NVIDIA AI Chips: The Complete Guide

NVIDIA's fiscal year ends in late January, so "fiscal 2027" is roughly February 2026 to January 2027. The quarter ended 26 July 2026 grew 106% year on year, Data Center grew 117%, and NVIDIA guided the next quarter to $108.0 billion, plus or minus 2% 2. Management told investors that supply, not demand, is the limit "at least through the end of fiscal year 2028" 3.

Two documents answer most questions about NVIDIA: the annual 10-K, which describes the business and its risks, and the quarterly 10-Q, which updates the numbers. Both are free on the US Securities and Exchange Commission's site and on NVIDIA's investor site 1 7. NVIDIA also publishes a "CFO Commentary" each quarter, which is where the plain-language explanation of each number usually sits 8.

NVIDIA's product rhythm is now one new platform a year: Hopper (announced March 2022), Blackwell (March 2024), Blackwell Ultra (March 2025), Rubin (unveiled January 2026 at CES and expanded at GTC in March 2026) 11 12 13 14 17. Each is announced at GTC, NVIDIA's spring conference, and each changes the rack as well as the chip. In December 2025 NVIDIA took a non-exclusive licence to Groq's inference technology, and Groq's founder Jonathan Ross and president Sunny Madra joined NVIDIA 18; a "Groq 3 LPU" is now one of the seven chips in the Vera Rubin platform 17.

The Vera Rubin platform is described by NVIDIA as seven chips: Vera CPU, Rubin GPU, NVLink 6 Switch, ConnectX-9 SuperNIC, BlueField-4 DPU, Spectrum-6 Ethernet switch and the Groq 3 LPU 17. The Groq 3 LPX system entered full production on 24 August 2026, with Nebius named as the first AI cloud to adopt it 26. NVIDIA also sells a Vera CPU rack of 256 Vera processors for work that needs processors rather than GPUs 17.

The supply chain is now moving partly to the US. In April 2025 NVIDIA said Blackwell chips would be produced at TSMC in Phoenix, packaged and tested by Amkor and SPIL in Arizona, and built into supercomputers by Foxconn in Houston and Wistron in Texas 32. The first Blackwell wafer made in the US came off TSMC's Phoenix line on 17 October 2025 33, and Wistron said the first US-built GB300 superchip was mass-produced at its Fort Worth plant in July 2026 34. Amkor's Arizona packaging campus is not expected to start production until early 2028 35, so advanced packaging still happens mainly in Taiwan today.

NVIDIA has not said which process node Rubin uses; its technical blog gives 336 billion transistors and two compute dies but no node 25. We do not repeat unofficial reports.

NVIDIA's own offices in India, from nvidia.com (the English India page lists the cities; the street addresses appear on its Asia contact page and may be older) 45 46:

• Bengaluru: C-1 "Jacaranda", Wing-A, Manyata Embassy Business Park, Outer Ring Road, Bangalore 560045
• Hyderabad: Plot 6A & B, IT Park Layout, Nanakramguda, Serilingampally Mandal, Hyderabad 500046
• Pune: Commerzone, Building No. 5, Samrat Ashok Path, off Airport Road, Yerwada, Pune 411006
• Mumbai: CNB Square No. 127, Andheri Kurla Road, Chakala, Andheri East, Mumbai 400093
• Gurugram and New Delhi: listed as NVIDIA office cities; street addresses not published on the pages we checked

In the US, NVIDIA says partner facilities in 43 states are involved in its American manufacturing push 47.

NVIDIA now designs at data-centre scale. A GB200 NVL72 rack has 18 compute trays and 9 NVLink switch trays joined by more than 5,000 copper cables, and was designed for 120 kW of cooling capacity 50; NVIDIA's DGX GB200 user guide puts the rack's power draw at about 120 kW 51. Copper is used inside the rack because optical links cost power: NVIDIA puts a pluggable optical transceiver at about 30 W per port against about 9 W for its co-packaged optics 52, which it began shipping in its Quantum-X and Spectrum-X photonics switches 53.

Low precision is the other lever. NVFP4, NVIDIA's 4-bit format, is how the headline numbers are quoted: a Blackwell Ultra GPU does 15 petaflops of dense NVFP4 against 10 for Blackwell 49. "Sparse" figures assume half the numbers are zero and are often close to double the dense figure; always check which one is quoted.

How to read these numbers without being misled:

• Sparse vs dense. Sparse assumes a model where half the weights are zero and can be skipped. Most real workloads are closer to the dense figure, which is often around half.
• Precision. FP4 numbers are roughly double FP8 numbers on the same chip. Comparing an FP4 figure on one chip with an FP8 figure on another is the most common mistake in chip comparisons.
• Inference vs training. NVIDIA quotes Rubin at 50 PFLOPS NVFP4 for inference and 35 for training 25; the training figure is the fairer comparison with older chips' dense numbers.
• Memory bandwidth is often the real limit for running large models, which is why each generation adds more HBM. Even US export rules have used memory bandwidth as the test: the April 2025 licence requirement covered the H20 and any chip reaching its memory bandwidth 1.

In India (rupees, per hour, before tax)

The Government of India said in January 2025 that the average cost of IndiaAI common compute worked out to about ₹115.85 per GPU-hour before subsidy 72. The gap between the government portal and open-market Indian prices is one of the clearest effects of the IndiaAI Mission.

A useful sense-check from NVIDIA itself: the software licence alone (AI Enterprise) lists at $4,500 per GPU per year, or $22,500 as a perpetual licence with five years of support 64.

NVIDIA is also extending its link technology to other companies' processors. In September 2025 it agreed to invest $5 billion in Intel at $23.28 a share, with Intel building custom x86 processors that connect to NVIDIA GPUs over NVLink 76. Networking is growing quickly: in the August 2026 quarter networking revenue rose 18% on the previous quarter and Spectrum-X Ethernet grew 2.6 times year on year 3. The practical point for buyers: the lock-in increasingly sits in the rack and the network, not only in the programming model.

NVIDIA does not name its largest customers, but it reports how concentrated they are. In the quarter to 26 July 2026 one direct customer was 16% of total revenue; over the first half of fiscal 2027, three direct customers were 16%, 15% and 13% 7. Customers headquartered outside the US were 38% of revenue in the quarter 7. Note that NVIDIA reports geography by where the billed customer is based, not where the chips end up, so a country's share shows where the billed customer is based, not where the chips are finally used 1.

11. US export rules name whole racks, not just chips: the October 2023 licensing requirements list products including H100, L40S, RTX 4090, GB200 NVL72 and B200 1.
12. The H20 rule was written around memory bandwidth, covering the H20 and any chip that reaches its memory bandwidth 1. The resulting $4.5 billion charge split into $1.9 billion of inventory and $2.6 billion of purchase obligations 81.
13. An NVL72 rack holds more than 5,000 copper cables and was designed for 120 kW of cooling 50.
14. A pluggable optical port burns about 30 W; NVIDIA's co-packaged optics about 9 W 52. Multiply by tens of thousands of ports and that is why optics moved onto the switch.
15. Groq is now inside NVIDIA's platform: after a December 2025 licence deal, a Groq 3 LPU is one of the seven Vera Rubin chips 18 17.

The supply chain is a risk in both directions. NVIDIA warns that long lead times force it to commit to supply before it knows demand, and that it may not be able to cut those commitments 1; those commitments reached $279 billion in July 2026 8. At the same time, memory prices rising faster than planned are expected to pull gross margin down from 75% to about 71–72% 8 3. The concentration of customers, suppliers and geography is the thread that runs through all of these.

In 2023 NVIDIA said India had about 60,000 experienced CUDA developers and around 40,000 CUDA downloads a month 83. NVIDIA does not report India revenue separately in its filings 1. For Indian buyers the practical numbers are the rental rates: the IndiaAI portal lists an H100 at ₹153 an hour on demand (subsidies of up to 40% may apply), against ₹255.55 at E2E Networks and ₹356 at Yotta 69 70 71.

After Rubin, NVIDIA has named Rubin Ultra and then Feynman. NVIDIA's own GTC 2026 summary lists Feynman-generation parts, including a Rosa CPU, an LP40 LPU and BlueField-5, but gives no year 88. Press coverage of the GTC 2026 roadmap slide reported Rubin Ultra for 2027 and Feynman for 2028 89; treat those years as reported, not confirmed by NVIDIA in writing.