Within a single procurement cycle, every private capital firm of any size will have access to frontier model capability on roughly the same terms. The same reasoning quality, the same context windows, the same enterprise controls, available to a twenty-person growth manager and a hundred-billion-dollar platform on the same day at a broadly similar price per seat.
A capability that everybody can buy simultaneously is a cost line, not an advantage. Yet a large share of the AI conversation inside funds is still a conversation about which model, as though the answer to that question were the thing that separates firms.
It is not. The separation comes from what the model is connected to, and that is the one part of the stack nobody else can sell you.
Model capability has already passed the threshold most fund work requires
The honest position on capability is that it matters enormously, and that it has stopped being the binding constraint for the majority of what a fund actually does. Reading a ninety-page questionnaire and reconciling it against three prior years of answers, summarising an investment committee memo, tracing a capital account balance back through its contributing transactions, drafting a first pass at a side letter comparison: these tasks sit comfortably inside what frontier models do well today. The marginal improvement from the next model release will be real, and it will be smaller than the difference between a system that can see your fund's records and one that cannot.
There is a genuine exception at the hard end, and we would not want to flatten it. Reasoning across thousands of pages of inconsistent documents without losing the thread, and doing so in a way where the chain of reasoning can be inspected rather than taken on faith, is still a real differentiator between models. This is the main reason we build on Anthropic's Claude: sustained quality over long, messy document sets, and reasoning you can actually read, which is what makes the output admissible in a setting where someone will eventually be asked to justify a decision. That choice matters. It is still not the thing that decides which firm gets more out of the technology.
What you own that no vendor can sell and no competitor can copy
Every established fund is sitting on a corpus that is genuinely proprietary, genuinely differentiated, and almost entirely unused:
Investment committee memos going back a decade or more, including the deals you passed on and, crucially, why.
Valuation judgements quarter by quarter, and the record of how each one aged.
The full correspondence history with every allocator who has ever looked at you.
Operating data and monthly reporting from every company you have ever held.
Side letters, precedents and the negotiating history behind each concession.
The archive of passed deals is the most interesting item on that list and the most neglected. It encodes the firm's taste: the things that make your partners uncomfortable, the patterns they have learned to distrust, the price discipline they hold to when a process is hot. That is the closest thing a fund has to an institutional edge written down. No vendor has it. No competitor can reconstruct it. And in most firms it is unreadable, because it lives in PDFs on a shared drive, in email, and in the memories of four people who will eventually retire.
The same is true, less romantically, of the operational record. Twelve years of quarterly reporting across forty portfolio companies is a benchmarking dataset specific to your sector focus and your hold period. It is not better than a commercial dataset in every respect, but it is about companies you actually owned, measured the way you actually measure them.
What makes this corpus valuable is also what has kept it unusable. It is unstructured, inconsistently filed and written by dozens of different hands over many years, with no common vocabulary and no reliable metadata. Those are precisely the conditions under which conventional data tooling fails and under which capable language models do their best work, which is why the asset has been sitting there for two decades without anyone being able to do much with it.
The mistake is treating this as a data warehouse programme
When firms accept the argument that their own context is the advantage, the next move is often wrong. Someone proposes a two-year programme to centralise everything into a new store, harmonised, cleaned and governed, after which the intelligence layer goes on top. These programmes are expensive, they slip, and by the time the foundations are laid the firm has learned nothing about what it wanted to ask.
The better pattern is to leave the records where they are and connect to them. The fund administrator's system stays the system of record for capital accounts. The document management system keeps the memos. The CRM keeps the relationships. What changes is that a model can reach each of them at the moment a question is asked, read what it needs under the permissions that already apply to the person asking, and answer with a traceable path back to source.
This is what the Model Context Protocol is for, and it is the reason we consider it more consequential for funds than any individual model capability. MCP gives a consistent way to expose a system of record to a model without copying its contents somewhere new. Practically, that turns a migration question into an integration question. It means permissions stay where they are already governed and already audited rather than being reimplemented in a second place. It means the corpus does not need to be finished before it is useful, because connecting one more system is an increment rather than a phase. And it means a firm can change its model, or use different models for different tasks, without rebuilding the connections underneath.
The case against proprietary data being much of an advantage
The strongest counter-argument deserves stating properly. A great deal of what sits in a fund's archive is stale, wrong, or about a market that no longer exists. A 2013 investment committee memo on a retail roll-up may encode reasoning that would now be actively misleading. Valuation marks reflect conventions that have since changed. The correspondence with an allocator whose entire investment team has turned over twice tells you about people who have left. Volume of history is not the same as relevance of history, and a firm that connects a model to a bad record will get confident, well-reasoned, badly-founded answers rather than obviously wrong ones, which is the more dangerous failure.
There is a second objection with real force. If proprietary data is the advantage, the firms with the most data should win, which points at the large platforms and against everything the mid-market has to hope for.
We think both objections are true and neither is fatal, for the same reason: the useful corpus is much smaller than the total corpus. The last three funds, the current portfolio, the live allocator base and the past two years of deal flow will carry most of the value. A firm of twenty people has enough of that to matter. What it needs is not more history but a clear-eyed decision about which history is still true, and the discipline to have a person answer that question rather than assume the system will. The data quality work is real, it is unglamorous, and it is where most of the effort genuinely goes.
Where this leaves the industry, and why we think it is a good direction
Across this series we have argued that a long list of familiar advantages is about to compress. Completeness of review, speed of close, responsiveness to investors, coverage of a market, the ability to read everything in a data room: all of these have been sources of differentiation largely because they were expensive, and all of them are becoming cheap. That can be read as a bleak forecast for an industry that has built a great deal of its self-image on effort.
We read it the other way. What survives the compression is the part of the job that was always the point. Judging a management team. Knowing which market will still be there in seven years. Having the relationship that gets you the call before the process starts. Holding price discipline when everyone in the room wants the deal. Deciding to walk away. None of that is being automated, and as the surrounding effort falls away it becomes a larger proportion of what partners actually spend their time on.
The last two decades of technology spending in private capital went into recording what happened: systems that captured the transaction, the document, the meeting, the mark. The next decade is about making that record usable by the people who need it, including the investors who have been asking politely for years and receiving folders. A firm's own accumulated judgement, finally legible to itself, is a better foundation for an industry than any model licence. It happens to be an asset that smaller, sharper firms can build as well as large ones, which should make the next ten years considerably more interesting than the last ten.
