What happened
A recent analysis on Tom Tunguz’s site argues that AI spending is becoming a major line item in software economics, with a sharp divide between frontier AI companies and the rest of the market.
The core comparison is straightforward: at Anthropic, the analysis says compute spending is about 2.3 times payroll. Using an estimate of roughly 5,000 employees and about $10 billion in inference and training spending in 2026, that works out to around $2 million of compute per employee per year. The same analysis places all-in compensation for a top AI lab employee at roughly $500,000 or more, meaning infrastructure can outweigh labor by a wide margin at the frontier.
Outside the frontier, the picture looks much more restrained. The analysis cites data showing that the top 1% of companies spend about $89,000 per engineer per year on AI, or roughly 40% of a fully loaded senior engineer salary estimate of $224,000. The median company, by contrast, spends only about $137 per engineer per year. That creates a large gap between frontier AI firms, elite software companies, and the broader market.
To frame where the market may be headed, the analysis lays out three scenarios for 2029:
- Bear: token deflation continues to reduce costs, keeping AI spend relatively contained.
- Base: the leading companies’ spending trajectory slows as adoption spreads but does not explode.
- Bull: the rest of the market converges toward Anthropic-like compute intensity by 2029.
The analysis argues that each scenario maps to a different annual AI bill per engineer, and that the difference could be large enough to reshape software margins and product strategy.
Why it matters
The main significance of the piece is that it reframes AI as more than a feature or a productivity tool. At least for some companies, AI may be turning into an operating cost that competes with headcount itself.
That matters for two reasons.
First, the economics are not uniform across the market. The source describes a huge spread between the most aggressive AI users and typical software firms. If a company is in the top tier of AI adoption, spending on model usage can become a meaningful portion of engineering cost. If it is a median company, AI may still be a relatively minor expense. That difference can affect whether AI is used broadly, narrowly, or only for specific workflows.
Second, the economics can change based on how AI is used. The analysis points to agentic workflows as a potential demand driver because they can consume many more tokens than chat-style usage. It also cites a Goldman Sachs projection of a 24-fold increase in token consumption by 2030, suggesting that usage intensity could rise quickly if software shifts from occasional prompts to automated, multi-step tasks.
At the same time, the source highlights countervailing forces that could keep costs down:
- token prices have fallen sharply over time
- open-weight models may narrow the quality gap at lower cost
- companies can ration usage by role or workload
That creates a tension at the center of the story. If AI capability keeps improving but prices keep falling, usage could expand without crushing budgets. But if demand for agentic systems grows faster than prices decline, AI could become a larger share of software operating costs than many teams expect.
The source also points to revenue-per-employee benchmarks at AI-native companies such as Anthropic and OpenAI, which it says are among the highest in the Forbes Global 2000. That comparison suggests the cost structure and revenue structure may both be moving upward together at the frontier. In other words, some companies are already organized around a business model where extraordinary compute costs are paired with extraordinary revenue generation.
For the rest of the market, the question is whether those economics are repeatable or exceptional.
What to watch
Several signals will help determine which path is more realistic over the next few years.
- Token pricing trends: If frontier model prices keep falling quickly, AI spend per engineer may remain manageable even as usage rises.
- Agentic workload growth: If companies move from simple chat use cases to automated workflows, token consumption could rise much faster than headcount.
- Open-weight model adoption: Cheaper alternatives could pressure proprietary model pricing and reduce dependence on the most expensive APIs.
- AI usage controls: Companies that restrict AI by role, team, or task may keep spending closer to the current median.
- Frontier lab unit economics: The relationship between compute spend, payroll, and revenue at leading AI companies will remain an important benchmark for the sector.
The article’s broader question is not whether AI is useful. It is whether AI spend becomes a normal software expense or a structural cost center that only the most productive companies can afford at scale.
If the bull case plays out, AI bills may rise to levels that rival or exceed the cost of the engineers using them. If the bear case wins, falling prices and cheaper models could keep AI embedded in software without breaking the budget.
The answer will shape how companies buy, build, and budget for AI between now and 2029.
