South Korean neocloud operator BesselAI unveiled one of the most technically detailed AI infrastructure roadmaps in the Asia-Pacific region on Thursday, announcing plans to scale from its current 7,000 GPUs to 200,000 by 2030 — and presenting production metrics to back up the software platform that CEO Ahn Jae-man says is the company's real competitive edge over capital-heavy global rivals.

The headline result: running inference for Upstage's Solar Pro 4 large language model, BesselAI's Fluid Computing platform raised the model's KV-cache hit rate from 46 percent to 94 percent, cutting output token latency from 35 milliseconds to 14 milliseconds — a 60 percent reduction — while sustaining up to 155 billion tokens per day at peak load.

Those are not pilot-program numbers. Upstage's Solar Pro 4 — released in August 2026 and among Korea's top-performing open-weight large language models — is a production inference workload running at commercial scale. The metrics BesselAI shared at its "2026 Neocloud Strategy Briefing," held at Dreamplus Gangnam in Seoul, put a specific, auditable face on what the company means by describing itself as a software-first neocloud rather than a GPU-rental business.

What Fluid Computing Actually Does: Real-Time Cross-DC Workload Routing

The technical architecture at the center of BesselAI's strategy is its Fluid Computing platform, a software control plane that abstracts the physical boundaries between data center facilities. Standard GPU cloud arrangements give a customer a fixed block of GPUs at a fixed location. Fluid Computing presents a unified GPU inventory across multiple geographically distributed sites—spanning GPU models from NVIDIA A100 through B300, plus Google TPUs and domestic NPUs—as a single resource pool that an automated system routes workloads across in real time.

CTO Jeon Ji-hwan explained the control plane's mechanics at the briefing: the system continuously monitors GPU availability, GPU type, pricing, and network latency across every connected facility. When a customer submits a job — specifying GPU count, start time, preferred region, and price — the control plane matches it to whichever GPUs in the global pool best fit the task characteristics. Average GPU delivery from pool to running job takes fewer than five days.

For inference workloads, Fluid Computing includes a prefix caching layer that raises KV-cache hit rates—the mechanism behind the Solar Pro 4 performance improvement. In a transformer-based large language model, generating each token requires computing attention scores across all prior tokens in the context. A KV-cache stores the intermediate key-value computation results from attention layers so they can be reused when multiple requests share the same prefix (such as a system prompt or a document chunk). When the cache hit rate rises from 46 percent to 94 percent, approximately 48 percentage points more requests are served from stored computation rather than full recomputation — which is why output latency falls sharply alongside it.

Beyond inference optimization, Fluid Computing incorporates predictive fault detection that the company says catches more than 80 percent of GPU hardware failures before they disrupt workloads. Jeon cited a demonstration in which BesselAI ran a cluster of more than 1,000 NVIDIA B200 GPUs for three consecutive weeks without a service interruption. As fleet size scales toward 200,000 units, Jeon said, human-in-the-loop failure response becomes impossible: "If humans have to respond to failures one by one, scaling becomes impossible." The company is building out agent-based automation covering GPU onboarding, performance validation, fault detection and isolation, recovery, and job migration.

What Is BesselAI Competing Against

BesselAI (also marketed as VESSL AI) was founded in 2020 as an MLOps platform — software for managing machine learning model development and deployment workflows. It pivoted into GPU cloud infrastructure with the February 2026 launch of VESSL Cloud, its commercial GPU cloud service. The World Economic Forum recognized the company as a Technology Pioneer for 2026, and a research paper on its DLLM-JEPA training efficiency approach was accepted to an ICML 2026 workshop, reporting 33 percent lower computational cost and up to 18.7 percentage points of improvement on math reasoning benchmarks.

The neocloud market it is entering is growing at triple-digit rates globally, but is also dominated by companies with substantially larger capital positions. CoreWeave, the US-listed market leader, reported $5.1 billion 2025 revenue and carries an order backlog of around $104.2 billion as of mid-2026. Nebius, the Netherlands-based competitor, posted $582 million Q2 2026 revenue — up 454 percent year over year. Both companies are building positions primarily through data center scale and capital deployment.

Ahn's counter-argument is that BesselAI's five to six years of MLOps software development constitutes an operational knowledge base that pure-capital entrants cannot replicate quickly. The Fluid Computing control plane, predictive fault-detection logic, and KV-cache prefix optimization layer all derive from that operational history. At H100 On-Demand rates, BesselAI says it charges roughly half the hyperscaler price — and up to 80 percent less depending on GPU type and availability conditions — by monetizing idle capacity in partner data centers rather than reserving capacity at full cost.

GPU Roadmap: From 7,000 to 200,000

The company's expansion plan spans four years and three phases. BesselAI currently operates approximately 7,000 GPUs, with more than 5,000 in active customer use. It is targeting 10,000 by year-end 2026, 50,000 by the end of 2027, and 200,000 by 2030. Of the 2030 target, approximately 150,000 — or 75 percent — are targeted for direct ownership through BesselAI's own infrastructure build. The remaining 50,000 will be secured through partnerships with third-party data center operators and suppliers.

The pivot toward direct ownership reflects demand dynamics Ahn described at the briefing: the company's identified customer pipeline already exceeds 30,000 GPU-units — more than 60 percent of the 2027 target — and that pipeline is 70 to 80 percent international. BesselAI currently supplies GPU compute to domestic AI model developers, pharmaceutical companies like SK Biopharm (128 NVIDIA B200 GPUs for drug discovery and foundation model pretraining), and Silicon Valley AI startups. Exporting Korean-hosted GPU capacity to overseas customers is described as a core strategic pillar.

Revenue targets track the hardware buildout: 250 billion won (approximately $186 million USD) in contracted revenue by the end of 2026, rising to 700 billion won (approximately $521 million USD) in 2027 and over 2 trillion won (approximately $1.49 billion USD) by 2028.

Financing the Buildout: The GPU-Backed Structure and Its Constraints

Acquiring 200,000 GPUs requires commensurate capital, and BesselAI's plan follows the financing template that CoreWeave pioneered in the US: GPU-backed securitization. In this structure, GPUs and the long-term customer contracts assigned to them are placed into a special purpose vehicle, which issues debt collateralized by both the hardware and the contracted cash flows. The effective collateral, as independent analysts have noted, is less the GPUs themselves than the contracted revenue stream — because the secondary market for used GPUs remains thin.

BesselAI has 400 billion won (approximately $298 million USD) in GPU-backed financing currently in progress, and plans to pursue 2 trillion won (approximately $1.49 billion USD) in total financing once data center contracts signed in 2027 are in place. The company is in discussions with eight financial institutions across the United States, Singapore, and Korea. On the equity side, it is targeting a Series C close between the end of 2026 and early 2027, with a Series D preparation underway for the first half of next year.

The structural risk in this model is real. Industry analysis of the neocloud financing market shows that borrowers without investment-grade customer contracts — as a pre-Series-C startup, BesselAI is not yet investment-grade — face reported financing rates of approximately SOFR+800 to 900 basis points, compared with SOFR+225 for pools backed by hyperscaler take-or-pay contracts. The 2023 and 2024 vintage contracts across the global neocloud industry are also reaching their renewal windows in 2026 and 2027, creating an industry-wide stress test for whether AI demand supports the current debt load. BesselAI's financing plan is well-structured for its stage, but the cost of capital and renewal risk are live variables the 2030 roadmap will need to navigate.

Strategic Partnerships and Next Hardware Milestone

BesselAI's near-term GPU supply runs through SK Telecom, which is collaborating with the company to secure and operate more than 1,000 NVIDIA B200 GPUs. Additional capacity flows through NHN Cloud under a GPU supply agreement, and a joint AI factory project with Samsung SDS uses Samsung's SCP cloud platform.

On the domestic hardware side, BesselAI has completed proof-of-concept testing integrating Rebellions' domestic NPU chips into its cloud environment. Rebellions — South Korea's first AI chip unicorn, following its December 2024 merger with SK Telecom's SAPEON Korea subsidiary — closed a $400 million pre-IPO round in March 2026 at a $2.34 billion valuation. Its second-generation Rebel100 chip, built on Samsung's 4-nanometer process in a chiplet design, delivers 2 petaFLOPS of FP8 performance with 144 gigabytes of HBM3e memory at 4.8 terabytes per second of bandwidth. BesselAI is targeting a 2028 deployment of an AI factory that combines Rebellions NPUs with NVIDIA GPUs — a hybrid compute architecture that would reduce dependence on NVIDIA hardware alone.

The next major hardware milestone is NVIDIA's DGX Vera Rubin NVL72, which BesselAI plans to bring online in the first half of 2027. Each NVL72 rack pairs 72 Rubin GPUs with 36 Vera CPUs, delivering 3.6 exaFLOPS of NVFP4 inference performance and 20.7 terabytes of HBM4 memory, connected by NVLink 6 at 260 terabytes per second of aggregate fabric bandwidth. NVIDIA has positioned the system as targeting inference token costs roughly one-tenth those of the prior Blackwell NVL72 generation. First deliveries to major cloud providers began in the second half of 2026; BesselAI's planned H1 2027 deployment would make it one of the earliest production operators of Vera Rubin in the Asia-Pacific market.

South Korea's National Infrastructure Backdrop

BesselAI's expansion plan arrives inside a broader South Korean government commitment to AI infrastructure at national scale. South Korea is targeting 18.4 gigawatts nationwide of AI data center capacity to be built by 2035, at an estimated total cost of approximately 1,000 trillion won (roughly $650 billion USD), anchored by a Phase 1 commitment of approximately 550 trillion won (roughly $357 billion USD) through 2029. SK Group, led by SK Telecom, is the largest committed developer under that plan, targeting 15 GW of that capacity.

To support its portion of that buildout, BesselAI is establishing data center facilities across eight regional locations in Korea — spanning the greater Seoul metropolitan area and the provinces — and is targeting sites in prepared industrial zones where power connections and land are already secured. The company is targeting 100 megawatts of total data center capacity by 2027 and 500 MW by 2028. The government's plan includes discounted electricity rates and streamlined permitting for qualifying AI data center operators — policy conditions favorable to BesselAI's direct-ownership transition.

Can Software-First Beat Capital-First?

The central question BesselAI's briefing raises, without quite answering, is whether a software-differentiated neocloud can sustain its edge at 200,000-GPU scale, when capital-heavy competitors like CoreWeave will also add software optimization layers to their already-vast hardware bases.

BesselAI's argument rests on the specificity of its operational experience: six years of production-grade GPU fleet management, beginning in MLOps and extending through two years of cloud operations, gives it a head start in predictive fault logic, workload routing, and inference optimization that a new data center entrant cannot easily replicate. The Solar Pro 4 case study — 60 percent latency reduction, 94 percent KV-cache efficiency, 155 billion tokens per day — represents that argument made concrete.

Whether the argument holds as the buildout scales, the financing matures, and competition intensifies remains the open variable. For AI infrastructure buyers evaluating options in Asia-Pacific, BesselAI has now given them a specific, auditable benchmark to measure against.

Exchange rates as of October 15, 2026; all KRW/USD conversions are approximate.

Frequently Asked Questions

What is Fluid Computing, and how does it differ from standard GPU cloud rental?

Standard GPU cloud rental assigns a customer a fixed block of hardware at a fixed location. BesselAI's Fluid Computing is a software control plane that presents GPUs across multiple geographically distributed data centers as a single unified pool. It continuously monitors GPU availability, type, pricing, and network conditions in real time, then routes workloads automatically to the best-matching hardware anywhere in that pool. For inference specifically, it incorporates KV-cache prefix optimization—reusing stored computation when multiple requests share the same context—which drove Upstage Solar Pro 4's latency from 35 milliseconds to 14 milliseconds. The practical result for buyers is pricing at roughly half the hyperscaler rate, access to spare capacity across multiple partner facilities, and automated fault recovery that doesn't require human intervention at scale.

What is KV-cache optimization and why does a jump from 46% to 94% hit rate matter?

In transformer-based large language models, generating each token requires computing attention scores across every prior token in the context. The KV-cache stores the intermediate key-value outputs from attention layers so they can be reused rather than recomputed when requests share the same input prefix—such as a system prompt or a repeated document chunk. When a request hits the cache, the GPU skips recomputation entirely, reducing latency and increasing token throughput. Raising the hit rate from 46 percent to 94 percent means roughly 48 percentage points more requests are served from cache. At 155 billion tokens per day, the compounding effect on compute efficiency is substantial — which is why output latency dropped 60 percent alongside it.

What are the risks in BesselAI's GPU-backed financing plan?

BesselAI is following the securitization model that CoreWeave pioneered: GPUs and their associated customer contracts are used as collateral for debt. The structural risk is that the secondary market for used GPUs remains thin, so the real collateral is contracted cash flow—meaning the financing depends on long-term AI demand and customer contract renewals. As a pre-Series-C startup, BesselAI does not yet have investment-grade credit, so its financing costs are substantially higher than hyperscaler-backed neocloud pools. The 70-to-80-percent international customer pipeline has also not yet been formalized into long-term take-or-pay contracts — converting that pipeline is the key execution variable for the 2027 financing round.

How does BesselAI's 200,000-GPU target compare to global neocloud competitors?

BesselAI's 2030 target of 200,000 GPUs would place it well behind current leaders by order backlog: CoreWeave's order book stands at approximately $104.2 billion as of mid-2026, reflecting a GPU fleet of a different magnitude entirely. Nebius posted $582 million in second-quarter 2026 revenue alone. However, BesselAI's target is regionally significant: as South Korea executes its 18.4 GW national AI data center plan, a domestically rooted neocloud with 200,000 GPUs and an operational software stack would occupy a position in the Korean and broader Asia-Pacific market that no global hyperscaler or US-based neocloud currently addresses with equivalent local infrastructure and optimization depth.