According to SaaStr, Anthropic's compute expenditure on reasoning and training is approximately $10 billion, translating to roughly $2 million in compute costs per employee per year; compared to Levels.fyi estimates, this is 2.3 times salary. But the Ramp AI Index for June shows that the top 1% of enterprise engineers spend $89,000 annually on AI, a 680x gap from the median engineer.
Current Data on Anthropic vs. Industry Median
Based on published data, the 2026 AI spending landscape is stratified as follows: Anthropic bears approximately $2 million in compute costs per employee, 2.3 times their compensation (over $500,000), a ratio unmatched in the entire software industry.
The Ramp AI Index for June 2026 shows that the top 1% of enterprises spend about $89,000 per engineer annually on AI (equivalent to 40% of a senior engineer's $224,000 salary); the median enterprise spends only $137 per engineer on AI, virtually zero; the gap between the top and median is about 680x.
Epoch AI research shows Anthropic generates $14 million in revenue per employee and OpenAI generates $6.5 million per employee, the two highest companies on the Forbes Global 2000 list.
Tunguz's Three-Scenario Model: Core Assumptions for Pessimistic, Baseline, and Optimistic
Renowned venture capital analyst Tomasz Tunguz, using a senior engineer's annual salary of $224,000 (assuming 5% annual growth) as a baseline, designed three scenarios to analyze the trajectory of AI token spending per engineer:
2026 (all three scenarios share the same starting point): $90,000 (40% of salary)
2027: Pessimistic $106,000 (45%), Baseline $164,000 (70%), Optimistic $258,000 (110%)
2028: Pessimistic $118,000 (48%), Baseline $259,000 (105%), Optimistic $444,000 (180%)
2029: Pessimistic $106,000 (41%), Baseline $363,000 (140%), Optimistic $596,000 (230%)
In the pessimistic scenario, the amount actually declines after 2028 because token price declines outpace salary inflation. The above are Tunguz's scenario analysis projections and do not constitute investment advice.
GPT-4 Token Pricing Historical Trend: A Known 90% Decline Over Three Years
According to public pricing data, OpenAI GPT-4 class model input pricing has dropped from $30 per million tokens at launch in March 2023 to under $3 in 2026, a ~90% annual decline over three years. Open-source models also exert competitive pricing pressure: DeepSeek-V3 and subsequent versions deliver performance comparable to top closed models at one-tenth to one-thirtieth of the API cost.
Goldman Sachs predicts token consumption will grow 24x by 2030, a core supporting argument for the optimistic scenario; but token price declines and the rise of open-source models are the main drags on the pessimistic scenario.
Frequently Asked Questions
What does the 680x gap in the Ramp AI Index mean?
According to the Ramp AI Index for June 2026, the top 1% of enterprises spend about $89,000 per engineer annually on AI, while the median enterprise spends only $137; the 680x gap means that heavy AI application is currently concentrated in only a very few top enterprises, with the vast majority of companies spending close to zero on AI. The core question of Tunguz's three-scenario model analysis is whether and how quickly the median enterprises will converge toward the top.
Why does the token spending amount in Tunguz's pessimistic scenario for 2029 end up lower than in 2028?
According to Tunguz's scenario design, the pessimistic scenario assumes that token prices decline faster than salary inflation (about 5% per year); under this assumption, the continued decline in token pricing outweighs the spending increase from demand growth, causing the annual dollar-denominated AI token spending to decline after 2028. The above are Tunguz's scenario analysis assumptions, subject to official data updates.
How much has the token pricing of GPT-4 class models declined over three years?
According to public pricing records, GPT-4 class model input pricing fell from $30 per million tokens in March 2023 to under $3 in 2026, an ~90% annual decline over three years; open-source models like DeepSeek-V3 further compress pricing with even lower API costs. Specific current pricing is subject to official announcements from each model provider.
