Markets have a way of settling arguments that op-eds can't. Thursday they settled one: the AI trade is no longer priced for a world where nothing can go wrong. Today's three items are really one question asked three ways — where does the value from all this spending actually land, and who gets to keep it?
The AI trade gets a haircut
The selloff was historic. Technology shares suffered their biggest one-day rout since the Covid pandemic, wiping out about $2.3 trillion in US market value on Thursday — the most on record. The Nasdaq Composite fell 5.52% into correction territory, its deepest one-day drop in 19 months, after Democrats' midterm sweep stoked fears that tighter AI regulation and a slower data-center buildout were coming. Nvidia bore the brunt among the megacaps, closing down 9.26% at $181.60 — its biggest one-day fall since the DeepSeek scare of January 2025. By Friday morning the dip-buyers were back: Nasdaq 100 futures climbed 1.4% and chip names like AMD, Broadcom and Micron bounced between 4.6% and 8.3%.
The operator's take: one strategist called it reactionary, not a fundamental reset, and that's probably right — but it's a useful warning shot. The market just told you it will no longer underwrite infinite capex on faith, and the trigger was politics, not technology. If your 2027 plan assumes cheap, unlimited GPU capacity priced the way it was last quarter, build a version that survives a 20% repricing and a permitting slowdown. The buildout's cost of capital just went up; yours should too.
Salesforce buys what it can't build fast enough
While public markets punished the infrastructure layer, the application layer kept consolidating. Salesforce signed a definitive agreement to acquire Cresta, a voice-first operating system for contact centers whose agents resolve complex calls end to end — reportedly for about $2.7 billion, and its third acquisition in two weeks. That followed its deals for defense-logistics startup Rune Technologies and a roughly $5.66 billion purchase of marketing firm Braze, part of a spree that now runs to six deals in under five months.
The operator's take: when the incumbent CRM is spending billions to bolt on voice agents it could theoretically have built, that's your clearest read on build-vs-buy in the agent era — the capability is moving faster than any internal roadmap, and even Salesforce has decided it's cheaper to buy time than to build it. For you, the lesson isn't "go shopping." It's that the glue between your systems of record and these new agent layers is where the lock-in — and the leverage — is forming. Negotiate for data portability now, before the vendor you rely on owns both the record and the agent acting on it.
A sharper blade, and a higher price
Security got a real upgrade and a real invoice on the same day. OpenAI launched GPT-6 Cyber, which solves 48.5% of tasks on the ExploitGym benchmark, up from 32.1% for the prior GPT-5.6 Cyber. Access is gated — the model is restricted through a program called Daybreak Red, with identity checks and organizational security requirements — and it isn't cheap: GPT-6 Cyber costs roughly 60% more for input and about 33% more for output than its predecessor.
The operator's take: read the benchmark carefully before you panic or celebrate. Those tasks measure exploiting a specified vulnerability, not discovering unknown flaws in your software — so this is a better blade, not a magic one. The honest move is to treat capability like this as dual-use and plan for both sides: use it to triage and patch your own code faster, and assume your adversaries get the same lift on the offense. The gating and the price tag are the tell that the vendors know it too. Patch cadence, not detection, is still your bottleneck.
Also on my radar
- Agents can shop, but they can't be trusted to buy yet. Visa's new Agentic Commerce Benchmark ran 22 models from 12 labs through 15 shopping tasks; they failed 42.8% of runs, and in 54% of those failures the agent bought something anyway — the angle for anyone deploying agents against money: decide now who eats the cost of a confidently wrong purchase.
- China keeps commoditizing the model layer. Baidu released ERNIE 5.5 Flash Base, an openly downloadable 280-billion-parameter mixture-of-experts model with just 14 billion active per token — a reminder that capable open weights keep arriving, though its license carves out the EU, UK and South Korea, so read the fine print before you build on it.
- The contact-center land grab isn't just Salesforce. Five9 agreed to acquire Level AI for about $325 million in cash — when two vendors race to buy the same category in the same week, the "voice agent" market is consolidating fast enough that waiting to pick a platform is itself a bet.
The throughline is a repricing, not a retreat. Public markets just marked down the people building the capacity, incumbents are paying premiums to buy the capability, and the model makers are charging more for the sharpest tools while open weights keep eroding the floor beneath them. For an operator, that sorting is the opportunity: value is migrating from raw compute toward the workflows and data you control, so spend there, contract for portability, and stop paying frontier prices for anything you can get commodity-cheap. The hype got repriced; make sure your plan did too. That's the Signal for today.
Paul Sapio is the CIO of Mikhail Education and a full-stack AI engineer. Open to contract work in security, networking, AI, and SaaS development — reach out.