TMT Brief · Friday, September 18, 2026

Koko: OpenAI is reportedly weighing a new funding round at a valuation of one point five trillion dollars. Days after Amodei and other CEOs stood up and asked the industry to slow down.

Max: So capital markets heard that speech and said, interesting, now here's another half a trillion dollars of valuation.

Koko: Fast around the rest. Salesforce and Google Cloud are deepening their alliance, migrating US customers onto Hyperforce and Gemini Enterprise starting Q4.

Max: That one's not new spend, that's consolidation. Worth noting the difference before anyone puts it in the AI capex bucket.

Koko: SK Hynix denied talking to Intel about a US chip tie-up, but didn't deny the Ohio fab leasing or a cloud-backed joint venture.

Max: A denial with three exits left open. Classic.

Koko: Here's why this week matters for the board deck: the gap between what AI is valued at and what it's proven to deliver just got wider on both ends, the frontier and the back office.

Max: Valuation's compounding. Deployment isn't. That's the whole tension today.

Koko: Let's actually sit with the OpenAI number, because one point five trillion is not a rounding error on the last valuation.

Max: It's roughly the size of the debate happening above it. Amodei's out there calling for a slowdown. Trump and Huang are in open safety disputes. And the market's pricing none of it in.

Koko: Which is exactly the disagreement worth having. Is that market discipline, or is that the market just not having a mechanism to price governance risk at all?

Max: I'd say the second. Nobody's shorting a foundation model over a security incident. There's no instrument for that yet.

Koko: And that's the CFO's actual exposure. If you're deepening dependence on one AI partner while that partner's political and financial volatility is rising, you own that risk whether you priced it or not.

Max: So the ask isn't 'watch the valuation.' It's get procurement to build a real multi-vendor exposure map before the next renewal, not a roadmap slide with a logo on it.

Koko: [pause] That's a CIO and general counsel conversation as much as a finance one.

Max: Correct, and it should happen before the next capital commitment, not after the incident.

Koko: New Auditoria.AI survey: sixty six point five percent of finance teams are raising AI investment. Only twenty one percent report measurable results.

Max: Say that ratio again slowly. Two thirds increasing spend, one fifth showing it worked.

Koko: And the same report says AI is speeding up individual tasks but not fixing the cash cycle. Faster reconciliation, same DSO.

Max: That's the token-legibility problem showing up inside the CFO's own house. You automated the busywork and left the actual bottleneck untouched.

Koko: Here's where we'd differ, though. I'd call this an adoption curve problem — twenty one percent this early isn't damning, it's just early.

Max: I'd call it a measurement problem dressed up as an adoption problem. If you can't show results after this much spend, the issue is you're not measuring the right thing, not that it's too soon.

Koko: Either way, the fix is the same. Demand the same ROI evidence standard for the next AI budget line that you'd demand for any other capital project.

Max: No exemption because it says AI on the label.

Koko: Quieter story, but it matters for anyone modeling AI margins. Berkeley research, cited by Tunguz: better agent harnesses cut inference cost seventy one percent at equal accuracy.

Max: Not a better model. Same model, better scaffolding around it.

Koko: Which relocates the margin lever. It's not sitting inside the model layer anymore, it's sitting above it, in orchestration.

Max: That's actually good news for anyone who isn't a frontier lab. You don't need to win the model race to win the cost race.

Koko: It also lines up with Nokia and Microsoft pairing telco data products with Fabric for agentic network operations this week.

Max: Vendors racing to monetize the operator's own network data before the operator builds the harness themselves.

Koko: So the CTO conversation this quarter isn't 'which model do we license.' It's 'who controls the orchestration layer, because that's where the seventy one percent lives.'

Max: Build that in-house and you keep the margin. Buy it from the vendor and you just paid someone else to find it for you.

Koko: Zooming out. Memory and chip supply deals keep forming in denials and side-doors — SK Hynix this week, same shape as the rest of the summer.

Max: Nobody announces the deal. Everybody denies the deal and confirms three adjacent things.

Koko: Infrastructure alliances like Salesforce-Google Cloud are consolidation dressed as partnership, not new capex.

Max: Watch which line item that lands in. It's not growth spend.

Koko: And the ERP layer keeps splitting in two: system of record stays put, the interaction layer moves to agents on top.

Max: Clean core, composable edge. Boring name, correct architecture.

Koko: So what actually goes on the board deck versus what needs a call this month.

Max: Board deck, next quarter: the vendor exposure map. Get procurement and legal to quantify concentration risk before the next AI capital commitment, not after.

Koko: This month: audit your own finance function's AI spend against the twenty one percent number. If you can't show measurable results, stop scaling that line item until you can.

Max: And put orchestration on the CTO's agenda now, not next cycle. Seventy one percent cost reduction from harness design alone is too big to leave sitting in someone else's product.

Koko: What would change our answer on OpenAI's valuation? A second security incident, or an actual regulatory move that prices the safety risk.

Max: [chuckles] Until then, the market's decided the rhetoric is just rhetoric.

Koko: Full breakdown of all of this at kokoknows dot A I.

Max: Open question for the coming weeks: does anyone's finance ROI number actually catch up to the spend, or does the gap just become the new normal.