Executive Summary
AI investment is at historic highs, but funded by debt, not cashflow.
A ~$600 billion annual revenue gap exists to justify today's capex levels.
95% of enterprise GenAI pilots fail to deliver mesaurable P&L impact.
GPU revenue recognition is circular by design, inflating demand.
Indicator
Reading
Read-through
Big-5 AI capex guidance (2026)
~$725 billion
Up 77% from $410B in 2025; largest in corporate history
Hyperscaler debt issuance (2025)
$121 billion
4× the prior decade's annual average
AI direct revenue vs. capex
~$50–60B vs. $725B
$8–10 of investment per $1 of current revenue
Sequoia revenue gap
~$600 billion / year
Capex requires end-user revenue that does not yet exist
Capex/revenue divergence vs. 2001
46% vs. 32%
Already exceeds the excess that preceded the telecom bust
Enterprise GenAI ROI (MIT, 2025)
95% see zero P&L impact
Adoption is broad; value realisation is not
Amazon Q1 FY2026 free cash flow
Down 95% YoY to $1.2B
Capex consuming all operating surplus
AI share of US GDP growth (H1 2025)
~92%
Macro story now depends on AI capex continuing
Datacenter capacity: announced/built
12GW / 5GW (2026)
5-year grid transformer backlog is the constraint
CAPE ratio (S&P 500, Jul 2026)
~41x
Second-highest in 125 years; dot-com peak ~44–45x
Source: Ametra Research
01
Key Takeaway
AI capex is now the primary engine of US growth. A slowdown threatens headline GDP, not just tech stocks.
● Before
● After
02
The $600 Billion Question No One Is Answering
$600B
Annual revenue gap to justify today’s capex — Sequoia
95%
Of organisations see zero P&L impact — MIT Project NANDA
46%
Capex/revenue divergence vs. 2001 (32%) — Allianz
$500M
Spent by one company in a single month on Claude Code
$5 / $30
GPT-5.5 token cost per million input / output tokens
03
The Circular Flow of GPU Revenue
Hyperscaler / Cloud Provider
Prepay / Secure Compute
Nvidia Ships GPUs
Revenue Recognised
More Capex Commits
04
Caution — priced for perfection
Nvidia, GPU-pure plays
Revenue real but circular; 2 customers = 39% of revenue; depreciation risk unpriced.
Caution — FCF collapse
Amazon, Alphabet, Oracle
Capex consuming 90–95% of FCF; credit repricing underway at Oracle.
Caution — IPO liquidity risk
Existing Mag-7 holdings
SpaceX/OpenAI/Anthropic listings absorb $200B+ from portfolios already long AI.
Watch — application layer
Enterprise SaaS with real AI revenue
Genuine end-user demand, not an infrastructure bet.
Relative caution — neocloud
CoreWeave, Nebius
GPU-backed debt; revenue lags capex 2:1; first maturities 2026–27.
Physical bottleneck
Grid, Transformers
5-year transformer backlog; only 5GW of 12GW building; 20% at delay risk.
Macro overhang
US equity market broadly
AI capex = 92% of US GDP growth; any slowdown is macro, not sectoral.
“
05
Ametra’s Read
We believe AI is transformative, but current valuations assume monetisation, infrastructure and deployment will scale together. They are not. Enterprise AI projects continue to see high failure rates, hyperscalers are committing $725 billion annually while power and transformer shortages delay capacity, and the industry still faces a $600 billion AI revenue gap. History offers a useful reminder: the fibre networks built during the dot-com era eventually became indispensable, but many of the companies that financed them did not survive. The technology may win—but today's valuations still need to prove they can.
