Skip to content

P1-C8 · What AI Applications Look Like Today

Core Takeaway

Applications determine the next 3 years of the industry chain — without ROI realization, capex is unsustainable. This is the biggest trigger for the 5th winter.

AI Industry Knowledge — History → Technology → Industry Chain → Business → Applications → Geopolitics

P1-C8 (Part 1, Chapter 8). After this chapter, you'll understand the 4 major AI application layers + current revenue reality, and judge whether the hyperscaler 2026 combined capex $600-725B (Yahoo / CreditSights / MUFG 2025-12; depends on AI-only vs total scope, ~75% AI-related) has ROI realization to support it.

1. The Problem: Hyperscaler 2026 Combined Capex $600-725B — Has the Application Layer Earned It Back?

In C7, you saw the industry chain value capture — NVDA / ASML and other true kings with 75% gross margins. But that's the shovel sellers making money.

Are the gold miners making money? If the application layer (OpenAI / Enterprise AI / Vertical SaaS) doesn't deliver revenue, the hyperscalers' 2026 combined capex $600-725B (Big 4 = MSFT + GOOGL + AMZN + META; 2025 ~$290-400B → 2026 accelerating, Yahoo / CreditSights 2025-12; ~75% AI-related) is burning cash as subsidies — unsustainable long-term. This is historical winter step 4 (expectation-reality gap → capital retreat).

C8 gives you the 4 major application layers + revenue reality for each + how to assess ROI realization in your thesis.

2. The Solution: 4 Major Application Layers + Revenue Reality

Layer Representatives 2026 Est. Revenue Realization Status
Foundation Models OpenAI · Anthropic · Google Gemini · Meta Llama · DeepSeek OAI ~$2B/month (~$24B annualized), enterprise >40% of revenue (per OpenAI official 2026/03/31 — R6 Fix 3) · Anthropic run-rate $30B+ (Apr 2026, Anthropic official/VentureBeat) Explosive growth, not profitable
Enterprise SaaS MSFT Copilot · Salesforce Einstein · ServiceNow · Adobe Firefly Microsoft AI business run-rate >$37B (FY26 Q3, official; Copilot-only ARR not separately disclosed) Steady growth, near profitability
Vertical AI Cursor (coding) · Harvey (legal) · Glean (search) · Clay (sales) · Notion AI $50M-500M / company Early stage, diversified
Consumer / Agentic ChatGPT · Claude · Perplexity · Character.ai · Claude Code · Devin ChatGPT >900M WAU + >50M subscribers, ~$2B/month (OpenAI official 2026/03/31) Moderate C-side stickiness, agentic new growth

Key insight: AI applications are still early stage. Similar to dotcom 1998-1999 — application layer revenue is growing, but far from enough to support capex. The next 2-3 years are the critical validation window.

3. How It Works: Detailed Breakdown of the 4 Application Layers

3.1 Foundation Models

  • OpenAI: ChatGPT >900M WAU + >50M subscribers (per OpenAI official 2026/03/31); ~$2B revenue/month (~$24B annualized), enterprise >40% of revenue (per OpenAI official 2026/03/31). Valuation $852B post-money (2026/03/31 OpenAI official, $122B committed capital — R6 Fix 3: supersedes 2025/10 ~$300B round)
  • Anthropic: Claude. revenue run-rate $30B+ (Apr 2026, Anthropic official/VentureBeat; was $14B run-rate in Feb 2026). Valuation $380B post-money (Feb 2026 Series G, led by GIC/Coatue; prior round $183B Series F Sep 2025)
  • Google Gemini: Internal + API. Embedded in Search/Workspace, not standalone revenue
  • Meta Llama: Open source, no direct revenue (strategic anti-incumbent)
  • DeepSeek: Open source + China market. V3 trained on H800 achieves near GPT-4 capability

Economics: Not profitable. Training one model costs $100M+, API prices are being commoditized (DeepSeek slashed API prices by 90%).

→ Investment implication: The foundation model layer may not be a good investment position (commoditization). **MSFT holds ~27% of OpenAI on an as-converted diluted basis (~32.5% ex-recent rounds), investment value ~$135B (per Microsoft 2025/10/28 disclosure — R6 Fix 2: the "49%" figure circulating earlier referred to historical capped profit share ≠ equity; after the 2025/10 PBC recapitalization the correct figure is 27% diluted; Note: contract terms amended 2026/04/27 v3 — license changed to non-exclusive / AGI provisions removed / IP through 2032 changed to non-exclusive, see LEDGER F-021 v3 + Microsoft blog 2026/04/27**) — this is the indirect play.

3.2 Enterprise SaaS

  • **MSFT Copilot: $30/seat/month embedded in Office 365. Microsoft's overall AI business run-rate >$37B (FY26 Q3, 2026/04/29 official earnings call; was >$10B at FY25 Q1); Copilot-only ARR not separately disclosed** (Microsoft describes Copilot as an ARPU-growth driver, doesn't break out ARR). Microsoft 365 Copilot paid seats 20M+ (2026/04 official). Enterprise penetration rising
  • Salesforce Einstein: CRM embedded with AI agent, 2025 Agentforce
  • ServiceNow: IT workflow + AI agent. Now Assist Q1 2026 ACV $750M (up from $600M 2025), 2026 full-year target raised from $1B → $1.5B (Bloomberg / Seeking Alpha 2026/05/04); avoid stating ARR — ServiceNow reports Now Assist ACV, not a separate AI ARR
  • Adobe Firefly: Image generation, embedded in Creative Cloud

Economics: High margins (60%+), sticky customers (enterprise SaaS is inherently sticky), AI is an upsell. Close to truly profitable application layer.

→ Investment implication: This is the most stable application layer thesis right now. MSFT / CRM / NOW / ADBE all have multi-year growth visibility.

3.3 Vertical AI

  • Cursor (Anysphere): AI coding IDE. valuation $29.3B post-money (Series D Nov 2025, CNBC; in talks for $50B+ Series E Apr 2026, TechCrunch). ARR $2B (Feb 2026 reported, TechCrunch/tech-insider). Claude Code is direct competition
  • Harvey: Legal AI. valuation $11B ($200M round Mar 2026, led by GIC + Sequoia, Bloomberg/CNBC; prior rounds $5B Series E Jun 2025, $8B Dec 2025)
  • Glean: Enterprise search. valuation $7.2B (Series F Jun 2025, $150M led by Wellington; ARR $200M by Dec 2025 — Fortune)
  • Clay: Sales enrichment. valuation $3.1B ($100M Series C Aug 2025, CapitalG-led, Crunchbase/BusinessWire; 2026 tender offer ~$5B reported)
  • Notion AI: Notes + AI

Economics: Early-stage SaaS economics. Each company $50M-500M ARR. Moderate customer stickiness (vertical switching costs are high).

→ Investment implication: Mostly private market. Once IPOs happen (Cursor possibly 2026-2027), this is a new IPO wave. Many are ROI validation for NVDA capex.

3.4 Consumer / Agentic AI

  • ChatGPT: >900M WAU + >50M subscribers (per OpenAI official 2026/03/31), ~$2B revenue/month (~$24B annualized) (OpenAI official 2026/03/31; R6 Fix 3). Plus $20/month + Pro $200/month
  • Claude (Anthropic): consumer MAU not officially disclosed by Anthropic — third-party estimate ~30M MAU (early 2026, with third-party ranges 18M-220M depending on methodology). Enterprise verifiable: Amazon Bedrock 100,000+ customers run Claude per AWS announcement; Anthropic Series F disclosure >300,000 business customers (Anthropic official). Pro/Max + Claude Code
  • Perplexity: Search. Valuation $20B ($200M round Sep 2025, TechFundingNews/PitchBook; early-2026 Series E-6 reported ~$21B). ARR ~$148-200M Sep-Dec 2025; 45M MAU late 2025
  • Character.ai: AI companionship, high stickiness but hard to monetize
  • Devin / Claude Code / Cursor: Agentic — LLM + tool loop, inference compute 10-100x normal chat

Economics: C-side (ChatGPT Plus) high margins but high churn. Agentic is the new growth curve — 1 user session uses 10-100x inference compute.

→ Investment implication: Agentic = inference compute compounding. NVDA Blackwell inference optimization, agentic growth = NVDA's 2nd growth curve.

4. vs C7 — What You Already Know

Dimension C7 Gives You C8 Adds
Value capture ✓ (who has moat) Doesn't explain how long the moat lasts
Application ROI ✗ 4 application layers + revenue reality
Investment implication Know who to hold long-term Know which layer realizes capex ROI fastest: Enterprise SaaS > Vertical AI > Consumer > Foundation

C7 = how profitable the shovel sellers are. C8 = whether the gold miners can keep paying for shovels. Without C8, you don't know how long the capex cycle can last.

5. Try It: Evaluate Whether a Vertical AI Company Can IPO Independently

Task (20 minutes): Choose 1 vertical AI company (recommend Cursor / Harvey / Glean), answer:

Assessment Criteria
Current ARR $100M+ = IPO ready
YoY Growth Rate 100%+ = strong; 50% = borderline
Customer Concentration Top 10 customers < 30% = healthy
Moat Data flywheel? Workflow lock-in? Ecosystem?
Competition Will OpenAI / Anthropic build the same product?

Cursor Example: - ARR $2B (Feb 2026 reported, TechCrunch/tech-insider; ~3x growth YoY from $500M Jun 2025) - Diversified customers (~10K teams) - Moat: VSCode fork + AI coding workflow lock-in - Competition: Claude Code (Anthropic direct) + Copilot (MSFT/GitHub) - Judgment: IPO possible 2026-2027, but facing commoditization from Anthropic / MSFT, long-term moat unproven

Self-check (3 items met → proceed to P1-C9):

  • You can distinguish "API revenue (foundation layer)" vs "application revenue (vertical layer)" — the latter is stickier + higher margin
  • You can explain in 1 sentence why MSFT Copilot is the best-positioned application layer thesis
  • You can identify application layer ROI failure signals (e.g., Salesforce cutting Einstein investment / Adobe Firefly revenue below expectations)

6. What's Next

You can now analyze applications. But application realization is also affected by geopolitics — DeepSeek's surprise attack / export controls / energy constraints all change application layer economics.

→ P1-C9 · US-China + Export Controls + Energy Geopolitics 3 geopolitical lines + case studies.

7. Deep Dive (optional): Foundation Model Economics / Agentic Compute Demand / 5th Winter Trigger Conditions

Click to open 3 deep dives

Foundation Model Economics Truth: OpenAI 2024 revenue ~$5B, cost $9B+ (historical); by 2025/06 annualized $10B (CFO Sarah Friar to Reuters), end-2025 crossed $20B annualized, and Mar 2026 annualized $25B+ (The Information). Main burn: compute (MSFT Azure) + talent + training. → If price commoditization continues (DeepSeek -90% API price), foundation layer doesn't make money. → Who survives: Those with application layer distribution (MSFT via Office embed / Google via Search embed) + strongest brand (OpenAI ChatGPT C-side).

Agentic Compute Demand: Traditional ChatGPT single query ~1K token inference. Claude Code writing 1 feature uses 50K-500K tokens + multi-step tool calls = 100-500x inference compute. → 1 agentic developer = 100 ChatGPT users' inference. → If agentic penetrates 5% of global developers = 10x current inference compute. This is NVDA's 2nd growth curve.

5th Winter Trigger Conditions (your thesis must monitor): 1. Any of MSFT/GOOGL/AMZN capex guide flat or down → AI chain de-rate trigger 2. OpenAI revenue growth rate < 80% YoY → foundation model thesis weakens 3. Microsoft AI business run-rate growth < 50% YoY (FY26 Q3 already >$37B, official 4/29/2026 earnings; Copilot-only ARR not separately disclosed) → enterprise SaaS thesis warning 4. Top Vertical AI companies (Cursor / Harvey) lose major customers → private market valuation reset 5. DeepSeek-like breakthrough + training cost -90% (DeepSeek-V3 official $5.576M training run only per tech report; excludes all-in R&D/data/failed-runs/hardware) → scaling laws thesis questioned

2 out of 5 triggered = warning. 3 out of 5 triggered = 5th winter likely begins.

8. Further Reading (this chapter — what AI applications look like today)

All free sources, aligned with P5 0-paid policy

Classic papers / primary sources:

Wikipedia (3-10 min):

Videos / public lectures:

Blogs / case studies:

Podcasts:

  • Acquired — Microsoft Volume IV (Satya era) — Copilot strategy
  • Lex Fridman — Sam Altman / Dario Amodei / Mira Murati past episodes — App-layer roadmap

Company IR (primary on revenue):

Books (library):

  • Reid Hoffman "Impromptu" (2023, free PDF at impromptubook.com) — Collection of GPT-4 application use cases
  • Ethan Mollick "Co-Intelligence" (2024) — Wharton professor, practical LLM uses at work

Pair with this chapter's self-check:

After OpenAI DevDay + several Cursor blog posts + 1 MSFT quarterly transcript, you should be able to answer "the 4 main application layers' revenue reality" and "can vertical AI IPO independently."