Reclaim the Performance Virtualization Has Been Stealing.
The moment you move to public cloud, up to 20% of your compute goes to the hypervisor. Velox eliminates this layer entirely. Run Kubernetes directly on bare metal — delivering container orchestration flexibility and raw physical performance together, provisioned in minutes.
General-purpose cloud was not designed for high-performance workloads.
General-purpose public cloud often makes GPU workloads expensive to run, while virtualization overhead and shared infrastructure can limit the performance enterprises expect. Velox is dedicated bare metal infrastructure built from the ground up for large-scale AI workloads and enterprise databases. Velox runs workloads directly on bare-metal xPU infrastructure and combines high-speed parallel storage with InfiniBand-based ultra-low-latency networking. Deliver higher performance for AI and high-performance workloads with a more cost-efficient cloud model.
Nothing between the hardware and your application.
Most general-purpose cloud services run on a VM architecture where multiple users share a single physical server. The hypervisor layer consumes up to 20% of available resources on its own and introduces significant network and storage I/O latency. Velox eliminates this layer entirely.
Velox Bare Metal
Standard Cloud (VM)
No Hypervisor Layer
CPU, memory, and NIC connect directly to your application. Nothing in between. Because the hypervisor overhead doesn’t structurally exist, performance loss cannot occur.
Single-Tenant Hardware
You occupy the entire physical server exclusively. There is no performance jitter from shared resources, and no other tenant’s traffic spike can affect your service.
API-Driven Provisioning
Deploy in minutes via the Thaki Cloud console and API. The days-long wait of traditional IDC provisioning is over. Enjoy cloud agility and physical server performance at the same time.
Hardware-Level Isolation
Physical isolation, not logical separation. VM Escape vulnerabilities don’t exist without a hypervisor. Structural security assurance for compliance-critical industries including finance, healthcare, and defense.
Kubernetes on Bare Metal
Kubernetes runs directly on physical servers — no hypervisor required. You get the full flexibility of container orchestration alongside bare metal performance.
Bare metal infrastructure built on four pillars.
COMPUTING
100% hardware performance, no virtualization overhead
Direct bare metal xPU access eliminates virtualization overhead entirely. Capture 100% of hardware performance for both foundation model training and ultra-low-latency inference. A single node with NVIDIA HGX B200 × 8 delivers 18 PFLOPS FP32 compute without compromise.
SECURITY
Physical isolation as standard
With a dedicated single-tenant server, VM Escape vulnerabilities structurally do not exist. Hardware-level complete isolation is provided as a baseline. CSAP low-to-mid grade requirements can be met.
STORAGE
Your GPU should never wait for data
High-speed parallel storage optimized for large-scale AI datasets. A 400Gbps data fabric based on NFS over RDMA eliminates data loading bottlenecks during training at the source. 10PB cluster-level storage is 2.5× the industry-recommended capacity.
AGILITY
Deploy physical servers like cloud
Days-long IDC provisioning is a thing of the past. Create and deploy bare metal servers to exact spec in minutes via the Thaki Cloud console and API. Run Kubernetes directly on physical servers for full container orchestration flexibility.
Reclaim 100% of your hardware performance — zero virtualization overhead
Contact UsThe numbers speak for themselves.
Velox standard node configuration. Custom configurations optimized for your specific workloads are available upon consultation.
Velox Standard Node
| Item | Specification |
|---|---|
| Form | 10U Rackmount |
| CPU | Intel Xeon 6900 Series (6960P) / 2-socket / 72-core |
| Memory | DDR5-6400 ECC RDIMM / 2.3TB Total (SK Hynix) |
| Storage | Samsung NVMe PCIe Gen4 / 30TB RAID 0 |
| GPU | NVIDIA HGX B200 × 8 / HBM3e 180GB (SK Hynix) |
| GPU Fabric | InfiniBand NDR400G / ConnectX-7 |
| DPU | BlueField-3 BF3220 / 200GbE Dual-Port |
| Power | 5,250W × 6 Redundant (3+3) |
| Certification | NVIDIA Qualified |
Datacenter — Gasan AI DC
| Item | Detail |
|---|---|
| Location | Gasan-dong, Geumcheon-gu, Seoul |
| Tier | Tier-III |
| Floor Area | 69,300㎡ (21,016 pyeong) |
| Power Capacity | 80MW (IT: 46MW) |
| Rack Power | Up to 44kW per rack (air-cooled) |
| Seismic Design | Richter magnitude 7 (intensity 9) |
| Opened | July 2021 |
Network Fabric
| Item | Detail |
|---|---|
| Compute | InfiniBand NDR / 400GbE per GPU |
| vs. RoCE | 5× lower latency |
| Storage | NFS over RDMA / 400Gbps |
| Management | NVIDIA UFM automation |
Performance Metrics
| Metric | Value | Note |
|---|---|---|
| FP32 Compute | 18 PFLOPS / node | 2.25× H100 |
| GPU Memory | 1.4TB / node | 2.25× H100 |
| GPU-to-GPU Bandwidth | 1.8TB/s | 2× H100 |
| Cluster Storage | 10PB | 2.5× industry recommendation |
B200 vs H100 Performance Comparison
| Metric | H100 | B200 (Velox) |
|---|---|---|
| FP32 Compute | 8 PFLOPS | 18 PFLOPS |
| GPU Memory | 640GB | 1.4TB |
| GPU-to-GPU Bandwidth | 900GB/s | 1.8TB/s |
| Training Performance | Baseline | Up to 3× improvement |
| Inference Performance | Baseline | Up to 15× improvement |
Source: NVIDIA, SMCI
Built for the workloads that can’t afford limits.
Velox provides the optimal answer for mission-critical enterprise workloads that general-purpose VM-based cloud simply cannot handle.
Foundation Model Training
Large-scale AI research clusters that demand 100% GPU and HBM utilization without loss. Velox on B200 delivers up to 3× training performance and 15× inference performance over the previous generation.
High-Performance Databases
Large-scale RDBMS and real-time NoSQL workloads where I/O bottlenecks are unacceptable. Removing the virtualization layer eliminates storage I/O latency at the source.
High-Concurrency Real-Time Services
Mission-critical infrastructure that must deliver a seamless experience to millions of concurrent users. Optimal for environments where unpredictable performance variance cannot be tolerated.
Finance, Healthcare & Compliance
Hardware-level physical isolation meets regulatory requirements. Infrastructure proven in rigorous compliance environments governed by CSAP and NIS certification standards.
HPC & Parallel Scientific Computing
Ultra-low-latency GPU fabric on InfiniBand NDR400G — 5× lower latency than Ethernet-based RoCE. Maximize parallel computing efficiency at scale.
Sovereign AI Infrastructure
For public sector and large enterprise environments where data sovereignty is non-negotiable. Full physical isolation within domestic datacenters with an independent operating framework.
The case for adoption, in numbers.
By internalizing bare metal infrastructure, you maintain compliance while fundamentally reducing operational costs. The only fully dedicated infrastructure for AI training, inference, high-performance databases, and real-time services.
FP32 compute per node — 2.25× H100
Hypervisor overhead — zero performance loss
Dedicated xPU occupancy — zero noisy neighbors
Provisioning time — days faster than IDC