A Korean AI startup is making the case that GPU scarcity is a software problem, not just a hardware one.
TL;DR: VIDRAFT, a Seoul-based deep-tech AI startup, has formally positioned its two proprietary inference engines — VKAE and VKUE — as a paired solution for AI data center operators. VKAE delivers kernel-level acceleration that can reach up to 23.4× the throughput of standard serving on the same GPU, while VKUE enables 35B-parameter-class frontier models to run on consumer GPUs, laptops, and even CPU-only environments. Together, the engines are designed to maximize both data center throughput and edge-site accessibility in a single deployment strategy.
Korean AI startup VIDRAFT announced on July 13, 2026, that it has completed a unified "set engine" strategy pairing its VKAE and VKUE technologies to address the operational efficiency challenges facing modern AI infrastructure data centers (AI IDCs). The Seoul-based company, led by CEO Minsik Kim, framed the announcement not as two separate product launches but as a deliberate, complementary architecture for organizations that need to squeeze more from limited GPU resources while simultaneously extending AI capabilities to sites where high-end hardware simply doesn't exist.
At the heart of the announcement is the VIDRAFT Kernel Ubiquitous Engine, or VKUE — a lightweight runtime designed to make 35B-class frontier models deployable far beyond the typical high-end server environment. VIDRAFT's own model, Ourbox-35B-JGOS, has 34.7 billion total parameters but is engineered so that only roughly 3 billion parameters are active during any given token generation step. This sparse-activation design is what allows the model to run across a wide spectrum of hardware.
VIDRAFT has demonstrated the model generating tokens at more than 18,000 per second in a high-end data center setting, dropping to 126 tokens per second on a single mid-range data center GPU card, 20 tokens per second on an 8 GB gaming laptop, and approximately 17 tokens per second in a CPU-only environment with no dedicated GPU at all. On the GPQA Diamond scientific reasoning benchmark, Ourbox-35B-JGOS scored 86.4% (maj@8), signaling that the efficiency gains do not come at the cost of reasoning quality.
Complementing VKUE is VKAE, VIDRAFT's kernel-level inference acceleration engine. Rather than focusing on portability, VKAE is built to maximize throughput on the GPUs already inside a data center. The company reports that VKAE has achieved up to 23.4× the inference throughput of standard serving frameworks on identical hardware. On the Qwen3.5-35B-A3B model, VKAE has demonstrated more than 601 tokens per second in single-stream mode and over 10,000 tokens per second under multi-request load. VKAE is delivered with an OpenAI-compatible API and a publicly verifiable leaderboard, allowing independent reproduction of results.
When the two engines are combined for a 35B-class deployment, VIDRAFT says peak performance in a high-concurrency scenario — 256 simultaneous users on a single high-end GPU — sustained more than 18,000 tokens per second continuously.
CEO Minsik Kim stated: "The competitiveness of an AI IDC depends not only on how many GPUs you own, but on how well you run the GPUs you have and how far you can extend AI to places that have no GPUs at all. VKAE handles speed; VKUE handles accessibility. Through both engines, VIDRAFT aims to grow into an infrastructure company that makes large AI cheaper, faster, and closer to where it's needed."
The AI data center market faces a structural tension: GPU supply is costly and constrained, yet demand for inference capacity keeps rising. At the same time, regulated sectors such as healthcare, finance, public administration, national defense, and manufacturing often operate in closed or air-gapped networks where sending data to a central cloud is not an option. VIDRAFT's paired engine strategy directly addresses both pressures simultaneously — improving the economics of centralized GPU farms while enabling on-premise, low-hardware deployments at the network edge.
VIDRAFT is a resident company at the Seoul AI Hub and has conducted its research and development through government GPU support programs and NIPA AI computing utilization projects. The company previously disclosed that its Darwin-398B-JGOS model achieved 90.9% on GPQA Diamond, establishing a track record in frontier scientific reasoning before this infrastructure-focused announcement.
The VKAE×VKUE strategy signals a broader ambition: rather than competing purely on model benchmarks, VIDRAFT is positioning itself as an inference infrastructure layer — a company whose value is measured in how efficiently its engines operate other people's AI workloads.
Q: What does VIDRAFT's VKUE engine actually do?
A: VKUE is a lightweight runtime engine that enables large frontier-class AI models — specifically at the 35B-parameter scale — to run on consumer GPUs, laptops, and CPU-only machines, not just expensive server hardware.
Q: How does VKAE differ from VKUE in VIDRAFT's strategy?
A: While VKUE focuses on broadening where AI models can run, VKAE is a kernel-level acceleration engine focused on maximizing inference throughput within a data center, achieving up to 23.4× higher throughput than standard serving on the same GPU.
Q: Who is the target customer for VIDRAFT's VKAE×VKUE set engine approach?
A: The primary audience is AI data center operators and organizations in regulated industries — such as healthcare, finance, defense, and public services — that need both high-throughput central AI processing and on-premise deployment in environments without cloud connectivity.
Source: 전자신문 (2026-07-13) — original article