A Korean AI startup's software layer may be the beginning of the end for GPU-only AI training.
TL;DR: VIDRAFT (Vidraft Lab) has demonstrated that inference-only NPUs can perform full-scale language model pre-training from scratch, achieving a 0.99 loss correlation with standard GPU benchmarks. The milestone centers on a 46-million-parameter model trained across 32 inference-dedicated chips using VIDRAFT's software orchestration layer, VKORE. This result directly challenges the long-held assumption that AI model training requires GPU-class hardware.
VIDRAFT, the Korean AI startup pushing the boundaries of hardware-agnostic machine learning, made headlines in mid-October 2026 when it disclosed a result that the broader AI industry had largely assumed was impossible: pre-training a language model entirely from scratch on chips that were designed purely for inference. Reported on October 14, 2026 by Mukta AI (글로벌), the disclosure has drawn significant attention from researchers and infrastructure architects alike.
At the center of the announcement is VKORE, VIDRAFT's software orchestration layer, which the company used to execute backpropagation and optimizer state management on hardware that was never built with those workloads in mind. Inference-optimized NPUs are designed to run the forward pass of a neural network efficiently — they are not architected for the heavy, bidirectional data flows that training typically demands.
Despite that architectural mismatch, VIDRAFT successfully pre-trained a 46-million-parameter language model using 32 inference-dedicated NPU chips. The critical validation metric: a 0.99 loss correlation when compared against equivalent training runs conducted on standard GPU hardware. In practical terms, this means the model quality produced by the NPU-based training run was statistically indistinguishable from what a conventional GPU cluster would deliver.
The implications of this are structural. For years, the AI training stack has been built around the assumption that high-end GPUs — with their combination of high-bandwidth memory, large VRAM, and CUDA-compatible compute — are a non-negotiable prerequisite for serious model development. VIDRAFT's disclosure suggests that prerequisite may now be a software problem, not a hardware one.
The GPU supply chain has been one of the defining constraints of the modern AI era. Demand for training-capable hardware has consistently outpaced supply, driving up costs and concentrating model development capacity among a relatively small number of well-resourced organizations. If VKORE can reliably abstract the training process across inference-optimized silicon, the cost of entry for model development could drop substantially.
This matters not only for startups and research labs working with limited infrastructure budgets, but also for organizations that already have NPU hardware deployed in edge or cloud environments for inference purposes. VIDRAFT's approach, in theory, would allow those static inference assets to be repurposed dynamically as training resources — converting underutilized hardware into productive capacity without new capital expenditure.
The editorial commentary from Mukta AI also points to a broader strategic implication: if a software layer like VKORE can bridge disparate hardware architectures, it erodes the proprietary advantage that dominant chip manufacturers have built around their ecosystems. Hardware vendors may face pressure to either embrace open abstraction layers or risk having their products commoditized by software platforms capable of migrating training workloads across any available silicon.
That said, Mukta AI's editorial is careful to flag what the original technical disclosure left unresolved. The 46-million-parameter model is a proof of concept, and the critical open question is whether this approach scales to billion-parameter models. Training at that scale demands sustained high-bandwidth memory access and intensive I/O throughput — workloads that inference-only NPUs were not designed to maintain over long periods. Thermal load and power delivery constraints on these chips also remain unaddressed in the current disclosure.
Q: What did VIDRAFT actually prove with its NPU training demonstration?
A: VIDRAFT showed that a 46-million-parameter language model can be pre-trained entirely on 32 inference-dedicated NPU chips, producing results that closely match GPU training outcomes with a 0.99 loss correlation.
Q: What is VKORE and why is it significant?
A: VKORE is VIDRAFT's software orchestration layer that enables backpropagation and optimizer state management on hardware not originally designed for training workloads, effectively decoupling model training from GPU-specific infrastructure.
Q: Does this mean NPUs can fully replace GPUs for AI training?
A: Not yet. The current demonstration covers a relatively small 46-million-parameter model, and scaling this approach to billion-parameter models raises unresolved challenges around memory bandwidth, I/O throughput, and thermal constraints on inference-optimized hardware.
Source: Mukta AI(글로벌) (2026-10-14) — original article