The AI data centre is the product: inside MangoBoost’s full-stack bet

For most of the past decade, the answer to slow AI was simple enough that nobody thought to question it. If performance lagged, you bought more GPUs. Jangwoo Kim spent nine years at a Seoul National University systems laboratory watching data centres behave in ways that made the answer look incomplete, and when he left to found MangoBoost in 2022 with Dongup Kwon and Eriko Nurvitadhi, he took the contrary reading with him.

The research had shown something the market was not pricing. Throughput in a live AI data centre was being lost between the components rather than inside them, in the movement of data across networks, storage and memory while expensive accelerators sat idle waiting for it.

“The prevailing assumption was that better AI performance simply meant more GPUs,” Kim said. “When we analysed real-world AI data centres, we found that performance and cost hinge on how seamlessly the network, storage, and AI system software work together.”

A university thesis only becomes a company when it survives contact with customers

Founding a systems company on that premise meant refusing the easier commercial path. Selling a faster component into an established category would have been a clearer proposition for early customers, and MangoBoost instead set out to optimise the data centre as a whole, unifying inference software, high-speed Ethernet fabric, storage and compute server platforms into one architecture.

The company built out of Bellevue and Seoul, drawing heavily on the laboratory’s research lineage. More than 30 patents now protect the core data processing unit technology, and the early work was hardened through joint validation with global technology partners and deployment into live customer environments rather than benchmarks alone.

The agentic turn in AI has since made the founding argument considerably easier to explain. Systems that plan, reason and coordinate across other systems place demands on infrastructure that GPU-centric design never anticipated. “Memory bottlenecks, network latency, low GPU utilisation, and high build and operating costs make it hard to support large-scale inference, long-context processing, and multi-agent workloads efficiently,” Kim said.

Benchmark records turned an architectural claim into commercial evidence

The proof arrived in April 2025, in a form the industry could not easily argue with. MangoBoost’s Mango LLMBoost AI Enterprise software recorded 103,182 tokens per second in the offline scenario and 93,039 tokens per second in the server scenario running Llama2-70B across 32 AMD Instinct MI300X GPUs, exceeding the previous best of 82,749 tokens per second recorded on NVIDIA H100 GPUs. It was the first multi-node MLPerf inference result on MI300X silicon, and it carried a 24% performance advantage over the strongest published result from Juniper Networks running 32 H100 GPUs.

The economics travelled further than the throughput number. With MI300X GPUs priced between $15,000 and $17,000 against $32,000 to $40,000 for H100 units, the configuration produced up to 62% cost savings and roughly 2.8 times more inference throughput per $1,000 spent. For enterprises committing to multi-year infrastructure, that ratio does more persuasive work than any single performance claim.

Kim locates the advantage in the co-design discipline the founding team carried out of the laboratory. “Orchestrating heterogeneous components into a single architecture, tuning the system as a whole to the workload, is what truly sets us apart,” he said. The in-house DPU sits at the centre of it, offloading network and storage data processing while accelerating GPU-to-GPU communication so that accelerators do computation and little else.

Selling components gave way to selling the whole rack

The commercial model has since expanded well past silicon. MangoBoost now offers the BoostX DPU, LLMBoost inference software, the Alphonso AI server and the Kesar storage server as a single tailored package, with the servers built on AMD CPUs and GPUs. The reasoning is that customers increasingly want infrastructure delivered whole rather than assembled in-house from parts sourced across a dozen vendors.

That shift is where the company’s revenue ambitions now sit. Kim has said MangoBoost aims to generate 10 billion won from the rack business this year, with revenue potentially rising tenfold in the year following, and the company is raising a new funding round against that trajectory. It has previously raised $65.5 million, including a $55 million Series A in October 2023 led by IMM Investment and Shinhan Venture Investment.

“Outside of Nvidia, MangoBoost is the only company capable of delivering everything from chips to software as a unified offering,” Kim said. Vertical control of the networking layer is the least common piece of that claim. “Designing and controlling the networking layer ourselves is a capability only a handful of companies can claim,” he said.

Momentum has accumulated on several fronts. The company demonstrated its 400G BoostX DPU, GPUBoost RDMA NIC and LLMBoost software at the OCP Global Summit and SC25, took the Next Unicorn Award at the LG Superstart Expo alongside a memorandum of understanding with LG, and co-developed an AI Cluster Benchmark Suite with SK hynix that identifies bottlenecks across compute, memory, storage and networking on live infrastructure without manual tuning. In March, it moved into data centre infrastructure directly with a Gangnam facility build-out.

Openness is the argument that carries the company into the Gulf

The next test is regional. MangoBoost was selected in June into Cohort II of the Presight AI Accelerator, the Abu Dhabi programme run by the G42 subsidiary, chosen among 12 companies from 376 applications across 62 countries by a jury drawn from Presight, the UAE Ministry of Industry and Advanced Technology, and the Mohamed bin Zayed University of Artificial Intelligence. Participants gain high-performance computing access, mentorship and commercial introductions across Presight and G42 to initiate proofs of concept with enterprise and government clients.

What makes the fit coherent is the open architecture rather than the accelerator badge. Because the stack is standards-based, customers retain freedom over which accelerators and models they run, which speaks directly to a market where sovereign capability has become the organising principle of AI investment. “Because the infrastructure is open and standards-based, free to run whichever accelerators and models customers choose, MangoBoost makes Sovereign AI possible,” Kim said.

Whether a company of MangoBoost’s size can sustain vertical integration across chips, servers, storage and software remains the open question, in a category that has consolidated around hyperscaler-backed players with far deeper capital. What the company has established is a measurable performance and cost position achieved on open hardware, without the pricing power of an incumbent behind it. “Our mission is to provide an open, optimised infrastructure for the next generation of AI,” said Kwon. “We focus on delivering solutions at the infrastructure level, ensuring you have the foundation you need to succeed.”