There Is Another AI Race We Could Choose to Run

Every few months, a new model arrives that can reason a little further, use more tools, and stay with a piece of work for longer. Governments answer with chips, data centres and national strategies. Companies answer with agents that remember a person and act across their software.

That contest is real. I want it to continue. A nurse who can ask for help in ordinary words, and receive a careful draft of a discharge plan, is in a better position than she was with a dozen disconnected systems and no time.

The question I cannot settle by watching model announcements is different. As these systems become more capable, does their capability have to grow by absorbing human and institutional capability? Or can it also grow by becoming better at discovering, reaching and assembling capability that remains with people, communities and institutions?

Both patterns can coexist, including inside the same vendor or the same hospital. The deeper proposition is that increasing machine capability does not necessarily require decreasing human or institutional capability. I do not know yet whether that second path holds. But I think it is worth taking seriously.

The hospital, and two ways capability can grow

Consider a regional hospital, because the abstraction is easier to see once it has a floor and a person in it.

A nurse is arranging support for a patient who is ready to leave. The patient lives alone. The wound needs care. There is a funeral on Thursday that he is determined to attend. Somewhere in the hospital there is a discharge policy, a delegation for who may vary it, a record of his admissions, an agreement with a community nursing service, and a physiotherapist who has treated him for years and is not listed in any system an outside agent would know to call.

Today, the nurse carries much of that map in her head, and in the heads of people she knows to ring. That is slow, uneven, and often unfair to the patient who does not have an experienced nurse on the day. It is also capability. It lives in the institution and around it.

Watch where similar capability is being asked to live in the current generation of AI products, and two trajectories appear.

Absorption. Useful knowledge, operating practice, interpretation or memory moves into the model, into a skill it has learned, into specialised weights, or into the platform that coordinates agents. OpenAI has spent two years giving ChatGPT a memory of the person it is talking to. In June 2026 it began rolling out a stronger background process it calls dreaming, which synthesises a fresher picture of preferences and constraints inside the product. OpenAI described the design as a way to cope with memory at the scale of hundreds of millions of users. The new architecture started with paying users in the United States, with a wider rollout planned after that.[1] Salesforce introduced Koa in September 2026, a reasoning model post-trained on synthetic scenarios drawn from nearly three decades of CRM processes. The company was explicit that customer data was not used. The asset being put into the model was Salesforce's own accumulated knowledge of how enterprise work of that kind is done.[4] Machine capability increases partly because more of the work now lives inside the machine environment.

Reach. The same market is also building machinery for the opposite move. Anthropic published Agent Skills in December 2025 as an open standard: a folder of instructions and reference files an agent can load when a task needs them.[2] A separate memory tool is designed so that the customer's application stores the files the agent reads and writes; the model requests an operation, and the application carries it out against storage the application owner controls.[3] Skills and client-side memory do not remove the model. They separate some expertise and memory from the weights. In the hospital, reach would mean the discharge policy, the delegation, the record and the nursing agreement stay as things the hospital owns. A model interprets what the nurse or patient is asking, establishes facts from authoritative sources where they can be established, and calls capabilities the hospital has chosen to make findable. The physiotherapist would still need an interface. Reach does not magic informal knowledge into view.

These are not rival races run by different companies for one prize. They are design choices that can sit next to each other in a product roadmap and in an institution's architecture.

Why absorption remains a serious path

I want to give the absorbing path its due, because a weaker version of this essay would pretend it is only a mistake.

Frontier training is expensive, and the expense is still rising. Epoch AI estimates that spending on the largest training runs has been growing at about 2.4 times a year, and that the most advanced models now cost hundreds of millions of dollars to train, on an accounting that spreads the cost of the cluster across the training run.[5] Those are estimates, not invoices. They are enough to show that the leading edge is not a commodity a hospital, a council or a smaller country can reproduce at will. The International Monetary Fund's 2025 modelling reached a related conclusion from the other end: on its estimates, the growth effect in advanced economies could be more than double the effect in low-income countries, and better preparedness is unlikely to offset that fully.[6] The World Bank's World Development Report 2026 locates the most advanced models, the chips they depend on, and the data centres that run them mainly with companies in a few economies, and warns that developing countries could become dependent on technology they do not control.[7]

There is also a serious technical argument for putting knowledge in the machine. Rich Sutton's bitter lesson, written in 2019, is that general methods which soak up computation have beaten carefully hand-built knowledge again and again.[8] I have written about the institutional limit of that lesson elsewhere.[9] If the goal is competence, there is a strong track record behind putting more knowledge into the model. A specialised model that has internalised a domain may be more reliable than a general model rummaging through documents it only half understands. Koa is a bet of that kind, and it may pay. World-model research pushes further toward internal accounts of how environments evolve; one 2026 example trains a compact internal model to predict physical dynamics and plan from that prediction.[10] A hospital's delegations and local agreements are a different kind of object, but the commercial and research pressure to absorb operational knowledge is not a misunderstanding of the technology.

The proposition I opened with can fail. It fails if competence really does require the knowledge to live in the weights. It fails if keeping capability outside the model costs so much that patients wait longer for a worse result. It fails if "local" becomes an excuse for a fragmented system no one can use. I am not willing to dismiss that failure in advance.

Reaching capability as ordinary engineering

The reason I have not dismissed the other trajectory either is that reaching external capability is no longer a thought experiment.

Patrick Lewis and colleagues showed in 2020 that a language model could be combined with a separate retrievable index so that knowledge-intensive answers draw on passages retrieved at run time rather than only on facts stored in the weights.[20] Retrieval-augmented generation is only a narrow precedent. It does not establish institutional authority, sovereignty or fairness. It does show that useful knowledge can remain outside model parameters and be pulled in when needed. That is the architectural fact the hospital example depends on.

This pattern is moving into mainstream AI infrastructure. Anthropic's Model Context Protocol lets models discover tools, data and services outside themselves; by late 2025 it had more than 10,000 public servers and had moved under Linux Foundation governance through the Agentic AI Foundation.[11] Google's Agent2Agent protocol addresses another connection: agents finding other agents and handing work across frameworks and organisations. It, too, was donated to the Linux Foundation in 2025.[12] Models can increasingly call capability that lives elsewhere, and agents can hand work to one another, without that capability having to be copied into the weights.

Neither protocol is a theory of sovereignty. Both are evidence that "the model reaches a capability it does not contain" is now ordinary engineering.

Making capability externally available is not the same as reliably discovering the correct capability. Most tool-use evaluations still pre-select a small set of relevant tools for each task. Shi and colleagues, in work published at ACL 2025, argue that real toolsets are large, similar tools abound, and retrieval quality limits what agents can do. On a pilot using ToolBench, they report that when the small annotated toolset evaluators usually rely on is replaced by tools a retriever must find, end-to-end pass rates fall sharply; in their figure, the pass rate with a pre-annotated toolset is 64.2, while a strong retriever reaches only about 27.3 recall at 10, and the pass rate drops by about ten points.[21] For the hospital, that is the difference between an open standard and a physiotherapist who still does not appear in the tool list.

One way to picture the relationship reach assumes is a short sequence:

Language → Intent → Context → Existing capabilities → Work → Human judgement → Accountable outcome

The hard part is keeping each step in its place. Finding a document is not the same as establishing that the document is the rule which applies. A model can help with language and search. The institution still has to determine what is authoritative before anyone relies on it. The operating model we have been developing around that distinction is Operational Intelligence.

The sentence I want the hospital example to earn is this:

A model does not need to contain the capability in order to help someone reach the capability.

That is an architectural proposition. RAG, the protocols and the skill format make it technically plausible. They do not make it true in a given ward, and they do not make it fair.

Language, intent, and who gets to the service

For decades, using an institution meant learning the institution: the right form, menu, team, professional vocabulary, software surface, process name, sometimes another spoken language. People who already spoke that language, in both senses, moved faster. Everyone else waited, gave up, or relied on someone who knew whom to call.

A person can describe a situation in language available to them. Interpretation can help discover capability. Authoritative sources still establish what can be established. That is the proposition behind the chain above, and it matters because the first two steps are where many people are excluded today. The patient who does not know the term "community nursing referral" may still know that someone needs to dress the wound at home. A model that connects that intent to a legitimate service is not the service. It is an interface.

A small 2025 study tried machine simplification on 14 preliminary decisions from a Brazilian court of accounts. The models produced plainer versions far faster than drafting by hand. Omissions of important information were common. When models scored the results, they tended to rate readability higher than human reviewers did.[13] Language can become an interface to capability without language-model interpretation becoming institutional authority. The gain and the risk sit in the same place: a fluent explanation that has quietly dropped the condition which would have changed the answer.

Local capability as distribution, not only survival

The familiar commercial path is easy to describe. Capture local knowledge. Encode it into a model. Scale the model. More of the value can then accumulate around the model, because that is where the knowledge can now be used.

The more interesting proposition is that AI may become a distribution mechanism for capability that remains elsewhere. A regional service, a specialist, a community organisation or a hospital practice may be hard to reach because people do not know it exists, what it is called, how it fits the process, or how to access it. If a model can interpret intent and discover that capability, its effective reach may increase. I would rather say it in the hospital's terms than as a slogan: make this hospital's discharge practice findable so a model can help this nurse and this patient reach it, rather than copying the practice into a global model so every hospital talks to the same discharge agent.

David Autor, in NBER Working Paper 32140, argues that AI's opportunity in the labour market may be to extend the relevance, reach and value of human expertise: by weaving information and rules with experience to support decision-making, it could enable a larger set of workers with complementary knowledge to perform some higher-stakes tasks now concentrated among elite experts.[22] Autor is explicit that this is an argument about what is possible, not a forecast. It supports the distribution proposition at the level of expertise. It does not show that inequality will fall, and I will not use it that way.

Whether discoverability increases the economic value retained by the capability holder is an open question. If AI increases the reach of a local capability, who captures the resulting value: the holder, the discovery or intermediation layer, or both? Local search is a sobering analogy. Michael Luca, Timothy Wu and colleagues ran experiments in which users saw either Google's usual local OneBox, filled only with Google's own review content, or a version that populated the box with listings from platforms Google's organic algorithm had already ranked as most relevant. Users were roughly 40% more likely to engage with the OneBox when it included that third-party content, which suggests Google's exclusive presentation was worse for users; it also shows how much power sits at the layer that decides what is shown.[23] Open protocols can reconcentrate at the point of discovery. A hospital that makes discharge capability findable still leaves someone choosing which model calls it, which tools are offered, and whose interests that choice serves.

The World Bank's report limits the proposition from another direction. AI works where the complements exist: electricity, connectivity, education, institutions, data in local languages. Adopting a tool trained elsewhere is not enough; the Bank's Nigeria example is a system trained on high-income-country data recommending far too many laboratory tests.[7] Where local capability does not exist, there is nothing to discover. A model that carries medical knowledge may be the capability. Many countries have too few doctors and advisers; the Bank argues that adaptation of existing models, including open models in local languages, is where most developing countries should put their money, not frontier training.[7] Where capability does exist, absorption and discovery are different events. Absorption can make the local practice less necessary. Discovery can make it more usable. Whether it becomes more valuable to the people who hold it is not settled.

Can people, communities and institutions retain, develop and benefit from their capabilities as AI becomes more capable of reaching them? That question is enough. I do not think we need another sovereignty noun to ask it.

Silent surrender

The ordinary case is not a machine acting alone. The nurse approves the discharge plan. The institution can point to the approval. It may no longer be able to point to the judgement.

Slow the decision down and it is not one moment. What the situation is. How it is interpreted. What matters. Which evidence counts. Which capabilities are visible. Which alternatives appear. What consequences are predicted. What action is taken. What follows.

A person who approves the action has touched one of those moments. The others may already have been settled by a model that decided this was a community-nursing referral, ranked clinical risk above the funeral, retrieved one paragraph of the guideline and not the ward note that complicates it, and offered three options that all assume the patient leaves today. The physiotherapist who is not in the tool list never appears. The nurse can still say no. She is saying no inside a space she did not build.

European law has noticed a version of this problem. Article 14 of the EU Artificial Intelligence Act requires that a person overseeing a high-risk system understand its limits, remain aware of the tendency to rely on it too readily, interpret its output, and disregard, override or stop it.[14] The Regulation does not tell an institution which earlier moments count as judgements it has delegated.

Silent surrender does not necessarily occur when a machine acts. It occurs when an institution ceases to know which judgements it has delegated.

The nurse's approval can be genuine. The loss, if there is one, is upstream: an institution that can no longer say which interpretations, which selections of evidence, and which closures of the option set were handed to a system, on what authority, with what chance of being contested. Formal accountability can survive that loss for a long time. Practical judgement does not.

Human sovereignty

From that loss follows a practical question about people, not a demand that humans manually approve every action.

If the model constructs more of the decision space, human approval occurs inside a space the human may not have constructed. The institution may cease to know which judgements it delegated. The person may lose the practical ability to understand, question, contest, reinterpret, refuse or choose.

Human sovereignty, as I mean it here, is the retained capacity to do those things and to remain answerable for consequential outcomes. It is not a checkbox at the end of a workflow. An institution that still knows its own meaning, its own capabilities and its own delegations gives people a place to stand. An institution that has forgotten which judgements it handed over can keep every approval step and still have lost that.

One practical response is to make delegated moments visible: what was interpreted, which evidence was used, which alternatives were in view, who was authorised to decide, and who remains accountable. That is the problem Judgement Governance™ is meant to address. Naming the problem does not solve it. An approval screen does not solve it either if the screen appears only after the decision space has already been built.

Institutional sovereignty, and the cost of being explicit

Institutional sovereignty asks whether the institution can still determine what its words mean, which information is authoritative, which capabilities exist, who may exercise them, what may be delegated, where judgement sits, and who remains accountable.

A hospital can host its models in-country, on infrastructure it trusts, and still lose that authority. The model can sit inside the building and still be the only place that remembers how discharge decisions are actually interpreted. Replacing it would mean teaching the hospital to itself again. Location is not meaning. Keeping authority with the institution, whichever model is helping, is the problem we describe as Sovereign Operational Intelligence.

There is an enabling side to the same idea. If an institution makes authoritative knowledge, capabilities, delegations, constraints and judgement boundaries explicit enough for a machine to find them, increasingly capable models may participate more effectively without having to absorb those things into themselves. The useful formulation may be: make the institution more machine-navigable while keeping it institutionally authoritative. I have written elsewhere about what happens when institutions become machine-navigable without that discipline: intelligence traverses the estate, but establishment, provenance and accountability are reconstructed inside the agent loop.[24] The same explicitness that helps legitimate AI reach the right capability can make an institution easier to traverse, exploit or misuse. Machine navigability is not an unqualified good.

Sovereignty layers, access, and what equality cannot mean here

Infrastructure sovereignty is where compute runs and who controls that supply. Model sovereignty is who trains, owns, inspects and can replace the model. Brookings Institution researchers, in February 2026, define AI sovereignty as a spectrum of ways to make independent decisions about critical infrastructure and adoption, not as self-sufficiency. Full-stack sovereignty, they find, is structurally out of reach for almost any country; managed interdependence, diversification and portability through standards and procurement are the practical stance. Sovereign AI can also serve protectionism, wasted spending, or tighter political control of information.[15] Interoperability stops a country being trapped in one supplier. It does not, by itself, keep the institution as the authority on its own rules.

Equality of access to capability does not require everyone to possess equal machine intelligence. It may require that people can reach, develop and participate in capability without being structurally excluded by language, wealth or dependence on one model provider. That is the proposition architecture might speak to. It is not evidence of income equality, wage convergence, equal economic outcomes, reduced capital concentration or universal access.

Acemoglu, Autor and Johnson have argued that generative AI can be built to complement workers or mainly to displace them, and which path is funded is a choice.[16] Acemoglu's macroeconomic work adds the caution that even when AI makes less expert workers more productive at existing tasks, that gain need not reduce labour-income inequality and is likely to widen the gap between labour and capital.[17] Autor's possibility argument and Acemoglu's inequality caution answer different questions. Both belong in the essay. The narrower question that survives is whether inequality in access to capability can be reduced even when access to frontier machine intelligence remains unequal. I treat that as an open question, not a finding.

Model agnosticism, in that limited frame, is more than vendor portability. It is architectural substitutability: institutional capability is not trapped inside one roadmap when prices, performance or politics change. It is a resilience property for smaller institutions that cannot buy, or are not allowed to use, the single model in which their operating knowledge now lives. Where the scarce capability exists outside the model, access to the strongest possible frontier model may not always be the only route to useful capability. Where the scarce capability is the model's own reasoning or learned expertise, that bounded claim does not apply. I will not argue that small models make frontier models unnecessary; benchmark contests on tool calling are not a hospital discharge plan.

A false choice for smaller economies

Australia's National AI Plan, released in December 2025, puts serious effort into infrastructure, domestic models, skills, adoption and safety, including GovAI hosting for agencies.[18] I have argued before that data centres and training capacity are not the same thing as national capability.[19] The Plan answers infrastructure and industry questions. It does not yet have to ask whether an Australian hospital remains the authority on its own meaning once models carry the daily work.

The World Bank's warning to countries with less room to spend is complementary: building frontier models is not a realistic near-term goal for most developing countries; adopt, adapt to local languages and problems, and only then decide where advancing the frontier is worth the money.[7] The clean choice is often presented as build frontier AI at home or accept permanent dependence. Alongside compute and adapted models, there is a further option the race framing leaves out: strengthen capabilities the country already has, make appropriate ones machine-discoverable, keep authoritative records and delegations outside model inference, insist on interoperability, support access in the languages people speak, and let local services participate in an AI-mediated economy without first donating their practice to a model they do not control. Labour law, education, competition policy and energy will still decide more of the distribution than any architecture document.

The fork that is already shipping

The absorbing trajectory is economically and technically credible. Training costs and global readiness diverge. Specialised models can hold a vendor's accumulated craft. Where local expertise is scarce, a model that carries knowledge can be a gain.

The reaching trajectory is also credible. Memory can sit with the application owner. Skills can sit in portable folders. Open protocols let models discover tools and agents discover agents. Language models can lower some navigation barriers, with a documented tendency to omit what a reviewer must catch. A country can host compute and still lose the ability to say what its institutions mean.

What has not been shown, and what I will not claim, is that the second trajectory delivers fairer economies, safer institutions or better care.

Do we build institutions around increasingly capable AI, or do we build an architecture through which increasingly capable AI can participate in increasingly capable institutions?

The second path assumes models become much more capable than they are now. The difference is where capability, knowledge, authority and judgement settle. AI could amplify and connect human and institutional capability that stays distributed. It could also become the place that capability gradually moves into. Both are being built. Leaving the choice unmade would itself be a decision, because the product decisions that express it are already shipping.

The argument does not depend on a product. In our own work we have been exploring one version of the second architecture, and building it in software through ifCEM, which is available today for guided evaluation and bounded organisational pilots. A bounded pilot can show whether the distinctions hold for a real piece of work. It cannot show that a society has chosen a trajectory, and it should not be read as if it had.

I do not know which path will dominate. But I am increasingly unsure that the model race is the only contest worth naming.

References

[1] OpenAI, "Dreaming: Better memory for a more helpful ChatGPT", 4 June 2026. https://openai.com/index/chatgpt-memory-dreaming/

[2] Anthropic, "Equipping agents for the real world with Agent Skills", engineering note, open-standard update 18 December 2025. https://www.anthropic.com/engineering/equipping-agents-for-the-real-world-with-agent-skills

[3] Anthropic, "Memory tool", Claude platform documentation. https://platform.claude.com/docs/en/agents-and-tools/tool-use/memory-tool

[4] Salesforce, "Announcing Koa: Salesforce's First CRM Reasoning Model, Built on NVIDIA Nemotron", 15 September 2026. https://www.salesforce.com/au/news/stories/koa-reasoning-model/ See also DataMPowered, The Knowledge We Buried Is Moving Again.

[5] Ben Cottier and Robi Rahman, Epoch AI, "Training compute costs are doubling every eight months for the largest AI models", data insight, 19 June 2024; underlying series updated through 2025. The page reports spending on large-scale training growing at about 2.4 times per year, with the most advanced models estimated in the hundreds of millions of dollars. https://epoch.ai/data-insights/cost-trend-large-scale

[6] International Monetary Fund, "The Global Impact of AI: Mind the Gap", Working Paper No. 2025/076, 11 April 2025. https://www.imf.org/en/publications/wp/issues/2025/04/11/the-global-impact-of-ai-mind-the-gap-566129

[7] World Bank, World Development Report 2026: The Promise of Artificial Intelligence, 2026, doi:10.1596/978-1-4648-2331-2. Overview booklet. https://www.worldbank.org/en/publication/wdr2026

[8] Richard S. Sutton, "The Bitter Lesson", 13 March 2019. http://www.incompleteideas.net/IncIdeas/BitterLesson.html

[9] DataMPowered, The Lesson After the Bitter Lesson.

[10] Lucas Maes, Quentin Le Lidec, Damien Scieur, Yann LeCun and Randall Balestriero, "LeWorldModel: Stable End-to-End Joint-Embedding Predictive Architecture from Pixels", arXiv:2603.19312, 2026. https://arxiv.org/abs/2603.19312

[11] Anthropic, "Donating the Model Context Protocol and establishing the Agentic AI Foundation", 9 December 2025. https://www.anthropic.com/news/donating-the-model-context-protocol-and-establishing-of-the-agentic-ai-foundation

[12] Google Open Source Blog, "A year of open collaboration: Celebrating the anniversary of A2A", April 2026. https://opensource.googleblog.com/2026/04/a-year-of-open-collaboration-celebrating-the-anniversary-of-a2a.html

[13] Karine Alves, Matheus Silva, Edney Santos, George Valença and Kellyton Brito, "A Generative AI approach for creating and validating simplified versions of government documents", Proceedings of the 26th Annual International Conference on Digital Government Research (dg.o 2025), published 20 May 2025. https://proceedings.open.tudelft.nl/DGO2025/article/download/968/1000/1373

[14] Regulation (EU) 2024/1689 of the European Parliament and of the Council, Article 14 (human oversight). https://eur-lex.europa.eu/eli/reg/2024/1689/oj

[15] Brooke Tanner, Cameron F. Kerry, Andrew Wyckoff, Nicoleta Kyosovska, Andrea Renda and Elham Tabassi, "Is AI Sovereignty Possible? Balancing Autonomy and Interdependence", Brookings Institution, February 2026. https://www.brookings.edu/wp-content/uploads/2026/02/20260217_AI_sovereignty_final.pdf

[16] Daron Acemoglu, David Autor and Simon Johnson, "Can We Have Pro-Worker AI? Choosing a path of machines in service of minds", MIT Stone Center on Inequality and Shaping the Future of Work, September 2023. https://shapingwork.mit.edu/research/can-we-have-pro-worker-ai/

[17] Daron Acemoglu, "The Simple Macroeconomics of AI", NBER Working Paper 32487, May 2024. https://doi.org/10.3386/w32487

[18] Australian Government, Department of Industry, Science and Resources, National AI Plan, 2 December 2025. https://www.industry.gov.au/publications/national-ai-plan

[19] DataMPowered, Australia Is Building AI Infrastructure. But What National Capability Are We Building?.

[20] Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel and Douwe Kiela, "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks", Advances in Neural Information Processing Systems 33 (NeurIPS 2020). https://arxiv.org/abs/2005.11401

[21] Zhengliang Shi, Yuhan Wang, Lingyong Yan, Pengjie Ren, Shuaiqiang Wang, Dawei Yin and Zhaochun Ren, "Retrieval Models Aren't Tool-Savvy: Benchmarking Tool Retrieval for Large Language Models", in Findings of the Association for Computational Linguistics: ACL 2025, 24497-24524. https://aclanthology.org/2025.findings-acl.1258/

[22] David Autor, "Applying AI to Rebuild Middle Class Jobs", NBER Working Paper 32140, February 2024. https://doi.org/10.3386/w32140

[23] Michael Luca, Timothy Wu, Sebastian Couvidat, Daniel Frank and William Seltzer, "Does Google Content Degrade Google Search? Experimental Evidence", Harvard Business School Working Paper 16-035, 2015. https://scholarship.law.columbia.edu/faculty_scholarship/1931

[24] DataMPowered, AI Didn't Create the Vulnerability. It Made the Institution Machine-Navigable.

---

Dakshan Pothuhera
Founder, DataMPowered®