RTX 5090 service-node power-supply and site-power decision#

Status: Current decision record Decision date: 2026-09-23 Scope: The Malaysian residentially hosted ComfyUI service nodes (PC1–PC3, planned to four or five), each built around a Ryzen 9 9950X and an RTX 5090 Electricity (2026-09-27): the site stays single-phase on the domestic tariff. Running cost, the verified TNB bill formula and supply headroom are in Electricity cost and supply headroom. Commercial constraint: Bootstrapped operation; prefer simple measures with measurable availability or capital benefit Current PSU decision (2026-09-25): the fleet standard is the Seasonic Vertex PX-1200. PC1, PC2 and PC3 (ordered 2026-09-25) all have it, and future nodes buy it too (why). This supersedes the 2026-09-24 HX1200i preference and its RM250 price gate. The Corsair HX1200i is the fallback if the Vertex is unavailable or more than about RM500 dearer. The Corsair RM1000x ATX 3.1 is the 1000 W fallback. Never use an Antec GSK1000 (1000 W class). Every node also gets the Pi Pico connector thermal watchdog (GPU-plug probe plus case-air reference, auto-shutdown). It adds no connection to the power path.

Fleet standard: Seasonic Vertex PX-1200 (2026-09-25)#

The connector-melt evidence (91 incidents) shows no PSU-brand difference in RTX 5090 16-pin melts once market share and cable path are accounted for. Brand is a tie, so the tie-breakers decide:

• One standard across the fleet. PC1, PC2 and PC3 are Vertex, so one spare PSU and one cable type cover every node. Modular cables must never be mixed across brands, and a mixed fleet makes that mistake possible.
• Native 16-to-16 cable. It has one fewer mated interface than the HX1200i's Type-4 2x8-pin cable, and it is interchangeable with any native-socket PSU or aftermarket cable. The Type-4 split does not lower per-terminal current: four Corsair Type-4 cables melted at the PSU plugs.
• The HX1200i's telemetry does not watch this failure mode. PSU-side power and voltage logging cannot see a failing contact. A temperature probe on the GPU plug can, and it can shut the machine down (Pi Pico connector thermal watchdog).
• Weak but consistent evidence: no Seasonic-cable melt in 34 PSU-cable incidents, a 12-year warranty against 10, and 1.4 ms more hold-up.
• Cost: about RM311 per node, roughly 2% of the GPU.

This is a tie-break, not a safety verdict. The controls that actually reduce melt risk are topology and monitoring:

• use the PSU's own cable, never the GPU's bundled adapter;
• seat it once;
• run a modest power limit for 24/7 inference;
• fit the Pi Pico connector thermal watchdog, with IR-thermometer checks of both plugs at commissioning and every few months;
• replace the cable on an imbalance alarm or when the 16-pin voltage sags below about 11.5 V.

Check each new cable once at full load when the node is commissioned. Sections below dated 2026-09-24 that prefer the HX1200i are kept as the evaluation record. Their electrical comparisons still hold, but their recommendation is superseded.

Why this document exists#

Historical framing (2026-09-23). This section and the "Decision in one page", "Price comparison", "What is and is not better value" and "Procurement instruction" sections were written before any node was ordered. The "existing worker" is PC1, and the build below (O11 Vision, a Gen4 OS drive) is not what was ordered. The ordered builds are in the node decision record. The electrical comparisons still hold; the site-power and UPS sections are current.

The PSU decision is not about finding the largest wattage number or the longest feature list. The machines will provide a paid ComfyUI cloud service, but the company is solo and bootstrapped. The useful question is:

Which power arrangement materially reduces downtime or hardware risk without spending money on rare failure modes, unused capacity, or branding?

One existing worker already uses a Seasonic Vertex PX-1200. The second worker is being quoted as part of an IdealTech promotional bundle whose historical Seasonic-based two-Samsung/1 TB-OS total was RM38,647. The storage plan now starts with one model SSD and a separate 2 TB OS SSD, and onboard 5GbE is retained, so a new system quote is required. This does not change the relative PSU comparison. The relevant build is:

Ryzen 9 9950X
Zotac RTX 5090 Solid OC, 575 W board power
64 GB DDR5
1x Samsung 9100 Pro 2 TB initial disposable model cache
1x separate 2 TB Gen4 chipset-connected system drive
optional second 2 TB Samsung only after the capacity/performance gate
360 mm AIO, pump and eight case fans
Lian Li O11 Vision case

The storage and PCIe decisions are documented separately in the complete hardware decision. This record covers the PSU, the incoming electrical supply, what electricity costs, and the UPS decision.

Decision in one page#

1. 1200 W remains the correct capacity class. Zotac specifies 575 W for the GPU and recommends a 1000 W system PSU. A good 1000 W unit can run the machine; 1200 W adds reasonable sustained-load, thermal, aging, and future CPU-concurrency margin. A 1500-1600 W label adds no ComfyUI performance and is not additional surge or outage protection.
2. Do not replace an already-owned Vertex PX-1200. It is a high-quality, independently tested PSU. A replacement would spend money and create downtime without improving throughput or materially changing risk.
3. Superseded 2026-09-25: buy the Vertex PX-1200 for new nodes as the fleet standard (see above). Items 3–4 below record the value analysis that led to the earlier HX1200i preference. The regular Corsair HX1200i is the best same-tier value in the 2026-09-22 list. At RM1,389 it is RM311 below the RM1,700 Vertex, while independent 230 V results are competitive or better in efficiency while improving measured average noise and load regulation with essentially equal ripple. It gives up 1.4 ms of AC-loss-to-PWR_OK time and two years of stated warranty. Neither difference is likely to matter during the useful life of this node.
4. Use a price gate because package prices can differ from component-list prices. Select the regular HX1200i if the final installed-system quote is at least RM250 below the Vertex and confirms the exact ATX 3.1/PCIe 5.1 revision, supplied 600 W 12V-2x6 GPU cable, and ten-year Malaysian warranty. Below RM250, retain the Vertex for consistency with the first worker.
5. The Lian Li SX1200P is the acceptable cost-first option, not a dangerous PSU. Use it if the actual package saves at least RM500 against the Vertex and the exact black SX1200P and local ten-year warranty are confirmed. The lab result is electrically suitable, but it is materially noisier and has higher ripple than the two premium choices.
6. Do not buy a PSU on an unsubstantiated reliability story. No statistically useful field-failure data were found for these exact revisions. A single review sample, component inspection, certification, and warranty are useful screening evidence, not proof that one model fails less often in this deployment.
7. Site protection has higher leverage than moving between good PSUs. Use one properly installed, dedicated circuit per worker, sound earthing, correctly coordinated breaker and residual-current protection, and a switchboard surge-protection device. Log power quality before buying UPS capacity.
8. Do not buy two large UPS units merely to regulate normal Malaysian mains. Add them when measured interruptions, recovery labor, lost jobs, or customer commitments justify the ownership cost.

Capacity: why 1200 W, not 1000 W or 1600 W#

The Zotac RTX 5090 Solid OC specification states 575 W maximum power and recommends a 1000 W system supply. The included Antec GSK1000 therefore meets the vendor floor. It is not automatically unsafe or incapable.

This worker also has a high-end CPU, three NVMe drives, a pump, memory, motherboard conversion losses, and many fans. Normal ComfyUI traces do not fully load the CPU and GPU simultaneously, but future preprocessing or concurrent jobs might. Moving to 1200 W provides 200 W of useful margin and keeps normal operation away from the supply's upper edge. ATX 3.1 power-excursion compliance is more relevant than purchasing a large nameplate solely to absorb brief GPU excursions.

A 1500 W or 1600 W PSU does not force the machine to consume that power, but the extra capacity does not:

• make the RTX 5090 faster;
• improve PCIe or storage performance;
• condition bad building voltage;
• absorb a lightning event that the installation and SPD do not control; or
• keep the machine running during a grid outage.

The capacity may be harmless when unusually cheap, but it is not a protection feature and should receive no economic credit for this one-GPU build.

Price comparison#

Prices below come from the local IdealTech extraction dated 2026-09-22. The promotion includes an RM799 Antec GSK1000. The node-total column preserves the historical RM38,647 storage baseline so the PSU deltas remain directly comparable; it is no longer a current whole-machine quote. Surcharges remain provisional until IdealTech supplies a written installed-system quote.

The RM38,337 Corsair total is RM310 rather than RM311 below the quoted Seasonic total because IdealTech rounded the Seasonic substitution from its RM901 list delta to RM900.

The whole-system percentage can be misleading in either direction. RM311 is only about 0.8% of a RM38.6k node, but a bootstrapped company should still keep it when doing so sacrifices no useful capability. On the other hand, delaying a deployable node, accepting an unclear revision, or giving up a valid local warranty to recover RM311 would be false economy.

Independent 230 V evidence#

The following results are from Cybenetics' public 230 V reports. Lower ripple, regulation percentage, noise, and inrush are generally preferable; longer hold-up is preferable; a 200% transient voltage closer to 12 V is preferable. These values should not be combined into an invented reliability score. Every shortlisted unit passed its applicable test regime, and several differences are operationally insignificant.

Important limitations:

• The current Seasonic Vertex product page specifies ATX 3.1, PCIe 5.1, native 12V-2x6, a 12-year warranty, and a 160 mm enclosure. The Cybenetics sample was an earlier ATX 3.0 revision received in June 2024. Its measurements are useful platform evidence but must not be presented as a test of every current retail revision.
• The current Corsair HX1200i page specifies ATX 3.1/PCIe 5.1, an included 12V-2x6 cable, 105 °C Japanese electrolytic capacitors, a ten-year warranty, and a 200 mm enclosure. Its monitoring software is not required for operation. The worker runs Linux, so Windows-oriented iCUE telemetry receives no purchasing value unless a supported monitoring path is demonstrated.
• The Corsair HX1500i is also ATX 3.1/PCIe 5.1 with a ten-year warranty, but its 300 W of additional capacity is unused by this design. Its specified 200-240 V input is compatible with Malaysian mains but less universal than the HX1200i's 100-240 V range.
• ASUS specifies ATX 3.1, native 12V-2x6, a ten-year warranty, and a compact 160 mm enclosure for the ROG Strix 1200P. Its independent report also measured 151.7 A inrush at 230 V. It remains a capable supply, but there is no reason to pay RM150 over the HX1200i for this machine.
• The current Prime TX-1600 Noctua Edition is ATX 3.1/PCIe 5.1, 210 mm long, Titanium rated, and warranted for 12 years. Its nearest public platform results are excellent, but it costs RM610 more than the HX1200i and buys silence and unused capacity rather than availability. The RTX 5090 and case cooling also limit the value of a quieter PSU under full compute.
• The O11 Vision officially supports ATX PSUs under 232 mm, so the 200-210 mm alternatives fit on paper. The shorter 160 mm Vertex or ASUS still leaves more cable-management room. Fit must include cable bend radius, not only the metal PSU body.

1000 W class for future nodes (2026-09-24)#

The worst case adds a 575–600 W RTX 5090, a fully loaded 9950X at its 230 W stock PPT (HWCooling), and 70–100 W for the rest of the system:

• about 875–905 W with a 575 W card (Zotac);
• about 900–930 W with PC1's 600 W GIGABYTE.

That is up to 93% of a 1000 W unit, against up to 78% of a 1200 W unit. Typical ComfyUI load is lower, since a job uses about two CPU cores: an estimated 650–750 W, or 65–75% of 1000 W against 55–63% of 1200 W.

A good ATX 3.1 1000 W unit can run the node, and both GPU vendors specify 1000 W. But these nodes run 24/7 for years. Sitting higher in the load range means a hotter PSU and faster capacitor ageing. A warranty replaces the unit but not the downtime. 1200 W is therefore the class for service nodes. The efficiency difference is negligible (about 0.3% between the RM1000x and HX1200i averages, roughly 2 W at 700 W), so it is not the reason. The 1000 W comparison below is kept as a fallback and to rule out the package's GSK1000.

Cybenetics 230 V results for the 1000 W units on the IdealTech list of 2026-09-22, with the Vertex for reference. Prices are the list prices; the difference is over the package's included RM799 GSK1000.

The Antec GSK1200 (RM999) shares the GSK1000's platform and weaknesses: 73.27 mV ripple, 37.65 dBA, 14.1 ms to PWR_OK.

Do not accept the package's GSK1000. It is efficient and passes ATX 3.1, which relaxed the required hold-up to 12 ms at full load, with 17 ms at 80% load recommended (Seasonic). But it is the weakest unit here in the two measures that matter for an unattended node on residential mains:

• Ripple is about three times the others'. Hardware Busters' verdict: "ripple suppression needs to get better at 12V" (review).
• Ride-through is the shortest. At 13 ms from AC loss to PWR_OK, it covers the least of a mains sag.

It is also the loudest, which is irrelevant for a service node.

Among 1000 W units without telemetry, the Corsair RM1000x (ATX 3.1, CP-9020271) at +RM90 is the pick. The HX1200i was preferred on 2026-09-24 for its Linux telemetry (next section); superseded by the 2026-09-25 Vertex fleet standard.

• Electrically it matches the Vertex: ripple 23.9 against 23.5 mV, and 18.6 against 20.1 ms to PWR_OK.
• It costs RM810 less per node than the Seasonic.
• It is one of the most widely deployed 1000 W units: CWT-built, Japanese 105 °C capacitors, 10-year warranty. That fits the fleet's popular-parts rule.

The Lian Li RS1000G at +RM40 has the best ride-through measured (21.2 ms to PWR_OK) and low ripple. It is less common, and it has a rotating AC inlet; take it if the RM1000x is out of stock. The SX1000P has the tightest regulation but is louder. The HX1000i costs RM440 more than the RM1000x for no service gain.

Whatever the unit, confirm the exact ATX 3.1 SKU and its own supplied 12V-2x6 cable on the invoice.

Linux telemetry and OCP (2026-09-24)#

The bench data above treats PSUs as equals once they are electrically sound. For an unattended fleet, one operational feature separates them: a digital interface that Linux can read.

Telemetry. The mainline kernel's corsair-psu hwmon driver supports the HX1000i, HX1200i and HX1500i (legacy, Series 2023 and Series 2025) and the RMi series. It exposes:

• AC input voltage;
• per-rail voltage, current and power;
• VRM and case temperatures and fan speed;
• OCP mode and uptime, through debugfs.

The kernel notes that flickering rail voltages can indicate a failing PSU. liquidctl adds estimated input power and efficiency. It lists both HX1200i ATX 3.1 part numbers (CP-9020281, CP-9020307). node_exporter collects hwmon automatically, so each node's power draw and mains voltage reach the existing Prometheus and Grafana stack with no extra hardware. That gives three things:

• per-node energy for run-cost accounting;
• a coarse input-voltage log toward the 30-day power-quality record the UPS section requires (it will not capture millisecond sags);
• early warning of PSU degradation.

The Seasonic Vertex, RMx and GSK units have no such interface.

OCP and the 12V-2x6 cable. The HX1200i defaults to multi-rail OCP, with a reported 40 A per rail (unconfirmed for the exact SKU). An RTX 5090 draws about 48–50 A through one 12V-2x6 cable. On the HX1200i that cable feeds from two Type-4 8-pin sockets at the PSU ("2x8-pin to 12V-2x6"; there is no native 16-pin socket). That does not lower per-terminal current: each PSU-side terminal still carries about 8 A, and the design adds one mated interface. Four Corsair Type-4 cables have melted at the PSU-side plugs on RTX 5090s (evidence). Native 16-to-16 cables (the Vertex) fail at the PSU end too. Neither path is demonstrably safer, and neither PSU brand shows a melt excess after normalising for market share. The 2025 series adds dual-colour 12V-2x6 plugs that show full seating. Three cautions:

• Verify no OCP trip. Run full GPU load with concurrent CPU load in acceptance. If OCP trips, set single-rail with liquidctl initialize --single-12v-ocp, applied at every boot.
• PSU OCP cannot prevent connector melting. The documented 5090 connector failures come from uneven current across the pins inside the connector, which PSU-side OCP (single- or multi-rail) cannot see. Correct seating, no sharp bends and the PSU's own cable remain the controls.
• Add per-pin monitoring for 24/7 inference. Power limits and undervolts did not prevent documented melts, and several failures happened under long AI loads. Automatic detection with shutdown is the control with evidence behind it. Here that is the Pi Pico connector thermal watchdog; a WireView Pro II can additionally check each cable's per-pin balance once at commissioning. Treat a sagging 16-pin voltage (below ~11.5 V at load) or a pin imbalance alarm as "replace the cable". Never use the GPU's bundled 4×8-pin adapter.

Fit and remaining Vertex advantages. The HX1200i is 200 mm deep; the Fractal North XL takes up to 290 mm, or 175 mm only with both HDD trays fitted (Fractal). The Vertex keeps a 12-year warranty (against 10) and 1.4 ms more AC-loss-to-PWR_OK time. On 2026-09-24 neither was judged to outweigh fleet telemetry and RM311 per node. That verdict was reversed on 2026-09-25: fleet commonality and the native cable decide, and PSU telemetry cannot see the connector failure mode (fleet standard).

What is and is not “better value”#

Corsair HX1200i: yes, at the current list spread#

This is the direct answer to the same-price-range question. The regular HX1200i has the same useful 1200 W capacity, Platinum class, current ATX/PCIe support, a supplied direct GPU cable, premium component claim, comprehensive protections, and a long warranty. The independent sample is quieter, has essentially identical full-load 12 V ripple, and regulates 12 V more tightly than the tested Vertex. It is RM311 cheaper.

The Vertex has meaningful but small advantages: two more warranty years, a 40 mm shorter body, 1.4 ms more AC-loss-to-PWR_OK time in the compared samples, and consistency with the first node. Those advantages are worth a modest premium, not an unlimited one. RM250 is the decision boundary:

• HX1200i at least RM250 cheaper, all-in: choose HX1200i for the new node.
• Difference below RM250: choose Vertex for commonality and the longer warranty.
• Vertex already installed or owned: keep it; do not create a replacement transaction.

The RM250 line is an operating rule, not a physical law. It prevents revisiting the decision for a RM50-100 quote movement while still capturing the present meaningful saving.

Corsair HX1500i: good hardware, weaker system value#

At RM1,579, the regular HX1500i is RM121 cheaper than the Vertex and has excellent measured ripple. It is therefore better specification per ringgit than the Vertex. It is not the best purchase for this node, because the HX1200i is another RM190 cheaper and already supplies all required capacity. The 1500 W label does not reduce grid risk or produce revenue. Buy it only if its live installed price is no higher than the HX1200i, or if a later configuration genuinely needs more than 1200 W.

ASUS and Seasonic Prime: no economic case here#

The ASUS Strix is competent but lands between the HX1200i and Vertex without producing a corresponding benefit. The Prime TX-1600 Noctua is an enthusiast/workstation luxury: very high efficiency, low noise, premium construction, and a long warranty, but no measurable workload or availability return in this node. Neither is the 80/20 choice.

Lian Li SX1200P: best capital efficiency, with cost-first trade-offs#

The SX1200P saves a provisional RM620 against the Vertex and RM310 against the HX1200i. Its measured output is suitable, and its current specification includes ATX 3.1/PCIe 5.1, 12V-2x6, operation at full load up to 50 °C, and a ten-year warranty. Its principal measured compromises are noise and higher full-load ripple, not evidence that it will damage the machine.

Choose it when conserving the additional RM310 matters more than premium acoustics and component margin.

Checked again against the HX1200i for future nodes (2026-09-24, from the Cybenetics report tested 2026-01-27 and the Lian Li page):

• For the SX1200P: single-rail 12 V with OCP at about 119 A, so it cannot trip on a 5090 the way the HX1200i's multi-rail default might. It is 140 mm deep against 200 mm. Inrush is 51.9 A, it uses Japanese 105 °C capacitors, and it has a ten-year warranty.
• Against it: no digital interface, so none of the hwmon telemetry above. The fan is loud under sustained load, averaging 31.3 dBA and reaching about 43 dBA near full load, against 18.3 dBA average for the HX1200i. Full-load ripple is 31.5 mV against 23.4 mV, both well inside the 120 mV ATX limit. The platform is a 2026 Zhong Yuan design with little field history, which goes against the popular-parts rule.
• What the telemetry is really worth: nvidia-smi and CPU RAPL already give roughly 90% of node power for cost accounting. What only the HX1200i adds is AC input voltage, PSU temperature and rail drift.

Verdict (2026-09-24, superseded by the Vertex fleet standard): the HX1200i remained the default for future nodes. Take the SX1200P when the vendor's installed price gap is above about RM300, or when the HX1200i is out of stock. Either is electrically sound for a 5090 plus 9950X. The regular HX1200i remains the balanced middle: it removes roughly half of the Vertex premium while keeping the premium unit's measured ripple and improving measured noise.

Reliability: what can honestly be concluded#

There is no evidence here that paying RM1,700 rather than RM1,389 produces a quantifiable reduction in annual failure probability. Manufacturer reputation, internal component selection, electrical measurements, and a 12-year warranty are positive signals. They are not a population failure-rate study under Malaysian heat, dust, voltage, and 24/7 GPU duty.

All serious candidates implement over-voltage, under-voltage, over-current, over-power, over-temperature, and short-circuit protection. Once a PSU is well designed, within specification, correctly sized, and used with its own fully seated cable, the next RM300-600 buys diminishing margins. It cannot economically insure the rest of the RM38k machine against every upstream or downstream failure.

Practical reliability controls are:

• use only the modular cables supplied for that exact PSU; never mix Seasonic, Corsair, Lian Li, or ASUS modular cables even when the plugs appear to fit;
• use the PSU-supplied direct 12V-2x6 GPU lead without the GPU's multi-connector adapter, seat it completely, avoid severe bends near the plug, and inspect it after transport;
• keep intake air and the PSU dust path clean and maintain reasonable room temperature;
• record model, part number, serial number, revision, invoice, and warranty owner;
• keep jobs retryable and the host reproducible so a failed node is an inconvenience rather than data loss; and
• when the fleet grows enough that downtime costs more than idle inventory, hold one known-compatible spare PSU and its unopened cable set. A shared spare across three or more standardized nodes will usually buy more availability than moving from one already-good premium PSU to another.

Incoming power: the higher-leverage protection layer#

Malaysia's declared low-voltage service is 230 V single-phase or 400 V three-phase. A domestic service is normally described as single-phase two-wire or three-phase four-wire, not “two-phase.” If two circuits are placed on different phases of a three-phase board, each PC remains connected phase-to-neutral at about 230 V; it must not be connected phase-to-phase at about 400 V.

One direct branch from the main distribution board to each machine is sensible. It reduces nuisance interaction with unrelated appliances but does not isolate the workers from the common incoming supply, neutral, earth, lightning exposure, or a whole-site outage. A Malaysian licensed electrical contractor should provide and test:

• one labelled final circuit per worker with no unrelated outlets or loads;
• correctly selected over-current and 30 mA residual-current protection, arranged so one worker fault does not unnecessarily disconnect the other;
• polarity, protective-earth continuity, earth-loop performance, and RCD/RCBO trip verification;
• a properly selected and earthed surge-protection device at the distribution board;
• a suitable outlet or isolator for the actual PSU/UPS input, with no daisy-chained strips or improvised adapters; and
• branch energy/power-quality logging for voltage range, interruptions, and peak current under real loads.

The relevant primary references are the Malaysian Energy Commission's residential wiring guide and TNB's supply description.

The site stays on single-phase supply (founder, 2026-09-27). A three-phase upgrade, reclassification to the commercial tariff, and solar are all ruled out for now: they are bigger changes than adding a few PCs at home, though the position may change. If the property ever does move to three-phase, spread the 230 V node circuits across the phases to balance the site; the PCs' PSUs don't need to change.

Electricity cost and supply headroom (2026-09-27)#

Canonical for what the fleet costs to run, and whether the supply can carry it. Capital limits the fleet to four or five nodes, so this section plans for three to five.

How the domestic bill is built (TNB RP4, verified)#

Our itemised August 2026 bill (816 kWh, RM396.37) is the primary check. It lists every line:

• energy at 27.03 sen;
• AFA split 447 kWh at 3.59 sen and 369 kWh at 3.80 sen;
• capacity and network;
• the RM10 retail charge in the taxed column;
• the rebate at −2.5 sen on all 816 kWh (−RM20.40);
• 8% sales tax on the 216 kWh above 600 (RM8.68);
• the 1.6% levy (RM5.47).

The formula above reproduces it to the sen.

Two findings from checking the formula against real bills:

• The 1,500 kWh tier is a cliff. Crossing it re-prices every unit, adding about RM164 in one step. This reproduces a published domestic bill: RM632.72 at 1,226 kWh and RM987.55 at 1,576 kWh, to within RM2–3. Charging the higher rate only on units above 1,500 would give RM825 for the second bill.
• The rebate uses one band rate for all units, not a sum across bands. The August bill states it outright (−RM0.025/kWh on all 816 kWh). A Gemini report's band-by-band method is RM85–105 out on our June–August bills.

Above the cliff, each extra kWh costs about RM0.636 all-in (≈15.6 US¢). This replaces the RM0.47 and RM0.58 assumptions used earlier, which left out the levy and sales tax.

Household base (our bills, before the nodes)#

That averages about 740 kWh and RM343 a month, on the lower tier.

Node draw (estimated until metered)#

• Per node: about 0.1 kW idle and about 0.9 kW flat out; the worst case in Capacity is 875–930 W.
• At the expected 50–70% utilisation: 0.50–0.66 kW average.
• Air-conditioning: removing that heat adds about 31% (an inverter unit at about COP 3.2).
• Result: 478–631 kWh per node per month.

Projected monthly bill (household at 740 kWh, September AFA)#

What the nodes add:

• The first three nodes add about RM1,020–1,320 a month. About RM110 of that is the household's own use moving to the higher rate once the premises crosses 1,500 kWh.
• Each further node adds a flat RM304 (50%) to RM402 (70%).
• Across the household's own range (650–820 kWh), five nodes cost RM1,917–2,512 a month.
• At December's forecast surcharge, bills are about 3% higher.

This is the running cost each node has to earn back. The ROI the next fleet waits on is measured against it.

Supply headroom on single-phase#

• Load: five nodes flat out draw about 4.6 kW, or about 6 kW with cooling.
• Supply: TNB's single-phase domestic scheme is about 12 kVA (hardware record); a Gemini report says 63 A (~14.4 kW), unverified.
• Verdict: four or five nodes fit alongside the household, with little margin when air-conditioners, a water heater and cooking coincide.
• Before adding nodes: confirm the main breaker rating and TNB's supply rating on the meter, then check the branch logging (above) for real peak demand once the nodes run.

Power limits buy headroom, not a smaller bill. At 450 W the 5090 ran about 12% slower warm (overview). Node draw falls about 14%, so the energy per job barely changes (about −2%). Lowering the limit only helps if peak demand ever threatens the supply; the fleet otherwise runs at 575 W or more.

Levers that don't need site changes#

What replaces these estimates: the first TNB bill with the nodes running, and per-node branch metering. Record observed kWh per node and per completed job, and re-run the table.

UPS decision#

A UPS can bridge a short interruption, provide automatic voltage regulation, and give Linux time to drain jobs or shut down. It does not replace correct earthing or surge protection, and batteries are consumables.

Do not use a common 1500 VA/900 W desktop UPS: its watt limit may be below a heavily loaded worker even if the VA label looks large. A reasonable per-node starting class is roughly 2200 VA with at least 1800 W continuous output, pure-sine output, AVR, USB or network shutdown, and replaceable batteries. The Eaton 5PX 2200i G2 is a reference design, not a mandatory model. A recent M-Link price of RM3,790 each means two units cost RM7,580 before batteries and conversion losses—about 9.8% of two nodes at the historical RM38,647 configuration.

That purchase is not justified merely by fear of a rare event. Use this sequence:

1. Install and test the dedicated circuits, residual-current protection, earthing, and switchboard SPD.
2. Make the OS and service reproducible; keep durable data off the local model cache; make jobs retryable.
3. Log at least 30 days of voltage, sag, swell, and interruption data, including storms.
4. Record actual recovery labor, failed jobs, customer impact, and lost gross margin from every event.
5. Buy one appropriately sized UPS per worker when expected interruption cost or service commitments exceed the monthly ownership cost. Protect the ONT, router, and switch as well, because an online server without its network path has little service value.

A central online double-conversion UPS introduces continuous loss, noise, distribution work, and a shared failure point. Revisit it when this becomes a real equipment room with several workers or logging proves the supply is persistently out of tolerance. It is not the current 80/20 answer.

Procurement instruction#

Closed for PC1–PC3 (all ordered with the Vertex; see the node decision record). The decision tree at the end of this section is the live rule for a future node. The IdealTech three-PSU quote request is kept as history.

Ask IdealTech for one written table showing the final installed-system price, exact part number/revision, included 600 W 12V-2x6 GPU cable, stock status, and Malaysian warranty owner for:

1. Corsair HX1200i regular ATX 3.1/PCIe 5.1;
2. Seasonic Vertex PX-1200 current ATX 3.1/PCIe 5.1 revision; and
3. Lian Li SX1200P black.

Then choose the operating preference and apply its gate without another open-ended brand discussion:

Already own a Vertex PX-1200?
└─ Yes: keep it.

Buying a new node? (fleet standard since 2026-09-25)
├─ Vertex PX-1200 in stock, current ATX 3.1 revision, within ~RM500 of HX1200i: choose Vertex PX-1200.
├─ Otherwise: choose HX1200i (exact revision and warranty confirmed).
└─ Either way: add the Pi Pico connector thermal watchdog; use only the PSU's own 16-pin cable.

Do not accept a model-family name alone. The invoice and build record must identify the actual SKU and revision, and the builder must use only that unit's bundled cables.