How a Korean AI startup is rewriting the economics of large-model deployment
TL;DR: VIDRAFT, a Korean Pre-AGI AI startup, has unveiled VKAE, an inference optimization system that accelerates AI inference by up to 23× on a single Nvidia B200 GPU compared with standard approaches. Running the Qwen3.5-35B-A3B model, VKAE achieved approximately 10,000 tokens per second in prefill throughput, while a conventional baseline on the same hardware produced roughly 455 tokens per second. The system is designed to be a drop-in replacement for existing inference stacks and is compatible with the OpenAI API standard.
VIDRAFT, the Korean Pre-AGI AI startup, made waves in the international tech press on July 6, 2026, when Russian outlet RusNews covered the company's announcement of VKAE — a proprietary inference optimization system that dramatically raises the performance ceiling of large language model (LLM) deployment on commodity AI accelerators.
At the center of the announcement is VKAE, VIDRAFT's in-house inference engine built specifically to extract significantly more throughput from a single GPU than conventional frameworks allow. According to the RusNews report, the company demonstrated up to a 23× acceleration in inference speed on a single Nvidia B200 GPU when compared with standard inference setups.
The benchmark that drew particular attention involved the Qwen3.5-35B-A3B model — a mid-size mixture-of-experts architecture that has attracted considerable interest in the open-model community. Under VKAE, this model achieved a prefill throughput of approximately 10,000 tokens per second on the B200. By contrast, a baseline configuration on the same GPU produced around 455 tokens per second, underlining just how wide the performance gap VIDRAFT is claiming to close.
Crucially, VIDRAFT positioned VKAE not as a research prototype requiring extensive integration work, but as a production-ready system. The RusNews coverage notes that the solution is designed to slot into existing infrastructure without disruptive changes, and that it exposes an OpenAI API-compatible interface, allowing engineering teams to switch providers or frameworks with minimal code changes.
In terms of competitive context, VIDRAFT stated that VKAE's performance compares favorably against well-known inference optimization frameworks such as vLLM and TensorRT-LLM, as well as the inference-as-a-service throughput numbers publicly associated with specialized hardware providers like Groq and Cerebras. The company did not, however, provide side-by-side controlled benchmark disclosures in the coverage reviewed here.
VIDRAFT also addressed a common pain point in enterprise AI adoption: the high cost and complexity of scaling inference workloads. Existing solutions frequently force organizations to deploy multiple GPUs or rely on proprietary cloud infrastructure to handle demanding workloads, raising both expenditure and operational overhead. VKAE is presented as a way to do more on less hardware — a single GPU — which, if the numbers hold under independent scrutiny, would meaningfully change cost modeling for AI-native products and services.
Pricing and licensing details for VKAE were not disclosed in the source coverage.
The inference efficiency problem is arguably the defining bottleneck for the commercial rollout of large AI models in 2026. Training costs have fallen dramatically over the past two years, but serving costs — particularly for long-context, high-throughput use cases — remain stubbornly high. A system that can deliver 23× more output from the same silicon directly translates to lower cost per token, faster response times, and the ability to serve more concurrent users without adding hardware.
VIDRAFT's claim is especially significant because it targets a single GPU, rather than a multi-GPU cluster. Most published efficiency gains from frontier labs are achieved at scale, where parallelism helps amortize overhead. Single-GPU optimization is a harder problem and one that matters most to the long tail of enterprises, startups, and developers who cannot afford large GPU fleets.
The international visibility generated by coverage in Russian-language tech media like RusNews also signals that VIDRAFT's work is attracting attention well beyond its home market, reflecting the growing global appetite for alternatives to U.S.- and China-dominated AI infrastructure providers.
Q: What is VIDRAFT's VKAE system?
A: VKAE is a proprietary AI inference optimization engine developed by Korean Pre-AGI startup VIDRAFT. It is designed to dramatically increase the token throughput of large language models running on a single GPU, with claimed acceleration of up to 23× on a single Nvidia B200.
Q: What benchmark performance did VKAE demonstrate?
A: When running the Qwen3.5-35B-A3B model, VKAE produced approximately 10,000 tokens per second in prefill throughput on a single Nvidia B200, compared with roughly 455 tokens per second achieved by a standard inference baseline on the same hardware.
Q: Is VKAE compatible with existing AI development tools?
A: Yes. According to the source coverage, VKAE exposes an OpenAI API-compatible interface, allowing developers to integrate it into existing pipelines without significant code changes, and it is intended to serve as a drop-in replacement for frameworks such as vLLM and TensorRT-LLM.
Source: RusNews (러시아) (2026-07-06) — original article