Amid growing concerns over the increasing costs of artificial intelligence, investors are directing their funds toward innovative startups designed to mitigate the financial impact on businesses. These startups, often referred to as AI-routing firms, assist developers in choosing from various AI models for their tasks while also keeping a close eye on overspending and quickly addressing outages. The demand for such solutions is witnessing remarkable growth.
In late May, OpenRouter announced a successful funding round of $113 million, which elevated its valuation to $1.3 billion. Recently, Concentrate AI has made its presence known by securing over $5 million in funding, as reported by Business Insider.
"The current landscape of models is highly fragmented, making it challenging to manage," said Ari Jacoby, CEO of Concentrate AI, in a conversation with Business Insider. "Our solution brings all this under a unified platform."
The timing is advantageous for both firms, as the surge in demand for AI coding resources has driven up the use of tokens – the fundamental units for AI input and output. Companies like Anthropic and OpenAI structure their pricing according to token consumption, leading some businesses to experience significant cost increases.
To aid customers in selecting the most economical AI solutions, these routing startups offer access to leading models from notable labs, alongside more affordable alternatives from providers such as Google, DeepSeek, MiniMax, and Xiaomi.
Significantly, major tech corporations also provide AI routing tools. Established giants like Amazon Web Services, Microsoft, and Google Cloud have their own systems for directing tasks to suitable AI models. However, startups like OpenRouter and Concentrate AI claim to cater specifically to developers and smaller teams, presenting a wider variety of models than traditional cloud providers. According to Adam Swick, strategy head at OpenRouter, the platform features over 400 models, with demand from developers experiencing explosive growth in the past six months.
Vercel, a cloud application startup valued at $9.3 billion, has also developed its own AI routing solution after finding it beneficial for internal operations. Harpreet Arora, who leads AI infrastructure at Vercel, noted that the tool gained popularity quickly as users began recognizing its advantages.
Arora describes it as a "centralized hub," allowing users to navigate fluctuating costs and the availability of models as they enter the market.
As the landscape changes, there’s an observable shift in demand toward cheaper AI models. Both OpenRouter and Vercel have noted DeepSeek, a Chinese laboratory, gaining traction this spring following the release of its advanced V4 models. These models performed exceptionally well in capability benchmarks while offering competitive pricing. For instance, while Claude's least expensive model, Haiku, costs $1 per million input tokens and $5 per million output tokens, the most costly version of DeepSeek's V4 model is priced at just 43 cents per million input tokens and 87 cents per million output tokens.
Data from mid-May suggests that DeepSeek’s models were consuming a greater share of tokens on both OpenRouter and Vercel as more developers gravitated toward these cost-effective solutions. Zach Moskow, a co-founder of Concentrate AI, remarked that some users have expressed concerns regarding the security of Chinese models, although many of these are hosted on AWS within the U.S. Jacoby emphasized that when companies discover many high-quality models available at lower prices, budgeting becomes increasingly straightforward.
In addition to these startups, others are capitalizing on the growing anxiety surrounding AI token costs. Lanai, an AI observability startup, has recently introduced a tool called Token Tuner, intended to evaluate the efficiency of clients' AI expenditures. Mohit Mehta, the chief product officer, noted a rise in complaints about the unpredictable pricing associated with high-end AI and anticipates a growing trend toward more economical, simplified models.
"You'll need to tailor your AI spending to its value, just as you would manage your workforce," Mehta stated.
