Key findings
- Treat power-ready delivery evidence as part of GPU inventory verification. Nscale lost a Microsoft anchor deal after grid delays pushed supply into the 2030s, while separate reporting says AWS has been receiving GB300 racks before some sites can power them. For any purchase or lease, require site-level energized MW, rack acceptance, commissioning sequence and RFS evidence—not only GPU allocation.
- China inference demand is spilling into workstation-class Blackwell. Reuters reported orders for more than 3 million RTX PRO 5500 cards from ByteDance, Alibaba and Tencent at a combined sticker value above $35 billion, but Nvidia reportedly warned partners that allocation could be only 5–10% of requested quantities. This is a demand signal, not a confirmed shipment.
- Use the public cloud prices below as an upper-bound negotiation benchmark, not a direct long-term cluster comparator. Published on-demand rates cluster around $4–$4.50/GPU-hour for H100 SXM and $8–$9.50 for B200/B300, but network topology, storage, SLA, region and commitment differ materially.
- Token economics are tightening. A public token-price index fell below $1 per million tokens and 54% from its May peak, while business token spend reportedly stayed flat for two months. Rising usage alone no longer proves rising model-provider revenue.
Material transactions and commitments
| Announced | Parties and status | Hardware / scale | Location / timing | Commercial meaning |
|---|---|---|---|---|
| 20 Oct | ByteDance, Alibaba and Tencent orders; reported by Reuters, not confirmed by buyers | >3m Nvidia RTX PRO 5500; ~$11,400 sticker/card; >$35bn aggregate sticker value | China; samples expected Dec 2026, commercial delivery Q1 2027 | Strong inference demand under export constraints, but Nvidia reportedly cannot guarantee fulfilment and may allocate only 5–10%. |
| 19 Oct | Nscale–Anthropic discussions; not signed | Proposed 460 MW second building using Vera Rubin; potential ~$45bn cloud contract | Monarch campus, West Virginia, US; timing undisclosed | Financing-scale demand remains large, but it is a negotiation, not contracted revenue or deployed compute. |
| 22 Oct | Nebius–Citi financing; committed finance, not customer revenue | $3.5bn, 364-day facility at term SOFR + 2.75% | Supports near-term data-centre and GPU procurement | Adds procurement capacity but also shows that neocloud expansion depends on continual capital access. |
| 19 Oct | Utho Cloud opens reservations | Nvidia B300; quantity and price undisclosed; 1/2/3-year terms | India; planned deployment Feb 2027 | A local-currency, data-resident reservation product. Treat as planned supply until GPU allocation, powered site and RFS evidence are disclosed. |
| 18 Oct | Microsoft withdraws from Nscale Loughton anchor arrangement after grid delay | Planned 23,040 GB300 GPUs / 50 MW | Loughton, UK; supply reportedly pushed into the 2030s | A negative transaction signal: grid timing can invalidate otherwise credible GPU contracts. |
Sources: Reuters — China RTX PRO 5500 orders, 20 Oct; Reuters — Nscale/Anthropic talks, 19 Oct; Reuters — Nebius facility, 22 Oct; ET Telecom — Utho reservations, 19 Oct; Reuters — Nscale power delay, 18 Oct.
Deployment and rentable supply
Available or publicly listed
- Runpod lists B300, B200, H200 and H100 pods across 30+ regions. Availability remains console- and region-dependent.
- Lambda lists self-service B200 and H100 instances and 16–2,000+ GPU clusters. Cluster prices include materially different fabrics and service scope from single-instance pricing.
- Nebius lists H100/H200/B200/B300 and launched public live spot-price visibility on 21 October. Spot rates may change as often as every 15 minutes and can be interrupted.
Planned / reservation stage
- Utho's India B300 capacity is targeted for February 2027; the public announcement does not disclose quantity, site, network, power readiness or reservation price.
- Nscale's proposed second West Virginia facility is under discussion. The UK Loughton delay is a clear reminder that announced GPU count is not deliverable capacity until grid and commissioning milestones are verified.
Reported staging risk
Business Insider reported that AWS's “Early Rack Receipt” process was set to accept roughly $5 billion of GB300 racks at US facilities before final deployment and around $3 billion more monthly, with some equipment waiting for power. Amazon has not provided a rack-count or site-level public reconciliation, so this is a credible reported risk signal rather than verified live capacity.
Sources: Runpod pricing; Lambda pricing; Nebius spot-pricing announcement, 21 Oct; Business Insider report, 20 Oct.
Price baseline — 23 October 2026
| Provider / product | Access type | USD/GPU-hour | Conditions and limits |
|---|---|---|---|
| Runpod B300 | Pod | 8.99 | 288 GB; public page updated 8 Oct; region/availability may vary |
| Runpod B200 | Pod | 7.99 | 180 GB |
| Runpod H200 | Pod | 5.29 | 141 GB |
| Runpod H100 SXM | Pod | 3.99 | 80 GB |
| Lambda B200 | Self-service instance | 6.69 (8-GPU) / 6.99 (1-GPU) | Tax extra; local SSD/CPU/RAM included |
| Lambda H100 SXM | Self-service instance | 3.99 (8-GPU) / 4.29 (1-GPU) | Tax extra |
| Lambda B200 | 1-Click Cluster | 9.86 (16) / 9.36 (64) / 8.87 (256+) | 2 weeks–1 year; production cluster scope |
| Lambda H100 | 1-Click Cluster | 6.16 (16) / 5.85 (64) / 5.54 (256) | 2 weeks–1 year |
| Nebius B300 | On-demand | 9.50 | Spot shown “from $0.99”; dynamic/preemptible |
| Nebius B200 | On-demand | 8.50 | Spot shown “from $0.99” |
| Nebius H200 | On-demand | 5.40 | Spot shown “from $0.79” |
| Nebius H100 | On-demand | 4.50 | Spot shown “from $0.79” |
This is a baseline, not a same-spec ranking. Published prices exclude or vary in tax, region, network, storage, SLA, CPU/RAM allocation and interruption risk. Long-term 128+ GPU RoCE/InfiniBand leases should normally price below fully managed public cluster rates only after matching scope and credit terms. No reliable comparable public B300/GB300 server purchase transaction price was disclosed in the reporting window.
Token market
- Silicon Data's volume-weighted LLM Token Expenditure Index fell below $1 per million tokens, down 54% from its late-May peak of $2.07. Ramp's business-spend index was down 48% from its late-February peak, while business token spend reportedly remained flat for two consecutive months even as token usage rose. The signal is mix shift, routing and provider price compression—not a fall in physical GPU consumption by the same percentage.
- Google Gemini 4 Argon remains an important current benchmark following its 13 October launch: $2/million input and $10/million output through 30 November, cached input 95% cheaper; standard pricing is scheduled to double to $4/$20 afterward.
- Operational implication: lease underwriting should model revenue per delivered token, cache ratio, batch/latency mix and utilization. A fixed GPU-hour rate cannot be justified by token-volume growth alone.
Sources: Axios token-price analysis, 22 Oct; Google Cloud — Gemini 4 Argon pricing.
East–West market comparison
| Market | Current signal | Access and economics | BD implication |
|---|---|---|---|
| North America | Very large cloud commitments coexist with power delays and rising financing scrutiny. | Deepest transparent GPU cloud market; premium managed-cluster prices; capital availability increasingly differentiates suppliers. | Verify energized capacity and lender/customer-contract dependencies before accepting RFS claims. |
| Europe | Nebius introduced public live spot visibility; UK grid delay removed a 23,040-GB300 anchor deployment. | Growing price transparency, but grid connection is a binding delivery constraint. | Europe can provide quality managed supply, but delivery evidence must include grid milestones. |
| China mainland | >3m reported RTX PRO 5500 orders reflect scarcity of permissible inference compute. H200 is case-by-case licensed; B300 is not permitted for Chinese buyers under current reported rules. | Workstation-class cards can command scarcity value; domestic accelerators gain adoption, but migration and software costs remain material. | Do not route controlled data-centre GPUs indirectly. Focus only on formally classified, licensed and end-user-verified structures. |
| India | Utho opened B300 reservations with rupee billing and local residency, deployment targeted Feb 2027. | Local commercial packaging is improving; public scale and price remain undisclosed. | Potential enterprise channel, subject to allocation, site and network proof. |
| Southeast Asia | No material new verified public transaction in this week's global scan. | Attractive deployment geography, but diversion controls and site-level proof remain central. | Maintain OEM/dealer payment and KYC/EUC discipline; do not treat regional location as proof of end use. |
| Japan/Korea/Taiwan | No material new verified transaction within the window. | Mature infrastructure and supply-chain roles; public rental comparables remain fragmented. | Track new rack-scale launches separately from announced ecosystem partnerships. |
| Middle East | No new signed deployment in the window; large 2027 plans remain the structural backdrop. | State-backed capital and power access can support scale, but many projects remain forward commitments. | Require phasing, customer allocation and power/commissioning milestones. |
Alternatives to Nvidia
AWS released TorchNeuron in public beta, enabling native PyTorch operation on Trainium with a configuration change rather than model-code rewrites. AWS says more than 30 early customers are using it. This matters because software migration cost—not peak silicon performance—is the main barrier to non-CUDA adoption. It is an adoption signal, not evidence that Trainium has achieved broad fungibility with Nvidia clusters.
No material new AMD Instinct deployment was verified in the 16–23 October window. Existing 2027 MI400/MI450 commitments remain relevant but should not be presented as new weekly supply. No new independently verified commercial Huawei Ascend deployment was found in this window; China demand should not be inferred solely from vendor roadmaps.
Source: AWS TorchNeuron announcement, 20 Oct.
Compliance and ITAD
BIS's public update page showed no new advanced-computing rule issued during the reporting window. The operative baseline remains: H200 and AMD MI325X applications for China are reviewed case by case under the 13 January 2026 policy when security requirements are met; B300 sales to Chinese buyers remain prohibited according to Reuters' current reporting. Every transaction still requires item classification, end-user/end-use screening, destination and ownership checks, and diversion controls.
For ITAD, Reuters reported that discontinued RTX 6000D prices in China had doubled. This suggests residual-value support for inference-capable workstation GPUs under scarcity, but it is not a globally transferable resale benchmark because export eligibility and local access determine realizable value.
Sources: BIS semiconductor licensing policy, 13 Jan; BIS News and Updates; Reuters, 20 Oct.
Decision view
For procurement and leasing, the best near-term negotiating anchor is a matched-scope public baseline of roughly $4–$4.50/H100 SXM-hour, $5.3–$5.4/H200-hour, $6.7–$8.5/B200-hour, and $9–$9.5/B300-hour. Production clusters can price higher because networking, orchestration and guaranteed capacity are included. For a long-term B300 cluster, accept a premium only against evidence of powered racks, commissioning milestones, fabric/storage scope and enforceable SLA.
For token businesses, assume continuing price compression and make utilization plus inference efficiency the core underwriting variables. For ITAD, inference-oriented GPUs may retain stronger value in constrained regions, but only where resale is legally deliverable and serviceable.