VSORA, a French AI chip company, has emerged with ambitious plans to disrupt the market dominated by industry giants such as NVIDIA.
As the AI hardware market continues to evolve rapidly, VSORA is positioning itself as a serious challenger with an innovative architecture focused on solving the cost per query challenge. The company represents Europe's hopes to compete in the high-stakes race to power the next generation of AI applications.
Following its €40 million fundraising round, the company is now focused on bringing its high-performance AI inference chip to market later this year. While NVIDIA has a 3-decade headstart, VSORA's founder believes the company can make a dent in that lead by solving one of the most pressing challenges in the AI industry today: the cost of deploying large language models and generative AI at scale.
"OpenAI turned on the light for LLMs and generative AI," said VSORA founder and CEO Khaled Maalej. "And the big issue with generative AI and LLMs is to reduce the cost per query for the inference."
The Journey from DSP to AI Powerhouse
Maalej has deep roots in the semiconductor industry stretching back almost three decades. His journey includes co-founding a business unit for an American company in France and later starting DiCom, a silicon company that was sold in 2012.
The seeds of VSORA were planted during his time at DiCom. "During this period, we started really seeing the limitation of the memory wall," said Maalej, referring to how memory bandwidth limits processing power in silicon designs.
VSORA was officially founded in 2015, initially focusing on solving the memory bandwidth challenge for 5G applications. The company started with an IP business model, offering its technology for licensing.
They discovered their architecture had significant advantages for AI applications. "People who understand a little bit about our architecture, they told us, 'Guys, you should look at AI as well, because AI is even suffering more from this limitation, '" said Maalej.
From Autonomous Driving to Generative AI
A pivotal moment came when a German automaker approached VSORA about autonomous driving technology. The automaker wanted to build a solution for level four autonomous driving. That collaboration led to the development of a powerful AI processor.
"We worked with them for six quarters, and we ended up releasing a processor IP in 2020, going up to petaflop in terms of processing power," said Maalej. "To put that in perspective, with Nvidia, that time, this probably was 60% more than the NVIDIA A100 performance."
The company's strategy evolved further with the explosion of generative AI in 2022. "OpenAI turned on the light for LLMs and generative AI," Maalej said. "And the big issue with generative AI and LLMs is really to go and reduce the cost per query for the inference."
Solving the Cost Per Query Challenge
VSORA's focus has shifted primarily to addressing the high costs associated with deploying large language models at scale. "The market is looking for a solution, for a technical solution, to reduce the cost per query in order to deploy this massively," Maalej explained. "If you take OpenAI today, they are handling roughly 5,000 queries per second. It's significant, but this number is to put in perspective what Google is doing today. Google is running 100,000 queries per second."
The difference, according to Maalej, comes down to cost: "Why the market is not going from 5k to 100k because of the cost of the query. No one can afford the investment in this."
VSORA discovered that the primary cost driver wasn't just energy consumption, as they initially believed. "We thought that it's coming from energy. Very quickly, we realized that it is not energy that is important. The main cost is coming from the utilization of silicon. And the cost is high because of the utilization rate of the architecture of the silicon today."
The Jotunn8 Chip and Fabless Model
VSORA's flagship product, the Jotunn8 (J8) chip, promises to dramatically improve AI inference efficiency. The architecture leverages chiplet technology to create customizable solutions. "We develop a chiplet at a time. And we play Lego more or less, build a solution for a data center with very high processing power, or for the car market with low processing power," Maalej said.
In 2023, VSORA received a significant boost when TSMC, the world's leading semiconductor foundry, validated its technology. This relationship has enabled VSORA to advance rapidly. "We went from a seven-nanometer design to a five-nanometer design, having access to the very advanced packaging technologies and having access to the latest HBM employees with other vendors."
The result is a powerful chip that can go up to 3.2 petaflops in terms of processing power, and to embed up to 288 gigabytes.
Business Strategy and European Sovereignty
Rather than continuing with an IP licensing model, VSORA has embraced a fabless semiconductor approach where the company will produce and sell the silicon.
While VSORA is positioning itself in the European semiconductor ecosystem, Maalej emphasized that the company's primary value proposition is universal: "Our main bet is not really on this sovereignty and on the public domain. Our main focus, again, is really to reduce the cost per query and offer a solution that will go worldwide."
However, he acknowledges the European dimension: "Europe as well, is looking for a technology as an alternative to US technology today...We are the only company in Europe capable of building this kind of solution."
Funding and Future Plans
The recent €40 million fundraising will support three key milestones. "The first one is to put the silicon in the fab. The second one is to build the servers. And we're going to build as well the small data center... for evaluation, for customers," said Maalej.
The third critical milestone is "to go to the benchmarking platform, which is called the ML Perf data center inference, so way we can really compare the different solutions today and see how we perform regarding the bigger players in this area."
VSORA plans to have its processors in the commercial market later this year. One advantage VSORA has is its experienced team.
"VSORA is really the continuation of the previous company mentioned before," said Maalej. "Most of the people that we were working with before at DiCom are joining VSORA."
This gives the company confidence in its ability to execute. "We have done this fundraising, and we are structuring the company from the financial point of view," he said. "We have done this exercise before. We know how to grow a company, what to do, and what not to do."
