The Only Thing You Can Own

Executive Summary

In August 2026, Thomson Reuters finished a project that had been running quietly for two years: its own large language model, named Thomson, built on open weights from Alibaba's Qwen family, trained on the company's own legal corpus by the company's own domain experts, at a reported cost of around $40 million (Business Insider, 2026). The company's chief technology officer, Joel Hron, explained the reasoning with an analogy worth sitting with:

Renting a house, you still have a roof over your head, and somebody's taking care of it, and it's great. But you're not building any equity that compounds into something valuable for you long term.

The analogy is exactly right. The conclusion most readers drew from it is exactly wrong.

The obvious reading is that you should own your model. That reading fails for almost everyone, including, on close inspection, Thomson Reuters itself, which kept using Anthropic's models after Thomson shipped and continues to evaluate outside providers for the work its own model does not do best (Business Insider, 2026). A $40 million model did not end the renting. It never could, because the model is not the part that compounds.

This paper argues something more specific and, we think, more durable. In an AI system there is exactly one layer an individual or a small firm can actually own, and it is not the weights, the compute, the interface, the roadmap, or the price. It is the memory: the accumulated record of what you did, what you decided, what you meant, who you know, what you corrected, and what turned out to be true. Everything else in the stack is rented, on terms you do not set and cannot see. Memory is the only equity available at your scale, the only layer that appreciates instead of depreciating, and the only asset in the entire arrangement that is unique to you by construction.

Most people are renting the house and storing everything they own in the landlord's basement, on a lease that says the basement is his.

1. What $40 Million Actually Bought

1.1 The Decision

Thomson Reuters is not a company that stumbled into this. It owns Westlaw and Practical Law. Its business is a corpus: decades of primary law, editorial annotation, and expert commentary, maintained by thousands of attorney-editors. When it built CoCounsel on frontier models, it was, in its CTO's framing, renting the roof while owning the furniture.

Two structural problems pushed it to build. The first is economic: high-volume repetitive work such as document review runs a metered bill forever, and past sufficient volume a smaller specialized model you own is simply cheaper per unit than a frontier model you rent. The second is strategic: fine-tuning someone else's frontier model leaves you inside their roadmap, their pricing, their deprecation calendar, and their judgment about which capabilities matter next.

Neither problem is unusual. What is unusual is being large enough that $40 million and two years is a rational answer to them.

1.2 The Part Everyone Skipped

The detail that matters most in that coverage is the least quoted. After shipping Thomson, Thomson Reuters kept Anthropic. Its executives describe frontier labs as vendors and say they continue to evaluate outside models for the use cases where those models win.

So the company that spent $40 million to stop renting still rents. What changed is not the vendor relationship. What changed is which part of the system the company treats as strategic and which part it treats as a utility. The frontier model became a supplier. The corpus, the expert judgment, and the model trained on them became the balance sheet.

That is a far more interesting lesson than build your own model, and unlike that one, it survives translation to any scale.

1.3 The Equity Was Never the Weights

Read the equity claim carefully and notice what is actually compounding. It is not the parameters. A set of weights is a snapshot: a lossy compression of a corpus, frozen on a date, obsolete on a schedule. Weights do not compound. They decay.

What compounds at Thomson Reuters is the loop underneath the weights. Every product update sends work through attorney-editors. Every expert review produces a labeled judgment. Every labeled judgment becomes training data for the next version. The corpus is not static; it grows with each cycle of expert attention. The model is just the current rendering of it, the way a printed edition is the current rendering of a manuscript that keeps being revised.

Thomson Reuters did not buy equity in a model. It bought the right to keep converting its own accumulated judgment into capability, on its own schedule, without asking permission. The weights are the receipt. The memory is the asset.

2. Why You Cannot Own the Model

The obvious reading of the analogy fails at your scale for three reasons, and it is worth being precise about all three, because the failure is not merely financial.

2.1 The Price of Entry Is the Least of It

Forty million dollars over two years is the visible barrier, and for the individual professional, the ten-person firm, and most mid-sized businesses it settles the question immediately. But suppose that price fell by two orders of magnitude tomorrow. The other two reasons would still hold, and they are the ones that matter.

2.2 Weights Are the Fastest-Depreciating Asset in Technology

A flagship model has a useful commercial life measured in months, not years. The frontier moves, price per token falls, context windows widen, and the model that felt remarkable in the spring feels merely adequate by the winter. Stanford's AI Index has documented both the compression of the performance gap between leading models and the collapse in inference cost for a given capability level (Stanford HAI, 2025). Epoch AI's tracking of training compute shows the same pressure from the other direction: the frontier is a treadmill, and standing still on it is identical to moving backward (Epoch AI, 2025).

Nothing else in your professional life depreciates like this. A building lasts decades. A brand lasts generations. A trained colleague gets better with time. Model weights are closer to a phone: excellent, expensive, and quietly obsolete before the warranty ends. Owning a depreciating asset is not equity. It is inventory.

Illustrative: share of original usefulness retained after twenty-four months. Weights get superseded, chat archives go stale and stay unsearchable, prompts survive a provider switch but do not grow on their own, and the knowledge graph is the only line that is worth more at the end of the period than at the start.

2.3 Models Converge; Memory Diverges

This is the deepest of the three reasons and the one most often missed.

Frontier models are trained on overlapping data, by overlapping talent, toward overlapping benchmarks. They are becoming more alike, not less. That is why the leaderboard reshuffles every few months without changing anyone's daily experience much, and it is why the honest answer to which model is best is usually whichever one is cheapest this quarter for this particular task.

An asset that everyone can buy an equivalent of is not equity. It is a commodity input, and the correct strategy for a commodity input is to rent it competitively and never get attached.

Your memory runs in the opposite direction. No two professionals accumulate the same one. Your record of clients, decisions, corrections, drafts, dead ends, obligations, and half-formed opinions is unique by construction, and it gets more unique every week. Uniqueness that increases with time is the actual definition of a compounding asset. It is also the only thing in the arrangement that no amount of money can buy an equivalent of, because the thing that made it was your own attention, spent once and unrepeatable.

3. The Ownership Test

Ownership is a word that survives a great deal of abuse in software marketing. So here is a test with four questions, all of which have to answer yes.

1. Possession. Is it on hardware you control right now, without a network call?
2. Portability. Could you take it, whole and structured, to a competitor this afternoon?
3. Survivorship. If the vendor died tonight, would you still have it tomorrow?
4. Erasure. Can you make it genuinely gone, everywhere, and verify that it is?

Apply the test honestly to each layer of a working AI setup.

Four of the six layers fail every question and always will. That is not a scandal; it is what renting is, and renting a commodity is the right call. The fifth is a coin flip decided by whether your prompts live in a text file or inside somebody's web app.

The sixth layer is the whole argument. Memory is the only place where the answer is set by a design decision rather than by the economics of frontier research. It can be genuinely owned. It usually is not.

4. Memory Is the Only Ownable Layer

4.1 It Is the One Input Intelligence Cannot Regenerate

Here is the asymmetry that makes memory different in kind from everything else in the stack.

Ask any competent model to draft a contract clause, summarize a filing, or write an apology, and it will produce something good. Ask it what you promised that client in March, which of your two positions on the pricing question you actually settled on, why you walked away from the second vendor, or what your co-founder said the week before she changed her mind, and it has nothing. Not a weaker answer. Nothing.

Capability is regenerable and increasingly close to free. Context is neither. A model can reproduce any output it has the inputs for; it cannot reproduce inputs it never had. As raw intelligence approaches commodity pricing, the entire value of an AI system migrates to the only scarce input in it, which is the accumulated particulars of your life and your work.

That is why memory is the only thing worth owning. It is the only thing in the arrangement that is actually scarce.

4.2 It Appreciates on the Same Curve Weights Depreciate On

A knowledge graph gains value in three ways at once, and none of them require you to do anything except keep working.

It gains volume: more entities, more documents, more history. It gains connection, which matters more, because the value of a graph rises with its edges rather than its nodes, and edges grow superlinearly with nodes. And it gains verification: a fact that has survived being cross-checked against later evidence is worth more than a fact that has not, and time is what does the checking.

Meanwhile the model reading that graph gets better and cheaper every year at somebody else's expense. This is the arrangement you actually want. Rent the part that improves for free; own the part that improves only because of you. Our companion paper The Appreciating Brain works through the growth curve in detail, and The Memory Trap covers what happens when the graph is held by someone else.

4.3 It Is What Thomson Reuters Actually Bought

Notice that this reading resolves the apparent contradiction in the story. Thomson Reuters spent $40 million and still rents from Anthropic, and both facts are consistent, because the $40 million never bought independence from frontier models. It bought a durable mechanism for converting owned memory into capability.

The individual version of that mechanism does not cost $40 million. It costs a decision about where memory lives. The mechanism itself is identical: accumulate judgment, keep it in a form you hold, and point whatever intelligence is currently best and cheapest at it. Thomson Reuters had to train a model because its memory is a corpus far too large and too specialized to fit in any context window. Yours fits. That is not a smaller version of their advantage. At your scale it is a larger one.

5. What You Own Right Now

Run the four-question test against what you have today. For most professionals the result is uncomfortable.

Chat history is not memory. It is a transcript archive: linear, unstructured, searchable by keyword at best, and stored on infrastructure you have no claim to. It fails possession and survivorship outright. It technically passes portability, in the sense that most providers offer an export, but what comes back is a pile of conversations rather than a structure, and pointing a different provider at that pile restores nothing. A transcript is a receipt of thinking, not the thinking.

Provider-side memory features are the landlord's basement. The memory settings shipped by the major assistants are genuinely useful and genuinely not yours. They live in the provider's system, in the provider's schema, under the provider's retention policy and product decisions, and they do not travel. They also stop at the provider's walls: memory formed in one assistant is invisible to every other tool you use, including the next assistant you will use.

Your documents are records, not memory. Files in Drive and threads in Gmail are the raw material. They are not connected, not resolved into entities, and not carrying the interpretive layer that makes memory useful: that these two names are the same person, that this decision reversed that one, that this commitment came due last week.

Which leaves most people in a specific and avoidable position: paying for intelligence monthly, generating unique context daily, and accruing equity in none of it. The lease has been running for three years and the basement is full.

6. The Architecture of Owned Memory

If memory is the only ownable layer, the architecture that decides whether you own it deserves stating plainly. Four properties, each mapping to one question in the test.

It lives on your disk. Kent's knowledge graph is a local database on your own machine, encrypted at rest. Possession is not a policy commitment from a vendor; it is a file path. If our servers vanished tonight, your graph would be exactly where it was this morning.

It is structured, not archived. Entities, relationships, provenance, and time, rather than a pile of transcripts. Every skill execution, dropped file, connector query, and correction feeds one graph, and the graph resolves them into people, organizations, commitments, and claims with sources attached. That structure is what makes memory portable in any meaningful sense: what you carry forward is a model of your world, not a log of your typing.

It is provider-agnostic by design. Six providers, Anthropic, OpenAI, Gemini, HuggingFace, DeepSeek, and local models through Ollama, read the same graph. Switching model is a dropdown, not a migration, because the model was never where the value lived. This is the individual's version of what Thomson Reuters built: intelligence as a supplier, memory as the balance sheet. In private mode the supplier is your own hardware and nothing leaves the machine at all.

It can be genuinely forgotten. Ownership without erasure is custody, not ownership. Kent's forgetting is a durable tombstone applied across every data surface rather than a hidden row, and it survives restarts, syncs, and reindexes. The right to delete is the part of ownership that most systems quietly omit, and it is the part regulated professionals are asked about first.

Our papers The Second Brain That Thinks and The Connector Moat cover graph construction and connector coverage in depth. The point here is narrower: these are not features. They are the specific properties that determine whether the word own is doing any work in a sentence.

7. What Ownership Costs

An argument for ownership that lists only benefits is a sales pitch, so here is the other column.

Ownership is a job. The analogy cuts both ways: renting really does mean somebody else fixes the roof. Owning your memory means owning backups, disk encryption, device security, and the consequences of losing a laptop. Kent automates most of that and encrypts at rest by default, but the obligation does not disappear. It moves to you. That is what ownership is.

Owned memory is a liability as well as an asset. A graph that remembers everything about your clients is exactly as sensitive as it sounds. The security burden scales with the value, and anyone claiming otherwise is selling something. Our paper The Leak Was Never Storage is the honest accounting of that surface.

Memory does not substitute for intelligence. A perfect graph paired with a weak model is a beautifully organized library with no reader. Owning memory is what makes rented intelligence worth renting; it is not a replacement for it, and any pitch implying otherwise deserves suspicion. It is precisely why we ship six providers instead of pretending a small local model is a frontier model.

Wrong memories compound too. Accumulation is not neutral. An error absorbed into a graph propagates into everything downstream of it, which is why provenance, verification, and correction are load-bearing rather than decorative, and why we have written more about claim verification than about almost anything else we build.

Nobody should own everything. The consolidation argument stops where the specialists begin, and the ownership argument stops in the same place. Your systems of record can stay exactly where they are. What has to be yours is the layer that interprets them.

Conclusion

The Thomson Reuters analogy deserves to outlive the news cycle it arrived in, because it names the thing correctly. Renting really is fine. The roof holds, somebody else maintains it, and for a commodity that keeps halving in price while doubling in quality, renting is not a compromise. It is the correct financial position.

What is not fine is renting the house and keeping everything you own in the landlord's basement, on a lease that says the basement is his, in boxes only he can open, for a term he can end.

Thomson Reuters spent $40 million to move its boxes. That is what the money bought: not freedom from landlords, which nobody has, but a place of its own to keep the one thing that was ever really its own. The frontier lab is still on the invoice, and always will be, and that was never the problem.

You will not spend $40 million, and you do not need to. At your scale the asset is small enough to hold in your hands. Your memory fits on your own disk, and every model on earth will read it for you at falling prices. The only decision is whether it sits somewhere you can hold, move, outlive, and erase.

Intelligence is becoming free. Memory never will be, because yours costs a life to make and cannot be bought secondhand. Rent the roof. Own the boxes.

References

1. Business Insider. (2026). "Thomson Reuters Builds Its Own AI Model to Rely Less on Anthropic." August 2026.

1. Stanford Institute for Human-Centered Artificial Intelligence. (2025). AI Index Report 2025: Technical Performance and Cost Trends.

1. Epoch AI. (2025). Trends in Machine Learning Training Compute and Inference Cost.

1. Thomson Reuters. (2026). CoCounsel Product Announcements and Model Documentation.

1. European Parliament and Council. (2024). Regulation (EU) 2024/1689, Artificial Intelligence Act.

1. Kent. (2026). Internal Usage Analytics: Knowledge Graph Growth, Provider Routing, and Retention.

Kent Research, August 2026. Kent runs 13 built-in skills and unlimited custom skills against six AI providers, cloud or fully local, and stores its knowledge graph as an encrypted local database on your own machine. Connectors cover Gmail, Google Drive, Google Calendar, Notion, PostgreSQL, MySQL, SQLite, MongoDB, REST APIs, and MCP servers. External figures are as reported by the cited sources; illustrative charts are labeled as such. Nothing here constitutes financial or legal advice. mykent.app