AI Companies Grow Fast and Spend Less—Who's Getting the Bill?
Deep thoughts on AI and aspirations —— ByteDance Deep Thought Circle
a16z partner David George recently shared a striking data point: the top-performing tier of AI companies saw year-over-year revenue growth of 693%. His team triple-checked before going public. Even more counterintuitive than the growth rate is the adjacent figure: these companies spend less on sales and marketing than traditional SaaS companies.
Fast growth, low spend. By old-school financial analysis standards, this is near perfect. But costs never disappear into thin air—saved money always ends up somewhere. My answer: the money changed line items. Understanding this reallocation is more valuable than memorizing any growth rate.
The Growth Engine Moved from Sales to Inference
Traditional software companies have a default P&L structure. After building the product, the bulk of spending goes to sales and marketing: educating customers, persuading customers, overcoming procurement inertia—all through headcount and budget. The sharpest SaaS companies were essentially the best at buying growth.
This cohort of AI companies upended that structure. Products create value on first use, and demand grows organically. According to a16z, AI companies overall grow two and a half times faster than non-AI companies, and the fastest tier actually spends less.
Where did the saved sales spend go? Into gross margin. AI companies generally have lower gross margins than traditional software, for essentially one reason: inference costs. Every time a user engages with the product, tokens get burned. Sales spend is a one-time acquisition cost; inference cost is an ongoing operational expense that scales with usage. The growth engine shifted from buying traffic to burning compute.
This chain goes deeper. Tokens come from cloud providers’ compute capacity, which comes from capital expenditures. By 2030, the cumulative capex of a few hyperscale cloud providers is expected to approach $5 trillion. The money that disappeared from P&Ls ultimately lands on someone else’s balance sheet.
| Cost Item | Traditional Software | Current AI Companies |
|---|---|---|
| Acquisition Spend | Sales and marketing, scales with growth | Drastically reduced, relies on product virality |
| Delivery Spend | Servers and R&D headcount | Inference costs, burned with each use |
| Heavy Assets | Essentially none | Tied up in cloud providers’ capex |
Low Gross Margins Should Actually Be a Plus This Time
Following this reallocation, many metrics should be read in reverse.
Gross margin is the prime example. Traditional software’s 80%+ gross margin is a faith-level number, but at AI companies, a16z reads it inversely: low gross margin is a badge of honor. High inference costs indicate two things: users are actually using the product, and inference costs will continue declining with model improvements and competition. Conversely, if an AI company has particularly attractive margins, they grow suspicious—wondering whether AI features aren’t actually why customers are paying, just a slide in the fundraising deck.
The same logic applies to revenue per employee. These companies generate $500K to $1M in annual recurring revenue per person, compared to about $400K for the previous generation of software companies. The numbers look good, but the mechanism matters more than the numbers: one founder had two engineers deeply fluent in Claude Code and Cursor rebuild a product from scratch with generous tool budgets—10 to 20 times faster, with tool bills so high he started rethinking how to structure the company. Revenue per employee is an outcome; the changed delivery model is the cause.
However, the “low margins as a badge of honor” reading has boundaries—let’s be clear upfront. Inference costs prove usage, not business model viability. If users develop habits on subsidized low-priced tokens, both usage and margins will collapse when costs normalize. The key judgment is who captures the benefit of declining inference costs: when models get cheaper, do you capture the price difference in your margin, or do competitors immediately use it to compete on price? The former is a business, the latter is just traffic.
Who Pays in the End
Zooming out to the macro level, this bill has a clear due date.
In 2025, all public software companies combined added $46 billion in new revenue. Just OpenAI and Anthropic alone added operating revenue approaching half that figure. A16z estimates that in 2026, new revenue from model companies could represent 70-80% of the entire public software industry’s growth. By this estimate, the new pie in software is being claimed mostly by a handful of model companies, and the discussion about how application and model layers split the pie needs recalculation.
The validation on the capital side is even harder. Nearly $5 trillion in cumulative capex, aiming for a 10% hurdle rate of 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 around $50 billion, most of which emerged in the past year and a half. From $50 billion to $1 trillion requires another 20x growth, in an environment of declining compute costs and price competition.
The enterprise picture must also be factored into the total ledger. Fortune 500 CEOs all talk onstage about full AI transformation, 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 validation pressure ultimately depends on these transformation cases shifting from outliers to the majority.
A widely circulated bull case argument goes: 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 utilized and having pricing power are different things. Public H100 rental rates dropped from $8 per hour to around $2, and full capacity utilization can coexist perfectly with collapsed returns. When everyone can access cards, cards cease to be a moat—price becomes the sole allocator. The higher cloud providers pile capex, the more validation pressure shifts to inference demand growth rates. Debt has also entered the picture, with Oracle’s credit default swap costs rising to around 2%—signals like these are worth monitoring continuously.
For people building companies and investing, this data set has very concrete implications. When evaluating AI businesses, stop using old software line items—the reading direction for both gross margin and sales spend has reversed. At the same time, stay clear-eyed: your efficiency is built on upstream balance sheets, and any shift in model pricing, token costs, or cloud provider capital discipline will directly rewrite your unit economics. Finally, that $1 trillion threshold in 2030 is the ultimate validation of this reallocation. If it’s cleared, this becomes the new normal; if not, the bill will flow back from cloud providers to every P&L.