Sunday, and the frontier chatter is loud as ever — but look past the demos and today's stories are all about the invoice. The AI buildout is settling into megawatts and long-term leases, the companies reselling these models are quietly re-shopping their supplier bills, and the boring security appliance in the corner is still where the real intrusions start. Three reminders that someone, eventually, pays for all of this. Here's what landed on the operator's desk.
Anthropic signs a 20-year bet on Canadian power
The AI land grab is now a power grab. Anthropic has signed a 20-year lease with EverWind Fuels for capacity at a new campus in Point Tupper, Nova Scotia — its first Canadian data center — with EverWind committing more than C$10 billion, roughly C$8 billion for the data center and C$2 billion for supporting onshore wind. The structure is the tell: a model company isn't just renting GPUs anymore, it's underwriting the electricity to run them for two decades.
The operator's take: when your AI vendor starts signing 20-year energy leases, that's a signal about where your costs are heading. Compute pricing is a proxy for power, and power is getting locked into long contracts because it's the actual constraint. If your roadmap assumes inference gets cheaper every quarter forever, pressure-test that assumption — the people building the capacity are planning in decades, not discount cycles. Model your AI spend like a utility bill, because increasingly that's what it is.
Replit drops Anthropic as its default to save money
The resellers are voting with their invoices. Replit, the AI coding platform popular with founders, is making OpenAI its default model provider over Anthropic, with CEO Amjad Masad citing cost among the reasons. Masad later clarified that Claude remains available to customers on the platform, but the default is now OpenAI — and since most users never change a default, that quietly moves a lot of spend.
The operator's take: this is build-vs-buy playing out one layer up, and it's a preview of your own negotiation. The tools you buy are themselves reselling a model, and they'll swap that model the moment the margins demand it — which means the "best model" powering your favorite SaaS can change without a press release. Two takeaways: don't hard-wire your workflows to a single model's quirks, and remember that default settings are a business decision made for someone else's P&L, not yours. If model choice matters to your output, make it an explicit setting you own, not one you inherit.
A Chinese APT burns a BeyondTrust zero-day
The privileged-access tool became the privileged-access path. SecurityWeek reports that a Chinese APT tracked as Armory Castle exploited a BeyondTrust vulnerability, CVE-2026-102764, as a zero-day for initial access, then deployed custom tools and deleted logs and artifacts to cover its tracks. The ugly part isn't just the entry — it's the cleanup. An attacker targeting the system that brokers your privileged sessions, then scrubbing the evidence, is building for persistence, not a smash-and-grab.
The operator's take: your privileged-access management tool is one of the highest-value targets you run, because it holds the keys to everything else. A zero-day there with active log-wiping means "no alerts" is not the same as "no breach." If you run BeyondTrust, patch on the vendor's emergency timeline and then go hunting under the assumption the logs may already be incomplete. The lesson generalizes: the tools that hold your most privileged access deserve your most paranoid monitoring, ideally logged somewhere the thing itself can't reach in to delete.
Also on my radar
- The breach that didn't need a zero-day. CenterPoint Energy disclosed a data breach affecting roughly one million Texans, exposing names, addresses, Social Security numbers, and driver's license numbers — a reminder that for most victims, the attack vector is still basic, and the fallout still lands on real people.
- A loud warning on AI concentration. "Big Short" investor Steve Eisman argued on a podcast that OpenAI has become so embedded in the US economy that a sudden collapse could trigger a recession, flagging cheaper open-weight rivals squeezing its pricing — a tail risk worth naming in your vendor-concentration planning.
- The frontier keeps moving the goalposts. Sam Altman confirmed OpenAI is building a humanoid robot and expects a demo in 2027, arguing the hard part is the brain, not the body — ambitious, and a reminder to judge your AI roadmap on what ships, not what's promised.
The through-line: the AI era's real scoreboard isn't benchmark scores, it's balance sheets and power contracts. Anthropic is locking in electricity for 20 years, Replit is swapping suppliers to protect its margin, and a nation-state is spending a zero-day on the one tool that unlocks everything else. The operators who do well here won't be the ones with the flashiest model — they'll be the ones who treat AI as a cost center to manage, a supply chain to diversify, and an attack surface to defend. Watch the invoices, own your defaults, and guard the keys. 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.