APAC AI Infrastructure Weekly

Reporting period: 26 October–1 November 2026. Report date: 2 November 2026. Compared with the 26 October baseline. “Verified” means supported by a public primary source; it does not establish physical stock, available-to-promise quantity, vacant data-centre capacity or a complete commercial offer unless expressly stated.

Most actionable changes

1. Singtel RE:AI announced the first APAC deployment of NVIDIA DGX GB300, but the initial GPU-as-a-Service capacity was fully committed before deployment. This is evidence of an operator-backed GB300 deployment and demand in Singapore, not available capacity. Singtel did not disclose GPU count, service date, network topology, contract term or price. Impact: do not present RE:AI’s initial GB300 tranche as a near-term rental option; use it as a demand and market-entry signal. Next: track the next tranche and require quantity, service date, region, interconnect, storage, operations and commercial terms before benchmarking. Evidence: high for deployment and full commitment; none for spare capacity or price. Singtel, 28 October 2026
2. STTGDC secured a JPY93 billion green loan for the fit-out and operation of its Greater Tokyo TYO1/TYO2 campus, which is designed for 70.8 MW of combined IT capacity. TYO1 is operational; TYO2 is under development for high-density AI workloads with provision for future liquid cooling. The announcement provides no vacancy, tenant allocation, TYO2 RFS date or lease price. Impact: financing reduces a delivery-funding risk for the campus but does not create verified leasable capacity. Next: qualify TYO1 vacancy separately and obtain a dated TYO2 RFS, liquid-cooling envelope and commercial schedule. Evidence: high for financing, operating status and design capacity; none for vacancy or rent. STTGDC, 27 October 2026
3. Singapore’s Strategic Goods (Control) Order 2026 takes effect on 1 December 2026. The official amendment table, checked after its 1 November resource-page update, says the 2026 list replaces the 2025 list and aligns with the 2025 Wassenaar and EU dual-use lists. The published key change under 4A003 is to technical-note definitions; the table does not identify a new GPU-specific hardware entry. Impact: this is a classification and process deadline for Singapore export, transhipment, transit, brokering and technology-transfer workflows, not evidence that every B300/GB300/H200/H100 transaction becomes controlled. Next: refresh item classifications and permit decision trees before 1 December. Evidence: high for commencement and published amendments; transaction applicability remains item-specific. Singapore Customs amendment table

GPU supply, pricing and lead time

No new physical inventory, OEM available-to-promise quantity, APAC transaction price or dated delivery lead time for B300, GB300, H200 or H100 was verified this week.

Cisco’s previously reported end-October ordering milestone has now been reached. Its current page says the relevant configurations are orderable through Cisco-authorized channel partners and labels the neocloud GB300 NVL72 offer “available today.” These are public orderability statements, not evidence of inventory, regional allocation, price or delivery date. A procurement opportunity remains unqualified until the authorized seller and OEM document the exact configuration, production or allocation evidence, delivery basis and support route. Cisco

Singtel’s GB300 announcement is deployment evidence, but the initial capacity is already committed. It should therefore be recorded as installed/committed operator capacity, not open GPU supply.

OEM and authorized channels

No new OEM production figure or authorized-channel APAC allocation was published. The Cisco/Supermicro route remains the most concrete authorized-channel path for B300 and GB300 in the public record, with the same limitations on stock and lead time.

Facility qualification remains part of procurement: Cisco states that NVL72 rack-scale systems can exceed 200 kW per rack and require liquid cooling, while the 8RU HGX B300 option addresses facilities that cannot yet support liquid cooling. Orders should be gated by the site’s rack power, cooling architecture, CDU and water conditions, network, redundancy and operations envelope.

Asian data centres and leasable capacity

New Greater Tokyo records

• STTGDC TYO1, Inzai: operational; the operator’s location page states more than 32 MW of IT capacity. The financing announcement states annualised design PUE below 1.25 and closed-loop cooling designed for no ongoing water consumption in normal operation. No vacant MW, rack count, rack density, minimum commitment, price, deposit, cross-connect, carrier list or SLA was published. It is a near-term candidate only for enquiry, not verified supply.
• STTGDC TYO2, Inzai: under development; the operator’s location page states more than 38 MW upon completion, while the financing release says it will support high-density AI and accelerated computing with provision for future liquid cooling. No RFS was published, so it is conservatively treated as a long-term pipeline record until a dated milestone is available.

The 70.8 MW combined figure is designed IT capacity, not vacancy. The JPY93 billion amount is financing, not a lease rate.

Rolling shortlist

• Near term, through 2 February 2027: add TYO1 as operational but vacancy-unverified. Continue to verify DayOne SG1’s historical Q4 2026 target and current availability at AirTrunk JHB1, STT Bangkok 1 and PDG TY1.
• Medium term, 3 February–2 November 2027: STT Bangkok 2’s Q1 2027 target and Equinix JH2’s Q3 2027 target remain unverified as allocable capacity.
• Long term, after 2 November 2027: add TYO2 conservatively because no RFS was published; retain Chief Telecom Huwei H2 2028 and DayOne Imari 2030.

No decision-grade public colocation rate in USD/kW/month was found for these sites. Electricity prices, project investment and campus power were excluded from rent comparisons.

GPU leasing economics

RunPod updated its public pricing page on 28 October, but the target rates remain unchanged from the previous baseline:

• B300 288 GB: USD 8.99/GPU/hour
• H200 141 GB: USD 5.29/GPU/hour
• H100 SXM 80 GB: USD 3.99/GPU/hour

Nebius remains unchanged:

• B300: USD 9.50/GPU/hour
• H200: USD 5.40/GPU/hour
• H100: USD 4.50/GPU/hour
• GB300 NVL72: contact sales

RunPod pricing · Nebius pricing

These public on-demand rates do not normalize region, sustained utilization, InfiniBand/RoCE topology, storage, managed operations, power, deposits, financing cost, residual value or guaranteed capacity. No payback calculation is presented because there is no current verified hardware acquisition price or field-complete APAC term quote.

Verda’s live page again did not expose numerical rates in the directly fetched page, while search-index snapshots contained values. Verda therefore remains outside the current comparable set; no price movement is asserted.

Cross-border compliance

• United States: no new BIS advanced-computing rule was identified in the official Federal Register list through 1 November. Existing ownership, destination, end-use/end-user, reexport and diversion controls remain the baseline.
• Singapore: SGCO 2026 becomes effective on 1 December. Refresh classifications and workflows before the commencement date; do not infer permit status from GPU model name alone.
• China: the 16 October interim registration requirement for listed AI-technology exports remains in effect, while the proposed catalogue revision remains open for comment through 15 November. Hardware, model, software and technical-service transfers require separate analysis.
• Malaysia, Indonesia and Thailand: no new official implementation change affecting the current GPU transaction baseline was verified this week.

ITAD

No verified H100/H200 completed secondary sale, physical lot, warranty-transfer policy change or decision-grade residual-value datapoint was found. Used DGX diligence should continue to require serial-level ownership, support entitlement, firmware baseline, management access, sanitization evidence and cross-border eligibility. Public cloud list-price changes do not establish secondary-market residual value.

Coverage and limitations

Primary-source checks covered NVIDIA/OEM and authorized-channel material, Singtel, STTGDC, Singapore Customs, BIS and official pricing pages. Public announcements rarely prove physical stock, vacancy or complete commercial economics. Designed IT capacity, committed GPU capacity and financing were not converted into available supply.