HyperAccel has nearly finished assembling investors for a 200 billion won Series B, with closing expected next month, ChosunBiz reported on September 10. The South Korean startup designs specialized chips for AI inference: the work of running an already-trained model to produce an answer. Data-center operators are a central target market. The report describes a financing approaching completion, not money already received.
Its Bertha 500 design addresses the repeated movement of information between a processor and memory. HyperAccel uses LPDDR, a type of lower-power memory, alongside processing cores tailored to language-model workloads. The company aims to keep useful work flowing through those cores rather than relying only on more expensive, higher-bandwidth memory. A cloud operator serving many users' AI requests is the intended buyer; the claimed cost and energy advantages are company benchmarks, not independently established results across every model.
The chip needs a software path into an actual service. HyperAccel provides tools for compiling model operations, assigning data across devices and running the resulting instructions on its hardware. It describes support for familiar developer tools including PyTorch and vLLM, and shows a Naver HyperCLOVA X chatbot running on its HX-F55X server system. That demonstration makes the workflow concrete: a model serves a user's request through the company's computing hardware. It does not by itself establish the size of a commercial contract.
The company is also developing a smaller Bertha 100 offering for local AI processing in connected devices and appliances. Its product page lists a fourth-quarter 2026 arrival, making that a roadmap item rather than a claim of broad shipment. In an earlier CTO interview carried on its site, HyperAccel described proof-of-concept work with Naver Cloud and LG Electronics. Those relationships illustrate two routes to market—data-center equipment and components for device makers—without establishing either partner's future purchase volume.
ChosunBiz puts the proposed post-money valuation at about 800 billion won. Its report identifies Korea Investment Partners and Company K Partners as leading financial investors and says Krafton is contributing 50 billion won as a strategic investor, within the overall 200 billion won round. The planned funding supports Bertha manufacturing and development of an edge-device accelerator. Captables is retaining the full round in its reported-financing category; the Krafton portion is not another round to add on top.
For this business, a successful chip design is only part of the product. Customers also need stable software, reliable hardware and a supply chain that can deliver repeatable production. HyperAccel's lower-cost-memory strategy gives it a specific proposition to test. The practical question is whether the complete system can make an AI service cheaper to operate after deployment and integration costs, not simply look efficient in a chip-level comparison.