HyperAccel, a Seoul-based artificial intelligence chip designer, is raising 200 billion won ($150 million) in a Series B round, with closing scheduled for next month after payment and other procedures, vaulting its valuation to 800 billion won and positioning the young company to challenge Nvidia's grip on AI hardware. Korea Investment Partners and Company K Partners led the financing, with gaming publisher Krafton supplying a quarter of the total, according to ChosunBiz.
The fundraise comes as HyperAccel began manufacturing Bertha, its data-center inference accelerator, on Samsung's 4-nanometer process on September 8. That production start, announced by manufacturing partner SEMIFIVE, arrived years ahead of the company's original 2027 roadmap. The acceleration suggests HyperAccel has found early traction in a market where Nvidia's GPUs still command overwhelming share, though specialized inference chips are gaining ground as generative AI models move from research labs into production environments.
A Bet on Memory Economics
Bertha's architecture hinges on a contrarian wager: that LPDDR5X memory, not the high-bandwidth memory stacks favored by Nvidia and most rivals, can deliver competitive performance at a fraction of the cost. HyperAccel claims twice the throughput of Nvidia's H100 accelerator, with 19 times better cost efficiency and 12-fold improvements in power consumption, according to specs on the company's website, though these vendor-provided claims await independent validation.
"We are trying to be a more affordable provider... that's why we chose LPDDR, which is only one-tenth of HBM's bandwidth," chief strategy officer Yongwoong Jung told EE Times earlier this year. The company compensates for lower memory bandwidth by eliminating much of the traditional memory hierarchy, achieving roughly 90 percent utilization compared to around 30 percent on GPUs running large-language-model inference, Jung said.
Whether those company-provided claims hold up under real-world workloads remains an open question. SEMIFIVE said the chip occupies more than 500 square millimeters of die area on Samsung's SF4 process, a substantial footprint that could complicate manufacturing yields. DIGITIMES confirmed the production start in a brief the following day, noting Bertha targets LLM inference in cloud data centers.
Krafton's Strategic Calculus
Krafton's 50 billion won investment reflects the gaming giant's broader push into AI-driven services, ChosunBiz reported. The publisher behind PlayerUnknown's Battlegrounds has been vocal about an "AI-first" transformation, and lower inference costs would directly benefit game mechanics that rely on real-time model responses.
Other participants in the round included KB Investment at roughly 14.3 billion won, Korea Investment Partners at about 12 billion won, and LB Investment, AJU IB Investment, and iMarketKorea each contributing around 10 billion won, according to The Bell. Mirae Asset Venture Investment, BonAngels, NH Venture Investment, SBVA, KDB Capital, We Ventures, and LG Electronics also joined.
The investor lineup tilts heavily toward Korean institutional capital, a pattern common among the country's chip startups. HyperAccel's post-money valuation of approximately 800 billion won represents a fourfold increase from its December 2024 Series A, when it raised approximately 55 billion won ($40 million) at a 200 billion won valuation.
From Microsoft Research to Startup Founder
CEO Joo-Young Kim launched HyperAccel in January 2023 after leaving an associate professor role at KAIST, the Korea JoongAng Daily reported. Kim spent nearly a decade at Microsoft between 2010 and 2019, working on chip projects in Microsoft Research and Azure, according to the same interview. CTO Lee Jin-won came from Samsung System LSI, where he focused on application-processor design, ChosunBiz reported.
The startup employed around 80 people as of March this year, including 65 chip engineers, Kim told the Korea JoongAng Daily. That headcount likely grew with the new funding, though the company has not disclosed recent figures.
Cloud Trials and Edge Ambitions
HyperAccel is running a proof-of-concept deployment with NAVER Cloud, Korea's dominant cloud provider, CTO Lee told ChosunBiz. Co-founder Seungjae Moon said NAVER's interest originated with Orion, an earlier prototype that preceded Bertha, according to EE Times.
The company also signed a joint development agreement with LG Electronics to create Bertha-Edge, an on-device accelerator for smartphones and consumer electronics. HyperAccel expects to deliver engineering samples of the edge chip by late next year, ChosunBiz reported, though timelines for such early-stage silicon often slip.
HyperAccel demonstrated a live inference system at the Hot Chips conference in August, engaging with what ChosunBiz described as "global big tech and semiconductor corporations." The company declined to name specific prospects.
Navigating a Crowded Field
The startup enters a competitive landscape already occupied by Korean peers Rebellions and FuriosaAI, both of which have raised larger sums and secured partnerships with major Korean conglomerates. ChosunBiz drew comparisons to Rebellions, which has worked with Samsung and SK Telecom and raised more substantial rounds.
Globally, specialized inference startups like Groq and Cerebras are targeting the same market, betting that purpose-built chips can undercut Nvidia's H100 and upcoming Blackwell GPUs, which handle both training and inference. Analysts expect the inference segment to outpace training hardware as AI models mature and deployment scales, though Nvidia still captures the vast majority of data-center AI spending.
HyperAccel plans to use the Series B proceeds for production scaling, initial contract manufacturing of Bertha-Edge, and international expansion, ChosunBiz reported September 11. Seoul Economic Daily wrote in July that the company has joined South Korea's IPO pipeline, with a potential enterprise value exceeding 1 trillion won should it pursue a listing following the fundraise.
The path from prototype to production has compressed in recent months, though mass-market adoption of Bertha depends on factors beyond technical specs. Cloud providers and enterprise customers have shown reluctance to bet on unproven architectures when Nvidia's ecosystem offers mature tooling and broad software support. HyperAccel's LPDDR gambit may prove prescient if cost pressures mount, or it may remain a niche solution in a market still defined by memory bandwidth and raw compute power.
