"$0 cloud spend" sounds like a slogan. It's actually a math problem, and the math is more interesting than the slogan.
Every token generated on this fleet runs on hardware sitting on a desk in Augusta, Georgia — an Intel Arc Pro B60, roughly $500, doing the work an API call would otherwise bill for. That's not free. It's a fixed cost that was paid once, against a variable cost that never stops. The question worth answering honestly is: at what point does the fixed cost win?
the fixed cost: one card, paid once
The Arc Pro B60 lists at roughly $500 and, per Intel's own data sheet, draws up to 200W under load — a single 8-pin power connector, no exotic cooling required. At a residential electricity rate of about 18 cents per kWh (the current U.S. average), running that card flat-out for an hour costs:
// the electricity math 0.2 kW × $0.18/kWh = $0.036 per hour of GPU time at full load. Even at a generous 8 hours a day of genuinely heavy inference, that's about $0.29/day — roughly $8.60 a month, or a bit over $100 a year, for the card alone.
That's the entire recurring cost. No per-token meter. No surprise bill after a long agent run chews through context. The $500 purchase is the whole story after that, plus a rounding error in electricity.
the variable cost it's replacing
The production model on this fleet — a 30B-class MoE, quantized IQ4_XS — generates at roughly 40 tokens per second sustained. Run that flat out for an hour and it's producing about 144,000 output tokens. Run cloud APIs at the same rate and here's what an hour of equivalent output actually bills, at published September 2026 rates:
- Claude Haiku 4.5 — $5 per million output tokens → $0.72/hour
- Claude Sonnet 5 — $10 per million output tokens → $1.44/hour
- GPT-5.6 Terra — $12 per million output tokens → $1.73/hour
Compare that to the $0.036/hour the B60 is actually costing in electricity. That's a 20x to 48x gap, and it's not a one-time gap — it repeats every hour the fleet is generating tokens, forever.
the crossover point
Here's the number that actually matters: how many hours of equivalent cloud generation does it take before the $500 card has paid for itself — net of the electricity it's spending the whole time it's earning that back?
At the cheap end (Haiku-equivalent pricing, $0.72/hour gross minus $0.036/hour in power, $0.684/hour net), that's about 731 hours — roughly a month of 24/7 use, or a few months of a normal workday. At the pricier end (GPT-5.6 Terra, $1.73/hour gross minus power, $1.694/hour net), it's closer to 295 hours — under two weeks running around the clock.
// the honest caveat This assumes generation-bound, output-heavy workloads — long agent runs, drafting, code generation — where the token count is real and sustained. A chatbot that idles most of the day doesn't hit these numbers, because the comparison is against tokens actually generated, not tokens available. Self-hosting wins fastest for workloads that keep the card busy.
For a fleet that runs agents continuously — the whole premise of Sentinel Prime planning and executing work in the background — the card breaks even within a month or two of real use, and every hour since has been running at a 20-to-48x discount versus the API alternative.
what this doesn't include
Being honest about the math means being honest about what it leaves out. The B60 is one box on a desk that also runs a mini PC host, a NAS, a router and a small standby watchtower — all drawing their own power, all part of the real setup described on the rig page. None of that shows up in the $0.036/hour figure above, because that number is the GPU alone, doing the GPU's job. A fair whole-desk accounting would add some amount for the always-on supporting hardware — real, but small next to a single API bill for equivalent output, and nowhere near enough to change which side of the comparison wins.
It also doesn't include the time spent building the thing. Self-hosting isn't a checkbox; it's a stack you maintain. That cost is real and it's not zero — it's just not a line item that shows up on a monthly bill, which is exactly the point. The build itself is the one-time cost. After that, the meter simply doesn't run the way it does for everyone paying per token.
the actual takeaway
"$0 cloud spend" isn't a marketing number — it's a description of what happens after a fixed cost is paid off, for a workload that runs enough to make the comparison meaningful. A $500 card, an electricity bill you'd barely notice, and a break-even point measured in weeks, not years. Everyone paying by the token is still paying by the token. This fleet just isn't.
Sources: Intel Arc Pro B60 data sheet (TDP, power connector), U.S. residential electricity rates, September 2026, and published per-token pricing from Anthropic and OpenAI's API pricing pages as of September 2026.