Software-layer optimization, not new hardware, is the efficiency lever Korean AI startup VIDRAFT is betting on.
TL;DR: VIDRAFT, a Korean AI startup developing language models and AI infrastructure, has published benchmark results showing its VKAE inference acceleration system delivers up to 23× throughput gains on a single NVIDIA B200 GPU without any hardware changes. Testing on the Qwen3.5-35B-A3B model recorded aggregate throughput exceeding 10,000 tokens per second under high concurrent load, with no observed degradation in output quality. The system is designed to be reproducible via a ready-made container.
VIDRAFT, the Korean Pre-AGI AI startup, made international headlines in early July 2026 when Russian technology authority iXBT.com reported on the company's VKAE inference acceleration system — a software-layer solution that claims to multiply the effective throughput of existing GPU hardware by as much as 23 times in select scenarios, all without touching the underlying silicon.
VIDRAFT published benchmark data for VKAE, its proprietary inference acceleration system, tested on a single NVIDIA B200 GPU. According to results covered by iXBT.com, the system achieved up to 23× performance gains relative to baseline serving configurations for certain model architectures, while maintaining output quality and model accuracy throughout all measured runs.
One of the headline demonstrations involved the Qwen3.5-35B-A3B model. Under high concurrent load, VKAE delivered aggregate throughput of more than 10,000 tokens per second on a single B200 card. The team was careful to contextualize this figure: under more varied, real-world request patterns, the same model produced approximately 455 tokens per second — a distinction the developers openly flagged to avoid misleading comparisons.
The benchmarks measured single-stream throughput on a single NVIDIA B200 GPU, comparing a standard baseline mode against VKAE's optimized mode, using consistent measurement tooling across all runs. Crucially, no reduction in output quality or accuracy was recorded in any of the test configurations.
VIDRAFT also highlighted that the acceleration factor is not uniform across all models. Some architectures see dramatic multipliers; others improve by more modest amounts. The developers attribute this variance to differences in computational bottlenecks, memory organization, and the internal structure of individual model architectures — an honest acknowledgment that VKAE is not a one-size-fits-all solution.
A notable practical detail flagged by iXBT.com: VIDRAFT designed VKAE results to be reproducible via a pre-built container, lowering the barrier for independent verification and enterprise evaluation.
The economics of AI at scale have quietly shifted over the past few years. Training a large language model is a one-time, capital-intensive event. Inference — the continuous, real-time process of generating responses for users — is the ongoing operational cost that determines whether an AI service is commercially viable. As demand for LLM-powered applications grows, inference cost reduction has become one of the most strategically valuable problems in the industry.
VIDRAFT's approach treats the GPU as an underutilized asset and targets the software stack — low-level compute kernels and task scheduling mechanisms — to extract more throughput from hardware that enterprises and cloud providers already own. This positions VKAE less as a competitor to GPU manufacturers like NVIDIA and more as a force multiplier on top of them.
For organizations already running NVIDIA B200 infrastructure, the proposition is straightforward: potentially significant throughput gains and corresponding cost reductions without a new procurement cycle. The container-based reproducibility also speaks directly to enterprise buyers who require auditable, deployable solutions rather than lab-only benchmarks.
The iXBT.com coverage — reaching Russian-speaking technology professionals across one of the world's largest IT readership communities — also signals that VIDRAFT's ambitions are global, not limited to the Korean or East Asian markets where the company is headquartered.
Q: What is VIDRAFT's VKAE system?
A: VKAE is an inference acceleration system developed by Korean AI startup VIDRAFT that optimizes low-level software — including compute kernels and task scheduling — to increase the throughput of existing GPU hardware without requiring hardware upgrades.
Q: How much faster is VKAE compared to standard inference setups?
A: According to published benchmarks on a single NVIDIA B200 GPU, VKAE delivered up to 23× throughput gains over baseline serving configurations in certain scenarios, with no degradation in model output quality observed across any test run.
Q: Does VKAE work equally well on all AI models?
A: No. VIDRAFT's developers explicitly state that the degree of acceleration varies significantly depending on model architecture, owing to differences in computational bottlenecks, memory layout, and internal structure. Some models see large multipliers; others see more modest improvements.
Source: iXBT.com (러시아) (2026-07-06) — original article