Analyst predicts: In 2029, enterprise AI token spending may be more expensive than engineer salaries.

Anthropic has only 5,000 employees, yet its computing expenses are 2.3 times its payroll; the median company spends just $137 per engineer per year. This 680x gap is the puzzle this analysis aims to unpack, and whether 2029 leads to one outcome or another remains unknown.
(Prerequisite: From Forcing Employees to Use AI to Fearing Too Many Tokens Burned: More and More Companies Are Tightening Internal AI Usage Quotas)
(Background: Oracle Rarely Reveals Data Center "May Not Be Able to Recover Costs," Oracle's Stock Plunged 40% in June)

Table of Contents

Toggle

• Computing Power Eats Salaries First
• Push and Pull Forces
• Every Company Is Placing Its Bets

Imagine that by 2029, an average company's annual AI bill for one engineer could very well exceed that engineer's salary. This is the conclusion drawn by renowned venture capital analyst Tomasz Tunguz using three scenario models.

When computing costs begin to approach or even surpass labor costs, AI spending is no longer a discretionary tool budget but a structural expense that must be compared with salaries on the same profit and loss statement.

Computing Power Eats Salaries First

This story starts with Anthropic's own books. According to SaaStr, the company currently has only about 5,000 employees but spent roughly $10 billion on inference and training in 2026. That translates to an average computing expense of about $2 million per employee per year. Compared to an estimated total compensation of over $500k from Levels.fyi, computing costs are 2.3 times salaries.

This ratio is unprecedented in the entire software industry; most companies live in a completely different world. According to the Ramp AI Index from June 2026, the top 1% of companies spend about $89k per engineer per year on AI, equivalent to 40% of the salary of a senior engineer earning $224k annually. The median company spends only $137 per engineer per year, virtually zero.

Between the top and the median lies a gap of nearly 680x, and this is the puzzle this analysis most wants to explain: will this gap widen or narrow in the future?

Will the remaining 99% of companies catch up to Anthropic's pace, and if so, how quickly? Tunguz frames the answer with three scenarios: the pessimistic scenario assumes token prices continue to decline until they offset demand growth; the baseline scenario assumes the growth curve for the top 1% gradually slows; the optimistic scenario assumes the overall market reaches Anthropic's current ratio by 2029. Each scenario converts the AI bill into a percentage of the baseline salary of $224k for a senior engineer, assuming salaries grow at about 5% annually:

• 2026: Pessimistic, baseline, and optimistic scenarios all at $90k, 40%
• 2027: Pessimistic $106k (45%), Baseline $164k (70%), Optimistic $258k (110%)
• 2028: Pessimistic $118k (48%), Baseline $259k (105%), Optimistic $444k (180%)
• 2029: Pessimistic $106k (41%), Baseline $363k (140%), Optimistic $596k (230%)

In the pessimistic scenario, the amount actually declines after 2028 because the rate of decline is faster than salary inflation.

Push and Pull Forces

The reason the optimistic scenario holds up lies in increasingly complex AI workflows—agentic workflows.

When you let AI autonomously execute continuous tasks and decide its next steps on its own, rather than simple one-question-one-answer exchanges, the number of tokens consumed is several orders of magnitude higher than in chat mode. Goldman Sachs predicts that token consumption will grow 24-fold by 2030.

On the other hand, according to Epoch AI research, Anthropic and OpenAI generate $14 million and $6.5 million in revenue per employee, respectively, the two highest among the Forbes Global 2000 list. Cost structures ultimately follow revenue structures; companies that can afford to spend are usually the ones that can earn it back.

But the forces pulling toward the pessimistic scenario are equally real—and have been happening for three years. The input pricing for OpenAI's GPT-4 class models has dropped from $30 per million tokens at launch in March 2023 to less than $3 in 2026, a tenfold decline each year for three years.

Additionally, open-source models are approaching frontier levels. DeepSeek-V3 and subsequent versions deliver performance comparable to top-tier closed models at one-tenth to one-thirtieth the API cost. This echoes why the open-source vs. closed-source debate is one of the most important political issues of the AI era: cheap open-source models directly determine whether the pessimistic scenario has a chance to come true.

Companies willing to actively limit usage based on role or workload can also push this curve down on their own, without passively waiting for prices to drop.

Every Company Is Placing Its Bets

What's truly noteworthy in this analysis is not the surface-level conclusion that "AI is expensive," but rather that AI spending is transforming from a discretionary tool budget into a structural expense comparable to labor costs. In the optimistic scenario, a single engineer's AI bill alone is enough to match the median revenue generated per employee by a public SaaS company (about $250k). This is no longer at the scale of a tool cost—it's the scale of another salary.

When computing costs begin to tug and pull against salaries on the same profit and loss statement, companies need to decide in advance which future they are willing to budget for.