The edge,
itemized.

The subscriptions

The predictable part: what renews every month no matter what. Note the plural on Claude — two separate Max subscriptions, because when your agents work in parallel all day, one seat's usage limits become the bottleneck. The second seat doubles the ceiling.

† Five of these bill a card without ever sending an email receipt — ChatGPT Pro sends none at all, Grok Ultra and Higgsfield bill on separate accounts, and Perplexity and PocketSmith only turned up on the card statement. Those amounts are the card charges or the plans' current billing prices; every other line is copied off an actual receipt. There's also a ChatGPT Team workspace for one of our operating businesses that isn't counted here.

The meter

Here's the part nobody puts on a pricing page: the subscriptions are only the entry fee. When the agents run hard, you blow through the plan limits and into metered extra usage — and the meter charges in small, relentless increments. In one nine-day stretch last month our inbox collected thirty-five separate $10 auto-recharge receipts.

The meter is why this page exists. It scales with how hard the tools actually work — which is the same reason inference is the bill that never stops for the whole industry, just at our scale instead of theirs.

The rack

A chunk of our AI doesn't rent compute at all — it runs on our own hardware in Las Vegas, serving local models with no per-token bill. Here's the fleet — priced at what the same machines are being offered for on the open market today, not what they cost when we bought them:

Market prices pulled from live eBay and Amazon listings on 13 Aug 2026 — medians of real unit listings, with obvious too-good-to-be-true outliers excluded. What we actually paid, from purchase records and list prices, was roughly $63,800.

Two things fall out of that ledger. First, the arithmetic: spread what we paid over three years and the rack costs about $1,770 a month — within a rounding error of what the cloud stack above billed in the last 30 days. Same money, opposite failure mode: the subscriptions scale up when the work gets heavy; the rack just gets busier. Second, the part nobody expects: the hardware has been appreciating. The same AI demand that runs our meter pushed our ≈$63,800 fleet to ≈$97,500 in open-market asking prices — the 512 GB Mac Studios alone nearly doubled. The cloud bill only ever spends; the rack, so far, has been holding its value and paying rent in VRAM.

That's the real shape of an edge stack in 2026: rented frontier models for the hardest work, owned local models for the endless work, and a meter in between that tells you when to move a workload from one to the other.

You don't need this bill

Let's be direct: this is an R&D budget, not a template. We pay it because knowing every tool first-hand is the product — it's how we can tell a client which $20 seat replaces which $200 one. A typical small business captures most of the value with a single assistant subscription, and — where privacy or volume justifies it — a local model on a machine it owns.

If you want help making that call for your business, that's literally what we do.

Common questions