D. AI spending cycle#

Status: Plan Last verified: 2026-09-26 Canonical for: the Gemini Deep Research prompt for the AI-capex workstream

Feeds decisions 1 (PC4 timing) and 2 (regret trigger). Paste everything in the block below.

CONTEXT
I run a small AI inference business in Malaysia (image and video generation). In September 2026 I bought three consumer PCs, each with one NVIDIA RTX 5090 (32 GB), a Ryzen 9 9950X, 64–128 GB DDR5 and a PCIe 5.0 NVMe drive, for about RM115,000 in total (roughly USD 27,000). They serve my baseline load; bursts and jobs that need more than 32 GB of VRAM go to serverless cloud GPUs. I do not need data-centre redundancy; brief downtime is acceptable because traffic fails over to the cloud.

Amortised over 36 months with local electricity, the owned cost works out to roughly USD 0.40 per GPU-hour when the nodes are busy. My planning horizon is October 2026 to September 2029.

The decisions this research informs:
1. Whether and when to buy a fourth node (now, after memory prices fall, or never).
2. The rental price at which owning becomes the wrong call. The relevant alternative is a dedicated GPU server with fast local NVMe, not per-second serverless.
3. What a used RTX 5090 will be worth in 18–36 months.
4. Malaysia-specific risks to the cost of owning.

TASK: THE AI SPENDING CYCLE
Research whether the AI infrastructure build-out keeps accelerating, plateaus, or enters a digestion phase within 36 months, and what that does to GPU supply and prices. Answer these questions:

1. What are the latest capital-spending figures and guidance for Microsoft, Alphabet, Amazon, Meta and Oracle, plus the large AI labs and sovereign programmes, for 2026 and 2027? How has guidance moved over the last two quarters?
2. What does NVIDIA's latest quarter and guidance say about data-centre demand, backlog, customer concentration and the Blackwell to Rubin transition? What do TSMC's monthly revenue and advanced-packaging (CoWoS) capacity plans say?
3. Is power now the binding constraint? Evidence on grid interconnection queues, GPUs delivered but not yet energised, and data-centre construction delays. Does a power bottleneck keep GPU prices high, or leave GPUs idle and push rental prices down?
4. Is AI revenue catching up with the spending? The latest evidence on AI revenue at the labs and clouds, inference demand growth, and the price per token trend.
5. What are the credible bear cases (for example SemiAnalysis, independent analysts, short sellers) and bull cases? What specific leading indicators separate them: order cuts, lead times, secondary-market prices for H100s, spending guidance revisions, financing stress?
6. Historically (for example the 2000 telecom fibre build-out and the 2018 and 2022 crypto GPU busts), how quickly did an infrastructure glut hit prices, and what happened to the hardware?
7. If a digestion phase comes, what is the likely timing, and which parts of the market feel it first: H100/B200 rental prices, used data-centre GPUs, consumer GPUs, or memory?

Tell me: (a) the probability and likely timing of a digestion phase within 36 months, (b) how it would pass through to consumer GPU prices and to dedicated-server rental prices, and (c) the three to five leading indicators to watch each quarter.

SOURCES
- Today is late September 2026. Prioritise sources published in the last 90 days. Use older material only as a historical baseline, and label it as such.
- Prefer leading and primary sources: independent semiconductor analysts (for example SemiAnalysis), market trackers (for example TrendForce), company earnings calls and transcripts, regulatory filings, price indices, marketplace price histories and sold listings, supply-chain reporting.
- Investment banks, consultancies and auditors are acceptable but usually lag. When you cite them, say so.
- Where credible sources disagree, show both positions with their reasoning. Do not average them.
- Give a URL for every sourced claim.

OUTPUT FORMAT
1. Answer: at most 10 bullets answering the questions above directly, with numbers.
2. Evidence table with columns: claim | number | as-of date | source (URL) | observed / forecast / rumour | leading / lagging.
3. Scenarios for October 2026 to September 2029: base, upside and downside for my position. For each: a rough probability, what happens to the prices that matter to me, and the indicator that would tell me early that this scenario is unfolding.
4. What would change the call: concrete thresholds (indicator, level, where to watch it, how often it updates).
5. Unknowns: what you could not find or could not verify.

Do not give generic investment advice or stock recommendations. Keep the focus on the decisions above.