Micro-datacenter side income: $30–40k capital decision

Update, Sept 19, 2026: official channels only

What would have to change for this to be worth doing: card price back under about $9,000 through a channel the operator accepts; or a signed contract that guarantees 85%+ billed hours at a host share of 75% or better; or a materially higher rental rate for 96 GB cards. Absent one of those, the correct decision is not to buy.

Buying it properly: official channels

Every line below was verified by the operator on Sept 19, 2026 at the retailer's own storefront. Central Computer is an official NVIDIA/PNY retail partner with five Bay Area stores: San Francisco (837 Howard St), San Mateo, Sunnyvale, Santa Clara and Newark/East Bay.

Bay Area hardware versus Edmonton power

This is now a genuine conflict, and it resolves in favour of buying in the Bay Area.

• The Canadian card is about $2,600 more per card est.CAD 23,999.99 × 0.725 = $17,400; − $14,799.99, so $5,200 more for two — and it is out of stock, so the comparison is academic until it isn't.
• Edmonton's power advantage on this smaller 1.22 kW load is $237/mo est.1.222 kW × 730 h × ($0.40 − CAD 0.19 × 0.71). Recovering a $5,200 hardware premium from that takes 22 months of power savings before it earns anything.
• So: buy at Central Computer either way. The remaining question is only where the machine physically lives.

If the machine goes to Edmonton, importing US-bought hardware reopens three things. Duty: GPUs classify under HS 8473.30 at 0% MFN^61,54, so nominally free, but Canada's United States Surtax Order (2026), effective Sept 8, 2026, names an electronics sector and CUSMA origin does not exempt goods from it^55,56 — check the tariff line before shipping. Tax: 5% GST at the border, recoverable through input tax credits once registered, and Alberta has no PST^61,58. Warranty: a PNY card bought from a US retailer and RMA'd from Canada means cross-border shipping and possibly a US-only service region; this was not verified and should be confirmed with PNY before purchase, because a single failed card is 38% of the build's capex. Broker and freight fees are unpriced.

If it stays in the Bay Area, the operator loses $237/mo of power margin and Canada's Class 50 100% expensing, but keeps warranty simplicity, no customs exposure, no Canadian corporation, and no cross-border tax filings — which is also where that $250/mo compliance placeholder goes away. On payback the gap is 52 months versus 62; neither clears the bar, so this choice does not rescue the project. It matters only if the card price falls.

The whole machine

A two-card host built from what is actually in stock. Verified prices are the operator's Sept 19 observations; every other line is an estimate, labelled, and should be replaced with a counter quote before committing.

Power

GPUs 2 × 300 W = 600 W, CPU 350 W, platform (board, RAM, NVMe, fans) 150 W est. → 1,100 W DC. At ~90% PSU efficiency that is 1,222 W at the wall est.1,100 ÷ 0.90. One 15 A/120 V circuit carries it with no headroom (1,222 W = 10.2 A, against a 12 A continuous limit); a dedicated 20 A circuit is the sane choice and still cheap. The 3.2 kW, 240 V/30 A problem from the earlier build is gone — a real benefit of the Max-Q part. Monthly power: $120 in Edmonton, $357 in San Francisco est.1.222 × 730 × CAD 0.19 × 0.71, and × $0.40.

Rack gear, if any

None of it is required for a single tower on a dedicated circuit, and the earlier switch price came from an auction listing, which is now excluded. If the machine is racked, the previously sourced figures still stand as a guide: enclosure $2,548.88⁶⁷, switched PDU $1,499.99⁶⁵, UPS $2,543 (2,700 W, ample for this 1.22 kW load, unlike the four-card build)⁶³. Adding all three pushes capex past $45k and past the ceiling. Recommendation: skip the rack, keep the UPS only if the site has poor power.

Prebuilt versus self-build

The Orion Rackmount starts at $15,548.91 and is orderable now⁷⁰. The equivalent self-built platform — everything above except the GPUs — is $5,698.98, of which $4,049 is estimated. At face value the prebuilt is $9,849.93 more15,548.91 − 5,698.98, and whether its base price includes a GPU or much more platform (more RAM, redundant PSUs, storage) was not verified. With two cards added at $29,599.98 the prebuilt route is $45,148.89 before tax, roughly $49,438 with tax est., which is 24% over the ceiling and pushes payback from 52 to 67 months at base assumptions.

Buy the prebuilt anyway if: the quote comes back including a GPU or a materially richer platform; or the operator values one vendor owning the whole warranty, since with two cards a single RMA is 38% of capex and self-built systems mean arguing with three suppliers; or rack integration, burn-in and a system-level return window are worth ~$10k, which at this margin they are not. The right next step is a quote: ask Central Computer to price an Orion Rackmount configured with 2× PNY RTX PRO 6000 Max-Q, and compare that single number against $38,652.36.

What RunPod actually pays

These are RunPod's published renter-side rates, from a page dated updated Sept 13, 2026⁷². They are what a customer pays, not what a host receives.

Monthly net and payback, two-card build

Net = 2 × 730 h × $2.09 × utilisation × host share − opex. Capex $38,652.36 est.. Opex is $544/mo in year 1 in Edmonton ($120 power + $103 internet + $71 insurance + $250 compliance placeholder) and $669 after the internet promo; $757/mo in San Francisco, where the compliance placeholder would in practice fall away with no Canadian entity.

Each cell: monthly net in year 1, then months to payback on operating cash. "never" means cumulative cash never reaches capex. Payback accounts for the post-promo opex step.

Is 12-month payback reachable?

No, and not by a small margin. Verified:

• Required net cash: $38,652.36 ÷ 12 = $3,221.03/mo.
• Theoretical maximum gross to the host: 2 cards × 730 h × $2.09 × 100% utilisation × 100% host share = $3,051.40/mo.
• $3,221.03 ÷ $3,051.40 = 106% of maximum gross, with zero opex and RunPod taking nothing. Including Edmonton opex the requirement is $3,765/mo, or 123%.

So 12 months is impossible at official pricing regardless of assumptions, which is a cleaner answer than the earlier build gave. For the record, the other bars: 24 months needs about 93% utilisation at an 80% host share in Edmonton; 36 months needs about 82% utilisation at 70%, or about 70% at 80%. At a realistic 65% and 70%, payback is 52 months in Edmonton and 62 in San Francisco, and the month-36 position including estimated resale is about −$3,900 est.−38,652 + 12 × 841 + 24 × 716 + 7,400 residual GPU value at 25%.

Verdict on the original (grey-market) basis

One configuration clears the 36-month bar on evidence: four RTX PRO 6000 Blackwell 96GB cards rented on Vast.ai/RunPod, sited in Edmonton, bought near the lowest observed listing of $8,500 per card²⁰. Its base case pays back in 26 months, which misses your 24-month "acceptable" line. No build reaches 12-month payback in any modelled case. At NVIDIA's $16,000 list price¹⁶ the base case does not pay back within 36 months.

All payback figures are the engine in the cashflow section with its default opex (Edmonton power at CAD 0.19/kWh, TELUS business fibre, insurance rider, $250/mo compliance placeholder, 0.71 USD/CAD). "If liquidated" adds estimated GPU resale on a straight line to 50% at month 24; non-GPU gear is assumed to resell for nothing.

Where this page departs from the research summary

Rebuilding the numbers line by line exposed errors in the earlier summary. The corrected figures are used everywhere on this page.

Location economics

Cross-border: US person, hardware in Canada

Every revenue model, ranked

Hard gates from your brief rank first: payback beyond 36 months is out, and a 24/7 support or on-call obligation is disqualifying. Composite score only orders models within the same gate. Scores translate the research's qualitative ratings onto a 1–5 scale (5 is best on every axis); they are judgement, not measurement. Click a column header to sort.

"Not observed" means the research collected no marketplace rate, filled listing or competitor price sheet for that model. Those rows are scored on the research's qualitative assessment only and cannot be shown to meet any payback target.

Top 3 builds and bills of materials

The research found one build that clears the bar, one fallback that fails once itemised, and the best consumer-GPU option, included to show why it fails. All three are paired with Edmonton: at equal revenue, Edmonton's power is $315/mo against $934 for an SF residence or ~$784 for SF colo for the same 3.2 kW, and Class 50 expensing requires the hardware to be in Canada. San Francisco figures are shown for comparison.

Per-build summary

Draw: GPU TDP × count + host overhead (0.8 kW for A and B, 1.4 kW for C's two hosts) est.. A100 PCIe 300 W is NVIDIA's published TDP and was not captured in the research sources. Power: kW × 730 h × rate. Edmonton net subtracts power, TELUS $103, insurance $71, and the $250 placeholder.

Build A: 4× RTX PRO 6000 Blackwell 96GB, Edmonton residence

Why Edmonton: power at $0.135/kWh USD against ~$0.40 in SF; hardware must be in Canada, in use by Dec 31, 2026, for 100% Class 50 expensing; free-air cooling for part of the year. Why this card: 96 GB VRAM rents at $1.69/hr on RunPod Community^43,44 and up to $3.00/hr managed^15,18, while consumer cards have collapsed. VRAM scarcity, not raw speed, sets rental income.

Build B: 2× A100 80GB, Edmonton residence (fallback that fails)

Why considered: 80 GB VRAM still rents, and used supply exists if PRO 6000 cards can't be found near $8,500. Why it fails: two cards at $14,150 buy half the rentable hours of Build A for 91% of its capex. Rental range observed $0.12–1.39/hr (research; per-listing URL not pinned). Even treating the top of that range as net at 80% utilisation, the build is still −$3,061 at month 36.

Scenario rates: pessimistic $0.50 est., base $0.755 est.midpoint of $0.12–1.39, optimistic $1.39 (top of observed range treated as net). Utilisation 40 / 65 / 80%.

Build C: 8× RTX 5090, Edmonton residence (shown to fail)

Why considered: the densest consumer-GPU option at this budget. Why it fails: real host earnings are ~$115/card/month on Salad (4× 5090 ≈ $460/mo)^13,14 and ~$175/card at 80% on Vast.ai¹¹, while 6 kW of draw costs $591/mo even in Edmonton. In SF, power alone ($1,752) exceeds the optimistic gross.

Scenario revenue per card per month: pessimistic $57.50 est.half of the Salad figure, base $115 (Salad), optimistic $175 (Vast at 80%). Without the $250 compliance placeholder, base nets $155/mo: 228 months to recover capex.

36-month cashflow, top pick (Build A)

Change any input and every table and the chart below recalculate. Scenario buttons set rate and utilisation to the research's cases; everything else is shared across the three scenario columns. Month 1 is assumed to be October 2026, so month 4 is January 2027, when EPCOR's fixed energy charge expires.

12-month payback verdict

No option hits 12-month payback in any modelled case. The closest, Build A, reaches 15 months only under the optimistic case.

12-month payback requires monthly net ≥ capex ÷ 12, so required gross = capex ÷ 12 + year-1 opex. Required rate × utilisation = required gross ÷ (GPUs × 730 h).

Against your other bars: Build A's base case (26 mo) sits between the 24-month acceptable line and the 36-month cutoff. It meets 24 months only in the optimistic case (15 mo), or in base if the $250 compliance placeholder turns out to be near zero (22 mo). Base meets 24 months at ≥$1.34/hr net at 65% utilisation, or ≥67% utilisation at $1.30.

Operating margin vs tax deferral vs resale

Most of the apparent return in the first two years is estimated resale value, not operating cash. The tax treatment adds timing, not dollars.

Don't do this

Kill criteria

Thresholds come from the Build A model at default opex (after the internet promo: $863/mo, 2,920 installed GPU-hours/mo). Two tiers. Stop adding capital: the investment can no longer pay back within 36 months. Liquidate: holding the hardware now destroys value, because monthly net cash is below the monthly fall in what the cards would sell for. The model's default resale decline is $708/mo est.$34,000 × 50% ÷ 24; replace it with the trend in eBay sold listings.

Risks, post-mortems and the arbitrage question

Staged plan and go/no-go

Sources

Sources 70–72 were verified by the operator on Sept 19, 2026 and override every earlier hardware price on this page. Sources 1–61 were retrieved by the research run, which recorded observation dates as Sep 17, 2026 (one day ahead of this conversation's Sep 16 date). Sources 62–69 were retrieved while building this page on Sep 16, 2026. Several URLs are vendor or retailer listing pages; treat single-listing prices as a point observation, not a market price.