Cross-border AI Compute Circuits and Multi-cloud GPU Scheduling

Challenge

The Shanghai R&D team trained on GPU clusters in Singapore and Tokyo, but moving 40 TB daily datasets over the public internet took 20+ hours and kept failing; China data-export rules required an auditable dedicated channel; GPU utilisation across the three sites sat at 35% with queues in one place and idle cards in another.

GSS scope of responsibility

GSS owned design and delivery of the cross-border compute circuits (domestic segments and GSS PoPs in-house, overseas last mile via partner carriers), transfer acceleration and the data-transfer audit gateway, infrastructure for the multi-cloud GPU scheduler and 24/7 operations. Training jobs, datasets and transfer approvals were owned by the customer R&D and compliance teams.

Solution

GSS delivered 2×1G Shanghai ↔ Singapore and 1G Shanghai ↔ Tokyo IEPLs (contracted and invoiced in China, overseas legs via upstream partners), built a cross-border transfer-acceleration and data-export audit gateway, and deployed a multi-cloud GPU scheduling platform unifying the three sites' queues — all operated by the GSS NOC.

Background

Headquartered in Singapore with a 300-person algorithm and data team in Shanghai, the company runs more than 1,200 accelerators across its own Singapore IDC, public cloud in Tokyo and public cloud in Shanghai. Each day the Shanghai team pushed roughly 40 TB of labelled datasets and intermediate artefacts to the overseas clusters over an internet VPN — more than 20 hours per run, with frequent interruptions and retries that directly slowed model iteration.

On compliance, the company needed to satisfy China’s data-export requirements: training data leaving the country had to use an auditable, traceable channel with a retained transfer register.

On resources, the three GPU pools were queued by three separate teams; overall utilisation was 35%, and “queue here, idle there” happened daily.

Solution

Cross-border circuits and export audit (traditional network business)

• IEPL circuits: 2×1G Shanghai ↔ Singapore (diverse routes) and 1G Shanghai ↔ Tokyo, contracted and VAT-invoiced by GSS in China with overseas legs delivered by upstream partners; Shanghai access via the GSS Shanghai POP, landing in the company’s Singapore IDC and Tokyo cloud on-ramp.
• Transfer acceleration gateways at both ends (multi-stream + forward error correction) lifting effective bulk-sync throughput from ~500 Mbps on the VPN to near line rate.
• Data-export audit gateway: every outbound transfer passes through the gateway, which records dataset name, hash, size, requester and approval ID and produces a monthly export register for legal and regulators; sensitive fields are de-identified in China before leaving.
• Three-site mesh: Singapore ↔ Tokyo interconnected over the upstream backbone, forming a triangle that re-routes automatically on any link failure.

Multi-cloud GPU scheduling (AI business)

• Unified compute queue: a Kubernetes-based multi-cluster scheduler manages all three GPU pools; researchers declare only compute spec and data location, and the scheduler picks a cluster by queue depth, data locality and circuit load.
• Data locality and prefetch: coordinated with the transfer gateway, datasets are prefetched to the target cluster so training starts with data already in place.
• Single entry for model APIs: commercial LLM API calls from R&D teams, overseas and domestic, pass through the GSS Unified AI API Gateway — overseas models via the Singapore / Tokyo nodes, four-tier quotas per team, end-to-end audit, no stray keys or duplicate bills.
• Cost and quota: per-team quotas and priorities, automatic demotion to pre-emptible jobs during idle periods, and monthly GPU cost / utilisation reports across the three sites.
• Operations: circuits, gateways and scheduler monitored 7×24 by the GSS NOC with bilingual support for the Singapore and Shanghai teams.

Results

Metric	Before	After
40 TB dataset sync	> 20 h (internet VPN)	~2.5 h
Shanghai ↔ Singapore latency	110–180 ms, high jitter	68 ms, stable
Transfer interruptions	3–5 per day	Near zero
Overall GPU utilisation	~35%	~78%
Model iteration cycle	~9 days	~4 days
Export audit trail	None	100% logged, monthly register

Delivered in March 2025; GSS continues to operate the circuits, gateways and scheduler, with staged upgrades to 10G planned as the business grows.