OpenAI Says NATO Used Up to 300 Billion Tokens a Month. What Does That Prove?
OpenAI disclosed that NATO has used up to 300 billion monthly tokens in cybersecurity work. That is evidence of substantial usage, not disclosed revenue or measured defensive success.
Wealth / The Compound journalPhoto: Romaine / Wikimedia Commons, CC0 1.0 public-domain dedication; 16 February 2018, adapted archival crop · Illustrative image.
The strongest business signal in OpenAI’s new NATO cybersecurity announcement is usage at scale, not a contract value. Governments have ample reasons to experiment with AI for cyber defence. The harder question is whether very high token throughput turns into faster, safer and more economical protection of systems that cannot simply be taken offline for testing.
On 17 November 2026, OpenAI said it works with NATO’s Communications and Information Agency (NCIA) to provide advanced cybersecurity models for NATO’s core institutions. It reported that NATO has used up to 300 billion raw input and output tokens per month as part of this work. That wording describes a peak or upper bound, not a stable run rate, an invoice, a disclosed annual contract, a number of users or proof of a specific security benefit. The company did not publish the start date or full procurement terms.
What 300 billion tokens actually measures
Tokens are units processed by a model, not unique incidents investigated or vulnerabilities fixed. A single software review can consume many input tokens as tools inspect code, logs and findings; complex work may involve repeated calls. A large count therefore signals considerable model activity but does not reveal how much of it was useful.
The 300 billion figure can be expressed as 300,000 lots of one million tokens. If all those tokens hypothetically cost an illustrative $1 per million, the arithmetic would suggest $300,000 in usage charges for such a month. At $10 per million, the same exercise yields $3 million. Neither rate is an OpenAI quotation for NATO, nor is the $300,000–$3 million range an estimate of its actual spend. Input/output mix, specialised models, enterprise pricing, caching, infrastructure, volume discounts and contract arrangements could alter it enormously. The reported volume is also up to, not necessarily a monthly average.
This is why multiplying the peak by twelve to announce a billion-dollar contract would be unsound. Neither a reliable annual run rate nor any OpenAI revenue from the arrangement was disclosed.
Why cybersecurity can justify high model usage
Cyber teams must triage alerts, investigate unusual events, review code and decide which issues need human escalation. A model may help an analyst connect evidence across a large incident history, propose detection logic or spot dangerous software patterns. OpenAI describes authorised activities including suspicious-activity review, vulnerability validation and developing or testing potential fixes.
The commercial appeal is not a faster paragraph of explanation. It is shorter time between a verified threat appearing and an effective, authorised defensive response. To assess that, a buyer would need a record of how many real findings were correctly confirmed, how much analyst time was saved, whether false alerts increased, and what percentage of recommended remediations worked without introducing new vulnerabilities.
OpenAI also says it supported the 2026 Crossed Swords exercise organised by NATO’s Cooperative Cyber Defence Centre of Excellence. Participants developed and evaluated AI agents for cyber operations, with an emphasis on human oversight. An exercise is valuable training evidence, but it is not equivalent to a published independent evaluation of operational NATO networks.
The cost that a token bill misses
For cyber teams, the expense of a bad result can dominate nominal inference costs. A false positive can absorb scarce specialist attention; a false negative can leave a dangerous flaw unaddressed. An agent that generates many plausible patches may create a larger validation workload than a carefully constrained tool that makes fewer recommendations.
An illustrative team receives 10,000 model-assisted findings per month. Suppose 5% demand an additional 20 minutes of specialist review. That would create 500 escalations and approximately 167 hours of additional review. At a hypothetical fully loaded labour cost of €80 per hour, the review capacity alone is about €13,333, before software and incident-response costs. These are editorial assumptions, not reported NATO staffing or accuracy figures.
If genuine detections shorten incidents or avoid losses, that expense may be a very good bargain. But a cost-per-token metric cannot establish it. A procurement scorecard should capture cost per independently confirmed, correctly handled finding alongside missed defects, response latency and safe rollback.
Authorisation and accountability do not disappear
An AI tool with access to internal software, identity systems or logs touches some of an organisation’s most sensitive assets. Safe deployment requires scoped credentials, classification rules, audit trails, separation of recommendations from consequential changes and a way to revoke access. For material changes, human approval and tested rollback procedures remain important.
This point connects with OpenAI’s separate Hacktron cross-integration security partnership. Testing a model’s refusal behaviour is not the same as testing the boundaries between identity providers, code repositories and automated actions. Neither announcement constitutes independent certification that every deployed AI security workflow is safe.
NATO-related deployments also differ from a typical enterprise chatbot: security teams must account for data location, procurement requirements, classified-material rules and allied institutions’ differing responsibilities. The new OpenAI statement does not establish which specific datasets or classifications were processed.
What this means for OpenAI’s business
OpenAI is pursuing a market where a supplier may be assessed not just on model accuracy but on contractual assurance, auditability, deployment flexibility and integration with existing government technology. That could make public-sector relationships strategically valuable; it also raises the cost of assurance and support.
The positive signal is that a major intergovernmental organisation has undertaken significant model usage in a demanding application. The limitation is that OpenAI’s published statement supplies no contract price, recurring revenue, audited ROI, quantified vulnerability reduction or publicly verified uptime guarantee. These remain material unknowns for anyone trying to value the privately held company.
A meaningful follow-up would disclose aggregate time-to-detection improvements, independently verified remediation accuracy, approval controls and recurring operating costs, without exposing sensitive defensive operations. Until then, the right conclusion is large-scale reported use, with commercial and defensive outcomes still unquantified.
Explore the OpenAI private-company research hub and our analysis of agent checkout authorisation for other ways governance affects the company’s business model.
Research checked 18 November 2026. OpenAI’s announcement is dated 17 November. All price and staffing calculations are illustrative scenarios, not NATO invoices or measured outcomes. Cover: Romaine, Wikimedia Commons, February 2018, CC0; archival photograph, unrelated to the reported cyber activity.


