Artificial intelligence is the defining technology investment theme of the 2020s. The release of ChatGPT in November 2022 triggered a wave of enterprise AI adoption, infrastructure spending, and startup formation that has reshaped every sector of the technology industry — and begun to affect finance, healthcare, legal, and media.
But “AI” is not one business. It encompasses foundational infrastructure (the chips and data centres that run models), foundation models (the large language models themselves), enterprise AI software (tools that let businesses deploy AI into their operations), and AI-native applications (products built from the ground up on AI capabilities). Each layer has different economics, competitive dynamics, and investment characteristics.
The global AI market is expected to reach $1.3 trillion by 2030, up from roughly $200 billion in 2024 — one of the fastest expansions of any technology sector in history.
The AI Value Chain
Layer 1: AI Infrastructure
The bottom layer consists of the chips, data centres, networking, and power infrastructure required to train and run AI models. NVIDIA dominates this layer with its GPU architecture and CUDA software ecosystem.
Training a frontier AI model requires tens of thousands of high-end GPUs running for months. Inference — serving the model to millions of users — requires even more sustained compute. This has driven an unprecedented wave of data centre capital expenditure:
- Microsoft committed $80 billion in data centre capex for calendar 2025
- Amazon Web Services, Google Cloud, and Meta have made similar multi-year commitments
The infrastructure layer benefits companies in the Semiconductors and Cloud Computing sectors more than pure-play AI companies.
Layer 2: Foundation Models
Foundation models are the large language models, image models, and multimodal models trained on internet-scale data. OpenAI (GPT-4o, o3), Anthropic (Claude), Google DeepMind (Gemini), and Meta (Llama) are the dominant players.
Most foundation model providers are either private (OpenAI, Anthropic, xAI) or divisions of large public companies (Google, Meta, Microsoft’s investment in OpenAI). Direct public investment exposure to foundation models is limited.
Layer 3: Enterprise AI Software
Enterprise AI software helps businesses deploy AI into workflows — sales automation, customer service, code generation, document analysis, and process automation. This is where most public AI pure-plays operate.
Companies like Palantir (AI-powered data analytics and decision software), C3.ai (enterprise AI applications), and SoundHound AI (conversational AI) operate here.
Layer 4: AI-Native Applications
AI-native apps are built entirely on AI capabilities — they could not exist without foundation models. Examples include AI coding assistants, AI-generated media tools, and AI-powered drug discovery platforms. Most are private or early-stage.
Revenue Models in Artificial Intelligence
| Model | Description | Examples |
|---|---|---|
| Infrastructure rental | GPU/compute time billed per hour | CoreWeave, Lambda Labs |
| API access | Per-token or per-call pricing | OpenAI API, Anthropic |
| Enterprise SaaS | Annual subscription for AI software | Palantir, C3.ai, ServiceNow AI |
| Platform + consumption | Base subscription + usage overage | Datadog, Snowflake (AI features) |
| Royalties / licensing | Per-use licensing of AI models | SoundHound AI |
The Unit Economics Challenge
Most AI companies face a fundamental tension: inference costs are high relative to willingness-to-pay. Foundation model providers spend enormous sums on GPU compute to generate each response — and the competitive pressure on pricing (with open-source models like Meta’s Llama available for free) makes profitability difficult.
The exception is infrastructure — NVIDIA, which supplies the GPUs, and the hyperscalers, which rent them out, capture the majority of AI economics today. Enterprise software companies capture value if they can achieve genuine workflow lock-in (Palantir) rather than commodity API access.
Key Companies in Artificial Intelligence
AI Infrastructure (chip-level):
- NVIDIA — see Semiconductors sector
- AMD — GPU competitor to NVIDIA
- TSMC — manufactures all leading AI chips
Enterprise AI Software and Data:
- Palantir — AI-powered decision intelligence for government and enterprise
- C3.ai — enterprise AI application suite
- Nebius Group — European AI cloud infrastructure
- BigBear.ai — AI analytics for defence and intelligence
AI-Native Applications:
- SoundHound AI — conversational AI for automotive and hospitality
- Rigetti Computing — quantum-AI hybrid computing
- D-Wave Quantum — quantum computing for optimisation
AI-enabled Hyperscalers:
- Microsoft (Azure OpenAI, Copilot) — see Cloud Computing
- Alphabet (Google Gemini, GCP) — see Cloud Computing
- Meta (Llama, Meta AI) — see companies coverage
Key Metrics for AI Companies
Revenue Growth Rate
AI software companies at early scale should be growing revenue 30%+ annually. Slowing growth below 20% signals limited market penetration or intensifying competition.
Gross Margin
AI infrastructure companies run 60–75% gross margins at scale. Enterprise AI software companies should target 70–80%. If gross margins are below 50% for a software company, it signals high compute COGS eating into profitability.
Rule of 40
The Rule of 40 (revenue growth rate + operating margin ≥ 40%) is a useful benchmark for SaaS and AI software companies. It balances growth investment against profitability. Best-in-class AI software companies score 50–70+.
Government vs Commercial Revenue Mix
For AI analytics companies like Palantir and BigBear.ai, the split between government contracts and commercial enterprise customers matters. Government revenue is sticky but slow-growing; commercial revenue is volatile but scalable.
Remaining Performance Obligations (RPO)
RPO is the total value of contracted but unrecognised revenue — a forward revenue indicator. Rising RPO signals strong bookings momentum ahead of reported revenue.
The Competitive Dynamics of AI
Open Source vs Closed Models
Meta’s decision to open-source its Llama model family has commoditised the foundation model layer. Any company can now deploy a capable large language model without paying OpenAI or Anthropic. This is deflationary for closed model providers and beneficial for enterprises.
AI as a Feature vs AI as a Product
The most successful AI companies will be those where AI is deeply embedded into workflows with high switching costs — not those selling AI as a standalone feature that can be replicated by incumbent vendors. Microsoft embedding Copilot across Office 365 has proven more durable than standalone AI chat products.
GPU Availability and Capex Arms Race
Access to GPUs remains a constraint for AI companies below hyperscaler scale. Companies that secured long-term GPU contracts (CoreWeave, Lambda Labs) or built their own silicon (Google TPUs, Amazon Trainium) have structural cost advantages.
Key Comparisons
Related Glossary Terms
- Gross Margin — the primary quality metric for AI software companies
- Operating Leverage — why AI software margins should expand with scale
- Stock-Based Compensation — typically high for AI-first companies
- Capital Expenditure — the massive infrastructure investment driving AI economics