Every foundational shift in computing history followed the same pattern. A proprietary platform achieved dominance, accumulated a developer ecosystem, and then became a structural barrier to the next wave of innovation — until a better-designed, more open alternative broke the dependency and reset the competitive landscape. Mainframe vendors controlled compute until the personal computer decoupled processing from centralized hardware. Microsoft’s Windows ecosystem defined software development until the web established a platform-agnostic runtime. x86 server architecture locked enterprise infrastructure until virtualization and cloud computing made the physical layer interchangeable.
The AI compute market is at an earlier stage of that same arc. NVIDIA’s CUDA platform has created a software dependency so entrenched that even technically credible GPU alternatives have struggled to reach their potential. Raja Koduri, founder of Oxmiq Labs and a career GPU architect who has spent more than two decades inside the organizations that have most directly confronted that problem, believes the mechanism for the next transition is now buildable. His company is building it.
The Anatomy of a Platform Transition
Platform transitions do not happen because a new technology is technically superior. They happen because a structural condition changes — specifically, when the cost of remaining on the incumbent platform begins to exceed the cost of switching to the alternative.
In the mainframe era, IBM’s hardware and operating systems were deeply integrated and technically capable. The transition to personal computing did not happen because PCs were immediately more powerful — they were not. It happened because the economics of centralized compute no longer matched the distribution of computing needs, and because a new programming environment emerged that developers could access without IBM’s mediation.
The same pattern repeated with Windows. The developer ecosystem built on Win32 APIs was enormous, commercially entrenched, and technically deep. The web did not displace Windows by offering a better API. It offered a universal runtime — a way to write software once and have it run anywhere, without dependency on any hardware vendor’s proprietary execution environment.
The GPU compute market is structurally analogous to both of these moments. NVIDIA’s CUDA platform is deeply capable, commercially entrenched, and developer-preferred — not because it is the only option, but because it became the option around which the entire AI software ecosystem was organized. The transition away from CUDA dependency will not happen because a better GPU appears. It will happen when a credible compatibility layer exists that lets the existing developer ecosystem run on alternative hardware without modification.
That is what Oxmiq Labs is building.
A Career Spent at the Point of Failure
Raja Koduri‘s academic preparation positioned him for exactly this kind of systems-level analysis. His bachelor’s degree in electronics and communications engineering from Andhra University and his Master of Technology from the Indian Institute of Technology (IIT) Kharagpur gave him a foundation in the electrical and systems engineering disciplines that govern how hardware and software interact at the architectural level.
His professional career then placed him, repeatedly, at the exact organizational point where the theory meets the commercial consequence. At ATI Technologies, Koduri contributed to GPU development at a company competing directly — and credibly — for the graphics hardware market. When Advanced Micro Devices acquired ATI, his scope expanded to Senior Vice President and Chief Architect of the Radeon Technologies Group, where he oversaw GPU architecture across multiple product generations and directly experienced what it meant to ship technically strong hardware into a market where software ecosystem dynamics determined commercial outcomes.
At Apple, Koduri was embedded in a company that had solved the hardware-software integration problem differently than anyone else in the industry. Apple’s approach — treating the silicon and the software stack as a single co-designed system — produced a vertical integration model that removed ecosystem dependency altogether. The lesson was not replicable for most hardware companies, but the underlying insight was clear: developer productivity and software portability are not features to be added after hardware is designed. They are design requirements.
At Intel, as Chief Architect and Executive Vice President of the Architecture, Graphics and Software (IAGS) division, Koduri led the company’s most significant discrete GPU effort in decades. That program required building not only new GPU silicon but an entirely new developer toolchain — effectively asking developers to adopt a new execution environment while CUDA remained available, well-documented, and deeply integrated into the tools they already used. The commercial challenge of that effort clarified, in precise terms, what the actual barrier to GPU ecosystem diversification is and where the solution has to be targeted.
Why RISC-V Is the Right Foundation
Oxmiq Labs, founded in 2023 and headquartered in San Francisco, uses RISC-V as the architectural foundation for its GPU intellectual property. The choice is deliberate and structurally meaningful.
RISC-V is an open-source instruction set architecture — a hardware design language that is not owned by any commercial entity and not subject to the licensing agreements and architectural constraints that govern proprietary instruction sets. For a company building software compatibility infrastructure, the hardware foundation matters because it determines what constraints exist on the software layer. A RISC-V foundation places no proprietary hardware constraints on the compatibility software stack that sits above it.
This is not primarily a cost argument, though open architectures carry real licensing advantages. It is a design argument. Proprietary instruction sets carry embedded assumptions — about how compute operations are organized, how memory is addressed, how instructions are decoded — that reflect the business interests and historical decisions of the companies that created them. RISC-V carries no such legacy. Its design is clean, modern, and optimized for the kinds of workloads — parallel, high-throughput, matrix-intensive — that AI computation demands.
For a startup attempting to build GPU software infrastructure that is genuinely hardware-agnostic, starting on an open hardware foundation is the only architecturally consistent choice.
The CUDA Portability Problem, Stated Precisely
The problem Oxmiq Labs is addressing is specific. It is not that CUDA is a bad programming model. It is that CUDA is a good programming model that has been made artificially inseparable from a specific hardware vendor’s products.
The AI developer community writes software in Python, using machine learning frameworks — PyTorch, TensorFlow, and their derivatives — that were built with CUDA as the primary hardware execution target. Those frameworks contain optimizations, library integrations, and toolchain dependencies that assume CUDA is available. When a developer writes a training loop, a model serving routine, or an inference pipeline, they are almost certainly writing code that will run on CUDA hardware by default — not because they made that choice explicitly, but because every layer of the toolchain beneath them made it implicitly.
This is not a problem that can be solved by offering an alternative framework. Developers are not going to rewrite their codebases for a new runtime — the switching cost is too high, the productivity loss is too immediate, and the risk to production systems is too significant. A solution that asks developers to change their code will not achieve adoption at scale, regardless of its technical merits.
Oxmiq’s approach addresses this directly. The company’s work targets CUDA workload portability — specifically, enabling Python-based AI applications developed in CUDA execution environments to run on non-NVIDIA, RISC-V-based GPU hardware as written. The developer does not change the code. The infrastructure accommodates it.
The Advisory Network as Ecosystem Infrastructure
A technical solution, however well-designed, does not become a market standard through technical merit alone. Platform transitions require adoption infrastructure — the relationships, validations, and integrations that make a new approach the default choice rather than an available option.
Beyond his operational role at Oxmiq Labs, Raja Koduri serves in advisory and board capacities for leading semiconductor and AI companies. That network is not merely a professional credential. It is functionally part of the company’s product strategy. Establishing CUDA portability as a practical standard — something that chip manufacturers, cloud providers, and enterprise AI teams rely on in production — requires organizational relationships that allow Oxmiq’s technology to be tested, validated, and integrated at the infrastructure level.
Koduri’s two decades of senior technical leadership across ATI, AMD, Apple, and Intel produced exactly that kind of network. The engineers and executives who worked alongside him, evaluated his hardware, or built on his architectures are now distributed across the semiconductor and AI infrastructure landscape in positions that directly determine which GPU software standards gain adoption.
Where the Transition Leads
Compute platform transitions, once they begin, tend to move faster than incumbents expect and slower than advocates predict. The web did not immediately displace Windows — but within a decade, the browser had become the dominant application runtime for most enterprise software. Linux did not immediately displace proprietary Unix systems — but it now runs the majority of the world’s servers.
The AI GPU market will follow a similar trajectory. The conditions for transition are present: clear enterprise demand for compute cost reduction, documented supply chain risk from single-vendor dependency, and a developer community large enough that the switching cost of rewriting CUDA-optimized codebases is commercially prohibitive for any individual organization — but commercially worth solving for a company that addresses it at the infrastructure level.
Raja Koduri is building that infrastructure. The technical foundation — RISC-V-based GPU architecture with CUDA workload compatibility — is the product of a career spent at the precise junctions where each prior attempt at GPU ecosystem diversification met its limit. The transition he is working toward is not a prediction. It is an engineering project, grounded in a specific analysis of why previous attempts failed and what a viable approach actually requires.
The history of computing suggests that this kind of transition is inevitable. The question has always been who builds the bridge.
About Raja Koduri
Raja Koduri is an Indian-American computer engineer, technology executive, and founder with more than two decades of experience in GPU architecture and computing platform development. He holds a bachelor’s degree in electronics and communications from Andhra University and a Master of Technology from the Indian Institute of Technology (IIT) Kharagpur. Koduri has held senior roles at ATI Technologies, Advanced Micro Devices (AMD), Apple, and Intel, where he served as Chief Architect and Executive Vice President of the Architecture, Graphics and Software division. In 2023, he founded Oxmiq Labs Inc., a San Francisco-based GPU software and IP startup focused on enabling CUDA workloads to run on non-NVIDIA hardware through RISC-V-based designs and open software frameworks. He also serves in advisory and board capacities for leading semiconductor and AI companies.