Why APAC operators need APAC GPU capacity
Most published GPU pricing benchmarks come from US-based hyperscalers, but for operators serving end users in Mandarin, Japanese, Korean, Vietnamese, or Thai markets, US data center rental means 150–250ms RTT inference latency — unusable for real-time chat, audio, or interactive applications. APAC GPU capacity has historically been constrained, list prices run 2–3x higher than US, and onboarding to AWS Trainium or Google TPU often requires multi-month enterprise contracts. We aggregate H100 and H200 capacity across Tokyo (NTT, Equinix), Hong Kong (Sunevision, MEGA), and Singapore (Equinix SG3, Digital Realty) and resell hourly with no minimum commitment.
Inference deployment patterns
For Mandarin-first LLM serving (DeepSeek, Qwen, Yi, GLM-4) we deploy vLLM or TGI on H100 80GB SXM5 single-node 8-GPU configurations, fronted by token-aware load balancing. Throughput on Qwen2.5-72B at FP16: ~840 tokens/sec aggregate per H100 8-GPU node, p50 first-token-latency 280ms, p99 750ms. For longer-context workloads (>32K tokens) we recommend H200 with HBM3e expanded memory — first-token latency stays under 400ms even at 128K context. For multimodal (vision-language) workloads we run InternVL2-78B or Qwen2-VL-72B on the same H100 8x stack.
Training and reservation pricing
Effective hourly pricing: H100 8x SXM5 at $32–38/hour (Asia spot when available, on-demand otherwise); H200 8x SXM5 at $42–48/hour; B200 (early access, Tokyo only) at $58–65/hour. InfiniBand HDR 200Gbps multi-node clusters (16–64 GPUs) available in Tokyo and Singapore. Reserved capacity available: 1-month commitment drops H100 to $28–32/hour, 3-month commitment drops to $24–28/hour. Reservation fully refundable if we miss our 99.5% availability SLA.
Workload acceptance
We host LLM inference, fine-tuning, and small-to-medium training (up to ~64 GPU multi-node) for legitimate AI workloads: chat applications, content generation, code assistants, voice/video synthesis, recommendation systems, multimodal search, scientific research. We do not host workloads designed to circumvent upstream API safety filters that we don't operate, training runs with known unlawful training data exposure, or services intended to generate non-consensual depictions of identifiable real persons. Standard engagement begins with MNDA. Settlement options include wire, regional rails, and stablecoin where commercially appropriate.
Anonymized case outline — A Singapore-based AI startup serving a Mandarin-first conversational platform fine-tuned Qwen2.5-72B on a 4-node H100 cluster (32 GPUs total, InfiniBand HDR) over 6 days. Total billable hours: 4,608 GPU-hours at $34/GPU-hour = $156,672. Inference deployment landed on a 2-node H200 cluster in Hong Kong serving 4,200 concurrent users at p99 first-token-latency 380ms. Customer migrated from a US-based GPU rental that was charging 60% more and adding 220ms inference latency to Asian end users.
FAQ
Hourly billing — really hourly?
Real hourly. Spin up at 14:23, shut down at 14:51, you pay for 0.47 hours. No daily minimum. Provisioning takes 4–12 minutes typically.
InfiniBand availability?
Tokyo and Singapore: HDR 200Gbps within rack, NDR 400Gbps in newer racks. Hong Kong: HDR 200Gbps. Multi-node configurations require explicit reservation 24h in advance.
Storage for training data?
Local NVMe 30–60 TB per node, plus optional S3-compatible object storage in-region (Cloudflare R2 for Singapore/Tokyo, Tencent COS for HK, Alibaba OSS for Singapore).
Settlement options for AI workloads?
Wire transfer, SWIFT, USDT TRC20/ERC20, USDC ERC20. Monthly invoicing with prorated burn-down for reserved capacity.
Talk to our infrastructure team
MNDA standard. Multi-channel: email, scheduled call, or Telegram. We respond within 4 business hours.