For US allies, the combination of geopolitics, ever-expanding risk surface and an unsustainable dependence on hyperscale cloud providers has created significant excitement about the potential value of AI data centers and sovereign digital infrastructure. Many of those nations and critical infrastructure owners are now starting to realize that the token, not the data centre, is the atomic unit of AI value. Whoever prices the token actually controls the economics and value creation of every AI-dependent industry irrespective of sovereign cloud, data center or AI factory.

This growing realization will require a material shift in sovereign policy and capability development for those nations combined with short-term patience and recalibration from US partners. The shift will result in both greater sovereign benefit and a more productive, resilient Western Alliance for all.

Why Sovereign Infrastructure Means Little Without Controlling Tokens, Intelligence & Equity

Most Sovereign AI Frameworks are consistent in identifying common levers for success:

-Digital Infrastructure

-Skills and Talent

-Research, development and innovation RDI

-Industry development and commercialisation

-Governance and Equity

Having spent the last 15+ years helping allied nations design their compute, AI and security capabilities, I believe that the rapidly escalating global demand for sovereign digital infrastructure does starts to address where compute is controlled, however, it does little to solve what that compute costs and how value is created and equitably distributed.

Given that whomever prices the token, controls the economics and value creation of every AI-dependent industry irrespective of sovereign infrastructure, nations should treat token supply the way they treat energy supply - as a strategic reserve requiring stockpiles, contracts and dedicated domestic production capacity. It’s sadly ironic that some of these same nations most vocal about sovereign digital infrastructure have been less than diligent in developing and maintaining energy security.

The Rapidly Evolving Discipline of Tokenomics

While many leaders still think of LLMs in terms of infrastructure (more requests require more compute and therefore cost more), the reality is more complex. Users and use cases can create vastly different types of requests which have highly varied infrastructure and cost implications.

The smallest current production unit for LLMs are tokens – a fundamental unit of data that an AI model reads, generates or uses equating to roughly 4 characters per token. Over the last 3 years inference costs have reportedly fallen roughly 1,000-fold, with inference now accounting for two-thirds of all AI compute demand.

By any normal utility logic, falling unit prices should mean smaller monthly bills for users, however new models are both more token-intensive at increased prices while usage has skyrocketed, leading many to consume their entire annual AI and tech budgets in only a matter of months. This seems clear proof of Jevons Paradox where decreased unit cost is significantly outweighed by material and accelerating increase in use.

In addition, recent BCG research indicates that only 5% of their clients interviewed are creating substantial and sustainable value from AI.

BCG Build for the Future 2025 Global Study (n = 1,250).

Government and enterprise leaders have now seen more than enough exemplars and representative AI use cases – demand is accelerating to show a true return on investment “ROI” for those programs, which fundamentally means an ROI on token usage.

This discipline of financial accountability to AI use and governance has earned the name of “Tokenomics”, meaning the demand for ROI during a period of increased usage, risk and associated costs has created a triple whammy for policy makers, regulators and users.

Sovereign Clouds, Rented Intelligence & Lost Value.

What does all of this mean for a nation state, alliance or critical infrastructure owner?

While an esteemed British colleague and I recently considered the Sovereignty 2.0 cloud and data center location and legal control aspects for the World Economic Forum, we didn’t answer the pricing question ie. who sets the cost and terms of the thing produced (token) running on that stack which dictates the sustainable value proposition?

This is the sovereignty risk the digital infrastructure debate has largely missed but we expect to rapidly evolve over coming months.

A nation can host its own data centre, run its own hyperscaler partnerships, satisfy every metric in a Cloud, Data Center or AI Sovereignty Framework BUT still be entirely price-taking on the tokens flowing through it without a dedicated available reserve if, or when, the supply chain is disrupted. The subsidy era of new foundation models is coming to an end with Anthropic's 2026 enterprise pricing shift the most public example of consumption growth outpacing cost declines. Operational control of the rack means little if the marginal cost of intelligence itself is set outside of a sovereign legal and commercial jurisdiction.

Perhaps even more sobering are the concerns of US and allied intelligence communities that the increased costs of token usage are causing critical public and private sector users to move to highly capable “open weight” models predominantly developed and sourced outside of the trusted Western alliance. This week’s launch of the Moonshot Kimi K3 with comparable capability to recent Anthropic and OpenAI versions has further entrenched the systemic risk.

Greater independence of allied sovereign objectives at the infrastructure AND intelligence layers may require a shift for some policy makers, agencies and technology vendors, however, the strategic value from greater trust, intelligence, innovation and resilience would be significant.

Closing the Tokenomics Gap

While rightfully asserting digital sovereignty, America’s allies must stop treating sovereign compute and intelligence as a real estate problem and start treating it as a strategic commodity problem similar to the way we manage oil, grain or semiconductors.

That creates some meaningful challenges from both policy and practice perspectives:

-Create a national token reserve/s via prebuilt capacity insulating critical government and infrastructure workloads from price and availability shocks the way strategic petroleum reserves insulate against supply shocks

-Mandate transparent token cost disclosure in critical-sector procurement, so cost-per-outcome (not cost-per-CPU/GPU) becomes the sovereignty metric regulators actually score. “Commercial in confidence” wont be good enough for an agentic world.

-Create an active domestic AI model routing capability, so government workloads can shift between frontier and commodity models rather than being locked to a single provider's pricing curve.

-Plan for AI inference at the edge – while large urban Data Centers take up the headlines and capital, it won’t necessarily be how we consume and create value domestically and across the alliance.

-Demand a social license of all who participate – many nations have exported a significant portion of the value created during the internet era to hyper-scale foreign companies and demanded little in return other than occasional headlines about “strategic investment”. New strategies and structures must be created where the value created of this agentic token-driven world accrue directly to the communities that use them and not to be reallocated or misspent as a new form of taxation. Oil-driven Sovereign Wealth Funds may provide an effective blueprint fit for the Agentic Age.

A token reserve is the logical and necessary next stage for Sovereign AI

None of these policies replace the need for sovereign digital infrastructure. As token prices fall but total AI spend climbs, the value to nation states deriving from infrastructure alone is incomplete and does not achieve multiple core objectives. Nations that control where compute resides but not what tokens cost and deliver will remain intelligence and price-takers in the AI economy. For America’s partners, a sovereign token reserve is the next necessary step toward genuine digital autonomy and an even stronger Western Alliance.

The Cipher Brief is committed to publishing a range of perspectives on national security issues submitted by deeply experienced national security professionals. Opinions expressed are those of the author and do not represent the views or opinions of The Cipher Brief.

Have a perspective to share based on your experience in the national security field? Send it to Editor@thecipherbrief.com for publication consideration.

Read more expert-driven national security insights, perspective and analysis in The Cipher Brief

The future of disinformation is no longer about creating better content. It is about teaching machines what to retrieve, summarize and recommend. In the AI era, the editor matters more than the article.

A story no longer has to trend. It has to be retrieved when a large language model (LLM) constructs the answer. The most important audience in the information environment is no longer human. If an AI assistant becomes the primary gateway to information, influencing what it retrieves becomes more valuable than influencing what millions of people read directly.

The Hugging Face Model Hub, the top repository worldwide, tracks over 2.9 million total machine learning models. Many models are for wonderful uses, ranging from research universities to new private sector companies built to solve problems. Meanwhile, more than 130 active national sovereign initiatives exist in more than 60 countries, according to the Center for a New American Security (CNAS) Sovereign AI Index. Every one of these models becomes another editor with its own worldview, training corpus and retrieval strategy.

As Bob Dylan reminds us, “For the times they are a-changin’”.

Countries are in pursuit of a foundational model (cognitive sovereignty) that will provide its own historical context, experts, values, national interests and more.

Different versions of reality will emerge, not because people disagree, but because different models were taught to retrieve, prioritize and reason differently. Not unlike media outlets as they evolved, just with a completely different scale.

Our Focus

The old world was about content, distribution and amplification. The new world is about training, retrieval and reasoning.

We must have the expertise to explain the mental frameworks machines construct before we see an answer to our query.

Our slide decks will cross out the “attention economy” and replace it with the “cognitive economy.”

We will remind ourselves that during the social media era, an adversary would flood the zone with thousands of fake articles and accounts to amplify a narrative. In the AI era, the objective changes. Rather than convincing one person at a time, adversaries will increasingly seek to influence the system that answers everyone.

Perspective is also important. Printing presses made publishing a reality. Radio introduced the broadcast message. Television opened up reach to mass audiences. Social media democratized who could have a voice. And AI now changes who decides what we receive.

Our Preparation

We will need to expand our remit and add expertise in training data provenance, retrieval indexes, embedding systems, model guardrails, agent memory, citation chains and reasoning architectures. AI engineers will become important parts of our team, if not already so.

The decade ahead will introduce AI models and agentic systems that decide what billions of people see. Agents will continuously search, compare, negotiate, monitor and decide for us. Humans may never initiate the request, but the agent will know what to do. That’s a different information ecosystem. We must learn how to track its development accurately and efficiently, so we are in-step or a step ahead on each new innovation of importance.

AI models will cite other AI models who cite other AI models. Over time, the original source may disappear entirely behind layers of machine summarization. The citation survives, but the human reporting becomes increasingly distant.

The editorial model will change as quickly as it needs to. How do we keep up with changes in the perspective of a model on a key topic and why it occurred?

Bad actors will optimize less for search engine optimization (SEO) and increasingly for generative engine optimization (GEO), engineering content specifically to influence what AI systems retrieve and cite.

We will need a new intelligence platform that tracks all publicly accessible LLMs and all innovation in places like Hugging Face, so we can see patterns earlier across the world. Imagine tracking hundreds and then thousands of LLMs in real-time. We still care about what happened, who said it and the rest of the 5Ws, but increasingly, it will be meaningful to know how Claude, Gemini, ChatGPT, DeepSeek and other models summarize and frame key messages.

These platforms will help us as we develop skills to understand training data integrity, how retrieval systems are poisoned through Retrieval-Augmented Generation (RAG) attacks, and how agent memories are manipulated.

Invisible Persuasion

When the audience is the machine, our efforts shift from how to protect the population to how we analyze and influence the infrastructure that reaches us.

Media literacy taught us to evaluate what people published. Machine literacy teaches us how to evaluate how machines constructed the answer.

The era will have many names, I’m sure, but one that resonates with me is “invisible persuasion.”

Unlike propaganda, machine-led information thrives on invisibility. It must appear ordinary, mundane and just do its job.

The next generation of AI will continue to quietly remove the human being from both ends of the media system. It is becoming the audience, and it is becoming the editor. And it is doing both at once.

It is also succeeding in building trust in humans.

SparkToro and Datos Group found that 60% of US google searches ended without a click in the first four months of 2026.

The same person who once clicked through to read is increasingly staying put while a machine goes and reads for them. The Reuters Institute expects search referrals to nearly halve over the next three years.

A Pew Research Center report showed that users clicked a source cited inside an AI summary just 1% of the time (900 US adults, 68,879 google searches).

Trust is migrating from the publisher to the summarizer. Our learning used to include more friction – a competing headline or comments we disagreed with. Now, we get a clean answer without friction.

How this impacts our judgement is a question we’ll study for many years ahead.

Conclusion

The printing press democratized publishing. Search democratized discovery. AI is centralizing editorial judgement again, this time inside machines.

The new editors are not confined to newsrooms. They include model developers deciding guardrails, publishers licensing training data, platform owners determining retrieval rankings, governments building sovereign AI models, open-source communities releasing foundation models, and enterprises curating the knowledge bases their AI agents consult. Editorial power is becoming distributed across the AI stack rather than concentrated in traditional media organizations.

The organizations that understand how machines learn, retrieve, reason and remember will shape the next information environment.

The Cipher Brief is committed to publishing a range of perspectives on national security issues submitted by deeply experienced national security professionals. Opinions expressed are those of the author and do not represent the views or opinions of The Cipher Brief.

Have a perspective to share based on your experience in the national security field? Send it to Editor@thecipherbrief.com for publication consideration.

Read more expert-driven national security insights, perspective and analysis in The Cipher Brief