The current wave of AI chip development has produced more GPU-adjacent startups than any prior period in semiconductor history. Billions in venture capital have flowed into companies promising faster training, cheaper inference, and hardware that will finally break NVIDIA’s grip on AI compute infrastructure. Most of them are building chips. Nearly all of them are missing the point.
The problem enterprise AI buyers face is not that no fast GPU chips exist outside NVIDIA’s product line. The problem is that their software cannot run on those chips without significant engineering investment — and that investment, in most organizations, is too expensive to justify the hardware savings. Competing on silicon specifications does not resolve that constraint. Raja Koduri understands this distinction precisely, and Oxmiq Labs, the company he founded in 2023, is built around it.
The Hardware-First Assumption Underlying Most GPU Challengers
The typical AI chip startup argument runs as follows: NVIDIA GPUs are expensive, supply-constrained, and power-hungry. A new chip architecture — designed from scratch for AI workloads rather than adapted from graphics rendering pipelines — can deliver better performance per watt, better performance per dollar, and better throughput on the specific matrix operations that dominate modern neural network computation. Build the chip. Win the market.
That argument is technically coherent. Several companies pursuing it have produced impressive silicon. Chips designed specifically for transformer-based inference workloads can, under the right conditions, outperform general-purpose GPU hardware on those specific tasks. The hardware engineering is real.
What the argument omits is the question of software. Enterprise AI teams do not select compute infrastructure based on peak hardware benchmarks. They select it based on whether their existing software stack will run on it reliably, without modification, at production scale. A chip that outperforms a GPU on a synthetic benchmark but requires weeks of software migration work to support a production workload is not, in practice, a viable substitute for that GPU. It is a project.
Most GPU challengers treat this software gap as a solvable second-order problem — something to be addressed through migration tools, compatibility libraries, and engineering support engagements after the hardware is validated. Koduri’s diagnosis is that the software gap is the primary problem, and that treating it as secondary is precisely why well-funded hardware challengers have failed to meaningfully displace CUDA-based infrastructure despite years of effort.
What the Market Is Actually Buying
Enterprise procurement decisions in GPU compute are not purely technical evaluations. They are risk assessments. The decision-maker — whether a CTO, a VP of infrastructure, or a machine learning platform team lead — is evaluating not just whether a piece of hardware performs well on a benchmark, but whether adopting it creates new operational risks that the organization cannot absorb.
Switching to a non-NVIDIA GPU platform in an environment where all production AI software was written for CUDA introduces multiple categories of risk simultaneously: compatibility risk, performance regression risk, toolchain support risk, and talent risk (the engineers who know the new platform are fewer and harder to hire). Each category is individually manageable. Together, they compose a switching cost that most organizations rationally decline to pay, regardless of the hardware price differential.
What the market is actually buying, when it buys NVIDIA GPUs at premium prices, is certainty. Certainty that the software will run. Certainty that the team knows how to use the tools. Certainty that something going wrong in production will have a documented solution path.
A hardware challenger that does not address this certainty equation is competing on the wrong dimension. Oxmiq Labs is competing on the right one.
The Diagnostic Advantage of Koduri’s Career
Raja Koduri‘s professional background is not conventional startup founder material. He spent more than two decades at the top of the GPU industry — as a chip architect, as a division head, and as a corporate chief architect — before founding Oxmiq Labs. That career is directly relevant to the problem his company is addressing, because it gave him direct operational experience of every significant prior attempt to break CUDA’s market position.
At Advanced Micro Devices, where he served as Senior Vice President and Chief Architect of the Radeon Technologies Group, Koduri oversaw AMD’s GPU programs through multiple hardware generations. AMD’s Radeon hardware has, at various points, been technically competitive with NVIDIA’s offerings in certain workload categories. That competitiveness did not translate into AI compute market share at scale, because the software ecosystem — CUDA-optimized frameworks, CUDA-tuned libraries, CUDA-fluent engineering talent — remained the dominant procurement criterion.
At Intel, where he served as Chief Architect and Executive Vice President of the Architecture, Graphics and Software division, Koduri led both a discrete GPU hardware program and the oneAPI software initiative — Intel’s attempt to build a heterogeneous compute framework that would abstract workloads across different hardware targets. The effort produced real technology: Xe-architecture GPU hardware reached market, and oneAPI established a legitimate programming model for heterogeneous compute. It did not shift enterprise AI infrastructure spending in material volumes.
The pattern across both experiences is the same: technically capable hardware, supported by real software frameworks, failing to displace an incumbent ecosystem not because the new offering was inadequate but because it asked developers to change their behavior. Koduri’s diagnosis, drawn from running those programs from the inside, is precise: any approach that requires developers to modify existing code will encounter adoption friction that hardware quality and software tool quality cannot fully overcome.
The Oxmiq Differentiation: Portability Without Migration
Oxmiq Labs, founded in 2023 and headquartered in San Francisco, is a GPU software and intellectual property company. Its core technology enables Python-based AI workloads written for CUDA to run on non-NVIDIA, RISC-V-based GPU hardware without modification to application code.
This framing — portability without migration — is the precise inversion of what previous GPU challengers offered. Those challengers offered migration: move your workloads to our platform, use our tools, rewrite the parts that do not transfer, and you will eventually reach parity with what you had on NVIDIA hardware. The effort was real and the support was genuine. But the ask remained an ask: developers had to do something they would not otherwise do.
Oxmiq’s architecture eliminates the ask. The developer’s code does not change. The frameworks do not change. The training scripts, the inference pipelines, the evaluation tooling — none of it is touched. The layer that changes is the execution infrastructure underneath the code, not the code itself. For an enterprise AI team evaluating whether to route workloads to non-NVIDIA hardware, that difference changes the risk calculus entirely. There is no migration project to manage, no compatibility regression to validate, no new toolchain for the engineering team to learn.
The hardware substrate for Oxmiq’s approach is RISC-V — an open-source instruction set architecture governed by an open consortium and unencumbered by the proprietary licensing constraints that apply to architectures like x86 or Arm. RISC-V carries no legacy architectural assumptions from pre-AI computing eras, no vendor-controlled design restrictions on the software layers that sit above it, and no single-vendor supply chain exposure. For a company building software portability infrastructure, a hardware foundation that imposes no constraints on the compatibility layer it supports is architecturally coherent in a way that proprietary ISA alternatives are not.
Why the Advisory Network Matters
Beyond his role at Oxmiq Labs, Koduri serves in advisory and board capacities for leading semiconductor and AI companies. That positioning is not incidental to Oxmiq’s strategy — it is part of the execution infrastructure.
Technology infrastructure companies succeed not just by building the right product but by building the right relationships with the organizations that will evaluate, pilot, and eventually deploy that product at scale. The decision-makers at cloud providers, hyperscalers, and enterprise AI platform teams are not searching for unknown vendors. They are evaluating solutions brought to them by people whose technical judgment they already trust.
Koduri’s career built exactly that trust, across exactly the organizations that matter most as early adopters of GPU portability infrastructure. The technical credibility required to have those conversations — the ability to speak fluently about GPU architecture, CUDA internals, software execution environments, and enterprise compute economics — is inseparable from the product itself. An engineer who has not built GPU programs at scale cannot credibly represent a portability solution to teams that have.
The Market Condition That Makes This the Right Moment
The timing of Oxmiq Labs’ founding is not incidental. The acceleration of large language model deployment between 2022 and 2024 exposed the GPU supply concentration problem in ways that years of academic discussion had not. Enterprise AI teams that had accepted GPU procurement constraints as a background condition suddenly found those constraints translating into multi-month hardware lead times, allocation competition with larger customers, and capital expenditure volatility that made infrastructure planning genuinely difficult.
The supply problem is, structurally, a portability problem. Organizations with no viable alternative GPU platform have no alternative supply chain. Every AI chip startup building novel hardware is, in theory, part of the solution to that supply problem — but only if the software gap is resolved. Hardware that requires a migration project to use is not an available alternative for a team operating under production pressure. It is a future project.
Oxmiq Labs is building the mechanism that converts available alternative hardware into actually deployable alternative hardware. The market condition that makes that mechanism valuable — concentrated GPU supply, concentrated software ecosystem dependency, and enterprise demand that exceeds available capacity — is not temporary. It is structural, and it is not resolved by building faster chips.
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.
