Seoul-based SEMIFIVE announced Tuesday it has signed a contract worth approximately KRW 70.3 billion (USD 52 million) with a U.S.-based AI fabless company to develop a next-generation AI inference accelerator. The company calls it the first "Spec Hand-off" engagement in the North American market — and its largest single contract to date.

The scale is what stands out. Per SEMIFIVE's announcement, the single deal represents more than 40% of the company's total orders of KRW 168.4 billion in 2025, and roughly 60% of the KRW 118.9 billion in new orders it secured in the first half of 2026.

From spec to silicon#

The "Spec Hand-off" framing is a competitive pitch as much as a contract structure. The customer supplies only key performance requirements and specifications; SEMIFIVE then takes full responsibility for the entire development process — detailed design, software development, packaging, testing, and mass production. That's a deliberate contrast with conventional turnkey engagements, where the customer typically finishes the chip design before outsourcing the rest. SEMIFIVE is bidding to move upstream, into the design work itself.

Macro photograph of a processed semiconductor die showing individual chip blocks
Photo: ceramics.net.

The chip: LPDDR6, PCIe Gen5, Big Die#

For this project, SEMIFIVE says it will integrate advanced technologies to hit the high memory bandwidth and data processing performance that large-scale AI inference demands, while cutting power consumption and total cost of ownership. The accelerator adopts LPDDR6, next-generation low-power DRAM, to reduce memory-related power draw, and PCIe Gen5 to minimize bottlenecks in high-volume data transfers. Drawing on experience with multiple Big Die projects of up to 800 mm², SEMIFIVE also plans an architecture built for high compute throughput and operational stability.

"As the AI inference market continues to expand rapidly, demand is accelerating for custom silicon optimized for each customer's services and workloads," said Brandon Cho, CEO and co-founder of SEMIFIVE. "By successfully commercializing this next-generation inference accelerator and improving the economics of large-scale AI models, we aim to establish SEMIFIVE as an essential custom silicon development partner for the global hyperscale infrastructure market."

Close-up of semiconductor chip circuitry on a processed wafer
Photo: ceramics.net.

Why it matters#

The deal is another data point in the custom-ASIC surge. As SEMIFIVE's own release notes, specialized ASIC companies are increasingly engaged from the requirements stage — Google and Amazon have developed custom AI silicon with Broadcom and Marvell, and inference is where the economics now bite hardest, since every model run has a cost line. A U.S. fabless betting its next accelerator on a Korean design house — from spec, not just manufacturing — says something about where the shortage has moved: not foundry capacity, but design-throughput.

The timeline is aggressive but concrete: tape-out in the first half of 2027, then global mass production beginning in 2028, targeting global hyperscalers and cloud service providers. One contract this large doesn't make SEMIFIVE a peer of Broadcom — but it does put a new flag on the map for where custom inference silicon gets built.

The bottom line#

Inference has become the AI race's cost center, and cost centers get custom silicon. SEMIFIVE's $52 million hand-off — spec in, chip out, aimed at hyperscalers in 2028 — is the custom-ASIC playbook moving one level earlier in the stack.

Sources