The AI investment boom of 2023–2025 created and then partially destroyed significant wealth. NVIDIA is up over 1,000% from 2022. The AI pure-plays that went public in 2024 peaked in late 2024 and have since given back much of their gains as investors priced in the uncertainty of whether capability leads to sustainable revenue. OpenAI reached a $852B valuation. Anthropic is at approximately $60B. The question for 2026 is not whether AI is real — it is which AI investments will produce durable returns.
The Three AI Investment Categories
Infrastructure winners — the picks-and-shovels plays. Companies that sell to every AI company rather than competing with them. NVIDIA is the canonical example. The infrastructure winner category also includes TSMC (advanced chip manufacturing), cloud hyperscalers (AWS, Azure, GCP), and power infrastructure companies. AI data centers consume enormous electricity — companies that solve the power constraint profit from AI growth regardless of which AI company wins.
Application winners — companies that use AI to deliver outcomes customers will pay for. The winners in 2026 are companies that had existing customer relationships and distribution, then added AI capabilities efficiently.
The incumbents — existing large companies adding AI capabilities. Microsoft (Copilot in every Microsoft 365 product), Google (Gemini in Search and Workspace), Amazon (Q in AWS and Alexa), Meta (AI across the ad stack), Apple (on-device intelligence).
The Infrastructure Bet: Still the Most Durable
Infrastructure investing in 2026 still looks like the most defensible play. NVIDIA’s position in AI training and inference hardware has a structural moat: CUDA has accumulated 15 years of optimized libraries, frameworks, and developer knowledge. A competitor that matches NVIDIA’s hardware at a lower price still has to replicate the software ecosystem. That takes years.
TSMC is the most underappreciated infrastructure play. Every AI chip — NVIDIA, AMD, Google, Apple, and the custom silicon from every hyperscaler — is manufactured by TSMC. TSMC’s advanced packaging (CoWoS, SoIC) is a bottleneck that no competitor can replicate in the near term.
Application Companies: Separating Signal From Noise
The application layer is where most creative destruction will happen — and where most investment losses will occur. Questions to ask about any AI application company:
What is the retention rate? AI features that drive initial adoption need to show 6-month retention cohorts.
Is there a sustainable price premium? Customers will pay more for AI-powered outcomes if the outcomes are measurably better.
What is the cost of inference? A company selling AI-powered services at prices that sound high but are being eroded by API costs is not a durable business.
Does the company have proprietary data? The best AI application companies have data assets that competitors cannot replicate.
The Specific AI Companies Worth Watching in 2026
Palantir — one of the most operationally successful AI application companies. Their AIP platform combines government and commercial data infrastructure with AI capabilities. Revenue growth and deal sizes have been consistently beating expectations.
Microsoft — Copilot is inside every Microsoft 365 product. Azure OpenAI service gives Microsoft a revenue stream from every company that wants to build on GPT models without going direct to OpenAI.
Salesforce — Einstein AI within the Salesforce CRM is one of the most underappreciated AI integration stories. The data advantage Salesforce has is a genuine moat for AI features that competitors cannot replicate.
AMD — the credible alternative to NVIDIA for cost-sensitive AI deployments. MI350X chips are genuinely competitive at significantly lower cost.
Anthropic — the $60B question mark. Claude Opus 4.7 is the most capable model in many benchmarks. Anthropic’s path to profitability depends on converting model capability into enterprise revenue.
The Model Releases That Changed the Investment Landscape in 2026
GPT-5.5 (April 23, 2026): OpenAI’s most capable model to date, with 1,050,000 token context. GPT-5.5 Instant became the default ChatGPT model globally on May 5. OpenAI’s path to profitability — reportedly burning $7B+ annually against $4-5B revenue — depends on enterprise adoption at scale.
Claude Opus 4.7 (April 16, 2026): Anthropic’s flagship with the highest documented reasoning capability. The June 15 deprecation of Sonnet 4 and Opus 4 means all production systems need to migrate now.
Gemini 3.1 Pro (February 19, 2026): Google’s answer to the cost and capability pressure. The 2M token context and significantly improved reasoning make it the tool for enterprise-scale document processing.
DeepSeek V4-Pro / V4-Flash (late April 2026): The cost disruptor. DeepSeek V4-Flash is the cheapest frontier-class API available, significantly changing the economics of high-volume AI applications.
The Risks That Are Not Fully Priced In
GPU oversupply: The amount of AI infrastructure being built is enormous. Hyperscalers building new data centers at the current pace could result in significant compute oversupply by 2027-2028, pressuring NVIDIA’s pricing power.
OpenAI’s profitability timeline: OpenAI is reportedly burning $7B+ per year. If GPT-5.5 doesn’t meaningfully accelerate enterprise adoption, the $852B valuation faces pressure.
EU AI Act enforcement: Conformity assessments and compliance costs are a real burden for AI companies targeting the EU market.
The Honest Framework for Individual Investors
If you want AI exposure without stock-picking risk: a combination of NVIDIA (infrastructure), a cloud ETF, and an enterprise software ETF gives diversified exposure to the AI ecosystem without single-company risk.
If you want to pick individual companies: apply the same framework as any investment. Does the company have a durable competitive advantage? Is management competent? Is the valuation reasonable relative to actual earnings power?
Do not invest in AI pure-plays at peak valuations. Early-stage technology trends create speculative bubbles. The lesson of 2024-2025 is not that the technology is fake — it is that peak valuations disconnected from business fundamentals get corrected.
→ What Are AI Agents? A Plain-English Guide to Autonomous AI in 2026
→ The Best AI Models in May 2026: Definitive Rankings for Builders
Disclaimer: Not financial advice. Do your own research. Last updated May 12, 2026.