Investors are starting to scrutinize how AI startups optimize the cost of using models offered by OpenAI, Anthropic or Google, considering it an indicator of founder discipline and warning that such expenses can surge and hurt how they build, scale and monetize their products.
According to investors, excessive usage of AI tools without a clear strategy often signals inefficient product design or a lack of infrastructure ownership and reveals whether a company has built proprietary systems or is merely reselling another company’s intelligence layer.
A startup at pre-seed funding stage that cannot articulate its usage economics could be an exemption because its product is still being defined. However, those at Series A stage without a view on computational expenses incurred when a trained AI model processes data and generates output as they scale up have a structural problem, said Natasha Malpani Oswal, founder of Boundless Ventures, an AI-focused investment fund.
“The discipline we look for at pre-seed and seed is simple: can the founder tell me what it costs per active user per month and what that number needs to be for the unit economics to work at scale?” said Oswal.
These costs, say investors, signal whether the company can understand and differentiate where AI is actually required and where simpler systems, APIs or smaller models can do the job.
“The best founders break workflows into smaller tasks and even subtasks, understand the requirements of these tasks/subtasks and ensure that instead of using the largest model for everything, things can be super-optimized while giving acceptable results,” said Som Pal Choudhary, co-founder and partner at Bharat Innovation Fund.
As AI adoption deepens, token economics is likely to become a defining factor in how startups build, scale and monetize their products. Token usage cost refers to the fee that AI providers charge startups and is based on the number of tokens, or the basic units of text or data, processed by an AI model.
Impact assessment
Startups, meanwhile, are still in trial mode as they evaluate token usage costs across models.
“It is difficult to fully assess the impact because we use different models,” said Sarthak Shrivastava, co-founder of YC-incubated startup Floworks. “The variation in token costs depends on how frequently you use them, how much you use them and which pricing tier you are on.”
To manage costs, some startups use both expensive frontier AI models and cheaper open-source alternatives.
“I’ve invested in a startup that uses Google Gemini as an official partner, which means Google subsidizes the tokens. Yet, even after that, because the cost is so high considering the volume, they are also using DeepSeek or Qwen for low-brain work, simply speaking, wherever no chain-of-thought reasoning is required,” said Kalyani Khona, an angel investor and AI researcher.
Some startups use cheaper models during the planning phase. Once everything is decided, they move to legacy models for execution, where they are comfortable spending more.
“When we are just testing a few things internally, we prefer using Deepgram, which is a real-time voice bot AI model and much cheaper. We make sure that less important tasks are not wasting expensive tokens,” said Additi Upadhyay, founder of AI startup Noveum.ai. “When it is in production, we use ElevenLabs. But during internal testing, which, honestly, burns a lot of tokens, we avoid using extremely expensive models.”
The growing investor scrutiny of token costs, especially for AI startups, stems from the fact that almost every company today uses AI for coding and software development-related work, which can become expensive very quickly.
A startup may have planned to hire 100 engineers, but with AI tools, it may now need fewer people. Instead of spending ₹5 crore a month on a large engineering team, a company may spend ₹1 crore a month on AI tokens.
However, even a ₹1 crore bill can rise to ₹3 crore or ₹6 crore as usage scales up. Initially, token costs appear manageable, but they tend to increase steadily with product growth, higher user activity and deeper AI integration across workflows.
Hard to justify
This is not a concern only for startups. Even large technology companies are monitoring AI infrastructure and token-related spending. Uber president and chief operating officer Andrew Macdonald recently said skyrocketing AI token costs are becoming difficult to justify as high consumption has not yielded proportional productivity gains.
“If you are spending ₹10 lakh or ₹80 lakh, investors compare that against the amount of work being done,” said Pranav Pai, founding partner and chief investment officer at 3one4 Capital, an early-stage venture capital firm based in Bengaluru. “So, if you are paying the equivalent cost of two people but getting the bandwidth of 10 headcount, that’s considered a good ratio. Founders have been asked to maintain oversight and make sure the ratio remains worth the cost.”
Investors also point out that startups, especially in India, have raised less capital and do not have unlimited access to GPUs or compute infrastructure, making “AI slop”, or the lazy overuse of expensive AI models, financially unsustainable.
AI startups in India including Sarvam and Uniphore have raised about $721 million so far in 2026, according to data platform Tracxn, compared with $242 billion raised by AI startups in the US.
“Efficient orchestration and disciplined model usage are becoming strong indicators of long-term viability,” said Pal of Bharat Innovation Fund.
However, investors don’t advise startups to aggressively cut costs in this segment because that can damage the product itself.
“The damage comes when we respond by cutting the wrong things. Fewer model calls, simpler prompts, more aggressive caching – these feel like cost management. They often experience management, and users notice immediately. You are solving a unit economics problem by breaking the product,” said Oswal of Boundless Ventures.