Why the Blackwell Public Catalog Runs Thin (And Why It's About to Get Thinner!)
June 26, 2026. Last Sunday, SpaceX signed a $6.3 billion compute deal with Reflection AI for GB300 chips at Colossus 2 — and none of that capacity will ever appear in a public price catalog. The published NVIDIA Blackwell and AMD Instinct catalog you can see across the nine major clouds runs thin for three structural reasons. Here's what that means for any team planning a frontier-silicon procurement cycle in the next 12 months.
The trade that wasn't in any catalog
On Sunday, June 22, 2026, SpaceX signed a multi-billion-dollar compute deal with open-source AI startup Reflection AI: $150 million per month starting July 1, payments totaling roughly $6.3 billion through 2029, for access to Nvidia GB300 chips at the Colossus 2 data center near Memphis. The contract carries a 90-day exit clause after an initial three-month lock-in. Reflection — funded in part by Nvidia itself ($800 million in October 2025 as part of a $2 billion round at an $8 billion valuation) — is the latest in a string of anchor tenants at Colossus that already includes Anthropic, Google, and Cursor. Committed revenue at Colossus from outside customers now reportedly exceeds $80 billion through 2029.
None of that GB300 capacity will appear in a public on-demand pricing catalog. It does not have a published hourly rate. It is not listed in any provider's REST API or pricing portal under a public SKU. The deployment is real, the silicon is being installed, and the customer is paying — but the entire transaction is structurally invisible to the public-catalog tracking that powers most cross-cloud pricing comparisons.
The Reflection deal is the largest recent example of a pattern that is about to get more visible — and more painful for procurement teams that haven't started the right conversations. Frontier model training in 2026 is anchored on GB300-class accelerators, and the slots available on the open market for that class of silicon are shrinking as the largest AI buyers lock in multi-year capacity. As July 2026 begins and the next wave of anchor commitments hits the books, procurement leverage on frontier silicon RFQs is likely to tighten faster than the public catalog will tell you.
This is the structural reality the public Blackwell and AMD Instinct catalog reflects. Here is what AIForge Works can show you about it.
What the June 29 published catalog actually shows
AIForge Works tracks the NVIDIA Blackwell line (B200, GB200, B300, GB300) and the AMD Instinct line (MI300X, MI325X, MI355X) across the same nine-provider catalog that powers the workhorse-family floor analysis. Coverage as of the June 29 harvest, in distinct catalog SKU configurations and canonical AIForge Works market presence (MapIt taxonomy):
NVIDIA Blackwell family
| Provider | B200 | GB200 | B300 | GB300 |
|---|---|---|---|---|
| SKUs / Geos | SKUs / Geos | SKUs / Geos | SKUs / Geos | |
| AWS | 1 / 1 | — / — | 1 / 1 | — / — |
| Azure | — / — | 20 / 17 | — / — | — / — |
| GCP | 3 / 5 | — / — | — / — | — / — |
| OCI | 26 / 26 | 26 / 26 | 26 / 26 | 26 / 26 |
| CoreWeave | 1 / 23 | 1 / 23 | — / — | — / — |
| Lambda | 4 / 14 | — / — | — / — | — / — |
| Vultr | 1 / 0 | — / — | — / — | — / — |
| Nebius | 4 / 2 | — / — | 2 / 1 | — / — |
| Crusoe | — / — | — / — | — / — | — / — |
AMD Instinct family
| Provider | MI300X | MI325X | MI355X |
|---|---|---|---|
| SKUs / Geos | SKUs / Geos | SKUs / Geos | |
| AWS | — / — | — / — | — / — |
| Azure | 34 / 17 | — / — | — / — |
| GCP | — / — | — / — | — / — |
| OCI | 26 / 26 | — / — | 26 / 26 |
| CoreWeave | — / — | — / — | — / — |
| Lambda | — / — | — / — | — / — |
| Vultr | 1 / 0 | 1 / 0 | 1 / 0 |
| Nebius | — / — | — / — | — / — |
| Crusoe | 1 / 1 | — / — | — / — |
Coverage tables show published catalog presence only — they do not represent actual deployed GPU inventory, which is not published by providers. CoreWeave's 23-geo footprint reflects uniform global availability per published SKU rather than per-region catalog entries. Vultr "0 geo" entries indicate "Market Unlisted" — the SKU is published but does not yet map to a canonical AIForge Works market.
Three patterns worth reading
OCI's published catalog is the broadest declared frontier-silicon availability among any hyperscaler. OCI lists 26 SKUs across all 26 regions for every Evolving Market family in the table — B200, GB200, B300, GB300 on the NVIDIA side, MI300X and MI355X on the AMD side. Whether that translates to actual deployed capacity in each region is a separate question — and one the published catalog does not answer — but the catalog signal itself is real: OCI is positioning aggressively across both vendors' frontier stacks, more publicly than any other hyperscaler.
Azure leads where production-scale deployment has reached the catalog. 20 GB200 SKUs across 17 markets and 34 MI300X SKUs across 17 markets confirm both the Azure-NVIDIA NVL72 rollout and the long-standing Azure-AMD partnership in published form. Azure is the only hyperscaler with GB200 published at scale outside OCI's 26-region listing, and the only one with MI300X at meaningful published catalog scale.
AWS, GCP, and most neoclouds show thin or absent published Blackwell catalog. AWS publishes 1 B200 SKU and 1 B300 SKU. GCP publishes 3 B200 SKUs across 5 markets. CoreWeave, Lambda, and Crusoe — all of which have documented Blackwell deployment activity in industry reporting — show limited or no published Blackwell rows in the catalog as of the harvest. The pattern matters because it tells you something important about the difference between what providers are deploying and what they are publishing prices for.
Why the catalog runs thin: three structural reasons
The catalog-vs-deployment distinction is not a measurement gap or a data limitation. It is the structural shape of how frontier GPU capacity actually moves into the market. Three reasons drive the pattern, and each one is procurement intelligence in its own right.
1. Pre-sold anchor capacity (and the open-market tightening that follows)
A substantial share of the initial Blackwell deployment on the hyperscalers and at scaled facility operators is sold to anchor tenants under long-term contracts — frontier model labs, large enterprise AI organizations, and the hyperscalers' own internal AI workloads. That capacity never enters the public on-demand pool and therefore never gets a published catalog SKU. The deployment is real; the public price is not.
The SpaceX–Reflection deal is the largest recent and most explicit example, but it is not an outlier — it is the most public instance of an industry-wide pattern. Anthropic, Google, and Cursor each separately signed for capacity at Colossus 2 ahead of Reflection; the facility's reported $80B+ in committed outside-customer revenue through 2029 represents an enormous block of GB300 (and other Nvidia) capacity that will be running production workloads and will never have a published hourly rate. The same pattern, at varying scale, runs through Azure's GB200 capacity (much of it pre-sold to OpenAI and other anchor tenants), AWS's Blackwell deployment (largely pre-sold to anchor enterprise contracts and AWS's own AI services), and GCP's Blackwell rollout (substantially backing Google's own AI workloads before any open public pricing). The provider list in the coverage tables above is a list of who publishes catalog SKUs publicly — not a list of who has Blackwell capacity.
The pattern extends beyond hyperscalers. CoreWeave — the largest publicly-traded neocloud, whose entire business model is renting cataloged GPU capacity — reported a $99.4 billion contracted revenue backlog at Q1 2026, with more than 75% of 2027's expected $30 billion+ annualized run-rate already committed under long-term contracts. Even the operator whose catalog you can most easily see has effectively pre-sold most of what it will deploy over the next 24 to 48 months. Separately, OpenAI's Stargate joint venture with Oracle represents another block of frontier-class capacity — 4.5 gigawatts across multiple US sites, over two million chips, roughly $30 billion per year to Oracle starting in 2028 — none of it structured for public on-demand pricing. NVIDIA itself has reportedly sold out Blackwell production allocations through 2027 to a small set of very large buyers. Across the entire frontier stack, what appears in the public catalog is the residual after enterprise allocation, not the primary market.
The procurement consequence is timing-sensitive. As July 2026 begins and Reflection's monthly $150M commitments start flowing alongside the existing Anthropic, Google, and Cursor anchor positions, the slots available on the open market for GB300-class accelerators are shrinking. Frontier model training in 2026 is structurally anchored on GB300, and the pre-sold pattern is moving fast enough that procurement leverage on open-market frontier-silicon RFQs is likely to tighten meaningfully over the next 60-90 days. Teams planning their next deployment cycle on the assumption that GB300 capacity will be available through the public catalog in Q4 2026 are likely to be wrong. The right conversations — direct provider, anchor-tenant-class — need to happen sooner than the public catalog will signal.
2. Rack-scale NVL72 architecture doesn't fit single-instance catalog patterns
GB200 and GB300 are designed around dense, liquid-cooled, megawatt-scale rack systems — the NVL72 architecture. They are deployed as monolithic, interconnected supercomputers, managed as cluster-pool infrastructure for specific enterprise workloads. The single-instance public on-demand SKU pattern that works for H100 or L40S does not naturally apply to NVL72 rack-scale capacity.
The physical shape of the hardware is a large part of why. A single GB200 NVL72 rack integrates 72 Blackwell GPUs into one high-bandwidth NVLink domain, draws roughly 120 kilowatts, and requires purpose-built coolant distribution units, custom liquid manifolds, and megawatt-class power provisioning per row of racks. The 72 GPUs are wired together at the memory and interconnect layer — they are designed to run as one logical system, not as 72 independent units. Fragmenting an NVL72 into 72 separately-billed hourly on-demand SKUs would break the NVLink domain that justifies the architecture in the first place, and would surrender the workload-shaped scaling that makes the design economically defensible for the buyers who need it. This is architecturally different from H100 or H200 SXM systems, where an 8-GPU node can meaningfully stand alone as a saleable unit — the SXM systems were built around 8-way scaling, and the on-demand catalog model tracks that. NVL72 was built around 72-way scaling, and the on-demand catalog model does not.
That is why Azure's 20 GB200 SKUs across 17 markets reads as a meaningful catalog signal but CoreWeave's footprint-projected 1 GB200 SKU across 23 markets reads differently — both providers have GB200 in the catalog, but the underlying deployment architecture is being expressed through different SKU patterns. Procurement teams reading the catalog for GB200 need to recognize that the unit of comparison is not always the same.
3. Sandbox and proof-of-concept rows
The small AWS B200 (1 SKU / 1 market), AWS B300 (1 / 1), Vultr B200 (1 / 0 Market Unlisted), and Nebius B300 (2 / 1) entries in the table are best read as developer-seeding nodes, not procurement inventory. Their purpose is to let early partners test CUDA compiler paths, validate framework performance, and adjust API orchestration wrappers on Blackwell silicon before at-scale capacity becomes buyable at all. A single-node SKU cannot support a training cluster; it is not designed to. Providers seed these sandbox nodes deliberately so that the ecosystem is ready when production capacity is eventually released — or, more likely, when it is offered under commitment structures that never touch the public catalog in the first place. The same single-SKU pattern appears on Vultr's AMD Instinct rows (1 / 0 across MI300X, MI325X, MI355X): public catalog presence at sandbox scale, useful for early evaluation and framework compatibility work, not for capacity planning. A procurement team that reads these entries as capacity indicators will materially misjudge the deployment landscape.
How AIForge Works treats this in the data
Catalog presence, cluster-pool deployment, and sandbox seeding are tracked separately. The four-family workhorse aggregate — the 49.3% floor delta from the companion post — uses only fully-cataloged, single-instance on-demand rows under the strict inclusion set, exactly the rows that have stable public pricing and represent capacity that procurement teams can actually buy through standard catalog APIs.
The Evolving Market tables above show all published catalog entries because surface presence itself is procurement intelligence — whether a provider has chosen to publish a public catalog SKU for a frontier-silicon family tells you something real about that provider's market posture, even if it doesn't tell you the full deployment story. Sandbox rows are flagged in MapIt and Cloud Advisor to prevent them from skewing capacity-aware comparisons. Pre-allocated and cluster-pool capacity is not tracked in the catalog at all — by definition, AIForge Works can only see what providers choose to publish.
This is the boundary of public-data-first methodology, stated honestly: AIForge Works tells you what is in the published catalog, exactly and consistently across nine providers, refreshed weekly. It does not tell you what is in the cluster-pool, the anchor-tenant contracts, or the not-yet-published roadmap. For specifics on exact SKU counts, inventory versus purchase-order sourcing, reservation windows, or confirmed delivery dates on frontier silicon, you need direct account-level confirmation from the relevant NVIDIA and provider teams — those details are not what the public catalog is built to surface, and they are exactly what an active procurement cycle on Blackwell capacity needs. The public catalog gets you to the right shortlist of conversations to have. It does not have those conversations for you.
What this means for procurement teams evaluating frontier silicon
Three actions, in order:
1. Use the public catalog to identify who is positioned, not to estimate capacity. OCI's 26/26 presence across the Evolving Market families tells you OCI is publicly committing to broad Blackwell + AMD availability. Azure's 20 GB200 SKUs and 34 MI300X SKUs tell you Azure is operationalizing both stacks in the published catalog at scale. AWS's 1 B200 and 1 B300 SKUs tell you AWS is seeding the ecosystem but not yet publishing meaningful Blackwell at-scale catalog. These are real signals about provider market posture, and they belong in the procurement evaluation early.
2. Start the direct-provider conversation earlier than your current procurement cycle assumes — meaningfully earlier. The single most expensive mistake on frontier silicon procurement in 2026 is assuming the public catalog represents the market you can buy from. The SpaceX–Reflection deal, the Anthropic and Google commitments at Colossus, the OpenAI anchor positions on Azure GB200 — all of these were structured months before the silicon was installed. As of late June 2026, capacity for the next deployment cycle is being committed weekly in private conversations that never touch a catalog, and the pace is accelerating. If your procurement timeline assumes public-catalog availability for GB300 twelve months out, it is likely already wrong. If it assumes leverage in an RFQ to be issued in August or September, it is more likely wrong than right.
3. Build the catalog-vs-deployment distinction into your evaluation rubric. A provider with published B200 SKUs at scale (Azure, OCI) is offering a different commercial structure than a provider with sandbox-only B200 SKUs (AWS, Vultr) or a provider with documented deployment but no published B200 SKUs at all (CoreWeave outside their footprint-projected single entry, Crusoe, Lambda). Treating them as comparable on the catalog signal alone will produce the wrong vendor shortlist. AIForge Works MPA and Cloud Advisor surface both the published catalog and the workload-fit signals; the public-catalog read is the starting point, not the answer.
The frontier-silicon market is shifting fast — faster than the public catalog can capture, and faster than any single-provider conversation will tell you. Reflection's $150 million monthly check to SpaceX is one data point. The $80 billion in already-committed Colossus revenue is another. The structural pattern they reflect — anchor-tenant pre-sold capacity, NVL72 cluster-pool architecture, sandbox-tier catalog seeding — is the third. Each tells you something the other doesn't, and none of them are visible from a single SKU lookup.
Data sourced from the AIForge Works June 29, 2026 cross-cloud normalization across 9 providers: AWS, Azure, GCP, OCI, CoreWeave, Lambda, Vultr, Nebius, Crusoe. Coverage tables show SKU variant counts (distinct catalog instance configurations) and canonical AIForge Works market presence per the MapIt taxonomy. CoreWeave's 23-geo count is footprint-projected uniformly per published SKU; Vultr "0 geo" entries indicate "Market Unlisted" status pending canonical market mapping. NVIDIA Blackwell line covered: B200, GB200, B300, GB300. AMD Instinct line covered: MI300X, MI325X, MI355X. These families are tracked in the AIForge Works dataset for catalog availability and market-structure observation; they are not part of the four-family workhorse floor aggregate (49.3% as of June 29, 2026) covered in the companion post, and per-SKU floor comparison on these families requires a deeper inclusion-set treatment than this post covers. Refreshed weekly; externally validated weekly against each provider's own published infrastructure pages. The SpaceX–Reflection deal, related Colossus 2 anchor-tenant information (Anthropic, Google, Cursor), and the $80B+ committed-revenue figure are sourced from public reporting by CNBC, TechCrunch, Data Center Dynamics, MLQ News, and Cryptopolitan published June 22–24, 2026; deal terms cited from those reports. CoreWeave's $99.4 billion contracted revenue backlog and the "75%+ of 2027 ARR already committed" figure are sourced from CoreWeave's Q1 2026 earnings report (May 7, 2026) and related coverage by CNBC and The Globe and Mail. OpenAI–Oracle Stargate figures (4.5 gigawatts, ~2 million chips, $30 billion per year starting 2028) are sourced from OpenAI's Stargate announcement, Data Center Dynamics, and Reuters reporting. NVIDIA's Blackwell allocation posture through 2027 is referenced from industry commentary; not a formal NVIDIA statement. Capacity-tightening interpretation reflects the structural pattern observable in published catalog plus the documented anchor-tenant pre-sold trend; specific timing claims are framed as procurement-cycle implications, not as predictions about specific provider behavior.