GPUs, CPUs, NVLink and Spectrum-X networking, rack-scale systems and CUDA
NVIDIA’s AI ecosystem: what has shipped and what is still a plan?
Sources reviewed: · A selected research view, not a complete relationship inventory. Source dates are separate from review dates.
NVIDIA spans five layers of AI infrastructure: chips, networking, systems, software and cloud. It owns no fabs. TSMC makes its wafers, SK hynix, Micron and Samsung supply its memory, and builders such as Dell and Foxconn assemble its racks. Treating every connection as an “order” mixes up the strength of the evidence: an annual-report disclosure, a first delivery, racks running at a partner and a letter of intent are different things.
Key companies
Wafer foundry for NVIDIA, including Blackwell wafers in Arizona
Memory supplier; multiyear memory partnership announced for Vera Rubin
Memory supplier; HBM4 in volume production for Vera Rubin
Memory supplier and wafer foundry; HBM4 designed for Vera Rubin
Delivered the first GB300 NVL72 and Vera Rubin NVL72 racks to CoreWeave
Neocloud that brought up the first Vera Rubin NVL72 racks
Azure named by NVIDIA as running Vera Rubin racks
Model lab with up to 1 GW of planned Grace Blackwell and Vera Rubin capacity
NVIDIA’s position across five layers
| Layer | NVIDIA’s position | Source |
|---|---|---|
| Chips and memory See connections → | NVIDIA is fabless. It designs GPUs and Arm-based CPUs (Grace, now Vera) and relies on foundries, memory suppliers and contract manufacturers for every manufacturing step. | |
| Networking See connections → | NVLink interconnects and switches, and end-to-end InfiniBand and Ethernet platforms: network adapters, cables, DPUs, switch chips and systems, plus software. NVIDIA reports networking inside its Compute & Networking segment, whose revenue grew 18% from the prior quarter and 114% from a year earlier in Q2 FY2027. | |
| Systems See connections → | Data center products are typically delivered as rack-scale systems, subsystems or modules. OEMs and ODMs such as Dell, HPE, Supermicro, Foxconn and Wistron build the racks customers receive. | |
| Software / CUDA See connections → | CUDA runs on all NVIDIA GPUs and sits beneath hundreds of domain-specific libraries, frameworks and SDKs. NVIDIA describes its developer base and installed base as strengthening the ecosystem. | |
| Cloud and customers See connections → | Data center revenue was $89.0 billion in Q2 FY2027 (quarter ended July 26, 2026), up 117% from a year earlier. Most GPUs reach end users through hyperscalers, neoclouds and model labs, often indirectly via OEMs and ODMs. |
Product roadmap: shipped vs announced
Shipped / in use
Shipping in volume
Blackwell reached volume production; NVIDIA marked the first Blackwell wafer produced at TSMC Arizona.
Shipped / in use
Shipping; NVIDIA cites its ramp for Q2 FY2027 growth
Announced at GTC in March 2025 for the second half of 2025. The first GB300 NVL72 reached CoreWeave in July 2025, and NVIDIA attributes Q2 FY2027 data center growth to the Blackwell Ultra ramp.
Announced / planned
Ramping; first racks delivered, broad shipments from fall 2026
Six chips (Vera CPU, Rubin GPU, NVLink 6 Switch, ConnectX-9, BlueField-4, Spectrum-6) announced in January 2026. NVIDIA said production shipments begin in fall 2026. Early racks are already running at CoreWeave and several clouds.
Suppliers, customers and partners
| Connection / status | Sources and limits | Source and next step |
|---|
No curated connections match these filters.
What should you verify next?
The next Vera Rubin checkpoint is production shipments from fall 2026: watch for cloud instance launches, system-builder delivery announcements and Rubin’s revenue contribution in NVIDIA’s quarterly results. Letters of intent and “up to” commitments need a later purchase agreement or deployment announcement. This page preserves the scope of the cited evidence; it does not assume every announced plan has been fulfilled.