Artificial intelligence inference startup Fireworks AI Inc. said today it has raised $250 million in a Series C funding round that brings its valuation to $4 billion.

Lightspeed Venture Partners, Index Ventures and Evantic led the round, and existing backer Sequoia Capital was among the participants, as were Nvidia Corp., Advanced Micro Devices Inc. and Databricks Inc.

The company has built a high-performance platform that’s used by enterprises to deploy and fine-tune large language models to perform specialized tasks. It offers access to cloud-based infrastructure that’s designed specifically for running AI inference workloads at scale.

Today’s round comes at a time when the AI industry is rapidly shifting its attention to inference, rather than AI training. Inference is where trained AI models are deployed in production, making predictions or drawing conclusions based on new data they haven’t encountered before. It follows the training phase, where the model learns from large datasets.

Chipmakers including Nvidia and Qualcomm Inc., as well as cloud computing giants like Amazon Web Services Inc. and Google Cloud have all designed chips dedicated to AI inference, rather than training, to support the everyday use of tools such as OpenAI Group PBC’s ChatGPT.

With Fireworks AI, enterprises can access a dedicated platform for AI inference that provides on-demand access to specialized graphics processing units and other AI accelerators, with per second pricing. The startup also offers discounted access to compute resources for bulk inference workloads, which makes it more cost-effective for large-scale deployments.

In addition, Fireworks AI provides tools for companies that want to fine-tune their LLMs, with support for low-rank adaptation and reinforcement learning. The platform also enables simple LLM deployment with a single application programming interface call for rapid prototyping.

Fireworks AI co-founder and Chief Executive Lin Qiao told the Wall Street Journal that the company will use the funds to hire more than 150 new AI researchers and engineers, and scale its sales and marketing teams. It’s also going to purchase more GPUs for its cloud platform. Ultimately, it aims to build “the best tool chain to help application developers build towards the next level of quality, speed and cost,” Qiao said.

Qiao knows a thing or two about AI, for she was notably one of the creators of the open-source PyTorch framework for developing and training AI models during her time at Meta Platforms Inc. She told the Journal that the company currently has around 115 employees, and recently reached $280 million in annual recurring revenue.

Although Qiao is confident that she’ll be able to grow Fireworks AI, some analysts note that the company is catering to a fairly niche market at this stage, though there are good chances it will expand over time. The startup’s platform is used by AI companies such as Cursor Inc. to deliver AI-enabled coding agents to customers, but other enterprises may be slower to embrace it, because they lack the expertise required to perform the bespoke AI engineering work that Fireworks AI is meant to support.

Gartner Inc. analyst Chirag Dekate told the Journal that AI engineering requires extremely specialized skills and knowledge about how to stitch AI models together with new datasets and inference platforms, and such talent is in short supply. “Roughly 80% of enterprises are yet to get to this advanced stage,” he said.

In addition, Fireworks AI faces competition from other AI inference startups such as Baseten Labs Inc. and Together Computer Inc., along with the major cloud infrastructure platforms.

Still, Qiao expressed confidence that the required skills for AI inference will become more common over time. Of course, there’s also the possibility that some AI engineering tasks may one day be automated by AI itself. “Our mission is to enable every business to achieve automated product and model co-design to reach maximum quality, speed, and cost-efficiency using generative AI,” she said. ”

Image: Fireworks AI