A French startup has made a significant seed investment to redesign computing infrastructure for developers looking to build and train AI applications more efficiently.
FlexAI, as the company is called, has been operating in secret since October 2023, but the Paris-based company is officially launching on Wednesday with 28.5 million euros ($30 million) in funding while also announcing its first product: a On-demand cloud service for AI training.
This is a big change for a seed round, which usually means an extensive founder pedigree – and that’s the case here. Brijesh Tripathi, co-founder and CEO of FlexAI, was previously a senior design engineer at GPU giant and now AI darling Nvidia before taking on various senior engineering and architecture roles at Apple. Tesla (working directly under Elon Musk); Zoox (before Amazon acquired the autonomous driving startup); and most recently, Tripathi was VP of Intel’s AI and super computing platform offshoot AXG.
FlexAI co-founder and CTO Dali Kilani also has an impressive resume. He has held various technical roles at companies such as Nvidia and Zynga and most recently took on the role of CTO at French startup Lifen, which develops digital infrastructure for the healthcare industry.
The seed round was led by Alpha Intelligence Capital (AIC), Elaia Partners and Heartcore Capital, with participation from Frst Capital, Motier Ventures, Partech and InstaDeep CEO Karim Beguir.
FlexAI team in Paris Photo credits: FlexAI
The computer puzzle
To understand what Tripathi and Kilani are trying to do with FlexAI, it is worth first understanding what developers and AI practitioners face when accessing “computes”; This refers to the computing power, infrastructure, and resources required to perform computing tasks such as processing data, running algorithms, and running machine learning models.
“Using any infrastructure in the AI space is complex; It’s not for the faint of heart and not for the inexperienced,” Tripathi told TechCrunch. “You have to know too much about how to build infrastructure before you can use it.”
In contrast, the public cloud ecosystem that has developed over the past few decades is a good example of how an industry emerged from developers’ need to build applications without worrying too much about the back end make.
“If you’re a small developer and want to write an application, you don’t need to know where it’s running or what the backend is – you just spin up an EC2 (Amazon Elastic Compute Cloud) instance and you’re good to go “We are finished,” said Tripathi. “That is not possible today with AI calculations.”
In the AI space, developers have to figure out how many GPUs (graphics processing units) they need to connect together over what type of network, managed by a software ecosystem that they are entirely responsible for setting up. If a GPU or network fails, or something in that chain goes wrong, the responsibility falls on the developer to fix the problem.
“We want to bring AI computing infrastructure to the same level of simplicity that the general-purpose cloud has achieved – after 20 years, yes, but there is no reason why AI computing cannot see the same benefits,” Tripathi said. “We want to get to a point where you no longer need data center experts to run AI workloads.”
As the current version of its product is put through its paces with a handful of beta customers, FlexAI will launch its first commercial product later this year. It’s essentially a cloud service that connects developers to “virtual heterogeneous computing,” meaning they can run their workloads and deploy AI models across multiple architectures, paying on a usage basis rather than GPUs on a dollar basis -rent per hour basis.
GPUs are important cogs in AI development and are used, for example, to train and run large language models (LLMs). Nvidia is one of the standout players in the GPU space and a key beneficiary of the AI revolution sparked by OpenAI and ChatGPT. In the 12 months since OpenAI launched an API for ChatGPT in March 2023, allowing developers to integrate ChatGPT functionality into their own apps, Nvidia’s shares rose from around $500 billion to over $2 trillion -Dollar.
LLMs are now pouring out of the technology industry, and at the same time demand for GPUs is skyrocketing. However, GPUs are expensive to run, and renting them for smaller tasks or ad hoc use cases doesn’t always make sense and can be prohibitively expensive. For this reason, AWS has looked into temporary leases for smaller AI projects. But renting is still renting, which is why FlexAI wants to abstract away the underlying complexity and allow customers to access AI computing power on demand.
“Multicloud for AI”
The starting point of FlexAI is that most developers don’t do this Really They largely care about which GPUs or chips they use, whether Nvidia, AMD, Intel, Graphcore or Cerebras. Their main concern is to develop their AI and build applications within their budget constraints.
This is where FlexAI’s concept of “universal AI computing” comes into play, where FlexAI takes the user’s requirements and maps them to the architecture that makes sense for the task at hand, taking care of any necessary conversions across the various platforms, be it Intel’s Gaudi infrastructure, AMD’s Rocm or Nvidia’s CUDA.
“This means the developer only focuses on building, training and using models,” Tripathi said. “We take care of everything that lies underneath. We manage outages, recovery and reliability, and you pay for what you use.”
In many ways, FlexAI aims to accelerate AI, which is already happening in the cloud. This means more than just replicating the pay-per-use model: it means the ability to use “multi-cloud” by relying on the other advantages of different GPU and chip infrastructures.
FlexAI channels a customer’s specific workload according to their priorities. If a company has a limited budget for training and fine-tuning its AI models, it can do so within the FlexAI platform to get the most computational bang for its buck. That might mean turning to Intel for cheaper (but slower) computing power, but if a developer has a small run that requires the fastest possible output, that could go through Nvidia instead.
Under the hood, FlexAI is essentially a “demand aggregator” that rents the hardware itself in the traditional way and, using its “strong connections” with the folks at Intel and AMD, secures preferential pricing that it distributes to its own customer base. That doesn’t necessarily mean avoiding front-runner Nvidia, but it may mean there’s a lot of incentive for them to engage with aggregators such as FlexAI.
“If I can make it work for customers and get tens to hundreds of customers into their infrastructure, then they will do it [Intel and AMD] “I will be very happy,” said Tripathi.
This is in contrast to similar GPU cloud providers in the space, such as the well-funded CoreWeave and Lambda Labs, which focus exclusively on Nvidia hardware.
“I want to bring AI computing to the point where current general-purpose cloud computing is,” noted Tripathi. “Multicloud is not possible with AI. You have to choose specific hardware, number of GPUs, infrastructure and connectivity and then maintain it yourself. Today, this is the only way to actually receive AI invoices.”
When asked who the exact launch partners were, Tripathi said he could not name all of them as some of them did not have “formal commitments”.
“Intel is a strong partner, they definitely provide infrastructure, and AMD is a partner that provides infrastructure,” he said. “But there is a second level of partnerships with Nvidia and some other silicon companies that we’re not ready to share yet, but they’re all in the mix and the MOUs [memorandums of understanding] are currently being signed.”
The Elon Effect
Tripathi is well prepared for the challenges ahead, having worked in some of the largest technology companies in the world.
“I know enough about GPUs; I used to build GPUs,” Tripathi said of his seven-year stint at Nvidia, which ended in 2007 when he moved to Apple when the first iPhone was released. “At Apple, I focused on solving real customer problems. I was there when Apple started developing its first SoCs [system on chips] for phones.”
Tripathi also spent two years from 2016 to 2018 as head of hardware engineering at Tesla, where he worked directly under Elon Musk for the last six months after two people above him abruptly left the company.
“At Tesla, I learned and adopted into my startup that there are no limitations other than science and physics,” he said. “The way things are done today is not the way they should be or need to be done. You should strive for what is right from the ground up and to achieve this, remove every black box.”
Tripathi was involved in Tesla’s transition to making its own chips, a move that has since been emulated by GM and Hyundai, among other automakers.
“One of the first things I did at Tesla was figure out how many microcontrollers there were in a car, and to do that we literally had to go through a bunch of these big black boxes with metal shielding and casing around them to find these really tiny little ones in there Microcontrollers,” said Tripathi. “And at the end we put it on a table, laid it out and said, ‘Elon, there are 50 microcontrollers in a car.’ And we sometimes pay a thousand times the profit margin for them because they’re shielded and protected in a big metal case.” And he says, “Let’s make our own.” And we did.”
GPUs as security
Looking to the future, FlexAI also aims to build its own infrastructure, including data centers. This, says Tripathi, will be financed through debt financing, building on a current trend in which competitors in this space, including CoreWeave and Lambda Labs, are using Nvidia chips as collateral to secure loans – rather than giving away more equity.
“Bankers now know how to use GPUs as collateral,” Tripathi said. “Why give away equity? Until we become a true computing provider, the value of our business will not be sufficient to provide us with the hundreds of millions of dollars we need to invest in building data centers. If we only had equity, we would disappear when the money was gone. But if we actually put it on GPUs as collateral, they can take the GPUs away and put them in another data center.”