Startups such as Fal and Fireworks AI that provide AI model and server invocation services are seeing continuously rising revenues. Developers use them to run models quickly, which has also triggered a new wave of investor enthusiasm.

According to two people familiar with the matter, Fal, which focuses on providing inference services for image and video generation models such as Google's Nano Banana, has already held talks with investors about a new funding round at a target valuation of $15 billion. A third person familiar with the matter said the negotiations are still at an early stage and the company could raise its target valuation to a range of $17 billion to $20 billion.

Whichever range the final valuation falls into, this round will nearly double the five-year-old company's valuation from the $8 billion round in the spring of this year. One person familiar with the matter said the funding talks come as its annualized revenue climbed to $800 million, double the level in March.

A third person familiar with the matter said Fireworks AI, the top-ranked inference service provider by revenue scale, is also considering a new funding round, and if it proceeds, the target valuation could reach $30 billion, nearly double the previous valuation announced in July.

Fireworks said on Friday that it had completed an employee share sale led by Atlaydes Capital at a valuation of $17.5 billion, in line with the previous funding round valuation announced in July. As of July, its annualized revenue had reached $1 billion, four times higher than a year earlier. Its latest revenue figures are not yet available.

According to reports, another inference service provider, Baseten, is also in talks with investors about a new funding round at a target valuation of $26 billion; peer service provider Modal is in funding talks at a valuation of about $15 billion, roughly three times the valuation of a round four months ago.

The密集 of funding talks reflects that such startups, which provide access to frontier models and help customers customize open-source models, are in an industry boom period. Thanks to advances in open-source AI technology, developers are making heavy use of such services, and open-source models often cost less than the closed-source models of OpenAI and Anthropic. Building hardware on their own to run open-source models is difficult to configure, and chip shortages push up costs, so many developers prefer to invoke models through simple application programming interfaces, which both Fireworks and Fal provide.

However, as closed-source model vendors such as OpenAI and Anthropic step up price cuts, inference service providers may come under pressure. Amjad Masad, chief executive of code startup Replit, said on Wednesday that because prices for some OpenAI models have fallen sharply, the company's use of open-source AI recently has decreased compared with the start of the year.

In addition, the cost for inference service providers to rent computing power from cloud vendors is rising, and there is a risk that corporate profit margins will be squeezed. Such inference companies have historically had gross margins of about 50%, lower than the 70%-plus gross margin level common among top software service providers. Inference service providers believe they can improve gross margins by optimizing how models run on chips. Together AI, which provides both inference services and AI server rental, is also purchasing chip servers and renting them out through its own data centers.