# Inference Stack — Taxonomy Draft (Aug 2026)

> Update, later Aug 2026: market share is now explicitly deferred (see README
> roadmap) — the near-term focus is the technical audience and mapping where this
> hardware is actually deployed, starting with AWS (see `deployments.aws` in the
> JSON and the "Where it runs" section of the page). Two products were added below
> (L4/L40S, A100/A10G) after the AWS mapping surfaced that several EC2 GPU families
> run on an inference/graphics tier that wasn't in the original 31.
>
> Update, Aug 21 2026: the AWS section is now an actual interconnected graph (product
> nodes ↔ instance-family nodes, tap either side to trace the connection) instead of a
> table-only view — see "Graph interaction model" below. Every product and deployment
> entry also got a `url` field ("Learn more" links). One naming correction: the Groq
> inference chip's official NVIDIA product name is **"Groq 3 LPX"**, not "Groq 3 LPU" —
> "LPU" is Groq's original/industry-generic term and is kept only as the `silicon_type`
> category label; the product's own `name` field now says LPX, per NVIDIA's own product
> page.

Family-level map of NVIDIA's product portfolio, organized by the two lenses that matter most: how NVIDIA itself reports the business (segment), and what kind of silicon/software each product actually is (type). This is the data behind the interactive layout — `data/products.json` has the same content in structured form, with a `market_share_pct` / `revenue_usd` field on every entry left as a placeholder for phase two.

## By business segment

### Data Center — 91.5% of FY26 Q4 revenue ($62.3B)
The core of the business, and the most structurally complex — NVIDIA now ships five distinct kinds of silicon here, not just GPUs.

| Product | Type | Generation | Status |
|---|---|---|---|
| Hopper (H100/H200) | GPU | Hopper | Legacy, still shipping |
| Blackwell (B200/B300/GB300 Ultra) | GPU | Blackwell | Shipping, high demand through 2027 |
| Vera Rubin (Rubin GPU) | GPU | Rubin | Launching H2 2026 |
| Rubin Ultra | GPU | Rubin Ultra | Roadmap, 2027 |
| Feynman | GPU | Feynman | Roadmap, 2028 |
| L4 / L40S (Ada Lovelace) | GPU | Ada Lovelace | Shipping — added after AWS mapping (G6/G6e) |
| A100 / A10G (Ampere) | GPU | Ampere | Legacy, still widely deployed — added after AWS mapping (P4/G5) |
| Grace CPU | CPU | Grace | Shipping |
| Vera CPU | CPU | Vera | Launching H2 2026 |
| BlueField DPU | DPU | BlueField-3 / -4 | Shipping / upcoming |
| Groq 3 LPX | LPU | LPU gen 1 | Launched GTC 2026 — non-GPU inference chip (NVIDIA's product name is "LPX"; "LPU" is the category) |
| NVLink / NVLink Switch | Networking | Gen 5–7 | Shipping / roadmap |
| ConnectX SuperNIC | Networking | ConnectX-8/9 | Shipping / roadmap |
| Spectrum-X Ethernet Switch | Networking | Spectrum-6/7 | Shipping / roadmap |
| Quantum InfiniBand Switch | Networking | Quantum-X800 | Shipping |
| DGX Systems | System | B200/GB200 | Shipping |
| HGX Baseboard | System | B200 | Shipping |
| GB200/GB300 NVL72 Rack | System | Blackwell gen | Shipping |
| Kyber NVL144 Rack | System | Rubin Ultra gen | Roadmap, 2027 |
| CUDA / cuDNN / TensorRT | Software | — | Shipping — the core moat |
| NIM Microservices / NeMo | Software | — | Shipping |
| NVIDIA AI Enterprise | Software | — | Shipping |
| DGX Cloud | Software | — | Shipping |
| Omniverse | Software | — | Shipping — also feeds Pro Viz & Robotics |

### Gaming — 5.4% of revenue ($3.7B)
| Product | Type | Generation | Status |
|---|---|---|---|
| GeForce RTX 50-series | GPU | Blackwell | Shipping, from $299 |
| DLSS | Software | — | Shipping |
| GeForce NOW | Software | — | Shipping — cloud gaming |

### Professional Visualization — 1.9% of revenue ($1.32B), fastest-growing segment
| Product | Type | Generation | Status |
|---|---|---|---|
| RTX PRO 6000 Blackwell (Workstation/Server/Max-Q) | GPU | Blackwell | Shipping |

### Automotive & Robotics — 0.9% of revenue ($604M)
| Product | Type | Generation | Status |
|---|---|---|---|
| DRIVE Thor | GPU (SoC) | Thor | Shipping |
| DRIVE Hyperion | System | — | Shipping — reference architecture |
| Jetson (Thor/Orin) | GPU | Thor / Orin | Shipping |
| Isaac (incl. GR00T) | Software | — | Shipping — robotics sim + foundation models |

### OEM & Other — 0.2% of revenue ($161M)
Smallest segment; legacy/misc silicon sold through OEM channel.

## By silicon type (the "not just GPUs" view)

- **GPU** (12): Hopper, Blackwell, Rubin, Rubin Ultra, Feynman, L4/L40S (Ada), A100/A10G (Ampere), GeForce RTX, RTX PRO, DRIVE Thor, Jetson, OEM legacy
- **CPU** (2): Grace, Vera
- **DPU** (1): BlueField
- **LPU** (1): Groq 3 LPX — the newest category, non-GPU inference silicon from the Groq deal
- **Networking** (4): NVLink, ConnectX, Spectrum-X, Quantum InfiniBand
- **System** (5): DGX, HGX, GB200/GB300 NVL72, Kyber NVL144, DRIVE Hyperion
- **Software/Platform** (8): CUDA stack, NIM/NeMo, AI Enterprise, DGX Cloud, Omniverse, DLSS, GeForce NOW, Isaac

## Deployments — AWS (new)

A separate top-level `deployments.aws` object in the JSON, not folded into `products`,
because a deployment is a relationship (product × cloud instance family) rather than a
product itself. Three parts: `instance_families` (11 EC2 GPU families mapped to the
NVIDIA silicon inside them and the matching product id), `platform_integrations` (5
software/networking integration points — DGX Cloud, NIM, Spectrum-X, NVLink Fusion,
cuVS), and `not_nvidia` (2 entries — AWS Nitro System, Trainium/Inferentia — kept short
and factual, not a competitive dig, just so the boundary is explicit). Every entry across
all three now carries a `url` field pointing at its official product page. Sourced from
AWS's own instance-type pages and recent AWS/NVIDIA blog posts as of Aug 2026; GPU
generations move fast enough that this data should be spot-checked before relying on it
for anything beyond a rough picture.

### Graph interaction model

The "Where it runs" section renders `instance_families` as a bipartite graph: NVIDIA
products on the left (one node per distinct `product_id` referenced by an instance
family, in `products` array order), AWS instance families on the right (one node per
`instance_families` entry), with an SVG line per instance family connecting it to its
product. Selection is intentionally asymmetric, matching how the relationship actually
works:

- Tapping a **product** node highlights every instance family whose `product_id` matches
  it (one-to-many — e.g. L4/L40S maps to both G6 and G6e).
- Tapping an **instance family** node highlights only that node, its one connected
  product, and the one line between them (one-to-one from the instance's side — each
  `instance_families` entry has exactly one `product_id`).

Both node columns are HTML `<button>`s (not pure SVG) so long labels like
"P6e-GB200 UltraServer" wrap normally; only the connecting lines are SVG, redrawn on
resize and on switching back from the table view. A plain `<table>` view is one tap away
(the `Graph` / `Table` toggle) for accessibility, printing, and anyone who just wants the
raw list. This is written specifically against `deployments.aws` right now — extending it
to a second provider means generalizing the renderer to take a `deployments.<provider>`
object as a parameter rather than reading `AWS` directly (see README roadmap).

## What's deliberately left out / simplified

- This is family-level, not SKU-level (no individual GPU model numbers like RTX 5070 Ti within GeForce).
- Software products that span multiple segments (Omniverse, CUDA) are tagged to Data Center for now since that's their primary reporting home — worth revisiting once we design the actual visual, since a strict single-segment tag undersells how much they cross-cut.
- No market share or revenue-per-product numbers yet — segment-level revenue is included only for scale/context.

## Open questions for the next pass

1. Should Omniverse/CUDA get a "cross-segment" tag instead of being forced into one segment?
2. Do you want SKU-level detail eventually (e.g., every GeForce RTX model), or is family-level the right permanent altitude?
3. ~~Market-share framing~~ — deferred; see README roadmap.
4. Which cloud/target is next for `deployments`? GCP and Azure would follow the same pattern (their own GPU instance families); on-prem/OEM server vendors (Dell, HPE, Supermicro building on HGX) are a different kind of "embedding" worth a separate shape.
5. ~~At what point does a table per deployment target stop being enough~~ — resolved
   sooner than expected: AWS itself became a graph (product ↔ instance family), not
   because a second cloud showed up yet. Worth revisiting once a second provider exists —
   would it be one graph per provider, or one combined graph with a provider filter?