A French startup has raised a significant seed investment to “reimagine computing infrastructure” for developers who want to build and train AI applications more efficiently.
The company, called FlexAI, has been operating in stealth since October 2023, but the Paris-based company officially launched on Wednesday with €28.5 million ($30 million) in funding, along with its first The company is announcing its product, an on-demand cloud service. For AI training.
This is quite a change for a seed round and usually means substantial founder pedigree, which is the case here. Brijesh Tripathi, co-founder and CEO of FlexAI, previously his GPU giant and now his AI darling, was a senior design engineer at Nvidia and later held various senior engineering and architect positions at Apple. I got the role. Tesla (working directly under Elon Musk). Zoox (before Amazon acquired the self-driving startup). And most recently, Tripathi was vice president of AXG, an offshoot of Intel's AI and supercomputing platform.
Dali Kilani, co-founder and CTO of FlexAI, also has an impressive resume, having held various technical roles at companies such as Nvidia and Zynga, and most recently at Lifen, a French startup developing digital infrastructure for the healthcare industry. I played the role of CTO.
The seed round was led by Alpha Intelligence Capital (AIC), Elaia Partners, and Heartcore Capital, with participation from Frst Capital, Motier Ventures, Partec, and InstaDeep CEO Karim Beguir.
FlexAI team in Paris
computing challenges
To understand what Tripathi and Kilani are trying to do with FlexAI, it's important to first understand what developers and AI practitioners face when it comes to access to “compute.” This refers to the processing power, infrastructure, and resources required to perform computational tasks such as processing data, running algorithms, and running machine learning models.
“Using infrastructure in the AI space is complex. It's not for the faint of heart and it's not for the inexperienced,” Tripathi told TechCrunch. “Using infrastructure requires a good knowledge of how it is built.”
In contrast, the public cloud ecosystem that has evolved over the past few decades is a great example of how an industry was born out of developers' need to build applications without worrying too much about the backend. .
“If you're a small developer and you want to create an application, you don't need to know where your application is running or what the backend is. You can just launch an EC2 (Amazon Elastic Compute Cloud) instance. It's over,'' Tripathi said. “He can't do that with AI computing today.”
In the field of AI, developers need to know how many GPUs (graphics processing units) they need to interconnect on what type of network, managed through a software ecosystem that takes full responsibility for the setup. . If the GPU or network fails, or if there's a problem somewhere in that chain, the onus is on the developer to fix it.
“We want to bring AI computing infrastructure to the same level of simplicity that general-purpose cloud has reached. After 20 years, that's true, but AI computing doesn't have the same benefits. There is no reason,” Tripathi said. “We want to get to the point where you don't need to be a data center expert to run AI workloads.”
FlexAI is working on its current product iteration with a small number of beta customers and plans to launch its first commercial product later this year. This is essentially a cloud service that connects developers to “virtual heterogeneous computing.” This means that instead of renting GPUs on a dollar-per-hour basis, you can pay on a usage basis to run your workloads and deploy AI models across multiple architectures.
GPUs are essential cogs in AI development, helping to train and run large-scale language models (LLMs), for example. Nvidia is one of the preeminent players in the GPU space and one of the main beneficiaries of the AI revolution sparked by OpenAI and ChatGPT. In the 12 months since OpenAI announced ChatGPT's API in March 2023, allowing developers to incorporate ChatGPT functionality into their own apps, Nvidia's stock has ballooned from about $500 billion to more than $2 trillion. It went up.
LLMs are leaving the technology industry, and with it, the demand for GPUs is surging. However, GPUs are expensive to run, and renting them from a cloud provider for small jobs or ad-hoc use cases doesn't always make sense and can be prohibitively expensive. This is why AWS is dabbling in short-term rentals for small AI projects. But a rental is still a rental. That's why FlexAI wants to remove the underlying complexity and give customers access to AI computing on demand.
“Multi-cloud for AI”
FlexAI's starting point is that most developers don't really care about the GPUs or chips they use, including Nvidia, AMD, Intel, Graphcore, and Cerebras. Their main concern is whether he can develop AI and build applications within budget constraints.
This is where FlexAI's concept of “universal AI computing” comes into play. FlexAI takes a user's requirements, assigns them to the right architecture for that specific job, and handles all necessary translations between different platforms, such as Intel's Gaudi. Infrastructure, AMD's Rocm or Nvidia's CUDA.
“What this means is that developers are solely focused on building, training, and using models,” Tripathi said. “We take care of everything underneath. Failure, recovery, and reliability are all managed by us, and you only pay for what you use.”
In many ways, FlexAI is trying to fast-track what's already happening in the cloud with AI. This means you can not only replicate a pay-as-you-go model, but also move to “multi-cloud” with a range of benefits. of various GPU and chip infrastructures.
For example, FlexAI handles a customer's specific workloads depending on the customer's priorities. If your company has a limited budget to spend on training and fine-tuning your AI models, you can set your budget within the FlexAI platform to get the most compute bang for your buck. This might mean going through Intel to get cheaper (but slower) computing, but if developers are doing small runs where they need the fastest possible output Alternatively he can be processed via Nvidia.
Internally, FlexAI is essentially a “demand aggregator,” renting out the hardware itself in the traditional way and leveraging its “strong connections” with folks at Intel and AMD to serve its entire customer base. We have secured preferential prices to spread the word. This doesn't necessarily mean bypassing the central player, his Nvidia, but perhaps in large part, as Intel and AMD fight over the scraps of GPUs left behind by his Nvidia. means there is a strong incentive to work with such aggregators. As FlexAI.
“If we can make this work for our customers and bring dozens to hundreds of customers into their infrastructure, they can [Intel and AMD] I will be very happy,” Tripathi said.
This is in contrast to similar GPU cloud players in the space, such as well-funded CoreWeave and Lambda Labs, which are focused on Nvidia hardware.
“We want to bring AI computing to the level of general-purpose cloud computing today,” Tripathi said. “You can't do multicloud on top of AI. You have to choose specific hardware, number of GPUs, infrastructure, connectivity and maintain it yourself. It’s the only way to make it happen.”
Asked who the exact launch partners are, Tripathi said he could not name them all as there was no “formal commitment” from some of the companies.
“Intel is a strong partner. They definitely provide the infrastructure, and AMD is also a partner that provides the infrastructure,” he said. “But we have a second tier 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 in the MOU.” [memorandums of understanding] Currently being signed. ”
elon effect
Tripathi has worked for some of the world's largest technology companies and is well-equipped to meet the challenges ahead.
“I know a lot about GPUs. I used to build GPUs,” Tripathi said of his seven years at Nvidia before 2007, when he jumped to Apple to launch the first iPhone. Ta. “At Apple, I became focused on solving real customer problems. I was there when Apple started building its first SoC. [system on chips] For telephone. ”
Tripathi also spent two years as head of hardware engineering at Tesla from 2016 to 2018, but the last six months were spent working under Elon Musk after two of his predecessors suddenly left. I ended up working directly at.
“What I learned at Tesla and am bringing to my startup is that there are no constraints other than science and physics,” he said. “The way we do things today is not and should not be the way we do things. We should seek right action from first principles, and to do that we need to remove all the black boxes.”
Tripathi was involved in Tesla's move to make its own chips, a move that has since been emulated by automakers such as GM and Hyundai.
“One of the first things I did at Tesla was figure out how many microcontrollers were in the car. We literally had to sort through it to find a really small microcontroller in there,” Tripathi said. “And in the end we put it on a table and lined it up and said, 'Elon, there's 50 microcontrollers in the car. And they're shielded and protected by a big metal case. , sometimes paying a 1,000x margin.'' And he said, “Let's build it ourselves.'' And we did it. ”
GPU as collateral
Looking further into the future, FlexAI also wants to build its own infrastructure, including a data center. The funding will be funded through debt financing, Tripathi said, following recent moves by rivals in the space such as CoreWeave and Lambda Labs to use Nvidia chips as collateral to secure loans rather than handing over more equity. It is said to be based on trends.
“Bankers now know how to use GPUs as collateral,” Tripathi says. “Why give away stock? Until we become a true computing provider, our value alone will not be enough to attract the hundreds of millions of dollars of investment needed to build data centers. Stock alone. If we run out of money, we're gone. But if we actually put it in GPUs as collateral, they can take them away and put them in other data centers.”