INVESTMENT THESIS 2026

The AI Infrastructure Supercycle

$650B in hyperscaler capex. A 4% monetization rate. Three investable layers across chips, power, cloud, and software — from structural winners to the IPO wave ahead.

$0
2026 Hyperscaler Capex
$0
NVIDIA Market Cap
0
AI Monetization Rate
Explore

The AI Spending Stack

The Big 5 hyperscalers will spend ~$650B in 2026, up from $241B in 2024. Three-quarters flows directly into AI infrastructure across four layers.

AI Market Layers (2026E)

Hyperscaler AI Capex (2026E)

Core Thesis

We are in a multi-year AI infrastructure supercycle with three investable layers — picks-and-shovels infrastructure (highest visibility), platform/cloud (high growth, leverage risk), and application software (longest duration, monetization proving out). Power availability is the #1 constraint.

NVIDIA Dominance & the Custom Silicon Threat

NVIDIA holds 86-97% discrete GPU share with $187B TTM revenue. But custom silicon is growing from 37% to 45% of AI chips by 2028.

AI Chip Market Share

NVIDIA Architecture Roadmap

Blackwell
Current — B200/GB200, 4-5x inference over H100
2024-25
Rubin
HBM4, next-gen interconnect, ~60% TSMC CoWoS allocation
Mid-2026
Feynman
Next paradigm shift, details TBD
2028

HBM Memory: The Structural Margin Upgrade

2025 Market
$38B
2026E Market
$58B
+53% YoY
SK Hynix Share
62%
Micron Share
21%
Best US-listed play
HBM Gross Margins
~70%
vs 30-40% commodity DRAM
HBM4 Timing
H2 2026
Rubin platform

Power Is the New Bottleneck

US data center power demand grows from 40 GW to 134 GW by 2030. Grid spending needs ~$720B. Transformer lead times stretch 2-4 years.

US Data Center Power Demand (GW)

Nuclear Deals for Data Centers

Microsoft / Constellation
TMI restart, 20-year PPA
837 MW
Amazon / Talen
Susquehanna through 2042
1,920 MW
Meta / Vistra
Nuclear fleet deal
2.6 GW
Meta / Oklo
SMR, first phase ~2030
1.2 GW

Power Equipment Backlogs

2026: The Inference Crossover

Inference spending exceeds training for the first time at 55% of AI compute. Token prices have fallen 200x since January 2024.

Training vs Inference Spend

Token Price Deflation

200x
price decline per year since Jan 2024
$60
GPT-4 equiv (2024)
$0.40
per M tokens (2026)
Revenue growth requires volume to massively outpace price compression

Hyperscaler AI Revenue

Azure AI Run Rate
$26B
Doubled in one fiscal year
AWS Trainium
Multi-B$
150% QoQ growth
GCP Vertex AI
20x YoY
Usage growth
CoreWeave Backlog
$55.6B
$863M GAAP loss, 6yr depreciation

The Monetization Proof Point

Foundation models are concentrating into 3-4 winners. Enterprise AI is growing 20-60% YoY. But the 4% problem looms large.

Foundation Model Valuations

Enterprise AI Traction

Palantir AIP
Guided $7.2B FY2026 (+61%)
189x P/E
Salesforce Agentforce
9,500+ paid deals
330% ARR
GitHub Copilot
42% market share, 20M users
$7.4B mkt
ServiceNow AI ACV
Targeting FY2026
$1B

4%

The Monetization Problem

Only ~$25B in direct AI service revenue vs $600B+ infrastructure spend. 42% of companies abandoned AI initiatives in 2025. The market is shifting from promise phase to proof phase.

The Investable Universe

From structural winners to speculative bets. Filter by conviction tier to explore the full stack.

All
Core Holdings
High Growth
Opportunistic
Private / IPO
Avoid

The AI IPO Wave

The largest AI IPO cycle in history is building. Private valuations at 25-30x revenue vs public SaaS at 6x — the gap must resolve.

Company Valuation Timeline ARR / Revenue Key Signal

Catalyst Calendar

Key dates that could move the sector, from chip launches to IPO filings and energy milestones.

What Could Go Wrong

Ten risks that could derail the thesis, from capex deceleration to the token price death spiral.

Recommended Watchlist

Build core infrastructure positions now. Evaluate the IPO wave at pricing. Watch for the 4% monetization ratio to improve — it is the single most important signal for the thesis.

Research compiled February 2026. Not investment advice. See sources below.

Sources & References

Company Filings & Earnings

  • NVIDIA quarterly earnings, 10-K/10-Q filings (FY2025-2026)
  • TSMC quarterly earnings calls, monthly revenue reports
  • Broadcom FY2025 earnings, AI revenue disclosures
  • AMD MI300/MI400 product announcements, quarterly filings
  • Micron Technology earnings, HBM revenue breakouts
  • ASML quarterly reports, order backlog data
  • Arista Networks quarterly earnings, 800G port data
  • Corning Incorporated earnings, Springboard plan updates
  • Ciena FY2025 earnings, FY2026 guidance
  • Vertiv Holdings earnings, backlog disclosures
  • Eaton Corporation earnings, Boyd Thermal acquisition
  • Constellation Energy earnings, Microsoft TMI PPA
  • CoreWeave S-1 filing, quarterly reports post-IPO
  • Palantir Technologies FY2025 earnings, AIP metrics
  • Salesforce Agentforce metrics, quarterly earnings
  • ServiceNow AI ACV targets, quarterly disclosures
  • Equinix and Digital Realty quarterly earnings
  • Quanta Services earnings, backlog disclosures

Industry Research

  • MarketsAndMarkets — AI chip, cooling, and data center market reports
  • Grand View Research — data center cooling, edge AI market sizing
  • Mordor Intelligence — liquid cooling, HBM market analysis
  • IDC — AI PC market forecasts, server market data
  • Gartner — enterprise AI adoption, agentic AI projections
  • Dell'Oro Group — data center capex, networking market share
  • Goldman Sachs — AI infrastructure supercycle research
  • UBS — HBM market share estimates, TSMC CoWoS analysis
  • MIT/RAND — AI project failure rate studies
  • Similarweb — chatbot market share data (Jan 2026)

Regulatory & Government

  • SEC filings (10-K, 10-Q, S-1) for all public companies listed
  • US Bureau of Industry and Security (BIS) — export control actions
  • CHIPS and Science Act disclosures, Intel government stake
  • NRC — NuScale SMR certification documents

Financial News & Analysis

  • Bloomberg, Reuters, CNBC — earnings coverage, IPO pipeline
  • The Information — private company valuations
  • Sequoia Capital — AI market analysis
  • J.P. Morgan, Bank of America — sector research notes