Opinion
The token trap: AI’s new pricing formula and the threat to startup margins
For early-stage startups, AI economics is not a financial issue that can be kicked down the road to growth stages, it is the core of product architecture from day one.
For the past decade, the tech world and capital markets lived in absolute harmony around one winning business model: seat-based SaaS (Software-as-a-Service) pricing. This model provided public companies and startups with Wall Street's most sacred metric: predictable and stable Annual Recurring Revenue (ARR), which the capital markets knew how to value at high multiples.
But then the Generative AI revolution arrived and reshuffled the deck. Suddenly, behind every user query or action, there isn't just static code, but expensive computing power and countless tokens. Capital markets are closely watching this shift because it forces a redefinition of business models. For early-stage startups, this shift is seismic-impacting them both as consumers of big tech and as vendors trying to bring their own products to market.
Beyond Tokens: Tech Giants Are Pricing Outcomes
Major public companies, from Microsoft and Google to OpenAI and Anthropic, understand that inference costs (the actual cost of running the models) are a bottomless pit if not priced correctly. As a result, we are witnessing an accelerated transition to usage-based models and token pricing.
However, in this dynamic market, this model is already evolving into its next phase. Software giants realize that while Wall Street wants to protect their gross margins, it dislikes extreme revenue volatility. Consequently, companies like Salesforce have recently shifted to pricing based on Agentic Work Units, and Zendesk is pioneering a pay-per-resolved-issue model using AI.
The game is changing from how many tokens the system burned to how much work the product actually saved the organization, shifting directly to an outcome-based model. The shockwaves of this transition land right on the doorstep of the entrepreneurial ecosystem.
The Startup as a Customer and the Hidden Runway Trap
For pre-seed and seed-stage startups, the shift by major vendors to token models creates a dramatic cash flow challenge. In the past, a young company's software budget was predictable and easy to manage: fixed licenses for tools like Slack, GitHub, or Salesforce. Today, when startups embed Large Language Models (LLMs) into their products, API usage costs have often become the second or third largest expense item in the company, right after headcount. A young company that does not enforce aggressive token optimization might find its financial runway cut in half just because its users spent a bit too much time playing around with the product.
However, this trap is not inevitable. Savvy founders already understand that the solution lies in the product architecture from day one. We are seeing more startups transitioning to Bring Your Own Tokens (BYOT) models, allowing enterprise customers to bring their own API keys, or integrating small, local models that dramatically lower compute costs and maximize hardware utilization.
The Startup as a Vendor: What Do Solutions Look Like on the Ground?
The other side of the coin is even more complex. When a startup develops its own AI product and goes to market, it faces a critical pricing dilemma that will determine its fundraising prospects. If the startup sells via a traditional, flat-rate SaaS model but pays cloud giants based on tokens behind the scenes, it could be on a fast track to eroding its profit margins.
Capital markets and investors are accustomed to software companies with gross margins of 70%-80%. A mispriced AI product could exhibit a gross margin of just 40%, a figure that makes it look more like a services company and drastically cuts its valuation. On the flip side, passing direct token costs onto the customer creates an adoption barrier with CFOs who hate billing surprises.
Israeli startups are already learning to crack this dilemma on the ground using creative, hybrid models:
Atera (an autonomous IT platform) cracked this with a smart combination: it retains the classic seat-based model but embeds a built-in AI quota, charging for overages based on specific actions executed by the agent. This model grants the client a predictable budget while protecting the company's gross margins.
An even clearer example of the shift toward the "Agentic economy" is Wonderful AI, which develops autonomous customer service agents. The company, which raised a massive round at a multi-billion dollar valuation, skipped the token trap entirely and moved to a pure pricing model based on successfully resolved tickets.
This breakthrough isn't exclusive to established growth companies; early-stage startups are applying these principles as early as the seed stage to protect themselves. For example, Overcut implements a BYOT model. It allows enterprise customers to bring their own API keys and existing cloud agreements, completely neutralizing inference cost risk from its financial model from day one.
The Bottom Line
Investors on Wall Street and in the VC world are no longer swayed by superficial AI hype; they reward companies that present clear unit economics and resilient business models with high multiples.
For early-stage startups, AI economics is not a financial issue that can be kicked down the road to growth stages, it is the core of product architecture from day one. The market leaders will not be those offering the most advanced technology, but those who know how to build a product that measures and bills based on the outcome and value it generates, rather than the amount of tokens it burns. Right there, at the intersection of computing efficiency and market-disrupting value creation, the next software giants are being born.
The author is an investor at lool ventures, an early-stage startup venture capital fund.
