Almost every senior dealmaker who spoke at this year's Milken Global said the same thing in slightly different words: AI has turned exit modeling into a blindfolded dart throw. They are not wrong. They are also not done paying for it.

Private equity is a long-term asset class. The standard hold is three to four years on the short end, six to seven on the long. For context, it has been less than four years since ChatGPT shipped. In that window, three different model generations have rewritten what enterprise software is, what a knowledge worker does, and what a SaaS company is worth. Any general partner claiming high confidence in the operating environment three and a half years from now is either lying or self-deluded. The veteran quoted by Axios got the metaphor exactly right. The dartboard moved. Then it caught fire. Then someone replaced it with a different game.

The signal has been hiding in the SaaS multiples for eighteen months. Public-market EV-to-sales for software has compressed to its lowest level in five years. Private credit, which underwrote much of the last decade's software buyout boom, started seeing redemption requests the week Anthropic shipped its first wave of agentic plug-ins. The chain of consequence is not subtle. Models eat features. Features were the moat. The moat was the multiple. The multiple was the LP return.

A buyout fund holding $40 billion of enterprise software, underwritten at 2022 multiples, is not running a portfolio. It is running an inventory of melting assets, in a market that is suddenly questioning whether the inventory was ever worth the price.

The "AI-resistant" sectors are not resistant

What is genuinely new in 2026 is that the modeling problem has crossed sector lines. A year ago, the consensus was that AI risk lived in software. Today the same Milken hallway conversations are happening about healthcare services, insurance, financial services, distribution, BPO, mid-market industrials, even rollups in fragmented physical-services categories. The question is no longer "is this business AI-exposed?" The question is "exposed in which direction, and how fast?"

That distinction matters because the deal model traditionally encodes the answer as a single number — the exit multiple. And the exit multiple is now a probability distribution with a tail wide enough to swallow the entire IRR. A nursing home rollup that looked like a 4x money business in 2024 might be a 6x in 2027 if AI-augmented care management compresses operating costs faster than reimbursement rates re-base. Or it might be a 1.8x if a competitor figures it out first and re-prices the regional market. The deal team that underwrote the median outcome is now responsible for a fund position whose variance has tripled while the cost of capital has not moved.

Three pains, one root cause

Pain 1: Diligence is now a moving target.

The traditional hundred-day diligence sprint assumed the technology landscape during diligence would resemble the technology landscape at exit. That assumption broke in 2024 and stopped being repaired. By the time the QofE is locked, a new model release has changed the unit economics of half the management team's assumptions. The Petras and Blueflames of the world are speeding up the analyst layer. They are not solving the deeper problem, which is that the underlying business model the analyst is analyzing is itself being rewritten in real time.

Pain 2: The "what good looks like" benchmark has no benchmark.

Operating partners used to compare a portco's gross margin against the relevant peer set and prescribe a hundred-day plan to close the gap. Today, the benchmark itself is contested. Is a 78% gross margin SaaS company a high performer or a sitting duck whose customer support function is about to be 90% automated by the buyer's in-house deployment of a foundation model? Both readings are defensible. Neither is fundable in a credit committee that wants a single number.

Pain 3: The exit multiple cannot be projected linearly anymore.

The 2018-to-2023 playbook — buy at 12x, fix at 11x, sell at 14x — assumed multiple stability across the hold. AI breaks that assumption from both ends. AI-native peers trade at 14x to 40x EBITDA today, while AI-displaced peers are quietly being re-rated toward 4x to 6x. The variance between "we executed the playbook and exited well" and "we executed the playbook and the category got disintermediated" is now larger than the variance between "we executed the playbook" and "we did nothing." That is a structural break. The model the industry uses to think about returns was not designed for a world in which the multiple is a function of the technology stack, not the operating performance.

The most uncomfortable line in the Axios piece

Buried inside the Milken reporting was the observation that most GPs are now doing deals against a multiple they do not believe in. The dry powder math forces it. LPs continue to commit despite the DPI drought. Funds need to deploy. Deals get done. But the underwriting memo and the IC conversation no longer match. The memo says 14x at exit. The IC conversation says "we hope." The gap between the memo and the conversation is the new alpha frontier in the asset class — and right now, the alpha is negative.

The funds that figure out how to close the gap between the memo and the IC conversation will outperform. The funds that don't will spend the next three years explaining unrealized markdowns to LPs who have a finite tolerance for narrative.

Why the OpenAI and Anthropic JVs do not solve this

In May, the two largest AI labs each announced multibillion-dollar joint ventures with the largest PE firms. Anthropic raised $1.5 billion alongside Blackstone, Hellman & Friedman, and Goldman Sachs. OpenAI raised roughly $4 billion of external capital alongside TPG, Brookfield, Bain, and Advent into a vehicle called The Deployment Company. Both deals are simultaneously offensive and defensive: offensive because the JVs are early bets on AI-native McKinseys; defensive because the labs have, in the same breath, been quietly decimating the value of the same firms' software portfolios.

These deals are real, and they are clever, but they do not address the modeling problem. They address the distribution problem the labs face when selling into mid-market. A consulting layer between a PE portfolio and a frontier model accelerates rollouts. It does not change the fundamental question every IC is now wrestling with: when we underwrite a deal at a 12x exit, what infrastructure has the portfolio company actually built between now and exit that justifies that multiple? Or are we again purchasing the right to find out?

What this series is going to do about it

I have spent twenty-five years building hyperscale platforms at Amazon, Microsoft, Meta, Visible (Verizon), and Property Finder. I now run BeHuman Capital and DouJou. I have watched the modeling problem from inside the operating company that the diligence team was modeling — and from inside the boardroom presenting the answer to the investor who paid for the answer. The two views rarely match. They need to start matching.

Over the next two installments, I will walk through the specific, calculable, CFO-verifiable cost that every one of your portfolio companies is paying right now while running AI initiatives that will not reach production. I call it the Hidden Year Invoice. Every portco has one. Most boards have never seen it. It is the single largest unrecognized line item in the modern PE operating model, and it is sitting between you and the multiple expansion you wrote into your underwriting memo.

Then we will get to what to do about it. Not at the model layer — the model layer is taken care of by the labs and the JVs. At the infrastructure layer. At the governance layer. At the layer where every portfolio company in your fund is currently spending two to six million dollars a year reinventing the same undifferentiated AI plumbing that another portco in the same fund just finished building. That number is the next markdown if you don't address it. It is the next multiple expansion if you do.

AI-proofing the portfolio is not a thesis update. It is an underwriting reset. The funds that get there first will reset the asset class. The funds that don't will explain why for the rest of the decade.

Part 2 lands next week. Bring the CFO of your largest portfolio company. We are going to do some math.

◆ ◆ ◆

— Himanshu Niranjani
Founder, BeHuman Capital · Architect, DouJou