Fireworks to Add European Inference Capacity on Nvidia Vera Rubin From 2027

Fireworks said the expansion will also use Nvidia's DSX AI factory platform and Spectrum-X Ethernet networking, and that it chose them for performance and throughput per watt. It did not give GPU counts, power capacity, data centre locations, customer names or the cost of the expansion.

"With NVIDIA, we're expanding our European capacity with the Vera Rubin and Grace Blackwell Platforms," Lin Qiao, Fireworks' co-founder and chief executive, said in the statement. Nvidia founder and chief executive Jensen Huang said in the same release: "I'm delighted to see demand for Fireworks skyrocketing."

Part of a Wider European Build-Out

The announcement builds on earlier European work. In an October 21 blog post timed to its GTC Berlin event, Nvidia said Fireworks and Era4 had partnered to serve open models to businesses in the UK and Europe. Fireworks supplies the training and inference software, and Era4 supplies liquid-cooled capacity running Nvidia Blackwell Ultra GPUs at its facility in Chesterfield. Nvidia said the two would target financial services, insurance and life sciences.

In the same post, Nvidia listed other partners expanding in the region. They included Nscale, which plans Vera Rubin deployments in five European markets, and Mistral, which plans Vera Rubin capacity for 2027. Abu Dhabi's G42 was also on the list of companies building Nvidia-based infrastructure across Europe, the Nordics, Africa and Central Asia.

New Inference Features

Fireworks also launched several products at the conference. Fireworks Ultrafast, which runs on Nvidia's Blackwell platform, is in public preview and is aimed at latency-sensitive interactive applications. Multi-day prompt caching is meant to reduce the cost of reusing large shared context across repeated model calls. A background mode cuts the price of work that does not need an immediate answer, which the company is pitching at long-running agents. Fireworks said it has also raised rate limits and reduced cold starts to under a minute.

The company says it processes trillions of tokens a day for teams that run, combine and fine-tune open-weight models on its platform.

The announcement adds to a busy week for independent inference providers. Fal, which hosts image, video and audio models, said on the same day that it had raised $1.3 billion at a $16 billion valuation. For European buyers, adding capacity inside the region matters because many regulated firms want their data processed locally. The useful details for those buyers, such as where Fireworks' Vera Rubin capacity will sit and how much of it there will be, have not yet been disclosed.