Field Note
The AI Boom Has a Math Problem
Enterprise AI demand is real, but the infrastructure buildout is now so large that the market has to prove the revenue, ROI, and usage can catch up.
![AI Boom Problem]
The most useful way to understand the current AI market is not “boom or bubble.” It is **demand versus infrastructure**.
On the demand side, enterprise AI adoption is scaling faster than almost any modern software category. On the supply side, hyperscalers are spending far ahead of current AI vendor revenue, betting that inference, enterprise workloads, and model usage will grow into the infrastructure.
---
1. Enterprise AI Demand
 _caption_
Enterprise generative AI spend rose from **$1.7B in 2023** to **$11.5B in 2024** to **$37B in 2025**. That is roughly **22x growth in two years**.
**Takeaway:** AI is no longer just a proof-of-concept market. Enterprises are allocating real budgets, buying real products, and moving use cases into production. The better question is which parts of the stack capture durable value.
---
2. Enterprise AI Spend Evenly Split
 _caption_
In 2025, enterprise generative AI spend was nearly evenly divided:
| Layer | 2025 Spend | |---|---:| | Applications | $19B | | Infrastructure | $18B |
This matters because the application layer is where users experience immediate productivity gains, while the infrastructure layer powers the models, APIs, training, data, retrieval, and orchestration underneath.
**Takeaway:** Apps are where ROI becomes visible. Infrastructure is where the strategic arms race is happening.
---
3. Infrastructure Buildout Ahead of Current AI Revenue
 _caption_
The infrastructure side is where the “AI bubble” debate gets more serious. Five major hyperscalers are expected to spend roughly **$660B–$690B** on 2026 capex, while pure-play AI vendor revenue is projected to be **less than $35B**.
This does not mean the spending is irrational. Hyperscalers are not only building for OpenAI, Anthropic, and model vendors. They are also building for their own AI products, enterprise cloud workloads, inference demand, and future AI-native applications.
But the gap is enormous.
**Takeaway:** AI demand is real, but the infrastructure investment is being built ahead of revenue. That creates execution risk.
---
4. Startups are winning AI applications, but incumbents still hold infrastructure
 _caption_
Startups are capturing a majority of AI application revenue:
| Category | Startup Share | |---|---:| | Finance + operations AI | 91% | | Sales AI | 78% | | Product + engineering AI | 71% | | AI applications overall | 63% | | AI infrastructure | 44% |
The pattern is intuitive. Startups win where workflows are changing quickly and incumbents are constrained by legacy product surfaces. Incumbents hold up better where trust, reliability, infrastructure depth, and existing enterprise relationships matter more.
**Takeaway:** The best AI startup opportunities are likely in workflow applications, not foundation models.
---
5. Coding is AI’s First Killer Enterprise Use Case
 _caption_
AI coding spend grew from **$550M in 2024** to **$4B in 2025**, roughly **7.3x year over year**.
This is the clearest category where AI has crossed from novelty to daily workflow. Developers can directly feel the productivity improvement through code completion, refactoring, repo-level context, test generation, pull requests, and autonomous coding agents.
**Takeaway:** Coding became the first killer use case because the feedback loop is fast, the ROI is measurable, and the user base is technical enough to adopt new tools quickly.
---
6. Copilots Dominate, Agents Still Early
 _caption_
Despite the agent hype, horizontal AI spend is still dominated by copilots:
| Category | 2025 Spend | |---|---:| | Copilots | $7.2B | | Agent platforms | $750M | | Personal productivity tools | $450M |
This is one of the most important reality checks in the AI market. Enterprises are willing to pay for AI assistance, but they are still more cautious about autonomy.
**Takeaway:** The near-term market is human-in-the-loop AI. The durable wedge is not “replace the employee,” but “prepare 80% of the work so the human can approve, edit, or decide.”
---
7. Healthcare leads vertical AI adoption
 _caption_
Healthcare is the clearest vertical AI proof point. It represents about **$1.5B** of vertical AI spend, or roughly **43%** of the vertical AI market.
The reason is not that healthcare is easy. It is that the pain is obvious: documentation burden, staffing shortages, margin pressure, and complex administrative workflows. Ambient scribes are a strong example because the ROI is easy to understand: less documentation time and more clinician capacity.
**Takeaway:** Vertical AI works best where manual work, regulation, and workflow complexity create obvious economic pain.
---
8. The AI infrastructure race is concentrated among a few hyperscalers
 _caption_
The largest 2026 capex plans are concentrated among Amazon, Alphabet, Meta, Microsoft, and Oracle.
| Company | 2026 Capex Plan | |---|---:| | Amazon | $200B | | Alphabet / Google | $175B–$185B | | Meta | $115B–$135B | | Microsoft | $120B+ | | Oracle | $50B |
This is not just a software race. It is a physical infrastructure race involving data centers, chips, networking, energy, land, permitting, and power availability.
**Takeaway:** AI is becoming an energy and infrastructure story as much as a model story.
---
9. Enterprises are buying AI, not building everything internally
 _caption_
Enterprise AI purchasing behavior shifted sharply:
| Year | Purchased | Built Internally | |---|---:|---:| | 2024 | 53% | 47% | | 2025 | 76% | 24% |
This is good news for AI application startups. Enterprises increasingly want finished workflows, not internal science projects.
**Takeaway:** The market wants products that solve business problems now. “We help you deploy AI” is less compelling than “we handle this painful workflow better, faster, and cheaper.”
---
In Plain Semantics
The current AI landscape is best understood as a race between **enterprise ROI** and **infrastructure overbuild**.
Enterprise demand is real. Companies are spending more, moving from pilots to production, and buying AI tools that solve real workflow problems.
But infrastructure spending is racing ahead of proven demand. Hyperscalers are committing hundreds of billions of dollars to chips, data centers, networking, and power on the assumption that AI usage will keep compounding.
Demand has to grow into the buildout.
If it does, today’s spending will look visionary. If it does not, the industry may end up with more AI infrastructure than the market can profitably absorb.
The question is not whether AI is useful. It is whether demand grows fast enough to justify what is being built.