Key findings
- GPU finance is separating by tenant credit quality. SpaceX entered a binding commitment for an approximately $40 billion master finance lease with an Apollo-managed entity. The lessor will acquire GPU pods and related equipment; schedules are expected to begin on delivery in H1 2027 and run about five to five-and-a-quarter years. This is a financing commitment subject to closing conditions—not current-week hardware revenue or live capacity. SpaceX, 26 Oct | SEC filing, event 24 Oct
- Commercial implication: investment-grade, diversified tenants can still fund next-generation fleets at exceptional scale, while sub-investment-grade neoclouds face tighter and more expensive capital.
- Next step: underwrite long leases on tenant cash flow, delivery tranches and residual-value protection—not headline contract value.
- APAC GB300 scarcity is now evidenced by pre-commitment, not just announcements. Singtel RE:AI said its initial NVIDIA DGX GB300 capacity in Singapore was fully committed before deployment and will be sold as sovereign GPU-as-a-Service. GPU count, price and RFS date were not disclosed. Singtel, 28 Oct
- Commercial implication: regulated buyers are willing to reserve high-end capacity before RFS where data residency is part of the product.
- Next step: pursue wait-list and expansion capacity, but require rack count, energized power and commissioning milestones.
- Public GPU-hour pricing was unchanged on the four same-source baselines. Runpod, Lambda and Nebius posted the same B300/B200/H200/H100 SKUs as last week. This is a verified 0.0% week-on-week list-price move, not evidence that negotiated cluster prices were flat. Runpod | Lambda | Nebius
- Commercial implication: the pricing signal this week comes from availability and financing terms, not list-price resets.
- Next step: negotiate on commitment, power acceptance and network/storage inclusions rather than expecting an immediate posted-price cut.
- Inference competition widened beyond GPUs. Oracle announced plans to scale NVIDIA Vera Rubin systems to multi-gigawatt capacity while also adding Google TPU and Cerebras inference offerings. The announcements do not disclose GPU counts, sites, price or RFS, so they are planned supply rather than rentable inventory. Reuters, 27 Oct | Oracle
- Commercial implication: token economics will increasingly be set by mixed accelerators and software portability, not GPU rent alone.
- Token pricing added new segmentation rather than a broad price war. OpenAI released GPT-6.1 Terra at $1 input / $5 output per million tokens on the standard tier, with cached input at $0.05; Anthropic kept 50% Batch API discounts; DeepSeek listed time-of-day peak/off-peak pricing and will begin charging $0.75 per 1,000 successful Search or Fetch requests on 1 November. OpenAI pricing | OpenAI changelog, 26 Oct | Anthropic pricing | DeepSeek pricing
- Commercial implication: routing, caching, batching and tool-call charges matter more than a single headline token price.
Material transactions and commitments
| Announcement / event date | Parties and roles | Status | Equipment / scale | Value / term | Location / RFS | Use and market impact |
|---|---|---|---|---|---|---|
| 26 Oct / 24 Oct | Apollo-managed entity: lessor and equipment buyer; SpaceX: lessee/end user; NVIDIA: supplier named by SpaceX | Binding commitment; master lease still subject to closing | Next-generation NVIDIA Vera Rubin GPU pods and related equipment; unit count undisclosed | Approximately $40bn aggregate principal payments; each schedule about 5–5.25 years; interest/other costs excluded | Data centers selected by SpaceX; sites undisclosed; delivery and lease starts expected H1 2027 | Training/inference and terrestrial compute buildout. Large demand signal, but no current live supply. |
| 28 Oct / 28 Oct | Singtel Digital InfraCo RE:AI: sovereign cloud operator; NVIDIA: OEM/platform; enterprise customers: lessees/users | Planned deployment; initial capacity fully committed pre-RFS | NVIDIA DGX GB300; count undisclosed | Undisclosed | Singapore; RFS undisclosed | Sovereign enterprise AI; tightens near-term APAC GB300 availability. |
| 27 Oct / 27 Oct | Oracle Cloud: cloud operator; NVIDIA: supplier; Google and Cerebras: alternative accelerator suppliers | Expansion plan / product commitment, not disclosed purchase contract | Vera Rubin NVL72 at multi-gigawatt target; Google TPU and Cerebras inference supply; counts undisclosed | Undisclosed | Sites and RFS not disclosed | Training, inference and agents; shows a mixed-accelerator cloud strategy. |
| 27 Oct / talks ongoing | Argentum AI: neocloud/borrower; private-credit firm: prospective financier; DL Invest: Poland site partner | Financing talks only; not signed | Additional AI chips for Poland site; model/count undisclosed | Up to $5bn financing under discussion | Poland; RFS undisclosed | Credit for non-investment-grade end customers; a lead, not verified deployable capacity. Reuters |
Deployment and rentable supply
| Location | Operator | Hardware | State at cut-off | Availability / constraints |
|---|---|---|---|---|
| Singapore | Singtel RE:AI | DGX GB300 | Pre-deployment; first tranche committed | No public spot inventory. Sovereignty is part of the offer; GPU count, network topology, liquid-cooling capacity and RFS not disclosed. |
| SpaceX-selected data centers, United States | SpaceX / Apollo lessor | Vera Rubin GPU pods | Financing commitment; delivery expected H1 2027 | Not externally rentable on current evidence. Site power and locations undisclosed. |
| Oracle global OCI estate | Oracle | Vera Rubin NVL72, Google TPU, Cerebras | Planned expansion/product offering | “Multi-gigawatt” is not a live-GPU count. No site-level RFS, capacity or public rate disclosed. |
| Poland | Argentum AI / DL Invest | Model undisclosed | Financing talks | Do not treat $5bn potential debt or a 300MW campus plan as available GPU inventory. |
No verified new live public rental deployment of AMD Instinct, Intel Gaudi or Huawei Ascend was identified during the observation window. Google TPU and Cerebras are accelerators, not GPUs. Intel reported supply disruptions affecting AI-server chips, while no model-level rentable capacity was disclosed. This absence of evidence should not be read as absence of deployments.
Pricing and week-on-week changes
Public list prices observed 30 October 2026; USD per GPU-hour. Taxes, egress, network, storage and managed services differ unless noted.
| Provider / region / service | SKU and condition | Current | Prior (23 Oct) | WoW | Classification |
|---|---|---|---|---|---|
| Runpod, multi-region Secure Cloud Pods | B300, on-demand | $8.99 | $8.99 | 0.0% | Public list; per-second billing available |
| Runpod, multi-region Secure Cloud Pods | H200 | $5.29 | $5.29 | 0.0% | Public list |
| Runpod, multi-region Secure Cloud Pods | B200 | $7.99 | $7.99 | 0.0% | Public list |
| Lambda, US, 8-GPU instance | B200 SXM6 | $6.69 | $6.69 | 0.0% | Public list; tax extra; storage included per SKU |
| Lambda, US, 8-GPU instance | H100 SXM | $3.99 | $3.99 | 0.0% | Public list; tax extra |
| Lambda, 16/64/256+ GPU clusters, 2 weeks–1 year | B200 | $9.86 / $9.36 / $8.87 | same | 0.0% | Public cluster list; not comparable to single-GPU pods |
| Nebius, on-demand | B300 / B200 / H200 / H100 | $9.50 / $8.50 / $5.40 / $4.50 | same | 0.0% | Public list, effective 1 Oct |
| Nebius, preemptible floors | B300/B200 from $0.99; H200/H100 from $0.79 | unchanged | unchanged | 0.0% | Spot floor; interruption risk; not a firm transaction price |
Secondary market analytics on 29 October showed medians of $10.12 B300, $8.50 B200, $5.40 H200 SXM and $3.94 H100 SXM on demand, but provider mixes and reservation periods differ; these are cross-sectional benchmarks, not same-condition weekly transactions. Cloud GPU analytics
Hardware purchase and ITAD: no verifiable public spot transaction price for B300/GB300/H200/H100, AMD Instinct, Intel Gaudi or Ascend was disclosed this week. SpaceX’s $40bn is lease principal across equipment tranches, not a per-GPU purchase price. Residual-value underwriting should use realized secondary-sale data by board/server condition, warranty, interconnect and export eligibility; no new public realized-sale evidence was found.
Token market
| Supplier / SKU | Standard list price per 1M tokens | Discounts / conditions | Weekly interpretation |
|---|---|---|---|
| OpenAI GPT-6.1 Terra | $1 input / $5 output; $0.05 cached input | Flex/Batch $0.50/$2.50; Fast $2/$10; regional processing +10% for newer models | New SKU on 26 Oct; establish baseline, not a price cut versus a same model. |
| OpenAI GPT-6.1 Sol | $2 / $10; $0.10 cached input | Flex/Batch $1/$5; Fast $4/$20 | Unchanged in this issue’s observation. |
| Anthropic Claude Sonnet 5.5 | $2 / $10 standard; $1/$5 Batch | Batch is asynchronous and 50% lower | Stable public structure; long-context and cache economics remain workload-specific. |
| DeepSeek Flash | Peak $0.006 cache-hit, $0.30 cache-miss input, $1.20 output; off-peak half | Peak 01:00–04:00 and 06:00–10:00 UTC weekdays; concurrency 2,500 | Time routing can halve token cost. Search/Fetch fees start 1 Nov, so agent cost is not token-only. |
Token price cannot be converted directly into GPU rent: model quality, quantization, batch size, latency, cache hit rate, accelerator type and utilization change the denominator.
East–West comparison
| Market | Supply and commercial pattern | Constraints / alternatives | APAC BD meaning |
|---|---|---|---|
| Mainland China | Advanced US-chip access remains license- and end-user-sensitive; no new verified Ascend rental price or deployment count this week | H200/MI325X licensing remains case-by-case under January policy; Ascend adoption cannot be inferred from launch claims | Structure compliant in-country accelerator offers separately from offshore NVIDIA capacity; require end-user and re-export screening. |
| Japan / Korea / Taiwan | Mature cloud and OEM ecosystems; no same-SKU weekly rate reset verified | Power, grid and sovereignty shape RFS; Taiwan is central to hardware supply but not equivalent to regional rentable capacity | Sell delivery certainty and managed clusters, not only nominal chip allocation. |
| Southeast Asia | Singapore GB300 demand is visibly pre-committed; sovereign service commands strategic value | Land/power cost and data residency; disclosed rate/count absent | Expansion capacity and regulated workloads are the clearest near-term lead. |
| India | Demand is cost-sensitive and often procurement-led; no new verified high-end rental transaction this week | Grid, network and financing vary by site | Pair lower-cost inference accelerators with software migration support. |
| North America | Deepest capital market and broadest posted GPU pricing; SpaceX demonstrates investment-grade finance access | Neocloud credit dispersion, power-ready sites and residual value | Use US public rates as reference, then adjust for network, sovereignty and financing—not as direct APAC parity. |
| Europe | Poland is a financing/development lead; data localization supports regional demand | Higher power and financing friction; planned MW is not GPU supply | Verify tenant credit and grid connection before treating projects as lease inventory. |
| Middle East | Independent capital-and-power hub; no material new disclosed transaction in this window | Export licensing, chain of custody and end-user controls remain decisive | Position compliance-ready capacity and documentation as product features. |
BIS’s 2026 register showed no new advanced-computing rule published during this observation window. The operative H200/MI325X China policy remains case-by-case licensing under specified security and supply conditions; no enforcement change should be inferred from silence. BIS policy, 13 Jan | BIS Federal Register notices, checked 29 Oct
Decision summary
- Procurement: keep B200/H100 competitive tenders open; posted prices were flat, so demand concessions on term, network/storage and power-acceptance remedies.
- Leasing: favor creditworthy tenants and tranche-level leases. Separate finance-lease principal from service revenue and require residual-value protection.
- Token business: optimize routing, cache, Batch/Flex and tool calls. Do not model margin from token list price alone.
- ITAD: do not mark down H100/A100 solely because next-generation financing headlines are large; use realized secondary bids and deployment timing.
Evidence notes
High confidence: official company releases, SEC filing, provider price pages and BIS pages. Medium confidence: Reuters reporting on financing talks and market structure. Low-confidence broker indications and private-channel quotes were excluded. All monetary amounts are USD unless stated.