Can a fabless chipmaker from Seongnam sell sovereign AI to Japan, Britain and the United States without fighting for the training-compute market Nvidia already owns? Rebellions's $2.34 billion valuation rests on a wager: inference economics, not raw model capability, will decide who controls the build-out.
The company closed a $400 million pre-IPO round on 30 March, led by Mirae Asset Financial Group and the Korea National Growth Fund, the first investment under Seoul's K-Nvidia programme. Total funding reached $850 million, and roughly $650 million of that arrived inside six months. The cap table now reads Aramco, Arm, Samsung, SK Hynix and SK Telecom.
Six months on, the capital has become contracts. On 15 September, Japanese operator ai& agreed to deploy up to 100 of Rebellions's rack-scale RebelRack systems at a Tokyo data center, the chipmaker's first large-scale infrastructure deal in Japan. In late August, Rebellions joined London-based Callosum to push more than 100 NPU racks into a U.K. government-backed research cluster.
Both deals point the same way. The buyer here isn't a hyperscaler chasing training throughput. It's a state, or a state-backed operator, that wants compute it controls.
Inference is where the meter runs
Rebellions valuation after March round
Total raised to $850 million, three quarters of it in six months. · Rebellions, 2026
Training gets the headlines. Inference pays the bills.
Every query a model answers burns compute, so the cost of serving a token sets the floor under what an AI product can charge. That is the seam Rebellions has aimed at since it was founded in 2020: accelerator designs built for inference from the outset, rather than training silicon adapted after the fact.
The pitch to buyers is deliberately narrower than Nvidia's, and more specific. Performance per dollar per watt. The flagship Rebel100 pairs a chiplet architecture with 144GB of HBM3E memory on Samsung's 4-nanometre process.
Sovereign procurement gives challengers a distribution channel Nvidia cannot fully close.
The open question is repeat orders, not flagship pilots.
Sovereign buyers want a second supplier
The company's most useful asset may be political rather than technical. Governments building sovereign AI — domestic compute they control — carry a structural reason to avoid single-vendor dependence. Japan and Britain both fund capacity that explicitly favours heterogeneous hardware, mixing GPUs and NPUs so no one vendor holds the keys.
The economics of inference directly influence how widely AI can be deployed and how much it can be used. Lowering the unit cost of serving tokens gives providers room to create new pricing tiers.— Sunghyun Park, co-founder and CEO, Rebellions
The open-source posture reinforces the pitch. Rebellions's software stack runs on Kubernetes and integrates with vLLM, PyTorch and Hugging Face, the same tools engineers already use on GPUs. David Bennett, ai&'s chief executive, framed the appeal as removing the friction that usually kills a second chip vendor: integration work. Callosum is designing a reference architecture that drops Rebellions NPUs into its routing layer without bespoke optimisation.
A challenger's balance sheet, not a winner's
The bullish case rests on a short track record and a long public-sector purchase cycle.
Nvidia's CUDA software moat still sets the default for most engineering teams.
Government pilots are not revenue; procurement can take years to convert.
At $2.34 billion, the valuation already prices in an IPO that has no date.
As we wrote in September, inference is turning into a market where no single architecture wins outright. Gimlet Labs raised $300 million on the premise that AI inference will stop picking one chip. Our September piece on Euclyd made the same argument from the other direction: a 19-person team can now challenge Nvidia's grip by choosing the right slice of the stack. Rebellions is the scaled-up version of that thesis, with government procurement as the distribution channel.
The IPO is the test. Marshall Choy, the chief business officer leading global expansion, declined to give a date; the $400 million round was raised explicitly to prepare for one. If the Japan and U.K. deployments convert into repeat orders, Rebellions becomes the rare inference challenger with public-sector revenue behind it. If they stay flagship pilots, $2.34 billion will look like the ceiling.