Safeera
A Sparks on the driveway of a suburban home at dusk, the ring at its base glowing; a man and his dog walk past; close-ups of the shell seam and the racks inside.
Sparks: ambient film, no narration.

Sparks.

Sixteen B300 GPUs for the office, with optional solar and battery.

Ownership path
Finance term

Est. monthly payment

$37,303/mo

Est., 7% APR, before the reservation credit. Reservation Agreement.

$264,000

16% minimum90% maximum

Direct purchase is reserved for existing Safeera investors.

Privacy

−$73,000 credit today

Decreases $1,000 a day. Locked at order.

Due today$350 order fee

We verify investor status and reach out to arrange the purchase.

Safeera confirms your site and schedules delivery. First deliveries planned from January 2027; timing subject to change.

Est. total $1,650,000 before credit · Final price confirmed before delivery.

Due to scarcity of compute components, the price may change in the event of memory or NVIDIA price increases.

Specification

An office that can power itself. Sixteen B300 Blackwell Ultra GPUs, with an optional solar array, battery and generator.

Compute
16 B300 Blackwell Ultra GPUs, 4.6 TB of HBM3e
Power
About 32 kW, grid-tied
Energy
Optional rooftop array, battery and generator
Size
About 2.6 metres tall
Data
Stays in the building

With the optional package, solar offsets a year's electricity under net metering; the grid carries the hours without sun.

1.75 m 2.6 m tall ≈ 2 m across ⌀ ≈ 2 m base ⌀ ≈ 1 m door ≈ 0.9 m ≈ 0.5 m floor ≈ 3.4 × 3.0 m
Sparks: 2.6 m tall, ≈ 2 m across. Stands on a ≈ 1 m base, anchored to the slab. Clearance ≈ 0.9 m at the door, ≈ 0.5 m on the other sides; floor space ≈ 3.4 × 3.0 m.

Economics

Capacity value is computed from public cloud on-demand list prices for equivalent GPU capacity, the median of the 4 providers below as of September 11, 2026, at 40, 70 and 90% utilization. It is gross revenue potential, not a forecast.

Capacity value

What the same capacity rents for in the public cloud: 16 B300 Blackwell Ultra GPUs, at published on-demand list prices. This is what an owner would otherwise pay, or could recover by making spare capacity available.

AWS
$284.80 per hour — P6-B300 (p6-b300.48xlarge, 8x B300), $17.80 per GPU-hour
Microsoft Azure
$432.64 per hour — ND GB200 v6 (Standard_ND128isr_NDR_GB200_v6, 4x GB200 Blackwell GPUs), $27.04 per GPU-hour
Oracle Cloud
$240.00 per hour — BM.GPU.B300.8 (bare metal, 8x B300), $15.00 per GPU-hour
CoreWeave
$137.60 per hour — B200 bare metal (8x B200), $8.60 per GPU-hour
Per hour
$262.40 — median of the providers above
Per year at 40%
$919,450
Per year at 70%
$1,609,037
Per year at 90%
$2,068,762

Method: median list price per GPU-hour × 16 GPUs × 8,760 hours × utilisation. Based on published on-demand list prices for equivalent GPU capacity as of September 11, 2026; excludes networking, storage and support; not a forecast of resale income.

Owner’s calculator

Ownership path and prepayment follow the order column: Lease 3.5 yr · prepayment $264,000 (16%). Change them in the order column.

Monthly gross revenue
$127,382 est.
Monthly electricity
$2,974 est.
Internet
$250 est.
Estimated monthly operating cost
$3,224 est.
Net before financing
$124,158 est.
Monthly payment
$37,303 est.
Net after payment
$86,855 est.
Payback on the prepayment
4 months est.
Payback on the full estimate
14 months est.
How this is calculated
  • Powered hours a month = 730 × (1 − downtime). Hours sold = powered hours × utilization.
  • Monthly gross revenue = sell price per GPU-hour × 16 GPUs × hours sold.
  • Monthly electricity = (32 kW × hours sold + 32 kW × powered hours not sold) × price per kWh. No published breakdown of the 32 kW, so the full draw is charged in every powered hour.
  • Estimated monthly operating cost = electricity + internet. Net before financing = gross revenue − operating cost.
  • Monthly payment = the est. total less the prepayment, amortised at 7% APR over 42 months on the lease or the finance term; direct purchase has no payment and the whole total is the prepayment.
  • Net after payment = net before financing − monthly payment.
  • Payback on the prepayment = prepayment ÷ net after payment; payback on the full estimate = est. total ÷ net before financing; both in whole months.

Defaults: utilization 70% and downtime 5% are assumptions to adjust. The sell price, $16.40 per GPU-hour, is the median of the 4 list prices above. Electricity, $0.134 per kWh, is the 2025 U.S. average retail price for the commercial sector (U.S. Energy Information Administration, Electric Power Monthly, Table 5.3; checked October 11, 2026). Internet, $250 a month, is a published business fiber list price for a 2 Gbps line, plus taxes and fees (checked October 11, 2026).

Revenue potential depends on demand and on the price Safeera’s distribution channels achieve. The figures exclude electrical service upgrades, internet upgrades, maintenance, insurance, taxes and time offline beyond the downtime entered. They are estimates, not a forecast or a guarantee.

The equipment price is an estimate and subject to change; the final price is confirmed before delivery.

Made for

  • Engineering firms

    Blackwell-accelerated Ansys, Cadence and Siemens solvers, at data-center speed, in the office.

  • Private assistants

    Agents and knowledge bases over confidential files that never leave the building.

  • Off-grid sites

    Clinics, farms and field stations; add the array, battery and generator for power and compute in one delivery.

Field notes

Small business, big models: engineers and creators bringing AI in-house

Ansys, Cadence and Siemens now run their simulation tools on Blackwell GPUs, so a small engineering team can run in hours what once took a weekend. For a firm whose models, footage or client files are the business, an owned appliance beats metered cloud rental. Read the note

AI beyond the grid: solar-powered intelligence for rural clinics, farms, and field sites

Qure.ai screens for tuberculosis at more than 2,600 sites; John Deere's See & Spray runs its vision entirely on the machine. Where the grid is thin, compute has to bring its own power. Read the note

Drones, cameras, and street lights: giving AI a body, safely

Traffic signals, building controls, inspection drones and camera networks carry AI that acts on the physical world. Seven governments' joint guidance asks that such systems fail gracefully under human oversight. That means an auditable control core on the operator's own premises, with an off switch. Read the note

Reserve Sparks

Also

Googie, for the homeCenturion, for the campusThe thesisPhysical AI on the factory floor