CFO Daily · Wednesday, September 30, 2026

Koko: PwC just told private equity that AI is repricing portfolio value right now, mid-cycle, not at year end. That's the freshest signal on my screen this morning.

Sam: Mid-cycle repricing on illiquid marks. That's a controller's nightmare and a valuation committee's Monday.

Koko: Same morning, Oracle ships Fusion Claw, a runtime that swaps per-step AI inference calls for deterministic policies to cap agent compute costs.

Sam: So one story says AI is repricing your assets upward, the other is basically an admission that AI spend was running out of control.

Koko: And OpenAI's ARR is nearing seventy billion, B2B revenue up over a hundred percent in a quarter, per Axios. Enterprise AI spend is scaling faster than budget cycles assume.

Sam: Here's the through-line for finance: cost architecture just stopped being a vendor-invoice line. It's a governance decision, and it's happening on a faster clock than your planning calendar.

Koko: Walk through the PwC piece with me. Their midyear PE outlook says AI is actively repricing portfolio value, and that's adding pressure on LPs who want liquidity.

Sam: Right, and deal count already fell sharply this year while average deal size surged. Sponsors are concentrating capital on fewer, higher-conviction bets.

Koko: So if you're a CFO sitting on a portfolio company, or you're a fund controller marking positions, the question isn't whether AI matters to valuation anymore. It's whether your marks already reflect it.

Sam: That's the part I'd push on. DPI has overtaken IRR as the real LP gating metric now, per the report. If your realized returns story is weak, fundraising dollars aren't flowing your way regardless of what your unrealized marks say.

Koko: Which makes the AI repricing question almost academic for some GPs. Doesn't matter how AI-savvy your portfolio company looks if you can't distribute cash.

Sam: I'd push back a little. It's not academic for the audit committee. If your valuation policy hasn't explicitly addressed how AI capability gets underwritten into a mark, that's a disclosure gap waiting to get found in due diligence.

Koko: Fair. So two separate jobs land on the same desk this week: pressure-test the marks for AI-driven repricing, and separately, get real about which GPs are going to miss first-close targets as LP patience runs out.

Sam: Both are finance function work now, not deal-team work. That's the shift worth flagging to your board.

Koko: Oracle's Fusion Claw is the concrete answer to a problem finance teams have only budgeted for in the abstract: agent compute cost creep.

Sam: It replaces repeated per-step model calls with deterministic policies. CIO Magazine's framing is important here — lower consumption doesn't automatically mean lower bills.

Koko: Say more, because that sounds like a contradiction. Isn't fewer inference calls the whole point?

Sam: It is, but someone has to encode the approval rules and verify the outcomes. That's governance overhead shifting onto your team, not disappearing. And Oracle's AI unit pricing has expiring pooled credits with only a thirty-day change notice.

Koko: So this is a cost shift dressed as a cost saving.

Sam: That's my read. I want per-outcome pricing commitments in writing before anyone signs off on scaling this.

Koko: I'll take the other side for a second. Even with governance overhead, an auditable decision record on every agent action is something most finance functions don't have today. That's real control value, not just cost noise.

Sam: Sure, but don't let the audit trail distract from the pricing risk. Pilot it on one high-volume, measurable process first — finance reconciliation is the obvious candidate — and measure cost per completed outcome, not cost per token.

Koko: That's the test every CFO should be running on every agentic vendor pitch right now, not just Oracle's.

Koko: OpenAI cut its agentic model price to one-fifth of its prior flagship. Any ROI model you built on last quarter's token pricing is already stale.

Sam: Sloan Review flags in-house teams pulling billable work back from outside providers using generative AI. Watch your own outside-advisory and audit-support spend for the same squeeze.

Koko: And AMD just paid eight point two billion for World Labs, Fei-Fei Li's spatial AI shop. Not a finance story today, but a capex signal for anyone tracking where compute demand goes next.

Koko: So here's the call. Valuation committees and fund controllers: don't wait for year-end to ask whether your marks already price in AI repricing. PwC just put that question on the table mid-cycle.

Sam: And for anyone piloting Fusion Claw or anything like it: get per-outcome pricing in writing, and run it on a bounded process before you commit pooled AI units. Controllers, that's your test case.

Koko: FP&A, rerun your agent pilot ROI with current token pricing, not last quarter's. OpenAI's price cuts make old assumptions worthless fast.

Sam: One more for the audit committee: ask whether your valuation policy explicitly addresses how AI capability gets underwritten into a portfolio mark. If it doesn't, that's a gap, not a footnote.

Koko: That's the show. Full brief's at koko knows dot A I.

Sam: Question worth watching: when Oracle's pooled AI credits expire under that thirty-day notice, who actually eats the true-up?