Oxmiq Raises $35M for GPU IP, Expands Focus to Data Center Design

Oxmiq Labs, the GPU startup founded by Raja Koduri, raised $35 million in Series A funding to expand its technology stack into custom data center-scale hardware and software, including bespoke silicon. The company added Tenstorrent CEO Jim Keller and Intel veteran Valluri Rao to its board. Oxmiq aims to offer a full stack from energy generation to AI factory design, with new offerings like OxFactory orchestration software and custom "Electron to token" appliances.

Oxmiq Labs, the GPU hardware and software startup founded by Raja Koduri, has raised $35 million in a Series A round, bringing its total funding to $60 million. This funding will cover an expansion of the company’s technology stack to include custom data center-scale hardware and software, including bespoke silicon. The company has also added Tenstorrent CEO Jim Keller and Intel veteran Valluri Bob Rao to its board. Technology stack Beyond silicon IP, Oxmiq wants to build a full technology stack from energy generation through to data center infrastructure design for the AI factory, Oxmiq CEO Raja Koduri told EE Times. The new technology stack includes custom design offerings at every layer, from GPU hardware IP the previously announced OxCore https://www.eetimes.com/koduri-unveils-gpu-hardware-ip-and-software-startup/ , through appliance design what Koduri calls “Electron to token machines” , to AI data center pod design. At the highest level, AI infrastructure design will include reference designs and custom designs for AI factories. The software side will include a new data center-scale orchestration software platform called OxFactory. View All https://www.eetimes.com/category/sponsored-content/ “OxFactory is a data center-scale version of OxCapsule, which is for edge and consumer hardware—we are continuing to develop OxCapsule, and we are having good traction there,” Koduri said. “But there is a difference between OxCapsule and a software stack that operates 10,000 GPUs and is resilient across hardware failures and network failures, which we are now calling OxFactory.” OxCapsule is Oxmiq’s GPU container software, partially based on software licensed from Intel, that abstracts away heterogeneous hardware and analyzes workloads to assign compute based on the desired speed and cost of execution. The challenge for data center orchestration software goes beyond CUDA, Koduri said, since many other layers of the stack only work with Nvidia or have been tested only on Nvidia hardware. This can make it tough to get non-Nvidia data center hardware up and running. “From day one, if your revenue counter is not ticking, it’s a massive loss,” he said. “The reason the software problem is such a big deal is you have to make money on the day the racks are up and running, or the whole financial equation doesn’t make sense.” OxPython, the company’s Python module, is designed to bridge the gap when porting from Nvidia GPUs to other hardware. Electrons to tokens Oxmiq’s Electron-to-token machines will be bespoke appliances designed for individual customers. This new offering will include the option to build custom silicon for customers based on the OxCore, Koduri said. “This business model has opened up in the last 12 months because infrastructure companies can afford to pre-pay for tapeouts, masks, wafers, and production,” Koduri said, noting that this model is a lot more efficient for a startup than spending hundreds of millions of dollars designing a chip and then trying to sell it. In the past, it wasn’t worth the risk for hyperscalers to spend $100 million to tape out a custom data center chip, Koduri said, but equipping a 100-MW data center with state-of-the-art GPUs now costs closer to $5 billion, which changes the economics. “The entire supply chain is margin stacked,” he said. “If you can buy wafers direct from the foundries and have an ability to construct your product, you can get the same compute, same token rate or higher, for $1 billion instead of $5 billion. So, you can save $4 billion by going direct. This is the reason hyperscalers do it.” This is an opportunity for Oxmiq’s OxCore GPU IP, Koduri said. “We have a couple of critical customer engagements to translate the IP into chips, and the operational software stack to run these AI factories has tremendous value,” he said. The OxCore, with its mix of CUDA-compatible GPU cores, tensor cores, and orchestration engines, is up and running on FPGA today, executing workloads together with Oxmiq’s software stack. The OxCore fits into the company’s chiplet architecture with its configuration tool, OxQuilt, which will allow bespoke designs by combining different ratios of compute, memory, and I/O chiplets. Customers have varied demands that a single pre-configured package could never be optimal for, Koduri said, so there is plenty of scope for customization down to the package level. “If you are selling tokens by throughput and need a large batch size, that’s one ratio,” he said. “If you’re selling low-latency tokens, it’s a different ratio, and there are customers who don’t care about either throughput or latency, but they can’t exceed a certain capex or megawatt budget.” Customer capex demands could mean custom designs on older process nodes, coupled with cheaper DRAM or emerging memories, he said. Training clusters would need more compute chiplets, while low-latency demands more memory chiplets, for example. “The beauty of it is, there is no change in the software stack, even for an SRAM-based design versus a DRAM-based one,” he said. Engineering partner Oxmiq has a couple of big customers at the table, Koduri said. The company recently announced a strategic partnership with AM Intelligence Labs, part of AM Group in India AM Group is also the parent company of Greenko, India’s largest renewable energy producer with 50 GW of energy capacity . Oxmiq will be the architecture and engineering partner for a 2-GW AM Intelligence Labs AI deployment, designing the systems architecture, hardware roadmap, and supply chain strategy that will underpin the facility. The idea is to co-design everything from the renewable energy generation through the compute architecture and data center design to optimize economics and power efficiency. Phase one is a 1-GW cluster in Noida, Uttar Pradesh, India, which will come online in late 2027. For this cluster, Oxmiq will help AM Intelligence Labs architect its data centers, including procuring compute, Koduri said. “Eventually, we’ll give them the ability to roll out something custom that gives them more control and more margin on those elements as well,” he said. Small team Oxmiq today employs 45 people; this is much, much smaller than most data center chip companies, especially given that Oxmiq is developing at every layer of the stack. “What’s fascinating is senior people with the breadth and depth of experience paired with state-of-the-art agentic tools are incredibly productive now,” Koduri said. “There is a scale factor of 10… agentic tools are very powerful in the hands of an expert who knows how to verify the results they produce. If you had asked me this question with this stack two years ago, frankly, there are things I would not have attempted. But today, we are confident we can pull this off.” Oxmiq will also work with partners. The company intends to license elements of its OxFactory software layer from a partner rather than starting from scratch, for example, and the company will also leverage silicon design services where necessary for the execution of custom chips. Koduri plans to grow the team to 60 or 70. “People ask me: What’s your five-year vision?” he said. “All I say is: keep the headcount in two digits. If I have a magic wand and a wish is granted to me, it will be that we can accomplish everything we want to with 99 people.” Read also: Jim Keller: ‘AI Still Obeys the Old Laws of Compute ’ Invoking Rent’s Rule and Amdahl’s Law, Keller argues that memory and communication, not bigger processors, will define the future of AI infrastructure. Koduri Unveils GPU Hardware IP and Software Startup Raja Koduri’s new GPU software and hardware IP startup, Oxmiq Labs, is emerging from stealth with $20 million in seed funding.