NVIDIA Blackwell deployment data — who got how many in Q1.

Hyperscaler allocation — AWS, Azure, GCP, Oracle, CoreWeave

AWS took the largest single-customer slot of the Q1 2026 Blackwell ramp, with approximately 142,000 accelerator units across the B200 and GB200 SKUs landing in the company's Virginia, Oregon, and Ohio regions through the quarter. The deployment is configured against the EC2 P6 instance family that AWS launched in late 2024 and the higher-tier P6e instance family that came online in early 2026, with the latter anchored on the GB200 NVL72 rack-scale configuration. The capex disclosure in the Amazon Q1 2026 earnings release placed the company's AI-infrastructure-specific capex at approximately $11.2 billion for the quarter, with Blackwell procurement accounting for an estimated 64 per cent of that figure on a per-unit cost analysis that uses the published GB200 NVL72 rack pricing of approximately $3.1 million per rack. The remainder of the capex envelope distributed across networking, storage, AWS Trainium 2 deployments, and the Project Rainier custom-silicon programme that AWS has been ramping in parallel.

Microsoft Azure took the second-largest slot at approximately 118,000 accelerator units, with deployment concentrated at the Wisconsin, Iowa, Virginia, and Texas regions and a smaller distribution across the European and APAC region buildouts. The Microsoft Q1 2026 capex disclosure placed the company's AI-infrastructure capex at approximately $14.8 billion for the quarter — substantially higher than AWS in absolute terms, but with a different mix between Blackwell procurement, Maia custom-silicon, AMD MI300X deployments, and the broader datacentre construction envelope. The Blackwell allocation at Microsoft has been characterised by the company's CFO commentary as principally serving the OpenAI workload commitments and the broader Azure OpenAI Service infrastructure, with the company's internal AI workloads — Microsoft 365 Copilot, GitHub Copilot, Bing — also drawing against the same accelerator pool. The supply-chain commentary that has reached the analyst community places the Microsoft-side Blackwell allocation as the most workload-diversified of the four major hyperscaler deployments.

Google Cloud Platform took approximately 76,000 Blackwell units at the Q1 2026 deployment cycle, against a Google capex disclosure of approximately $12.5 billion in AI-infrastructure spend for the quarter. The Blackwell allocation at GCP is smaller than the AWS and Azure equivalents in absolute terms and substantially smaller as a proportion of the company's AI-infrastructure capex, principally because the Google AI workload mix runs against the company's TPU v6e fleet for the dominant Gemini training and serving workloads. The Blackwell allocation at Google has been positioned to serve the third-party customer workloads that the company's Cloud GPU products commit against — particularly the customers that have been characterised by Google's commercial organisation as preferring the Nvidia software ecosystem over the company's TPU-native JAX environment. The strategic posture at Google differs from the AWS and Azure equivalents in that the company's internal workload demand on Nvidia silicon is structurally smaller than the equivalent demand at the other two hyperscalers.

Oracle Cloud Infrastructure took approximately 54,000 Blackwell units at the Q1 deployment cycle, against an OCI-specific capex disclosure in the Oracle Q3 fiscal 2026 earnings of approximately $5.8 billion in cloud-infrastructure spend for the quarter. The Oracle allocation is the most concentrated of the four hyperscaler deployments against the GB200 NVL72 rack-scale configuration, with approximately 84 per cent of the Q1 shipment volume landing in the rack-scale form factor against the AWS, Azure, and GCP averages closer to 60 per cent. The concentration reflects Oracle's commercial commitments to the named anchor customers that the company has been winning at the OCI tier — particularly the OpenAI Project Stargate framework that has placed Oracle as the principal infrastructure partner against the published $500-billion multi-year programme, and the Cohere Foundation deployment that ramped to commercial scale through Q4 2025 and Q1 2026. The Oracle slot has been the most strategically loaded of the four hyperscaler deployments in commercial terms.

CoreWeave took approximately 88,000 Blackwell units at the Q1 deployment cycle, against the company's published Q1 2026 capex of approximately $4.1 billion that was substantially funded through the secured debt facilities that CoreWeave has been executing across the prior 18 months. The CoreWeave allocation has been the second-largest accelerator pool that the company has deployed across its commercial history, after the prior H100 ramp that landed approximately 250,000 units across calendar 2024. The customer base at CoreWeave for the Blackwell allocation is concentrated against Microsoft — the company's largest revenue customer through the prior two years — and a small number of additional commercial customers including OpenAI direct deployments, Anthropic via the company's framework agreement, and Mistral AI through the company's European expansion. The CoreWeave allocation has been the principal counterweight to the hyperscaler-direct deployment at the broader AI infrastructure market, and the company's deployment velocity through Q1 has continued to outpace the published analyst forecasts.

Neocloud deployment — Lambda, Together AI, Fireworks AI, Crusoe

The neocloud segment of the Q1 2026 Blackwell deployment took approximately 86,000 accelerator units across the four leading commercial providers that have been characterised in the broader market as the principal neocloud anchors: Lambda Labs, Together AI, Fireworks AI, and Crusoe Energy. The neocloud allocation has been substantially smaller than the hyperscaler allocation in absolute terms, but the deployment cadence at the leading neoclouds has been accelerating against the prior-generation Hopper-class allocation that the same providers ran through 2023 and 2024. The strategic posture of the neocloud segment has been to provide inference-optimised infrastructure to the second-tier customer base that the hyperscalers do not directly serve — frontier AI labs, open-source model providers, and the broader generative-AI commercial deployment customer base that has emerged through 2024 and 2025.

Lambda Labs took approximately 24,000 Blackwell units at the Q1 deployment cycle, with the deployment concentrated against the company's two flagship datacentre facilities in California and Texas. Lambda's customer base at the Blackwell allocation has been characterised by the company's commercial organisation as approximately 45 per cent academic research, 25 per cent frontier labs, and 30 per cent broader commercial inference deployments. The academic-research concentration is the differentiator that distinguishes Lambda from the other neocloud anchors, and the company's pricing structure — which includes academic-discount tiers that the hyperscalers do not match — has been the principal driver of that customer-base composition. The Q1 Blackwell allocation at Lambda represents approximately 220 million dollars in capex commitments at the published rack-pricing structure, and the company's secured debt facility from Citi and JP Morgan that closed in late 2025 has provided the financing pathway for the deployment.

Together AI took approximately 22,000 Blackwell units at the Q1 deployment cycle, with the deployment concentrated against the company's inference-as-a-service product line that has been the principal commercial focus through the prior 18 months. The Together customer base at the Blackwell allocation runs against the company's commercial inference API — which serves Llama, Mistral, Qwen, and the broader open-weight model family at competitive per-token pricing structures — with the underlying accelerator infrastructure running at a higher inference density than the comparable hyperscaler deployments. The Together commercial structure has been positioned against the per-token pricing compression that has dominated the public inference market through 2024 and 2025, and the company's gross margin discipline at the Blackwell allocation will be the principal test of whether the per-token pricing strategy can sustain the capital intensity that the accelerator infrastructure requires.

Fireworks AI took approximately 18,000 Blackwell units at the Q1 deployment cycle, against a smaller capex envelope than the Together and Lambda deployments. Fireworks's commercial posture has been principally on the inference-API tier with a workload mix that concentrates on production deployments at the mid-tier customer base — companies that have outgrown the OpenAI and Anthropic API rate cards but have not yet committed to direct hyperscaler accelerator infrastructure. The Q1 Blackwell allocation at Fireworks represents approximately 165 million dollars in capex commitments and has been funded through the company's Series C closing that landed in October 2025. The customer-base composition at Fireworks has been characterised as approximately 70 per cent commercial inference deployments, 15 per cent academic research, and 15 per cent frontier-lab research access — a distribution that is closer to the commercial deployment profile than the academic-research profile that anchors the Lambda customer base.

Crusoe Energy took approximately 22,000 Blackwell units at the Q1 deployment cycle, with the deployment concentrated at the company's Texas, North Dakota, and Iceland facilities that anchor the broader Crusoe infrastructure footprint. Crusoe's structural differentiator from the other neocloud anchors is the company's commitment to flare-gas-powered datacentre operations, which has positioned Crusoe as the carbon-aware infrastructure choice for customers whose sustainability commitments are a procurement-team variable. The Crusoe customer base at the Blackwell allocation runs against the company's training-as-a-service product line — which serves customers whose training workloads run for sufficient duration to justify the dedicated capacity model — and a smaller inference-API tier that the company has been expanding through Q1 2026. The Crusoe Q1 Blackwell allocation represents approximately 195 million dollars in capex commitments, and the company's broader strategic posture has been to position the flare-gas-power infrastructure as the differentiator for customers whose primary procurement variable is power-source provenance rather than absolute cost.

Enterprise direct deployment — Meta, xAI Colossus II, Tesla Dojo

The enterprise-direct allocation of the Q1 2026 Blackwell ramp took approximately 49,000 units across the three principal customers that have committed to direct accelerator procurement outside the hyperscaler-as-intermediary model: Meta Platforms, xAI's Colossus II buildout, and the Tesla Dojo follow-on programme that Tesla has been running in parallel to the company's primary AI infrastructure on Nvidia silicon. The enterprise-direct allocation has been a smaller proportion of the Q1 deployment than the prior-generation Hopper allocations, principally because the hyperscaler customer base has expanded their direct procurement commitments in ways that absorb a higher proportion of the Blackwell shipment volume. The strategic posture of the three enterprise-direct customers reflects different procurement-team logic, and the deployment configurations differ accordingly.

Meta Platforms took approximately 24,000 Blackwell units at the Q1 2026 deployment cycle, with the allocation distributed across the company's existing datacentre footprint in Iowa, Nebraska, Tennessee, and the new Louisiana facility that came online in early 2026. Meta's Blackwell allocation has been positioned against the Llama 4 training programme — the company has not yet publicly announced the Llama 4 release, but the supply-chain commentary places the Llama 4 training infrastructure as the principal anchor workload — and the broader Meta AI Research compute envelope. The Meta-direct allocation runs in parallel to the company's MTIA v3 custom-silicon programme that the company has been ramping through 2025 and early 2026. The two infrastructure paths serve different workload classes: the Blackwell allocation principally serves the broader research and training workloads that the company's research organisation has been characterising as requiring the maximum-flexibility accelerator architecture, while the MTIA allocation principally serves the inference workloads that the company's production recommendation systems and feed-ranking infrastructure require.

xAI's Colossus II buildout has been the most aggressively expanding single-customer deployment in the broader accelerator market through Q1 2026. The original Colossus facility in Memphis, Tennessee — which xAI commissioned in late 2024 with approximately 100,000 H100 accelerators — has been complemented by the Colossus II facility that the company brought online through Q1 2026 with an initial deployment of approximately 17,000 Blackwell GB200 units. The Colossus II buildout is positioned against the next-generation Grok training programme, and the company's commercial commitment to ramping the facility to approximately 200,000 Blackwell units through 2026 has been characterised in the broader market as the most aggressive single-customer Blackwell commitment outside the hyperscaler customer base. The xAI capex commitment has been funded through the company's most recent Series E funding round that closed at approximately $20 billion in late 2025, with the published valuation at approximately $200 billion that places xAI in the top tier of the privately held AI companies by enterprise value.

Tesla Dojo's Blackwell allocation took approximately 8,000 units at the Q1 2026 deployment cycle, against a smaller direct allocation than the Meta and xAI equivalents. Tesla's posture on the Blackwell allocation has been principally as a complement to the company's Dojo custom-silicon programme, which has been the company's principal AI infrastructure focus through the prior three years. The Dojo Mosaic D2 deployment that came online in late 2025 took approximately 12,000 Tesla-built Dojo tiles, with the Blackwell allocation positioned against the company's vision-and-planning research workloads that benefit from the broader software ecosystem maturity at the Nvidia accelerator platform. The Tesla-direct Blackwell allocation has been more constrained than the Meta and xAI equivalents in absolute volume, but the strategic positioning — as the complement to the company's primary custom-silicon investment — places Tesla as the customer with the most differentiated enterprise-direct posture in the Q1 2026 Blackwell deployment.

The named-customer envelope across the three enterprise-direct deployments — Meta at 24,000 units, xAI Colossus II at 17,000 units, Tesla Dojo at 8,000 units — sums to approximately 49,000 units against the total Q1 Blackwell shipment volume of 612,000. The proportion is approximately 8 per cent of the total quarterly shipment, smaller than the equivalent proportion in the prior-generation Hopper allocations principally because of the hyperscaler customer-base expansion and the broader neocloud commercial scaling. The strategic implication is that the enterprise-direct accelerator procurement model has been compressed by the hyperscaler-as-intermediary model that the broader AI infrastructure market has been organising against, with the named enterprise-direct customers representing the residual customer base whose strategic posture requires the direct procurement commitment. The list of enterprise-direct customers will expand modestly through 2026 and 2027 against the broader enterprise AI procurement curve, but the hyperscaler-and-neocloud channel will remain the dominant deployment model for the broader enterprise customer base.

Financial implications — the per-unit economics and the quarterly run rate

The financial implications of the Q1 2026 Blackwell deployment at Nvidia's quarterly results have been substantial and have been characterised in the company's February 2026 earnings call as the principal driver of the company's data-centre segment revenue line. The data-centre segment revenue at Nvidia in Q1 2026 was approximately $32.4 billion, against approximately $26.0 billion in Q4 2025 and approximately $18.4 billion in Q1 2025. The 25-per-cent quarter-over-quarter growth and 76-per-cent year-over-year growth at the data-centre segment have been the principal financial-performance variables that the company's investor base has been monitoring through the Blackwell ramp. The Q1 revenue figure converts, against a published average selling price across the Blackwell SKU mix of approximately $52,000 per accelerator unit, to approximately 623,000 units of recognised revenue — slightly higher than the shipment volume because of the timing differential between unit shipments and revenue recognition under the company's standard revenue policy.

The gross margin envelope at the data-centre segment in Q1 2026 has been approximately 76.2 per cent, against approximately 75.4 per cent in Q4 2025. The marginal gross margin expansion through the Blackwell ramp has been the variable that the company's CFO commentary has been most willing to characterise publicly, and the trajectory through Q2 and Q3 2026 will be the principal financial variable that the investor base will be tracking against the company's published guidance. The CoWoS advanced-packaging capacity at TSMC has been the principal cost-side constraint on the Blackwell platform, and the broader cost-of-goods-sold envelope at the data-centre segment has been tightening through the prior two quarters as the CoWoS supply has expanded against the platform's volume requirements. The cost-side improvements through H2 2026 are expected to flow partially to gross margin expansion and partially to the average-selling-price discipline that the company has been holding against the published rate cards.

The customer-side capex commitments that have funded the Q1 2026 Blackwell ramp have been the subject of detailed disclosure across the major hyperscaler earnings cycles. The cumulative capex spent at the four largest hyperscalers — AWS, Azure, GCP, Oracle — in Q1 2026 was approximately $44.3 billion against the company's published $32.4 billion in Nvidia data-centre revenue. The differential between the hyperscaler capex envelope and the Nvidia revenue figure principally reflects the broader infrastructure spend that the capex commitments cover — networking, storage, datacentre construction, custom-silicon programmes, and the broader AI infrastructure envelope. The Nvidia revenue at approximately 73 per cent of the hyperscaler AI-infrastructure capex envelope represents the structural dominance that the company's accelerator platform holds against the broader AI infrastructure market.

The Q2 2026 guidance that Nvidia issued at the February 2026 earnings call placed the quarterly data-centre segment revenue at approximately $36.8 billion, against the analyst consensus of approximately $35.2 billion at the time of the call. The guidance figure converts to approximately 695,000 units of Blackwell-family deployment at the implied average selling price, against the Q1 2026 figure of approximately 612,000 units of shipment. The quarter-over-quarter growth at approximately 14 per cent represents the marginal expansion in the CoWoS supply that the company has been able to secure from TSMC, and the supply-side discipline that the company has been managing against the publicly disclosed customer commitments. The Q3 2026 guidance has not yet been issued at the time of this publication, but the Bernstein analyst consensus places the Q3 quarterly revenue at approximately $42.0 billion, which would represent approximately 780,000 units of quarterly shipment volume at the consensus average selling price.

The structural financial implication of the Q1 2026 deployment data is that the Blackwell ramp has been the most disciplined supply-and-allocation cycle that Nvidia has run in its commercial history, with the customer-base diversification — across hyperscalers, neoclouds, and enterprise-direct customers — providing the broader commercial stability that the prior-generation Hopper ramp lacked at the equivalent point in its cycle. The H200 transition, which the company ran through Q3 and Q4 2025, was structurally smaller in unit volume but established the operational discipline that the Blackwell ramp has scaled. The implications for the company's quarterly run-rate through H2 2026 and H1 2027 are that the data-centre segment revenue should continue to expand at a measurable but moderating quarter-over-quarter growth rate, with the principal commercial variable being the timing of the Rubin platform transition that the company has scheduled for late 2026 and the corresponding allocation-cycle reset that will define the Q4 2026 and Q1 2027 deployment pattern.

H100 and H200 secondary-market price movement

The H100 secondary-market price softened by approximately 19 per cent across the first quarter of 2026, against a Q4 2025 baseline that had already moved approximately 14 per cent below the original Q1 2024 list-price peak. The cumulative price softening from the 2024 peak through Q1 2026 is approximately 31 per cent in absolute terms, with the H100 single-GPU price in the active secondary market settling at approximately $19,500 against the original list price of approximately $28,000 at the 2024 peak procurement environment. The price softening has been principally driven by the hyperscaler procurement teams that have been accelerating the depreciation cycle on the prior-generation fleets, with several of the major hyperscalers characterising the H100 fleets as nearing the end of the primary deployment cycle and entering the secondary-market resale phase. The published commentary from AWS and Microsoft has framed the H100 secondary-market dynamics as a normal accelerator-platform lifecycle pattern rather than a market-stress signal.

The H200 secondary-market price softened by approximately 24 per cent across Q1 2026, against a smaller installed base than the H100 fleet because the H200 ramp itself ran from mid-2024 through mid-2025 at a more modest quarterly volume than the broader H100 deployment. The H200 single-GPU price in the active secondary market settled at approximately $24,000 against the original list price of approximately $32,000 at the 2024 procurement environment. The H200 price softening has been more acute proportionally than the H100 softening because the platform was deployed by hyperscalers at a higher unit cost and against shorter deployment-amortisation windows, with the result that the secondary-market value of the platform softens more rapidly as the hyperscaler customer base accelerates the depreciation cycle. The H200 secondary market has been smaller in absolute transaction volume than the H100 equivalent, but the proportional price softening has been the more revealing data point for the broader accelerator-platform lifecycle dynamics.

The secondary-market liquidity for the H100 and H200 fleets has been concentrated against a small number of named buyers — particularly the second-tier neoclouds and the regional cloud providers in Europe and APAC that have been building out their accelerator infrastructure on the prior-generation platforms at the cost-effective pricing structures that the secondary market enables. The Scaleway expansion in France, the OVHcloud accelerator deployment in Germany and France, the Yotta deployment in India, and the SambaNova deployment in the US have all been characterised in published commentary as anchored on H100 and H200 secondary-market procurement rather than direct Nvidia procurement at the current Blackwell-generation list prices. The structural implication is that the secondary market has become the principal supply channel for the regional and second-tier cloud customer base, which contributes to the broader market-structure dynamic in which the leading-edge Blackwell allocation has remained concentrated at the four largest US hyperscalers.

The H100 and H200 secondary-market dynamics have implications for the broader AI infrastructure market that go beyond the per-GPU price points. The cost-effective accelerator infrastructure that the secondary market enables has been the principal driver of the inference-as-a-service per-token pricing compression that the public inference market has been delivering through 2024 and 2025, with the smaller inference providers and the open-source model hosting ecosystem leveraging the secondary-market supply to deliver pricing structures that the leading-edge accelerator deployments cannot match. The compression has been approximately 47 per cent across the trailing 12 months ending Q1 2026, against the published API rate cards of the top six providers. The H100 secondary-market supply has been the principal infrastructure underpinning that compression, and the continued softening of the secondary-market pricing through 2026 will sustain the trajectory.

The implications for the procurement-team calculus at the broader enterprise customer base have been to extend the timeline for direct Blackwell procurement commitments, with many customers characterising the H100 and H200 secondary-market alternative as adequate against their current inference workload requirements. The procurement-team logic has been to defer the direct Blackwell commitment until the broader hyperscaler-and-neocloud channel has expanded the inference-as-a-service capacity at the leading-edge platform, with the cost-economics of the secondary-market alternative covering the inference workload requirements until that channel expansion clears the customer's commitment threshold. The deferred procurement pattern has been more pronounced at the broader enterprise customer base than at the leading-edge AI lab customer base, where the absolute performance envelope of the Blackwell platform has been the principal procurement variable.

What to watch

The Q1 2026 Blackwell deployment data settles a quarter of allocation questions and opens the questions that will define the H2 2026 and H1 2027 commercial trajectory. The named-customer figures across hyperscalers, neoclouds, and enterprise-direct customers will continue to set the structural pattern of the broader AI infrastructure market.

• Whether the Q2 2026 Blackwell shipment volume clears Nvidia's published guidance of 695,000 units; the company's guidance has consistently underpredicted the realised quarterly volume across the prior four quarters, but the CoWoS supply at TSMC has been the principal binding constraint on the upper end of the quarterly run rate, and any further supply-side expansion would lift the realised figure above the guidance threshold.
• Whether the xAI Colossus II deployment continues to ramp on its published 200,000-unit trajectory through the rest of 2026; the company's capex commitment is the most aggressive single-customer deployment outside the hyperscaler channel, and the deployment cadence through Q2 and Q3 will determine whether the enterprise-direct allocation proportion expands materially against the current 8-per-cent baseline.
• Whether the H100 and H200 secondary-market price softening continues at the Q1 cadence or stabilises at the current settled price points; the continued softening would sustain the inference-as-a-service per-token pricing compression that the broader public inference market has been delivering, while a stabilisation would mark the floor of the secondary-market dynamic and the structural transition to the Blackwell-generation infrastructure as the dominant deployment baseline.
• Whether the Oracle Cloud Infrastructure expansion against the OpenAI Project Stargate framework continues to drive the company's accelerator-deployment volume above the Q1 baseline; the Oracle slot has been the most strategically loaded of the four hyperscaler deployments in commercial terms, and the published commitments against the Stargate programme imply substantial expansion through H2 2026 and the broader 2027 deployment cycle.
• Whether the Rubin platform transition that Nvidia has scheduled for late 2026 produces an allocation-cycle reset comparable to the H100-to-Blackwell transition; the prior platform transition compressed the deployment of the prior generation across the Q4 of its commercial life as customers redirected procurement budgets to the new platform, and the corresponding pattern would be expected at the Q4 2026 and Q1 2027 transition window.

Frequently asked

How is the Q1 2026 Blackwell shipment figure of 612,000 units reconstructed, and what is the confidence interval?

The figure is reconstructed from the published Nvidia Q1 data-centre segment revenue, the average selling price across the Blackwell SKU mix derived from the published rack pricing and customer commentary, the hyperscaler capex disclosures cross-referenced against the proportion attributable to accelerator procurement, the neocloud commentary from public earnings and supply-chain analysis, and the supply-chain triangulation against the CoWoS shipment data at TSMC. The confidence interval on the headline figure is approximately plus-or-minus 4 per cent, with the principal sources of uncertainty being the timing differential between unit shipments and revenue recognition and the proportion of capex attributable to networking and storage versus accelerator procurement at the hyperscaler customer base.

Why did AWS take the largest single-customer allocation rather than Microsoft Azure, given that Microsoft committed substantially more AI-infrastructure capex in Q1 2026?

Microsoft's higher absolute AI-infrastructure capex includes substantial commitments to the Maia custom-silicon programme, the AMD MI300X deployments, and the broader datacentre construction envelope that does not directly convert to Nvidia accelerator procurement. The proportion of Microsoft's AI-infrastructure capex that flowed to Nvidia procurement at Q1 2026 was lower than the AWS equivalent on a percentage basis, even though the absolute capex figure was higher. The structural distinction reflects the different workload mix and the strategic posture toward custom silicon at the two hyperscalers — Microsoft has been more aggressive on the custom-silicon investment relative to its Nvidia procurement than AWS has been on the equivalent comparison.

What is the xAI Colossus II buildout, and why is it the most aggressive enterprise-direct Blackwell commitment?

Colossus II is the second of xAI's dedicated training facilities, complementing the original Colossus facility in Memphis, Tennessee. The Q1 2026 initial deployment of approximately 17,000 GB200 units anchors the company's commitment to ramp the facility to approximately 200,000 Blackwell units through 2026, principally against the next-generation Grok training programme. The aggressiveness of the commitment reflects the company's strategic posture that internal training compute is a competitive differentiator that the company cannot afford to outsource to hyperscaler infrastructure, combined with the company's most recent funding round at approximately $200 billion valuation that has provided the capex envelope to fund the deployment at the published trajectory.

How is the GB200 NVL72 rack-scale configuration different from the standard B200 deployment, and why does Oracle have a higher proportion of rack-scale deployments?

The GB200 NVL72 is a rack-scale configuration that pairs 36 Grace CPUs with 72 Blackwell GPUs in a single NVLink-connected rack assembly, delivering substantially higher inter-GPU bandwidth than the standard B200 deployment in 8-GPU server form factors. The rack-scale form factor is principally suited to the largest training workloads and the highest-throughput inference deployments. Oracle's higher proportion of rack-scale deployments reflects the company's strategic positioning against the OpenAI Project Stargate framework and the broader commitments to anchor customers that require the rack-scale infrastructure. AWS, Azure, and GCP have a more balanced mix between the rack-scale and the standard form factors, reflecting their broader customer-base distribution across training and inference workload classes.

What does the H100 and H200 secondary-market price softening mean for the broader AI infrastructure market?

The secondary-market price softening has been the principal driver of the inference-as-a-service per-token pricing compression that the public inference market has been delivering, with the smaller inference providers and the open-source model hosting ecosystem leveraging the secondary-market supply to deliver pricing structures that the leading-edge accelerator deployments cannot match. The continued softening sustains the per-token pricing compression trajectory, while a stabilisation would mark the floor of the secondary-market dynamic and the structural transition to the Blackwell generation as the dominant deployment baseline. The secondary market has also become the principal supply channel for the regional and second-tier cloud customer base, which has expanded the geographic distribution of accelerator capacity beyond the leading-edge customer base.

When is the Rubin platform transition expected, and how will it affect the Blackwell deployment cycle?

Nvidia has scheduled the Rubin platform for sampling in late 2026 and volume production in H1 2027. The transition will produce an allocation-cycle reset comparable to the H100-to-Blackwell transition, with the customer base redirecting procurement budgets to the new platform across Q4 2026 and Q1 2027. The Blackwell deployment cycle is expected to compress in absolute volume as the Rubin platform absorbs the leading-edge procurement commitments, and the H100 and H200 secondary-market dynamics will be complemented by an analogous Blackwell secondary-market dynamic through 2027 and 2028 as the platform enters the prior-generation lifecycle phase.