AI Companies Grow Fast and Spend Less—Who Gets the Bill?
Deep thinking on AI and aspirations —— ByteDance Deep Thinking Circle
a16z partner David George recently shared data: the top tier of AI companies achieved 693% year-over-year revenue growth. His team triple-checked before going public. More counterintuitive than the growth rate is the figure beside it: these companies spend less on sales and marketing than traditional SaaS companies.
Fast growth, low spending. By traditional financial analysis, this is near perfection. But no cost disappears into thin air—saved money has to go somewhere. My answer: the money changed line items. Understanding this relocation of costs is worth more than memorizing any growth rate.
The Growth Engine Moved from Sales to Inference
Traditional software companies’ income statements follow a default structure. After the product is built, the bulk of spending goes to sales and marketing: educating customers, persuading customers, overcoming procurement inertia—all relying on headcount and budget. The sharpest companies in the SaaS era were essentially those best at buying growth.
This cohort of AI companies upended the structure. Products create value on first use, and demand generates itself. According to a16z, overall AI company growth is two and a half times that of non-AI companies, and the fastest-growing batch actually spends less.
Where did the saved sales expenses go? Into gross margin. AI companies’ gross margins are generally lower than traditional software, for essentially one reason: inference costs. Every time a user uses the product, they burn tokens. Sales expenses are one-time acquisition costs; inference costs are ongoing operational expenses that grow with usage. The growth engine shifted from buying traffic to burning compute.
This chain can be traced further. Tokens come from cloud providers’ compute, which comes from capital expenditures. By 2030, the cumulative capital expenditure of several hyperscale cloud providers is expected to approach $5 trillion. The money that disappeared from income statements ultimately lands on someone else’s balance sheet.
| Cost Item | Traditional Software Companies | Current Generation AI Companies |
|---|---|---|
| Acquisition Spending | Sales and marketing, grows with growth | Significantly compressed, relies on product virality |
| Delivery Spending | Servers and R&D labor | Inference costs, burn per use |
| Heavy Assets | Barely touched | Hung on cloud providers’ capital expenditures |
Low Gross Margin—This Time It Should Be a Plus
Following the cost relocation, the interpretation of many metrics has reversed.
Gross margin is the most typical example. Traditional software’s 80%+ gross margin is a faith-level number, but for AI companies, a16z’s interpretation flips: low gross margin is a badge of honor. High inference costs indicate two things: users are actually using it, and inference costs will continue to decline as models improve and competition intensifies. Conversely, if an AI company has particularly attractive margins, they become suspicious, suspecting AI features aren’t the reason customers are paying—just a page in fundraising materials.
The same applies to per-employee productivity metrics. These companies generate $500K to $1M in annual recurring revenue per person; the previous generation’s standard was around $400K. The numbers look good, but the mechanism matters more than the numbers: one founder had two engineers who heavily use Claude Code and Cursor rebuild a product from scratch, gave them full tool budgets, and speed increased 10 to 20x—the tool bills got so high he started rethinking company organization. Productivity is the result; changed delivery methods are the cause.
However, the “low margin as badge” interpretation has boundaries—let’s be clear upfront. Inference costs prove usage, not business model. If users develop habits on subsidized low-price tokens, when costs rise, usage and margins will collapse together. The key judgment is who captures the benefits of declining inference costs: when models drop prices, do you capture the difference in your margins, or does a competitor immediately use it for price wars? The former is business, the latter is just traffic.
Who Pays in the End
Moving up to the macro level, this bill has a clear settlement date.
In 2025, all public software companies combined added $46 billion in revenue. OpenAI and Anthropic alone accounted for nearly half that figure in new operating revenue. A16z estimates that in 2026, model companies’ incremental revenue may account for 70-80% of the entire public software industry. By this estimate, the software industry’s incremental pie is being taken mostly by a very few model companies, and the discussion of how application and model layers divide the pie needs recalculation.
Settlement on the capital side is harder. Nearly $5 trillion in cumulative capital expenditure, to achieve a 10% hurdle return, requires AI annual revenue to reach around $1 trillion by 2030—equivalent to 1% of global GDP. The industry’s current annual revenue is roughly $50 billion, most of which grew in the past year and a half. From $50 billion to $1 trillion requires another 20x increase, in an environment of declining compute costs and competitive price wars.
The enterprise picture must also factor into the total bill. Fortune 500 CEOs all say they’re going all-in on AI, but actual moves are much slower—the bottleneck isn’t budget, it’s change management. But the few companies that have completed internal transformation show hard numbers: Chime cut customer service costs by 60%, Rocket Mortgage saved 1.1 million hours in underwriting annually, equivalent to about $40 million in operating costs. The $1 trillion settlement pressure ultimately depends on these transforming companies moving from cases to the majority.
In bull arguments there’s a widely circulated saying: the internet era laid dark fiber, the AI era has no dark GPUs—cards get used as soon as they’re installed. This statement is half right. Being used and having pricing power are different things. H100 public rental rates dropped from $8/hour to around $2, and full capacity utilization and collapsed returns can happen simultaneously. When everyone can get cards, cards are no longer a moat—price becomes the only allocator. As cloud providers’ capital expenditures pile higher, settlement pressure shifts increasingly to inference demand growth rates. Debt has already entered the picture—Oracle’s credit default swap costs rose to around 2%, signals like this are worth continuous monitoring.
For those building companies and investing, this set of numbers has very concrete implications. When evaluating AI businesses, stop using old software line item structures—the interpretation of gross margin and sales expense metrics has reversed. At the same time, be clear: your efficiency is built on upstream balance sheets, and any change in model prices, token costs, or cloud provider capital discipline will directly rewrite your unit economics. Finally, that $1 trillion hurdle in 2030 is the final settlement for this cost relocation. If it’s cleared, it becomes the new normal; if not, the bill will be returned from cloud providers back through every income statement.