Reporting period: 19–25 October 2026. Report date: 26 October 2026. Compared with the 19 October baseline. “Verified” means supported by a public primary source; it does not mean physical stock, available-to-promise capacity, or commercially available data-centre space unless stated.
Three decision-relevant changes
- Cisco says its NVIDIA/Supermicro AI POD configurations become orderable globally through Cisco-authorized channel partners at the end of October 2026. The portfolio includes air-cooled and liquid-cooled 8-GPU HGX B300 systems and GB300 NVL72. This creates an OEM/authorized-channel route aligned with controlled payment and support requirements, but Cisco’s page does not prove APAC inventory, allocation, price, lead time, or a precise order-opening date. Action: accept a quotation only after seller authorization, exact configuration, available-to-promise date and warranty/support route are documented. Evidence: high for stated channel milestone; low for stock and lead time. Cisco source
- Late-verified correction: China’s interim registration requirement has been in effect since 16 October 2026 for exports covered by the proposed AI-technology catalogue revision. The catalogue revision remains a proposal open for comment through 15 November, but the interim registration applies now to listed technology transfers, including certain licensing, overseas use and changes of ownership or actual control, including transactions already started but not completed. This can affect an APAC GPU deployment when China-origin model, agent, compute-cluster or information-processing technology is transferred with the project; it does not by itself make offshore GPU hardware a newly controlled product. Action: separate hardware, software, model and technical-service workstreams and obtain transaction-specific advice before transfer. Evidence: high for the interim measure and proposal status. MOFCOM announcement · MOFCOM Q&A
- Verda’s current public rate cannot be verified consistently. The live pages checked on 26 October did not expose numerical rates to the verifier, while cached/search-index snapshots disagree with each other and with the direct 19 October observation. Action: remove Verda from the current comparable-rate set until a dated, configuration-specific public quote or written offer is available. The 19 October figures remain historical observations only. Evidence: high for verification failure; no evidence for a current price movement.
GPU supply, pricing and lead time
No new, publicly verifiable physical inventory, OEM available-to-promise quantity, APAC transaction price or delivery lead time for B300, GB300, H200 or H100 was found for the reporting period.
Cisco’s end-October channel milestone is orderability, not stock. The public page identifies configurations but provides no unit count, regional allocation, price, Incoterms, tax treatment or lead time. A quoted opportunity should therefore remain unqualified until the authorized seller and OEM provide serial/configuration detail, allocation or production slot, delivery location and support coverage.
OEM and channel
Cisco’s published full-stack route is the only material channel change verified this week. Cisco states that it will be the first support contact for the full stack. The page also gives an important facility-design split: GB300 NVL72 can exceed 200 kW per rack and requires liquid cooling, while the 8RU HGX B300 option is designed for facilities without liquid cooling. This turns cooling readiness into an order-qualification gate, rather than a post-purchase facilities question. Cisco source
Asian data centres and leasable capacity
New long-term candidate: Chief Telecom held a groundbreaking on 20 October for the Huwei AI Data Center in Huwei, Yunlin, Taiwan, with formal operation planned for the second half of 2028. The official announcement states 16 MW of total facility power, estimated PUE 1.4, air and water cooling, and more than 24 hours of generator operation at full load. It references Tier IV design standards; it does not establish certification. No rentable IT-load MW, empty rack count, rack density, minimum commitment, lease price, deposit, cross-connect cost, supply/return-water conditions or SLA was published. The 16 MW figure must not be treated as available IT capacity. Impact: relevant for long-term liquid-cooled pipeline and network-led business development, not for near-term deployment. Action: track construction and obtain site-level IT-load, cooling-envelope and commercial evidence before treating it as supply. Evidence: high for project facts; none for vacancy or rent. Chief Telecom source
No verified change was found in the near-term or medium-term shortlist from the previous baseline. The highest-priority verification points remain:
- Near term, through 26 January 2027: DayOne SG1’s historical Q4 2026 target; AirTrunk JHB1, STT Bangkok 1 and PDG TY1 operating status. None has public proof of current vacancy or a complete lease quote.
- Medium term, 27 January–26 October 2027: STT Bangkok 2’s Q1 2027 target and Equinix JH2’s Q3 2027 target. Neither target proves allocable capacity or price.
- Long term, after 26 October 2027: Chief Telecom Huwei, planned for H2 2028, and DayOne Imari, planned for 2030. Both require milestone and available-capacity verification.
No public, decision-grade colocation rate in USD/kW/month was found for the shortlisted sites. Electricity, land rent, project investment and campus design capacity were excluded from rent comparisons.
GPU leasing economics
Public on-demand list prices verified on 26 October were unchanged from the 19 October baseline:
- RunPod Secure Cloud: H100 SXM 80 GB — USD 3.99/GPU/hour; H200 141 GB — USD 5.29/GPU/hour; B300 288 GB — USD 8.99/GPU/hour. The public pricing page says it was updated 8 October 2026. RunPod pricing
- Nebius AI Cloud: H100 — USD 4.50/GPU/hour; H200 — USD 5.40/GPU/hour; B300 — USD 9.50/GPU/hour; GB300 — contact sales. Nebius pricing
These list prices are not fully comparable with a dedicated lease. Public pages do not normalize term, sustained utilization, InfiniBand/RoCE topology, storage, managed operations, power, deposits, financing cost, residual value or exact region availability. No payback estimate is presented because there is no current verified acquisition price and no APAC contract quote with the necessary cost fields.
Cross-border compliance
The China measure above is a correction to the prior reporting baseline because the official 16 October publication was verified after the 19 October issue. The interim registration measure is already effective; the catalogue amendment remains proposed. MOFCOM’s Q&A distinguishes the technology itself from products made using it and indicates that public-domain open-source material and papers are outside the proposed scope. It also describes possible treatment of overseas model localization and certain self-operated overseas compute centres; those points should remain labelled as proposed until final rules are issued.
No new U.S. BIS advanced-computing rule or material Singapore, Malaysia, Indonesia or Thailand implementation change was verified for 19–25 October. Existing ownership, destination and end-use controls therefore remain the working baseline; no transaction should rely on the absence of a weekly update.
ITAD
No verified H100/H200 secondary-market sale, completed price, physical lot or warranty-transfer change was found this week. NVIDIA’s current DGX H100/H200 firmware guide routes firmware packages and the update tool through the Enterprise Support Portal. For used DGX diligence, verify support entitlement, firmware baseline, management-controller access, serial numbers and data-erasure evidence before valuing a lot. This is an operational check, not evidence that NVIDIA warranty transfers automatically. NVIDIA firmware guide
Coverage and limitations
Primary-source checks focused on NVIDIA, Cisco, OEM/authorized-channel material, facility operators and regulators. Public information rarely proves physical inventory, vacant data-centre capacity or a complete lease economics package. Marketing capacity, total facility power and historical targets were not converted into supply.
