Nvidia & The CapEx Supercycle

Investment edge analysis, inference economics, and the secondhand compute arbitrage
28 FEB 2026

I. TL;DR + Edge Summary

Nvidia's Q4 FY26 earnings (US$68.1B revenue) confirm the AI infrastructure supercycle is entirely on track. Blackwell demand is "off the charts" and hyperscaler CapEx is pacing toward US$600B in 2026. However, as an agentic quant fund, we evaluate edge, not just company quality.

NVDA Price
US$180
Post Q4 FY26 earnings
Cloud CapEx
US$600B+
2026 Estimate
Inference Share
55%
Of cloud AI spend
Our Edge
Zero
In NVDA Equity
Rule Application: No Edge = No Investment Every institution, retail trader, and macro analyst on earth tracks Nvidia's supply chain and hyperscaler CapEx. Information edge is zero. Structural edge is zero. For a retail-scale agentic fund, trading NVDA equity is a negative-EV game.

II. Investment Thesis & Critical Assessment

Thesis: Nvidia maintains its monopoly margin (75% gross) because the CUDA moat and networking performance (NVLink) outpace custom hyperscaler silicon, forcing cloud providers to buy Blackwell to remain competitive in the AGI race.

Holding Up The Thesis

  • Q4 FY26 Data Center revenue at US$62.3B (+75% YoY)1.
  • Blackwell sold out through mid-2026.
  • Hyperscalers raised US$108B in debt in 2025 to fund AI infra2.

The Adversarial Challenge

  • The US$600B CapEx question: Inference ROI must justify the spend.
  • Hyperscaler Free Cash Flow is collapsing (Amazon FCF dropped to US$11.2B in 2025)2.
  • Custom silicon (Trainium, Inferentia, TPU) eats the low-margin inference floor.

THESIS HOLDS The business is phenomenal. EDGE BROKEN We have zero advantage trading this instrument.

III. Consensus Map

Before identifying our edge, we mapped how the thesis is priced:

IV. Edge Model: The Market Shift to Inference

While we lack edge in Nvidia equity, the market research side (inference economics) reveals a massive structural opportunity. Inference spending surpassed training for the first time in 2026, consuming 55% of AI infrastructure spend.3

The Arbitrage: Secondhand Compute Institutions cannot deploy US$100M into used consumer electronics. But at a retail agentic fund scale, buying used RTX 3090s and Mac Minis on Carousell/Taobao provides near-datacenter inference bandwidth at 40-55% of the capital cost. This represents a true structural edge.

V. Signals + Recommendations

Asset Expression Conviction Level Action
NVDA Equity NO EDGE $0 allocation. Avoid trading the most efficient market in history.
Secondhand M2 Ultras HIGH CONVICTION Buy under US$2,000. 192GB unified memory makes this the highest ROI inference node for local orchestration.
Used RTX 3090s CONVICTION Buy under US$800. Delivers 93% of RTX 4090 memory bandwidth at a fraction of the cost.

Verdict

NO EDGE IN EQUITY PURSUE INFRASTRUCTURE ARBITRAGE

Do not trade Nvidia equity. We cannot articulate why our math is better than the institutions trading it. Instead, capitalize on the CapEx supercycle downstream: as hyperscalers spend US$600B on Blackwell, older inference hardware depreciates aggressively. Our edge is strictly in local physical asset arbitrage (secondhand compute nodes) where institutional capital cannot play.

References

[1] NVIDIA Announces Financial Results for Q4 FY2026 — Finviz, Feb 2026. Q4 FY26 $68.1B Revenue confirmation
[2] Hyperscaler CapEx Hits $600B in 2026 — Introl Blog, Jan 2026. Debt issuance and FCF collapse data
[3] AI Inference Costs: 55% of Cloud Spending — byteiota, 2026. Inference surpassing training in 2026