Hardware and compute market research, September 2026#
Status: Reference Last verified: 2026-09-26 Canonical for: the external market research behind the owned-node decisions: prompts, results and synthesis
Start with buy-now-first-principles.md, the analysis of whether to
buy more nodes now, built from our node's parts. synthesis.md covers the rest of the
Gemini material: owned versus rented cost, Malaysia risk, and how far to trust the reports. Raw Gemini Deep Research results, kept verbatim, are
in results/. Two of them (B and E) embed LaTeX-rendered images as base64.
Five Gemini Deep Research prompts, one per workstream. Each prompt is self-contained: paste it as is.
The decisions this research serves#
Reframed 2026-09-26, after the results came in: owning compute is permanent. Decisions 2 and 3 below
(a rent-versus-own regret trigger and a resale floor) are withdrawn. The research now feeds the
question of how many more nodes to buy now, given where prices are heading. See
synthesis.md. The original list is kept, because it is what the prompts asked.
- PC4 timing: buy now, wait for memory prices to fall, or never.
- Regret trigger: the rental price at which owning becomes the wrong call. The yardstick is a dedicated GPU server with local NVMe, against our owned all-in cost of about $0.40 per GPU-hour over 36 months.
- Resale floor: what a used RTX 5090 is worth in 18–36 months.
- Malaysia-specific risk: anything local that changes the owned-node cost.
The horizon is 36 months (October 2026 to September 2029), matching how the purchase is amortised.
Workstreams#
| File | Workstream | Feeds decisions |
|---|---|---|
a-memory-cycle.md |
DRAM, HBM, GDDR7 and NAND prices and capacity | 1, 3 |
b-consumer-gpu.md |
RTX 5090 supply, street and used prices, next generation | 1, 3 |
c-gpu-rental.md |
GPU rental prices, dedicated servers, neocloud health | 2 |
d-ai-capex-cycle.md |
Hyperscaler AI spending, power, digestion signals | 1, 2 |
e-malaysia.md |
Electricity tariffs, AI chip permits, duties, ringgit | 1, 4 |
Left out on purpose: image and video model efficiency. That's our own domain, and we answer it from Hugging Face directly.
How the prompts are built#
The context and output blocks every prompt repeats are kept in _shared-blocks.md.
- Anonymous framing. Each prompt describes the buyer as a small AI inference operator in Malaysia running consumer GPUs. It carries no company, product or invoice details.
- Leading sources first. Each prompt asks for data from the last 90 days and names independent analysts, primary price data, earnings calls and filings ahead of bank and consultancy outlooks, which lag.
- One output contract. Every prompt asks for the same shape of answer, so the five results merge
cleanly:
- a short answer first;
- an evidence table in which each claim is dated, sourced, and tagged observed, forecast or rumour and leading or lagging;
- three scenarios with rough probabilities;
- thresholds that would change the call;
- the unknowns it couldn't resolve.