hundreds of billions in contract value — is now effectively closed to Anthropic unless it changes its policy.

Reflection AI: What You Need to Know

Reflection AI is the least familiar name on the list and the most significant from a market-signal perspective.

Founded in 2024 by former Google DeepMind researchers, Reflection has been developing frontier models specifically designed for high-stakes deployment contexts — including defence, intelligence, and critical infrastructure — where commercial model constraints are not acceptable. The company has been explicit that it does not apply the same content restrictions that OpenAI and Anthropic use.

NVIDIA's $1 billion investment (with Jensen Huang personally investing $500 million) gives Reflection preferential access to Blackwell GPU allocations at a time when AI compute is genuinely scarce. The investment is both financial and strategic: NVIDIA wants a frontier model partner that can deploy on its classified government hardware infrastructure without Anthropic-style restrictions.

The $25 billion pre-money valuation at less than two years old reflects the classified government market premium: a company with unrestricted defence AI capabilities and Pentagon authorisation commands a valuation that pure commercial AI labs cannot match.

What This Means for the AI Industry

The Pentagon authorisation creates a structural bifurcation in the AI industry:

Companies willing to accept unrestricted defence use (OpenAI, Google, Microsoft, AWS, NVIDIA, SpaceX, Reflection): Access to the classified government market, which includes some of the largest and most stable long-term contracts in the US economy. Defence AI contracts are not subject to the commercial budget cycles and LLM price wars that govern the consumer AI market.

Companies with responsible use restrictions (Anthropic, others): Excluded from classified operations markets by policy choice. Continue to access civilian government, regulated commercial, and consumer markets where their safety positioning is a competitive advantage.

This is not a temporary state. Each company that builds out classified infrastructure (dedicated deployment, cleared personnel, IL5/IL6 audit compliance) creates increasing switching costs for the DoW. The seven companies selected now have a structural head start in the defence AI market that will compound for years.

Developer and Infrastructure Implications

If you work in defence contracting, cleared tech, or government IT:

• The seven approved companies will be hiring aggressively for engineers with security clearances (TS/SCI preferred) to build and operate classified AI infrastructure
• Cleared software engineers commanding 30-50% salary premiums over equivalent uncleared roles is the near-term outlook
• IL6 AI deployment requires architecturally isolated infrastructure — no commercial model endpoints, dedicated inference servers, airgapped or restricted-connectivity environments

If you work in commercial AI development:

• Anthropic's exclusion is the starkest demonstration yet that responsible use policies have commercial costs. The debate within AI companies about how far safety constraints should extend into government markets will intensify.
• The Reflection AI valuation ($25B at less than 2 years) sets a new precedent for what defence-aligned AI labs can raise. Expect more defence-first AI startups to emerge targeting the same market.

Key Takeaways

• Seven companies authorised for IL5/IL6 classified AI deployment: OpenAI, Google, Microsoft, AWS, NVIDIA, SpaceX, Reflection AI — effective May 1, 2026, operational military support contexts
• Anthropic explicitly excluded: declined to sign unrestricted operational use terms; Claude not deployed on classified networks; defence market closed to Anthropic under current policy
• Reflection AI is the signal: founded 2024, NVIDIA-backed ($1B round, Jensen Huang $500M personal), $25B pre-money valuation, Pentagon authorisation in under 2 years — fastest defence AI contractor authorisation on record
• What it enables: ISR analysis, C2 decision support, cyber operations assistance, intercepted comms processing — all at Secret-level classification and above
• Industry bifurcation: companies accepting unrestricted defence use vs. companies with responsible use restrictions now operate in structurally different markets with different growth trajectories
• Cleared engineering premium: TS/SCI-cleared AI/ML engineers, classified cloud infrastructure specialists, and IL6 compliance engineers are now the most sought-after roles in government tech

For the Oracle OCI classified government context, read Oracle OCI +84%, $553B Backlog: The AI Cloud Dark Horse. For the NVIDIA custom silicon that will run these classified workloads, read TSMC Q1 2026: 58% Profit Jump, 4.17M Wafers, HBM4 Sold Out. For the Iran war context that is driving classified AI demand, read Iran May 1: Trump Invokes 2002 AUMF, Ignores War Powers Deadline.