Nvidia books Korea's entire $30m NPU export ledger in about three hours

Take the low end of "tens of billions per quarter" at $20bn, divide by a 91-day quarter, and Nvidia's data centre line moves roughly $220m a day, which makes the combined overseas order book of Rebellions, FuriosaAI and Mobilint, more than $30m as of mid-2026, about 3.3 hours of it ^[1][2][15]. Which is why the ask is interesting, and why on its own it is probably insufficient.

A reference deployment is normally bought, not granted, and the buyer is usually the vendor: you sell the first quotable installation at or below cost because the logo prices higher than the gross margin on it. What the Korean fabless firms are asking Seoul for is not the discount but the counterparty, an organisation whose workload is large enough to survive a slide ^[5]. SK Telecom is the existing template, running Rebellions' ATOM-Max in data centre infrastructure that handles up to 50 million AI service requests a day ^[6], which annualises to roughly 18.3bn calls ^[16], and that is a number an export salesperson can say out loud.

The awkward part is that the template already exists and the order book is still $30m ^[6][1]. The two named contracts, $500,000 for a UK wheelchair platform and $2.5m for water monitoring across Vietnam and Taiwan, come to $3m, about a tenth of the total ^[14], and nothing in the record says the SK Telecom deployment converted a single overseas buyer. Cryptobriefing's own account of why Nvidia holds the accelerator market points at CUDA and more than a decade of developer accumulation around it ^[12], which is a different constraint from an empty reference list: references answer a buyer worried about operational risk, benchmarks and porting tools answer a buyer worried about the cost of moving a software estate. The K-NPU project, started in late 2025 to develop domestic NPUs and verify their performance against incumbent hardware including Nvidia's GPUs ^[8], is the instrument aimed at the second objection, and a published benchmark travels in a way a Korean utility's procurement record does not.

Meanwhile one arm of the state is spending while another adds friction. High-performance AI chips joined the strategic goods list on 1 September 2026, so overseas shipments now need government approval ^[11], and the source names the tension inside Seoul's own apparatus plainly: money to help these firms compete globally, paperwork in front of the exports that would prove it ^[17]. The cost is asymmetric by design rather than by intent, since Nvidia absorbs licensing across dozens of jurisdictions and a few-hundred-person fabless startup with $30m of total overseas contracts does not have the same overhead to spread ^[13].

The allocation question sits inside the 8.4 GW of AI data centre capacity national plans target ^[9], funded from the hundreds of trillions of won earmarked for the wider semiconductor push ^[10]: every megawatt reserved as a domestic showroom is a megawatt not filled with imported accelerators, and the material does not say what share, if any, is set aside. On this evidence the credibility framing is the vendors' explanation of their own lost deals, offered by the parties asking for the remedy, and the one data point available cuts against it.