Market research synthesis: owned RTX 5090 nodes, September 2026#

Status: Reference Last verified: 2026-09-26 against the five Gemini Deep Research results in results/ and three spot checks listed below Canonical for: price projections for the owned nodes, the owned-versus-rented cost, Malaysia risk, and the analysis of whether to buy more now

The five raw results are in results/; the prompts are listed in the README. This page uses our real numbers where the reports assumed wrong ones.

Framing corrected the same day (founder, 2026-09-26). Owning compute is a permanent goal. Renting is a bridge, resale value never enters a decision, and we invest now rather than cost-average. The first version of this page had a rent-versus-own "regret trigger" and a resale floor; both are withdrawn.

Should we buy more now?#

Superseded by buy-now-first-principles.md (2026-09-26).

This section used to blend the Gemini forecasts into node-level ranges ("flat to +15%" at 6 months, "±15%" at 12). Those ranges weren't computed from our node's parts. Their GPU decline assumed the next consumer generation (RTX 60) would launch in late 2027, and it has since reportedly slipped to 2028.

The first-principles analysis builds the projection part by part from our invoices, and explains how each part's price is set. It projects the node at +3% to +20% over 12 months, about −5% at 24 months (range −21% to +9%), and only clearly cheaper, likely −20% or more, at about 36 months.

The rest of the questions#

2. Owned cost against renting (a benchmark, not a trigger)#

A monthly dedicated server is paid whether or not it's busy, and so is an owned node. So compare per node per month, not per busy GPU-hour. Utilisation affects both sides equally.

Corrected 2026-09-27: this table first costed electricity at RM0.58/kWh and 0.5 kW with no cooling (about RM210), which understated it.

After 36 months an owned node costs only electricity, while a rental keeps billing.

The rentals we could verify, checked 2026-09-26, with ratios at about 4.1 ringgit per dollar:

Location matters less than the prompts assumed. The control plane is in the US and jobs pull their work, so a US or EU dedicated box is a valid alternative. The deciding factors are resident models on local NVMe and reliable availability.

Withdrawn 2026-09-26: this section used to set a trigger to freeze owned expansion once rentals reached about RM1,300 a month. Cheap rentals now only make lane 3 cheaper while the fleet grows. They're still worth checking once a quarter, to price the bridge.

Price outlook: what the reports agree on#

What the reports agree on:

• 5090 street prices are still rising. In the US the average went from about $3,100 in January to $4,900 in September 2026. Used cards trade at about $3,700–3,800. The reports see no sign of oversupply. Locally, the list prices we recorded are RM20.5–23.3k.
• Memory prices peak in 2027 and break in 2028. DRAM contract prices roughly tripled from late 2025 to mid-2026. New factories come online through 2027–2028: SK hynix M15X and Yongin, Micron Boise, Samsung P4 and P5. The reports forecast a 2–3% DRAM surplus in 2028 and roughly a 50% price fall from the late-2027 peak. Chinese memory makers are held back by equipment export controls and are not expected to flood the market within 36 months.
• The next consumer generation is late 2027 to 2028. The RTX 50 Super refresh is stalled because 3 GB GDDR7 chips cost about 3× more than 2 GB chips.
• AI data-centre spending is still accelerating. 2026 guidance from the big five is about $775–800B. The reports put a digestion phase at about 35% odds, most likely between mid-2027 and mid-2028. It would hit H100/B200 rental prices first and consumer cards last.

The buy-now analysis built on these facts is at the top of this page.

Removed 2026-09-26: a resale section. Resale value never enters a decision, because owning is permanent. The raw reports still contain their resale estimates.

4. Malaysia#

• Electricity (corrected 2026-09-27 after a dedicated research pass, result F and its re-verification).
  • The tariff is a cliff, not progressive blocks. Above 1,500 kWh a month, the 37.03 sen energy rate applies to every unit. With capacity and network charges, the fuel surcharge, the 1.6% levy and 8% sales tax, each extra kWh costs about RM0.636 (≈15.6 US¢). The household base is about 740 kWh, so any node load puts the premises over the cliff.
  • Checked against real bills. The formula reproduces a published bill and our own June–August bills; the full model and projected bills are in the site-power doc.
  • Outlook. RP4 ends on 2027-12-31. Gemini's base path for RP5 rises to about RM0.69 by 2029; its range runs from RM0.56 to RM0.79.
  • Like for like, Malaysia is mid-table. At our tier it's about the same as Texas or Virginia households. It's below the US residential average (18.3¢), and far below Europe, Singapore, Japan and California. It's above Thailand, Vietnam and Indonesia, and above US small-commercial rates in cheap states. Those comparison figures are reported, not verified.
  • Rooftop solar is cheaper than the grid. Panels alone cost about RM0.20/kWh over 25 years, and about RM0.40 with batteries. There's no immediate plan for solar (founder, 2026-09-27): single-phase caps it at 5 kWac.
• Report C's capacity-charge and power-factor penalties don't apply here. They belong to commercial C1/C2 medium-voltage accounts, not domestic supply.
• Chip rules. MITI Directive 1/2025 has a safe harbour for hardware not designed for data centres with TPP (a performance score) under 4,800. The 5090 qualifies, so buying, running and reselling within Malaysia needs no permit.
• Tax. Imports carry 0% duty and 10% sales tax, which is already inside local retail prices.
• Ringgit. The reports say it strengthened from above 4.70 to 4.15–4.25 per US dollar over the last year, a mild tailwind for buying later. I didn't verify this.
• Insurance. A householder policy likely excludes business equipment, so RM115k of hardware may be uncovered. Get an Electronic Equipment Insurance (EEI) or all-risks quote. This is cheap and worth doing now.
• Power quality. Voltage sags on residential feeders are already covered by the power and UPS decision.

How far to trust the reports#

Use their direction, not their exact numbers.

• Few truly leading sources. The prompts named SemiAnalysis, but the reports cite it only through secondary pages. Much of the evidence comes from SEO blogs, Reddit and price trackers.
• Wrong baselines. Report A made up a node parts list: a $3,200 GPU plus $4,100 for motherboard, PSU, case and cooling. Reports D and E based resale on a RM10,500 list price, when we paid about RM20k.
• Figures that failed spot checks. Report C listed GPU Mart at $449 (or $399) a month; the live page says $599 (or $479). C also said no Southeast Asian provider offers dedicated 5090s; IP ServerOne does, on quote. Report B's "MYR 3.05/hr" for IP ServerOne doesn't appear on their page.
• Same thresholds, different numbers. The break-even estimates range from $0.40 to $0.65 per GPU-hour, depending on whether a report credited resale value and SLA. The monthly comparison above replaces them.
• Mislabelled scenarios. Report E labels rising electricity prices plus cheaper cloud as its "upside" case.
• Unverified headline numbers. Micron fiscal Q4 guidance of about $50B revenue at an 86% gross margin, and Anthropic and OpenAI run-rates, appear consistently across the reports. I haven't verified them, and none of the decisions rests on them.

Spot checks (2026-09-26):

• IP ServerOne's GPU server page: prices as listed; the RTX 5090 is quote-only.
• GPU Mart's provider comparison, dated 2026-09-20: $599, or $479 on a 24-month contract.
• The bestvaluegpu 5090 price history returned HTTP 429, so its figures are unchecked.

Indicators worth watching#