How large will AI token spend become as a share of employee compensation costs?

Harry Stebbings · Jun 20, 2026

Enterprise AI spend needs to rise from around 4% of developer salaries to 18-20% for the trillion-dollar model-company valuations to be justified.

Benioff spends $300M a year on Anthropic, which is only about 3.8% of developer salaries — far below the required level.

18:10 20VC: Why Remote Work is White Collar Fraud | Why Revenge and Patriotism are the Best Founder Traits | Two Questions Every Founder Needs to Ask | The Wild Story of Raising $1BN from Masa Son in an Hour Long Meeting with Ryan Peterson, Founder @ Flexport

Harry Stebbings · Jun 22, 2026

Frontier lab valuations hinge on what share of developer salary spend goes to tokens: at Benioff's ~3.8% Anthropic and OpenAI are grossly overvalued, at 20% they are undervalued, and at parity with salaries grossly undervalued

Benioff spends $300M a year on Anthropic for his developers, which works out to about 3.8% of average developer salary spend

Scope: framework built on Benioff's disclosed numbers and Brandon at McCall's parity scenario

20:38 20VC: Nikesh Arora on the Frontier Model Problem: Breadth vs Depth | The Future of Token Costs | Memory Becoming the Moat | Where Value Accrues: Infra, Models, or Apps? | Why Enterprise AI is Not Ready & Systems of Record vs Systems of Intelligence

Harry Stebbings · Jul 4, 2026

If token spend stays at roughly 3.8% of developer salaries, many AI companies are grossly overvalued; if it reaches 20%, they are undervalued

Benioff's reported $30M/year Anthropic spend for dev teams works out to about 3.8% of developer salaries, which is less than the headline suggests

25:52 20VC: Open Models vs Frontier Models: Who Actually Wins? | The $100,000 Token Budget Every Engineer Will Need | Why Forward-Deployed Engineers Are the Future of Enterprise AI with Clay Bavor, Co-Founder of Sierra

Harry Stebbings · Jul 6, 2026

Whether Anthropic's exponential revenue growth continues hinges on the share of developer salary spend that goes to tokens — at 20% Anthropic is grossly undervalued, but if the ratio stalls or buyers migrate to open models it is a very different game

Salesforce spent $300M with Anthropic for its dev team, equating to roughly 3.8% of developer salaries spent on tokens

22:33 20VC: Why Now is the Time for the Application Layer | Why OpenAI & Anthropic Won't Win the App Layer | Why Startups Should be TokenMaxxing | Why VCs Should Reduce Weighting on Price & Ownership in an Age of AI with Mike Mignano, USV

Harry Stebbings · Jul 11, 2026

AI dev tools like Cursor are still dramatically underpriced.

Benioff's $300M Anthropic spend is only about 3.7% of developer salaries, which is relatively small.

Scope: specific to dev tools, not AI generally

30:23 20VC: Why OpenAI and Anthropic Won't Win the App Layer | Why Teams Will Get Bigger Not Smaller in a World of AI | Why AI Removes Incumbents Advantage of Bundling | China vs America: Who Wins the AI War with Arvind Jain, Co-Founder @ Glean

Harry Stebbings · Jul 11, 2026

If buyers consider 3.8% of developer salaries too much to spend on AI tools, the model providers are in serious trouble.

31:23 20VC: Why OpenAI and Anthropic Won't Win the App Layer | Why Teams Will Get Bigger Not Smaller in a World of AI | Why AI Removes Incumbents Advantage of Bundling | China vs America: Who Wins the AI War with Arvind Jain, Co-Founder @ Glean

Fred Turner · Jul 18, 2026

Frontier model pricing could double or even 5x and agentic contracting would still be economically compelling.

A contract done by people costs $1,500-$2,000 on average versus about $70 with the agent, so the labor-displacement arbitrage survives large price increases.

Scope: based on Curative's contracting workflow economics

54:02 20VC: $5BN in Revenue, 7 to 7,000 Employees in 9 Months, 206,000 Tests in a Single Day: The Craziest Story in Startups: Curative with Fred Turner

Brendan Foody · Jun 1, 2026 · hedged

Within five years the average enterprise will spend more on compute than on headcount

Models are becoming so capable that there is enormous ROI in having a model do a job for ~$100k a year, and model capability compounds exponentially in a way human intelligence does not

Scope: humans still play an important role at what models can't do

46:41 20VC: Mercor CEO on Why Application Layer Companies Have No Defensibility, The Model is the Product | Token Spend Will Exceed Headcount Spend in 5 Years | The True Cost of Hiring AI Researchers in the Valley Today with Brendan Foody

Matan Grinberg · Jun 13, 2026 · hedged

Within three years, median token spend per person will be on the same order of magnitude as their salary

Scope: order of magnitude only; given under protest that averages are the wrong frame

23:36 20VC: Who Wins the Model War: OpenAI, Anthropic or Open-Source | Token Maxing, AI Hangovers & The Coming ROI Reckoning | Labour Displacement Fears are BS & Overblown | From Physicist to Sequoia Founder with Matan Grinberg, Founder @ Factory

Fred Turner · Jul 18, 2026 · speculative

In three years, company spend on Anthropic will run at roughly 2 to 5 times total developer salary, and firms will deploy more agents than engineers while keeping engineer headcount flat

Workflows are shifting to one senior engineer managing many downstream agents, with mostly unsupervised agents implementing features and engineers acting as reviewers and architects — so companies will just build a lot more

80:33 20VC: $5BN in Revenue, 7 to 7,000 Employees in 9 Months, 206,000 Tests in a Single Day: The Craziest Story in Startups: Curative with Fred Turner

Aaron Levie · Apr 20, 2026

Jensen's point is directionally salient even if the exact number is off by half: companies will spend far more on compute per person in the future than they expect or spend today.

Scope: the specific figure may be off by half

22:35 20VC: Everyone is Wrong; We Will Have More Developers in Five Years | Why Frontier Labs Will Be Way More Valuable Than They Are Today | Are SaaS Companies Cooked: Which Thrive & Which Die with Aaron Levie, Founder at Box

Aaron Levie · Apr 20, 2026 · hedged

Box's compute spend per employee will be a larger number in five years than it is today.

Scope: declines to give a specific figure due to public-company disclosure constraints

23:27 20VC: Everyone is Wrong; We Will Have More Developers in Five Years | Why Frontier Labs Will Be Way More Valuable Than They Are Today | Are SaaS Companies Cooked: Which Thrive & Which Die with Aaron Levie, Founder at Box

Andrew Feldman · May 26, 2026 · hedged

Software engineering spend on AI tokens rising from ~3% to 15-20% of developer salaries is unproblematic, because hardware engineering already spends roughly that share on EDA tools

Historically software threw people at problems while hardware bought expensive tools because the cost of hardware bugs was so high; as AI tools become productive, software will converge on the hardware norm

Scope: I've never done this in any detail; I bet

31:57 20VC: Cerebras CEO on the Future of Data Centres, Token Costs and Memory | We are Not in an Infra Bubble & Dario Got a Bad Deal with Elon for Compute | Should US Companies Sell to China & Why Most Layoffs are AI Washed with Andrew Feldman

Clay Bavor · Jul 4, 2026

Token spend as a share of developer salary will settle far above 3.8% — much closer to 20%

The productivity gains in software engineering are unequivocal; even at 2x you have effectively doubled the size of your engineering team, which justifies far more spend

Scope: the exact multiple (2x, 10x, 20x) is debatable

26:28 20VC: Open Models vs Frontier Models: Who Actually Wins? | The $100,000 Token Budget Every Engineer Will Need | Why Forward-Deployed Engineers Are the Future of Enterprise AI with Clay Bavor, Co-Founder of Sierra

Harry Stebbings · Jul 11, 2026

Agents belong in the same budget conversation as labor because, unlike prior technology, they act proactively, hold opinions, make decisions, and replace the labor we used to pay for.

That substitution is exactly what technology is for.

32:16 20VC: Why OpenAI and Anthropic Won't Win the App Layer | Why Teams Will Get Bigger Not Smaller in a World of AI | Why AI Removes Incumbents Advantage of Bundling | China vs America: Who Wins the AI War with Arvind Jain, Co-Founder @ Glean

Andrew Macdonald · Aug 17, 2026

Companies should combine headcount and compute budgets into a single pool and let trusted leaders allocate between them as they see highest ROI

A CTO may rationally trade heads for inference spend or vice versa depending on whether throughput or compounding new product value is the bigger win; combining the pools forces that judgment

Scope: offered as 'an interesting approach'

38:43 20VC: Uber President on The Untold Uber Stories: Travis, China and Self-Driving | Why Autonomy Is Existential | How to Beat DoorDash to #1 in Food with Andrew MacDonald

Matan Grinberg · Jun 13, 2026

Token spend as a percentage of individual salary will vary enormously across people — from 0% to thousands or tens of thousands of percent — and any org enforcing a single target ratio across all engineers is painting with far too wide a brush

It depends on each individual's unique skills and how they create leverage; some deliver value by meeting customers or doing light data analysis, others by delegating to dozens of parallel agents

Scope: about individuals rather than a company-wide average

22:07 20VC: Who Wins the Model War: OpenAI, Anthropic or Open-Source | Token Maxing, AI Hangovers & The Coming ROI Reckoning | Labour Displacement Fears are BS & Overblown | From Physicist to Sequoia Founder with Matan Grinberg, Founder @ Factory

Arvind Jain · Jul 11, 2026

Token consumption follows a power law: a few employees spend $10–15k a month while others spend $20; everyone uses basic information-seeking AI but advanced use cases reach only about 5% of the employee base.

Observed across Glean and its customers.

Scope: based on Glean and its customer base

34:17 20VC: Why OpenAI and Anthropic Won't Win the App Layer | Why Teams Will Get Bigger Not Smaller in a World of AI | Why AI Removes Incumbents Advantage of Bundling | China vs America: Who Wins the AI War with Arvind Jain, Co-Founder @ Glean

Harry Stebbings · Apr 20, 2026

Jensen Huang's advice that every engineer should spend hundreds of thousands of dollars on compute is self-serving and therefore unpersuasive — it's the drug dealer telling you to buy drugs.

Nvidia sells the compute he is recommending everyone buy more of.

22:16 Vendor self interest makes the projection unpersuasive

Clay Bavor · Jul 4, 2026

Companies are heading toward per-employee token budgeting, with CFO capital allocation treating tokens as part of headcount cost alongside salary and SBC

Token spend per engineer is becoming a meaningful fraction of compensation, so it has to be allocated like headcount

24:20 Tokens become a per headcount budget line

Aravind Srinivas · Jun 15, 2026

OpenAI and Anthropic can be $10 trillion companies regardless of whether token spend reaches a full percentage of developer payroll

A large amount of non-developer agent work — finance, corp dev, sales, data science — will also be done with agents

18:54 Non developer agent work makes the payroll ratio irrelevant

Arvind Jain · Jul 11, 2026

Current AI technology is absurdly expensive relative to what it delivers.

Glean's engineering triage agent handles 95% of production issues but runs at roughly $1M a month, which is at or above the cost of the human on-call team it replaced.

30:10 Agent cost already equals the labor it replaces

Arvind Jain · Jul 11, 2026

Putting technology costs and labor costs in the same sentence is historically unprecedented and wrong; models are supposed to get cheaper and more affordable over time.

Open source already does the same work at a tenth of the cost, and technology has never before been traded off against headcount this way.

31:39 Reject the labor benchmark technology costs must fall

Alex Atallah · Aug 10, 2026

Employee cost is becoming a dynamic number driven by inference spend, replacing the static salary model

Companies are freaking out about inference costs; unlike a salary that's set and forgotten, each employee's AI usage costs a different, constantly changing amount depending on which models and tools they choose

53:53 Employee cost becomes dynamic not a static salary