A Korean AI startup's proof-of-concept could loosen NVIDIA's grip on the AI training stack.
TL;DR: VIDRAFT, a Korean AI startup, has successfully pre-trained a 46-million-parameter model — named VKORE-46M-NPU — using inference-only NPU hardware, with no high-end NVIDIA GPUs involved. The milestone, first reported on dev.to and analyzed by Mukta AI, challenges the long-held assumption that backpropagation-based training requires CUDA-capable GPUs. The result is an early but significant signal that hardware-agnostic AI development may be technically feasible.
VIDRAFT, a Korean AI startup, has completed what may be one of the more disruptive proof-of-concepts in recent AI hardware history: pre-training a neural network entirely on inference-oriented NPUs, sidestepping the CUDA ecosystem that has defined AI development for over a decade. The achievement, detailed in a dev.to post by the VIDRAFT team and subsequently analyzed in an editorial by Mukta AI published October 14, 2026, centers on a 46-million-parameter model the company calls VKORE-46M-NPU.
VIDRAFT publicly documented the pre-training of VKORE-46M-NPU — a 46-million-parameter language model — using hardware that was designed for inference, not training. The key distinction matters: inference-oriented NPUs are typically built for deploying already-trained models at low power and low cost, not for the computationally intensive, gradient-heavy workloads that training demands.
For years, the standard assumption across the machine learning industry has been that backpropagation — the core mathematical process by which neural networks learn — requires the massive parallel compute throughput and specialized memory hierarchies that NVIDIA's GPU lineup, and specifically its CUDA software platform, provides. VIDRAFT's VKORE-46M-NPU result directly challenges that assumption by demonstrating that a model can, in fact, be brought through a full pre-training run on silicon never intended for that purpose.
The original technical reporting appeared under the ML tag on dev.to. Mukta AI's editorial team subsequently provided broader industry context and critical analysis of what the achievement does — and does not — prove.
The implications of VIDRAFT's announcement extend well beyond a single 46-million-parameter model. The deeper significance lies in what successful NPU-native training would mean for the economics of AI research. High-end GPUs carry substantial costs — both in direct hardware purchases and in cloud rental fees — creating a significant barrier to entry for smaller research teams, independent developers, and startups operating outside the well-funded orbits of big tech.
If inference-grade chips, which are far more widely deployed and cost-effective — including in mobile and edge devices — can be repurposed for training, the addressable pool of hardware available to AI developers expands dramatically. As Mukta AI's editorial frames it, the bottleneck would shift from a physical hardware limitation to a software abstraction challenge, which is a fundamentally more solvable problem.
That said, Mukta AI's analysis is careful to note that a 46-million-parameter model is orders of magnitude smaller than the multi-billion-parameter models that define the current frontier of AI development. Open questions remain around thermal management, memory bandwidth, training speed relative to GPU alternatives, and whether NPU-trained models can match the convergence quality of their GPU-trained counterparts. The editorial specifically flags that inference chips are generally built for low-power, thermally constrained environments — conditions that may not scale cleanly to the prolonged, high-intensity demands of large-scale model training.
There is also the question of architectural trade-offs. Training typically requires high-precision gradient computation, whereas inference hardware often assumes quantized, lower-precision arithmetic. Whether VIDRAFT has found a principled solution to this mismatch, or a highly specialized workaround specific to their use case, remains to be seen.
Q: What did VIDRAFT actually build and train?
A: VIDRAFT pre-trained a neural network model called VKORE-46M-NPU, which has 46 million parameters, using inference-only NPU hardware rather than the high-end GPUs and CUDA software stack that AI training has conventionally relied upon.
Q: Why does it matter that training happened on inference-only NPUs?
A: Inference NPUs are far more widely available and cost-effective than high-end training GPUs. If NPU-native training proves scalable and economically viable, it could significantly lower the barrier to AI model development for smaller organizations and reduce dependence on a hardware ecosystem dominated by a single vendor.
Q: Is VKORE-46M-NPU comparable in scale to leading AI models?
A: No. At 46 million parameters, VKORE-46M-NPU is a proof-of-concept that is considerably smaller than the multi-billion-parameter models at the frontier of AI development. Demonstrating that NPU-native training scales to those larger sizes remains an open and unresolved challenge.
Source: Mukta AI(글로벌) (2026-10-14) — original article