US AI Company Product Matrix · Updated 11-18

NVIDIA

NVIDIA's play is"Full-stack compute toll": chips (Blackwell) + closed software stack (CUDA) + networking (NVLink / InfiniBand) in a triple lock-in — from training to inference, every layer of AI applications pays it a "compute tax." Data center revenue over 90%, gross margin 75%+ — it's the foundation of the entire matrix. Hugging Face ("the GitHub of AI") adds the open-source ecosystem piece: 3M+ models, 18M+ developers.

Q3 FY2027 earnings smash expectations: revenue $115.84B (+103% YoY), Q4 guidance $130B; Vera Rubin NVL72 shipping; chairman says the board has no intention of bringing Jensen Huang back (11-17)

Key numbersAt a glance

$5T
Market cap (broke through Oct 2025, first in the world, widely reported)
$96.2B
Q2 FY2027 revenue (earnings, 2026-08-26)
$89B
Data center revenue, Q2 FY2027 (~90% of total revenue)
75.0%
Gross margin (GAAP, Q2 FY2027)
$115.84B
Q3 FY2027 revenue (earnings, 2026-11-17, +103% YoY, ~7.2% above the $108B guidance midpoint); data center revenue $107.6B (+110% YoY); adjusted EPS $2.68 (expected $2.47–2.48); GAAP net income $65.3B (+105% YoY); gross margin 74.8%
$130B
Q4 FY2027 revenue guidance (±3%, includes an extra week), gross margin guidance 71–72%; FY2028 revenue growth outlook raised to ~75% (70% in August); Q4 guidance includes no China data-center compute revenue

Product layersAll entries

1 · Chips

Blackwell sold out; Rubin in mass-production shipments; Ultra (Kyber) arrives 2027
Blackwell B200 / B100Launched
Current flagship AI chip; the earnings call called sales "off the charts" — cloud GPUs sold out
GB200 NVL72 rack unit ~$3M (analyst estimate)
Hopper H100 / H200Launched
Last-gen workhorse; still shipping in volume for inference
H200 with HBM3e
RubinLaunched
Mass-production shipments in 2026; ~5× Blackwell's inference performance, ~10× lower token cost
CoreWeave opened its Vera Rubin NVL72 platform 10-01 with Cognition as the first production customer; Modal adopted the Vera CPU 10-19 (first publicly named third-party customer; early tests doubled throughput on comparable tasks); NVIDIA guides at least $20B in Vera chip sales this fiscal year; full mass production announced at GTC Berlin 10-21, supply chain locked (substrate capacity to grow 2.5× in 3 years, $1.1B KYEC test booking); Nov 17 earnings call: Vera Rubin NVL72 has begun shipping, driving nearly $20B in incremental revenue
RTX Spark (N1 / N1X)Launched
Local AI PC chip; RTX Spark PC lineup announced at Microsoft's Windows / Surface event Oct 7 (San Francisco; Satya Nadella, Jensen Huang, Pavan Davuluri on stage) with pre-orders open — theme: "local AI reshapes the PC"
Surface Laptop Ultra first to ship with N1X (Arm CPU + Blackwell RTX GPU, up to 128GB unified memory, 1 petaflop of AI compute), from $2,599, shipping Oct 16; Dell, HP, ASUS, MSI, Lenovo models in pre-order; RTX Spark Dev Box ships to the US in November; MarketWatch, wccftech cross-reported
Nemotron 3.5 Bolt 74B-A9BLaunched
Launched Oct 15: 74B-parameter MoE with 9B active, 2-bit quantized to 24GB, positioned as agentic AI on RTX Spark laptops; available at launch via llama.cpp / vLLM / Ollama / Unsloth / LM Studio
NVIDIA commits to fully open (training data, weights, recipes); announced Oct 13, covered by TechTimes, blockchain.news, alfw.ai and others Oct 13–14; downloadable from Oct 16 via Hugging Face, NVIDIA NIM and ModelScope under the OpenMDW-1.1 commercial license
Nemotron 3.5 Super VL 120BLaunched
Launched Oct 15: flagship multimodal model of the family, 120B total parameters / 12B active, 1M-token context, for always-on data-center agents
Downloadable from Oct 16: Hugging Face, NVIDIA NIM and ModelScope, open weights + datasets + training recipes under the OpenMDW-1.1 license (commercial use); covered by letsdatascience, blockchain.news, Unite.AI and others; the smaller Bolt E4B (4GB lightweight chat model) launched Oct 23 — see below
Nemotron 3.5 Bolt E4BLaunched
Launched Oct 23: open small language model, ~4GB VRAM, 64K-token context, 35 languages and multi-turn tool use, for local agents on RTX / Jetson / Spark
NVIDIA Robotics announced it on LinkedIn ("Available now"); on Hugging Face under the OpenMDW license

2 · Software stack

Closed ecosystem — a decade-deep moat
CUDALaunched
Parallel computing platform; the root of developer mindshare
Launched 2006; nearly 20 years of ecosystem accumulation
NIM microservicesLaunched
Pre-built AI model inference microservices; one-click deployment
Connects enterprise on-prem and cloud
NeMo / TensorRTLaunched
LLM training and inference optimization framework
Inference latency and cost optimization
PAIR (Personal AI Router)Launched
Personal AI router; left beta Oct 13 — free and open-source, with added llama.cpp support and direct downloads from Hugging Face
The routing layer for local model inference

3 · Systems & networking

Servers + networking, both covered
DGX / HGXLaunched
AI server systems and board solutions
The first DGX-1 (2016) went to OpenAI
NVLinkLaunched
High-speed GPU interconnect; proprietary protocol
The core of in-rack bandwidth
InfiniBand / Spectrum-XLaunched
Data center networking; InfiniBand came via the Mellanox acquisition
Mellanox acquisition completed 2020 for ~$6.9B
DGX Spark 64GBLaunched
New configuration launched Oct 2: same GB10 Grace Blackwell Superchip, unified memory halved to 64GB; claimed to run 100B-parameter models on one box, 200B with two linked (128GB pooled), with Sync Cluster Assistant auto-networking
From $4,999, on sale Oct 23 via Acer, ASUS, Dell, Gigabyte, HP and MSI; the 128GB version's official price rose to $6,950 the same day (The Register, citing memory price hikes)

4 · Enterprise & cloud

Reaching into every enterprise
DGX CloudDiscontinued
Formerly a cloud compute rental service; stopped accepting new customers end-2025, now an internal R&D platform
AI EnterpriseLaunched
Enterprise AI software suite
Commercial licenses incl. NIM and NeMo

5 · Emerging businesses

Small share; narrative-building
OmniverseLaunched
Industrial digital-twin and simulation platform
Visualization revenue $760M in FY2026 Q3 (+56%; folded into Edge Computing from Q2 FY2027)
DRIVELaunched
Autonomous driving platform
Automotive & robotics revenue $592M in FY2026 Q3 (+32%; folded into Edge Computing from Q2 FY2027)
Jetson / IsaacLaunched
Edge AI and robotics development platform
A beneficiary of the humanoid-robot boom

6 · Open-source ecosystem

Hugging Face: the GitHub of AI
Hugging Face HubPress Coverage
3M+ models, 500K datasets, 1M apps, 18M+ developers, 200K+ companies
Annualized revenue broke $100M in June 2026, ~$150M in August; 97% of users free
Transformers / DiffusersLaunched
Open model library; the root of developer mindshare
The transformers library is the de facto standard for open-model distribution
Inference Providers / EndpointsLaunched
Aggregated third-party inference API plus one-click dedicated endpoints
Pay-per-call; enterprise SLA
Enterprise HubLaunched
Private model hosting, SSO, security & compliance
2,000+ paying enterprises (press)
Nemotron Personas UKPress Coverage
Launched Oct 14: 1M-row synthetic UK-population persona dataset from Locai Labs in collaboration with NVIDIA, CC BY 4.0, joining the Nemotron Personas Collection (France, Japan, Brazil, US; Germany announced at GTC Berlin)
EINPresswire corporate release, single source, awaiting cross-verification
Kumo open-source modelsLaunched
Launched at GTC Berlin Oct 21: open models + Agent Knowledge Foundation
NVIDIA official release (GlobeNewswire) + coverage from Next Platform and others
Flora robot-video annotation modelsLaunched
Open-sourced by NVIDIA's GEAR lab: three models fine-tuned from Qwen3.5 (a 9B segmenter + two 27B describer and reward models) that segment robot and human operation videos, attach text instructions, and score process and outcome rewards — already used in GR00T training; claimed 512 GB300s can annotate 200K hours of video per day
Claimed annotation cost $0.72 per video hour vs $88.49 for GPT-6 Astra (modeled on a $10/hour GPU rental, not a measured bill); 57.1% quality on a 50-episode internal benchmark vs Astra's 53.7%, with mixed wins by sub-metric; non-commercial research license; paper and model cards dated 10-20, coverage wave 10-25; source: The Decoder, 10-25
Cosmos 3 world modelsLaunched
Open-sourced Nov 5: four post-trained weights of the world-model family, each a post-trained Cosmos 3 Nano with 16B trainable parameters, under the OpenMDW 1.1 commercial license: Transfer-Auto (multi-view driving data + LiDAR generation), Sim-Auto (streaming synchronized camera and LiDAR), Sim-Transfer (simulator control to photorealistic video), Sim-Bimanual (bimanual robot action video prediction)
Announced on the Hugging Face blog Nov 5; weights on Hugging Face

7 · Agent infrastructure

Announced at GTC Berlin, 10-21
DSX PODLaunched
Modular AI-factory reference design: container-prefabricated, on-site deployment cut from 6 months to 14 days, global delivery from early 2027
Partners Dell / HPE / Lenovo / Foxconn / Quanta and others; DSX closed-loop liquid cooling (45°C warm water) and DSX Flex dynamic power; announced at GTC Berlin 10-21, official release + Reuters cross-coverage
OpenShell reference architectureLaunched
Agent-security reference architecture: Rubin does the "thinking", Vera CPU runs the agent harness, BlueField-4 runs Sentry hardware-level monitoring
Anthropic (building OpenShell agents for Claude), OpenAI, Google, Meta, Microsoft, AWS, SpaceX Super Intelligence and others involved; announced at GTC Berlin 10-21

TimelineMilestones

Radar zone · Unconfirmed (not in the main layers)

2026 AI chip order bookings: $500B(Order visibility announced by Huang Oct 2025 — not confirmed revenue) · Feynman (2028) roadmap details(Limited official disclosure; Rubin already shipping — see above) · China export-control shifts(US conditionally allowed H200 sales to China from early 2026, but China restricted them and NVIDIA plans for zero China revenue; the bipartisan SAFE Chips bill of Dec 2025 proposed a 30-month ban on H200/Blackwell China export licenses; China considered releasing the RTX Pro 5500 workstation card in Sept 2026) · China's homegrown-chip acceleration(Long-term risk; DeepSeek shifted to Huawei for 160K AI chips in Sept 2026, press reports) · NVIDIA acquires Hugging Face: $12.93B signed, pending regulatory approval($11.9B to shareholders + ~$1B employee retention equity; expected to close H1 2027; NVIDIA commits to keeping the platform open and supporting other chip vendors; ~86× annualized revenue; OpenAI, AMD, Salesforce also bid) · Multiverse Comet 60B(GlobeNewswire corporate release, single source: a compressed Nemotron 3.5 Super VL — total parameters halved to 60B, active parameters unchanged at 12B; not cross-verified) · TSMC denies moving into 3D DRAM(Told Tom's Hardware Nov 7 the company "has always focused on its core business, professional integrated circuit foundry services" — read as a substantive denial of a mass-production 3D DRAM plan; the statement did not directly answer whether TSMC had drafted a plan at NVIDIA's request, and NVIDIA did not comment; commercial 3D DRAM roadmaps point to ~2030; The Information reported Nov 6 that TSMC was studying 3D DRAM production at NVIDIA's request) · FT: Jensen Huang is discussing a return as CEO with confidants(FT, Nov 17, citing people familiar: Huang is discussing a possible return as CEO with confidants, and Silicon Valley allies are lobbying the board to reinstate him; directly conflicts with chairman Harvey Jones's opening statement on the earnings call that the board has no intention of bringing him back; not cross-verified)

Verification: FY2026 Q3 earnings (revenue $57.006B, data center $51.215B, gross margin 73.4%) officially reported; latest disclosed is Q2 FY2027 (2026-08-26): revenue $96.2B, data center $89.0B, gross margin 75.0%, Q3 FY2027 guidance $108B±2%, gross margin 74%; Blackwell sold out and Rubin in mass production per Huang's earnings-call remarks; $500B is the AI chip order visibility Huang announced (not confirmed revenue); market cap over $5T widely reported; DGX Cloud halting new customers from The Information's Dec 2025 report, followed up by multiple outlets; Hugging Face platform data (3M models / 18M developers / annualized revenue ~$150M) from NVIDIA's Sept 2026 acquisition announcement — $12.93B signed, expected to close H1 2027, pending regulatory approval, filed in the radar zone; RTX Spark PC lineup launch + pre-orders confirmed by MarketWatch, wccftech and others; NVL72 platform opening confirmed by Barron's and CNBC's interview with the CoreWeave CEO; DGX Spark 64GB cross-confirmed by neoteo, groundtruth, aistockwire and other tech outlets; Nemotron 3.5 Bolt officially launched (Oct 15, fully open commitment) covered by TechTimes, blockchain.news, alfw.ai and others; Nemotron 3.5 Super VL launched (Oct 15, 120B total / 12B active parameters, OpenMDW-1.1 open license) cross-covered by letsdatascience, blockchain.news, Unite.AI and others; Nemotron Personas UK open dataset (Oct 14) is an EINPresswire corporate-release single source, not cross-verified; Nemotron 3.5 family downloadable from Oct 16, cross-confirmed by multiple tech outlets (OpenMDW-1.1 commercial license; Bolt E4B due Oct 23); the German industrial-AI joint declaration of intent was formally signed Oct 19 (Huang and Chancellor Merz on stage at the Invest in Germany summit — intent stage, not a final contract; seed funding of several hundred million euros; Siemens / Audi / Black Forest Labs and about a dozen other German companies involved; advancing physical-AI open models around the SOOFI sovereign open-model program; NVIDIA official blog + German government portal + Reuters, multiple sources). Modal adopting the Vera CPU (Oct 19, Reuters: first publicly named third-party customer, early tests doubled throughput on comparable tasks; NVIDIA guides at least $20B in Vera chip sales this fiscal year); GTC Berlin opening-day details on German cooperation (Oct 20: investment plans in four German startups, economics-ministry MoU, Siemens AI planning agent, Jülich JUPITER 104-qubit quantum-simulation record; madrobot.blog Oct 20 relaying Handelsblatt details, paktechstan coverage, cross-checked against the Oct 19 declaration). GTC Berlin keynote announcements on Oct 21 (DSX POD, OpenShell, Kumo) and Vera Rubin full mass production cross-confirmed by GlobeNewswire's official release, Reuters, Next Platform and Morningstar; the ~$10B European commitment and Agile Robots' 1,000–2,000 small humanoid robots per year target were disclosed in the keynote; Nemotron 3.5 Bolt E4B officially launched (Oct 23: NVIDIA Robotics LinkedIn "Available now"; on Hugging Face, OpenMDW license; ~4GB VRAM, 64K context, 35 languages, multi-turn tool use). Flora open-sourced (Oct 25 coverage, model cards Oct 20: The Decoder; cost and quality comparisons are NVIDIA internal benchmarks and modeled prices, not independent reviews; non-commercial research license). Cosmos 3 open-sourced, confirmed by the Hugging Face blog Nov 5. Fleet Intelligence requirement reported by Bloomberg Nov 6 (citing people familiar, NVIDIA did not respond); Brightspeed inference-data-center plan reported by The Information Nov 5 and widely re-reported Nov 6–7, no final agreement yet; ~600K H200s stranded in Hong Kong, per Reuters Nov 6. NYT H200 investigation from the Nov 7 original and multiple Nov 8 re-reports; TSMC's 3D DRAM denial from its Nov 7 statement to Tom's Hardware, widely reprinted. The board's independent investigation and the Rosen securities class action cross-reported by TechCrunch, YourStory and others Nov 9. The two-day share slide and class-action follow-up per Investopedia, Stocktwits, TradingView and others on 11-10; allegations untested in court, the FBI-investigation claim unverified; TSMC's board approval of up to $500M each in Anthropic / OpenAI per the board resolution report. H200 chips reaching a PLA-linked university per Reuters' Nov 11 exclusive investigation (26 procurement records); the records do not prove NVIDIA supplied directly or knew of violations; Baird's downgrade per Reuters on 11-11; no explanation has emerged for the after-hours slide, so it is not used. The CEO search update per the FT on Nov 13, relayed by Reuters (Reuters says it could not independently verify it); the board's same-day denial that it has formally approached any candidate — conflicting accounts; the share move and HSBC rating per Finnhub data and multiple Nov 13 reports; the Senate Banking Committee testimony summons per NBC, relayed (single-source update). Q3 FY2027 earnings confirmed by the company release and multiple outlets; the FT return-as-CEO report is single-source, filed in the radar zone. Last updated 2026-11-18. · Privacy · Terms · Contact
← Back to Collection