What the Published Blackwell Catalog Does—and Does Not—Show
On this page
June 26, 2026. SpaceX signed a reported $6.3 billion compute deal with Reflection AI for GB300 chips at Colossus 2. The June 29 public catalog snapshot below is a separate evidence set: it shows what providers published, not the terms, inventory, or availability of private transactions.
Correction and basis note (updated July 22, 2026): The coverage tables below reflect the June 29, 2026 harvest and are retained as a dated snapshot. Earlier copy inferred orderability, capacity, provider intent, and time pressure from catalog shape; those inferences have been removed. Public catalog visibility is not orderability, availability, a live quote, or a forecast of future supply. The July 6 harvest showed the same family-level publication pattern. Current weekly numbers live in the latest Weekly Pricing Pulse.
A reported private deal and a separate public catalog
On June 22, 2026, TechCrunch reported that SpaceX signed a compute agreement with Reflection AI at $150 million per month, with total payments of up to roughly $6.3 billion through 2029, for access to Nvidia GB300 chips at the Colossus 2 data center near Memphis.
No matching hourly listing appeared in the June 29 public catalog snapshot. That observation is limited to the published evidence: it does not establish the transaction's present status, the provider's inventory, or whether other commercial paths were orderable.
The reported deal demonstrates why public catalog data is only one input to a procurement process. Private agreements and published hourly listings are different commercial channels, and the existence of one does not let us quantify supply, demand, or negotiating leverage in the other.
Here is exactly what the dated public Blackwell and AMD Instinct snapshot can show—and where the evidence stops.
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 deployed inventory, orderability, or available capacity. CoreWeave's 23-geo value is a footprint projection applied to a published SKU, not evidence of per-region availability. Vultr "0 geo" entries mean "Market Unlisted": the SKU was published but did not map to a canonical on-demand market in our index.
Three patterns worth reading
OCI has the broadest published frontier-silicon catalog presence in this dated hyperscaler comparison. OCI lists 26 entries for six of the seven Evolving Market families in the table—B200, GB200, B300, GB300, MI300X, and MI355X. Those rows do not establish inventory, orderability, per-region availability, or OCI's commercial intent.
Azure has 20 published GB200 entries and 34 published MI300X entries in the snapshot. This is a catalog-count comparison only. It does not confirm current availability, inventory, or purchase terms in any listed market.
AWS, GCP, and several neoclouds have fewer published Blackwell rows in the snapshot. AWS publishes one B200 row and one B300 row. GCP publishes three B200 rows across five indexed markets. CoreWeave, Lambda, and Crusoe have limited or no published Blackwell rows in this snapshot. A low row count does not establish deployment level, orderability, provider intent, or future supply.
Context the catalog cannot prove
External reporting and system architecture can provide context for why a public catalog may differ from private procurement. They do not prove why any specific provider published—or did not publish—a row.
1. Reported private commitments
Reported long-term commitments show that some frontier compute is purchased through private contracts rather than published hourly catalogs. Public sources do not provide enough consistent detail to quantify what share of any provider's inventory follows each route.
Reflection's reported agreement is evidence about one named private transaction. It is not evidence about the providers in the coverage table, their unlisted inventory, or their current orderability.
CoreWeave's official Q1 2026 results report $99.4 billion in contracted revenue backlog and a $21 billion Meta commitment. These figures show that private contracting is material. They do not reveal which catalog listings are orderable, how much uncommitted inventory exists, or whether catalog rows are residual supply.
The practical consequence is an evidence requirement: teams should confirm current inventory, orderability, reservation windows, and quote terms directly with providers. The dated catalog alone cannot support timing or leverage forecasts.
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.
Azure's 20 GB200 rows across 17 indexed markets and CoreWeave's single footprint-projected GB200 row are not equivalent units of observation. The snapshot supports that data-model distinction; it does not disclose the providers' underlying inventory or commercial structure.
3. Sparse rows with unknown commercial purpose
The small AWS B200 (1 row / 1 indexed market), AWS B300 (1 / 1), Vultr B200 (1 / 0 Market Unlisted), and Nebius B300 (2 / 1) entries establish only that sparse catalog rows existed in the snapshot. The rows do not disclose whether they were intended for evaluation, production, reservation, or another commercial path. They should not be used as capacity indicators without provider confirmation.
How AIForge Works treats this in the data
The four-family workhorse comparison uses rows that meet a strict dated-public-pricing inclusion contract. Inclusion makes those rows comparable on the published basis; it does not certify current orderability or availability through a provider API. That comparison entered review after July 3 and is not used here as a current availability claim.
The Evolving Market tables show published entries as dated evidence. AIForge Works can report which rows a provider published and how they normalize; it does not infer market posture, unlisted inventory, or commercial purpose from presence or absence alone.
This is the boundary of public-data-first methodology: 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 a cluster pool, a private contract, or a not-yet-published roadmap. Inventory, reservation windows, orderability, live quotes, and confirmed delivery dates require direct provider confirmation.
What this means for procurement teams evaluating frontier silicon
Three actions, in order:
1. Use the public catalog to compare published evidence, not to estimate capacity or intent. The snapshot establishes the listed row counts and indexed-market mappings. It does not establish inventory, broad availability, deployment stage, or provider commitment.
2. Confirm commercial facts directly. Ask providers for current orderability, inventory, reservation windows, quote validity, and delivery dates. Do not substitute a catalog row—or its absence—for those answers.
3. Preserve the catalog-vs-deployment distinction in the evaluation rubric. Compare like-for-like published rows, then treat provider confirmation as a separate evidence lane. A footprint projection, sparse row, or missing row does not resolve a commercial question.
The reported private deals and the June 29 catalog snapshot are different evidence sets. Each can inform a procurement review, but neither should be used to infer facts the source does not publish.
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 published SKU-variant counts and canonical indexed-market presence. CoreWeave's 23-geo value is footprint-projected; Vultr "0 geo" means "Market Unlisted." Neither value establishes orderability or per-region availability. Private-deal context is supported by TechCrunch's Reflection AI report and CoreWeave's official Q1 2026 results. Private-deal facts and dated catalog observations are separate evidence sets.