Luminal Raises $23 Million to Take On Nvidia’s CUDA Software
Luminal raised $23 million in a Series A led by Standard Capital to help AI models run on chips other than Nvidia's. Backers, product and rivals.

Luminal AI Inc. has raised $23 million in a Series A round to build software that lets AI models run on chips other than Nvidia Corp.’s, according to SiliconANGLE. Standard Capital led the round, and the total Luminal has raised now exceeds $28 million.
The startup is going after the software that keeps most AI developers tied to Nvidia hardware. Its product launched four months ago and already processes a large volume of AI traffic.
Who put money in
Existing investors Y Combinator, SagaVC, Liquid 2 and Felicis took part in the round. It comes about a year after Luminal raised $5.3 million in seed funding. The round is a Series A, the stage after seed funding (see how startup funding rounds work).
Dalton Caldwell of Standard Capital said the firm backed Luminal because AI inference, the work of running a trained model to answer requests, is the fastest-growing market in the world. He pointed to the company’s technical depth and its record of getting strong performance from mixed hardware, and said getting more out of GPUs already purchased is high leverage.
The software barrier Luminal wants to remove
Nvidia has a market valuation of $5 trillion, which SiliconANGLE says makes it the most valuable publicly traded company in the world. The outlet argues that its lead rests on software as much as on chips. Nvidia’s CUDA framework has existed for almost two decades, and major AI frameworks including TensorFlow and PyTorch were built natively on it. Models written that way run on Nvidia GPUs without changes. Moving them to other chips means rewriting much of the underlying code, which takes time.
Rivals such as Advanced Micro Devices Inc., Google LLC and Amazon Web Services Inc. have built capable chips, according to SiliconANGLE, but have not matched the size of Nvidia’s software ecosystem.
Luminal’s answer has two parts: a search compiler and a managed inference cloud. Co-founder and Chief Executive Joe Fioti told SiliconANGLE that the compiler tries many ways of running a neural network on a target chip, benchmarks each one, picks the best and generates the code automatically. It then stores that implementation for reuse on repeated workloads.
The system is aimed at high-volume, always-on applications such as AI coding tools, image generators, video generators and real-time voice agents. Fioti said: “The cost of running AI shouldn’t limit who gets to build it.”
Early traction and competition
According to SiliconANGLE, Luminal now processes more than 45 trillion tokens per month, which it puts at roughly 10% of all tokens consumed so far by Together AI Inc., a company valued at $8 billion that has served the market for more than four years. Luminal has a partnership with AWS and runs a significant share of its workloads on Amazon’s Trainium processors.
It is not alone. Google and AWS have each launched alternatives to PyTorch that make it easier to build AI applications for their own processors. Spectral Compute Ltd. sells a toolkit and compiler that runs CUDA-based code on any AI-accelerated chip. The UXL Foundation’s OneAPI, an open-source effort, converts CUDA code into software that runs on other silicon.
What this means for the company
The new money goes toward expanding the engineering team, scaling go-to-market operations and tuning the compiler and inference infrastructure to support more chips and more AI models. Fioti said the company began by building in the open on GitHub as AI adoption took off and saw demand for a more efficient way to run models across different hardware.
Source: SiliconANGLE



