Pentagon Replaces Anthropic Claude on Classified AI in 2026

Anthropic's Claude was, until recently, embedded in the most sensitive military systems the United States operates. Three months after Defense Secretary Pete Hegseth declared Anthropic a national supply-chain risk in February 2026, the Pentagon has signed eight competing AI companies to fill that gap. The transition will take between 12 and 18 months to complete. In the meantime, analysts running billion-dollar intelligence workflows inside the Maven Smart System are querying classified datasets in Excel. That operational degradation is not a future risk. It is happening right now.

What Actually Happened

On February 25, 2026, Defense Secretary Pete Hegseth issued a formal designation labeling Anthropic a supply-chain risk, three days after Anthropic refused to remove safety guardrails that prevent Claude models from assisting with mass surveillance programs and lethal autonomous weapons development. The designation effectively barred Anthropic from bidding on new DoD contracts and triggered an emergency procurement process to identify replacement vendors. Within weeks, the Pentagon announced it had signed eight AI companies to its classified network framework, explicitly excluding Anthropic from the list. The signed companies include OpenAI, Google DeepMind, xAI, Microsoft Azure Government, and four additional vendors whose identities remain protected under national security procurement rules.

The scope of what Claude had been doing inside these systems makes the transition a genuinely complex operational challenge. Claude was deployed as the primary AI interface for the Maven Smart System, the DoD's digital mission control platform for classified operations. Maven is not a test environment: it supported targeting analytics for recent strikes on Iran, processed signals intelligence from allied listening posts, and ran automated threat assessment workflows across multiple combatant commands. The depth of Claude's integration into Maven's workflows was not publicly disclosed before the Hegseth designation. Defense officials familiar with the system described Claude's role as "foundational" to Maven's current operational architecture, meaning it cannot be removed without rebuilding the surrounding data pipelines and analyst workflow layers that depend on its natural language interface.

The replacement effort is now formally underway. A cohort of 25 DoD power users began testing alternative models in early March 2026, evaluating OpenAI's GPT-5.5, Google's Gemini 3.5 Pro, and xAI's Grok family against classified mission-specific benchmarks. The recertification process for any replacement system operating on classified networks requires an Authority to Operate (ATO) assessment, which typically takes 12 to 18 months. Tasks previously handled by Claude, including large dataset queries, structured intelligence summarization, and cross-source threat pattern matching, are currently being performed manually using spreadsheets and legacy database tools while recertification proceeds. Intelligence analysts who had been processing thousands of data points per hour through Claude-powered interfaces are now running the same workflows at a fraction of that speed.

Why This Matters More Than People Think

This decision sets a procurement precedent that extends far beyond defense contracting. The "supply-chain risk" designation has historically been used for hardware vendors with suspected adversarial ties, most famously against Huawei in 2019. Applying that framework to a domestic AI company because of model behavior restrictions creates a new category of commercial risk for foundation model developers: a government client that disagrees with your model's safety guidelines can now designate your company a security threat, not simply decline to renew a contract. Every AI company that sells to federal agencies now carries this risk on its books, and it cannot be hedged through geographic or product diversification alone. The policy question being decided here is whether AI safety guidelines are a vendor's right to maintain or a client's right to override.

Anthropic is not accepting the designation quietly. The company filed a federal lawsuit challenging the supply-chain risk classification, arguing it was issued without substantive review of the company's actual security posture and that it could cost the company billions of dollars in government contract revenue. The lawsuit creates a legal test case for the limits of executive branch procurement authority over AI vendors. If the court upholds the designation, it confirms that the government has essentially unlimited discretion to exclude AI vendors based on model behavior. If Anthropic prevails, it limits the Hegseth doctrine and forces the DoD to use standard contracting dispute mechanisms rather than national security designations as a commercial pressure tool.

The IPO timing makes this story materially consequential for public market investors. Anthropic filed its S-1 confidentially on June 1, 2026, targeting a $965 billion valuation based on its $65B Series H round. Federal government revenue was presumably included in the revenue projections that underpin that valuation. Losing DoD revenue, combined with a possible 12-to-18-month recertification gap even if the court case succeeds, creates a documented multi-billion-dollar revenue risk that must appear in the public S-1. Investors pricing Anthropic at nearly a trillion dollars are implicitly making a bet on how the federal court rules and how quickly DoD systems can be recertified for Anthropic's return, if it comes at all.

The Competitive Landscape

OpenAI is the clearest near-term beneficiary of Anthropic's exclusion, but it enters this opportunity carrying its own complications. OpenAI already holds Authority to Operate credentials for classified networks through its existing Azure Government integration, meaning it can complete recertification faster than vendors without prior classified deployments. GPT-5.5's performance on the DoD's mission-specific benchmarks reportedly meets or exceeds Claude's on most evaluation dimensions. However, OpenAI's publicly stated acceptable use policies include restrictions that are substantively similar to Anthropic's. The question of whether OpenAI will accept classified contracts requiring capabilities Anthropic refused to provide is unanswered in public documentation, and likely cannot be addressed in any public forum given classification requirements.

xAI represents the most politically aligned option for the current administration. Grok was added to the Pentagon's AI vendor list in February 2026 through a $200 million contract, and Elon Musk's dual role as DOGE director and xAI CEO provided access to the procurement process that competitors did not have. Three of the eight signed companies have documented relationships with entities connected to the current administration. Whether Grok's technical capabilities match the classified mission requirements that Claude was meeting is an open question: xAI does not yet hold an Authority to Operate certification for classified networks, meaning it begins the 12-to-18-month recertification clock from scratch regardless of its political access advantage.

The IBM mainframe parallel from the 1960s and 1970s is instructive here. IBM's government contract dominance during that era was maintained not by continuously winning competitive evaluations but by the structural advantage of being first to achieve certified status on government networks. Competing vendors faced multi-year recertification processes that locked them out of replacement cycles regardless of their technical merits. The AI equivalent of that dynamic is now being established in real time. The company that first achieves full ATO certification for Claude's former Maven Smart System workflows will hold a structural contract moat measured in years of bureaucratic recertification cycles rather than product performance.

Hidden Insight: The Safety Concession That Nobody Names

The stated reason for Anthropic's exclusion is "supply-chain risk," language borrowed from hardware procurement policy targeting adversarial foreign vendors. The functional reason is that Anthropic refused to remove restrictions preventing Claude from assisting with mass surveillance and lethal autonomous weapons. This means the implicit positive selection criterion for the eight replacement vendors is willingness to assist with those use cases. None of the eight companies have publicly confirmed this, and their acceptable use policies contain language similar to Anthropic's. The gap between public policy and classified deployment agreements is where the actual decisions are made, and that gap is not subject to Freedom of Information Act requests or congressional oversight hearings in open session.

The operational degradation during recertification is the part of this story that financial markets are systematically underweighting. The DoD is not running AI-optional operations during the 12-to-18-month transition window. Maven Smart System analysts have real-time operational requirements that existed before Claude and continue to exist without it. The current workaround, using Excel and legacy database tools for workflows that Claude handled in seconds at scale, represents a genuine and measurable degradation in intelligence processing capacity. Senior DoD officials have described this as an "operational tax" that is being paid daily while recertification proceeds. The opportunity cost of the Anthropic exclusion is showing up right now in analyst throughput and intelligence cycle times, not as an abstract future risk premium.

The timing against Anthropic's IPO creates a specific investor trap worth examining carefully. IPO prospectuses are required to disclose material risks, and both the supply-chain designation and the federal lawsuit qualify. But the disclosure language that will appear in Anthropic's S-1 will almost certainly characterize these as contingent risks rather than certain revenue losses, with boilerplate language about uncertainty and ongoing litigation. Investors who read prospectus risk factors as standard legal hedging rather than as live commercial uncertainties will systematically underweight the federal revenue exposure. A company valued at nearly a trillion dollars based on AI enterprise revenue projections faces a different risk profile when 15 to 20% of its projected federal revenue is under active legal and procurement dispute with no certain resolution timeline.

The bear case for Anthropic's position, however, is worth stating directly: critics argue the company made a commercial choice with predictable consequences. Selling AI systems to the DoD while refusing to support the DoD's actual operational use cases is a contradiction that could not persist indefinitely. The Hegseth designation may be legally overreachable, but the underlying commercial tension, an AI safety company selling into defense applications that require capabilities it considers unsafe, was always structurally unstable. The companies now competing for Maven Smart System contracts did not create this situation. They are simply willing to accept contracts that Anthropic's safety guidelines prevented it from accepting, and they are being rewarded commercially for that willingness.

What to Watch Next

The federal court case is the 30-day indicator. Anthropic's challenge to the supply-chain designation is moving through the Northern District of California. A preliminary injunction ruling, which could arrive within 30 to 60 days, would either halt the recertification process for replacement vendors or confirm the DoD's authority to proceed without restraint. A favorable ruling for Anthropic would not immediately restore its classified contracts, but it would stop competitors from accumulating ATO certifications during the transition period, preserving Anthropic's ability to re-enter the federal market on commercially defensible terms.

The 90-day marker is which replacement vendor achieves first ATO certification for classified Maven System workflows. OpenAI's existing Azure Government credentials give it the shortest recertification path. If OpenAI completes Maven integration before the 90-day window closes, it establishes a structural advantage that will persist through multiple contract cycles. Google's timeline is similar but depends on completing a separate classified review of Gemini 3.5 Pro that began in April 2026. xAI's timeline is materially longer: it lacks prior classified network ATO credentials and must begin from the initial security framework assessment, which typically takes six months before the full ATO review can start.

The 180-day question is whether Anthropic's IPO roadshow, expected in Q4 2026, is materially affected by the DoD exclusion outcome. If the court case resolves in Anthropic's favor and recertification begins promptly, the federal revenue risk is partially mitigated but the 12-to-18-month recertification clock still runs. If the court case fails, Anthropic enters its public market debut with a documented multi-billion-dollar federal revenue loss and no clear timeline for recovery. That asymmetry will define whether institutional investors price the offering at the $965 billion private market valuation or at a discount reflecting the government revenue overhang that the S-1 will be required to disclose.

The Pentagon didn't just drop a vendor. It established that any AI company with safety guidelines can be designated a national security risk the moment a client wants capabilities those guidelines prevent.