AMD AI Chips: The Complete Guide

AMD's financial year is the calendar year (it ends on the last Saturday of December) 1. Fiscal 2025 revenue was $34.6 billion, with Data Center at $16.6 billion 7. For the third quarter of 2026 AMD guided revenue of about $13 billion, plus or minus $300 million 3. At its November 2025 Financial Analyst Day it set targets of more than 60% a year growth for Data Center and more than 80% for Data Center AI over three to five years 8.

AMD's main documents are the annual 10-K, the quarterly 10-Q and the earnings press releases, all on ir.amd.com and the SEC site 11 12 3. AMD reports three segments: Data Center, Client and Gaming, and Embedded; in the June 2026 quarter they were $6.7 billion, $3.8 billion and $977 million 3.

AMD bought its way into whole-rack systems. It agreed to buy ZT Systems in August 2024 for $4.9 billion including contingent payments 23, completed it in March 2025 20, then sold ZT's manufacturing arm to Sanmina in October 2025 while keeping the design teams; Sanmina became its preferred partner for building racks 24 11. Software was bought too: Nod.ai (2023), Silo AI (about $665 million, 2024) and Brium (2025) 25 26 27.

The MI350X and MI355X share a design but differ in power: MI350X is rated at up to 1,000 W and offered with air or liquid cooling, MI355X at up to 1,400 W 33 34. AMD has published the roadmap after MI400: MI500 in 2027 and MI600 in 2028 4.

AMD is also paying to secure packaging. In May 2026 it announced more than $10 billion of investments in Taiwan's supply chain, including advanced packaging with ASE, SPIL and PTI, and substrates 40. In the US, AMD said it validated its 5th-generation EPYC processors at TSMC's Arizona fab and that its "Venice" EPYC was the first high-performance product taped out on TSMC's 2 nm process 41. AMD has not said that any Instinct AI chip is made in Arizona.

AMD's India offices (from amd.com) 45:

• Bengaluru, Technostar campus (largest AMD design centre): Building 1, Level 1, Prestige Technostar, Doddanekundi Industrial Area, Bengaluru 560037 46
• Bengaluru, Whitefield: Building II, Plot 100–103, Export Promotion Industrial Area, Whitefield, Bengaluru 560066
• Bengaluru, Koramangala: 2nd Floor, Kabra Excelsior, 6A, 7th Main Road, Koramangala, Bengaluru 560034
• Hyderabad, Knowledge City: 10th–17th Floor, Salarpuria Sattva Knowledge City, Hyderabad 500081
• Hyderabad, Mindspace: Building 11, Mindspace, Madhapur, Hyderabad 500081
• Gurugram: Spaze I-Tech Park, Tower B1, Sector 49, Sohna Road, Gurugram 122018
• New Delhi: A-25, 2nd Floor, Mathura Road, Mohan Cooperative Industrial Estate, New Delhi 110044

The MI300X has 304 compute units on eight compute chiplets 37 38; the MI355X has 256 compute units 34. Helios links its 72 GPUs with 260 TB/s of scale-up bandwidth and 43 TB/s of scale-out networking, using AMD's Pensando "Vulcano" 800 Gb/s network cards 47 32. AMD says Helios supports UALink, the open chip-linking standard it co-founded 6.

A quirk to watch: for MXFP4 on the MI355X, AMD lists the same 10.1 petaflops figure for dense and sparse 34, while its FP8 and FP16 figures double with sparsity. Compare the FP8 dense figures (2.61 for MI300X, 5 for MI355X) for a like-for-like view 37 34. For full-precision science work the MI300X does 163.4 teraflops of FP64 matrix maths against 78.6 for MI355X 37 34: AMD traded scientific precision for AI speed, and now offers MI430X for scientific work 5.

In India (rupees per hour)

The spread is wide: the same MI300X lists at $6.00 an hour on Oracle and under $2 on some specialist clouds 54 57. For comparison, the IndiaAI portal lists NVIDIA's H100 at ₹153 an hour on demand (see our NVIDIA guide).

Software is AMD's known gap, and it is buying help. Under its July 2026 deal with Anthropic, AMD adopts Claude internally and uses it to speed up ROCm development 31. On the August 2026 earnings call, Lisa Su said open-source contributions to ROCm had risen more than tenfold in a year and that more than 3 million models run "out of the box" 63 (call quotes are from a third-party transcript). AMD itself names NVIDIA's "proprietary software ecosystem" as a competitive risk 1.

AMD said in June 2025 that seven of the ten largest AI model builders run production work on Instinct 19. It does not name customers' shares of revenue in the filings we could read; it says only that a small number of customers account for a substantial share 1. International sales were 67% of revenue in 2025 11. On the August 2026 call AMD said it expects Data Center revenue to more than double in 2027 63.

11. More than $10 billion committed to Taiwan's supply chain, including packaging with ASE, SPIL and PTI 40.
12. AI chips earn less than average: the CFO said Data Center AI margins are slightly below the company average 63 (third-party transcript).
13. Full-precision maths halved from MI300X to MI355X (163.4 to 78.6 teraflops FP64) as AMD optimised for AI 37 34.
14. Two-thirds of sales are outside the US: 67% in 2025 11.
15. Samsung is AMD's primary memory partner for the MI350X and MI355X, and the two say they will align on HBM4 supply for the MI455X 35.

AMD's inventory build (62% work in progress at $8.47 billion) and $12.2 billion of purchase commitments are bets that the MI450 ramp lands on time 12 11. The warrants given to OpenAI and Meta dilute existing shareholders if the milestones are met 69 65.

AMD's MI300X is among the GPU types in the IndiaAI Mission's shared compute pool, offered through empanelled cloud providers 72; the portal lists MI300X at ₹168.20 an hour on demand 59. We found no official AMD India update after 2023 on headcount or investment.

The first big test is delivery: 1 gigawatt for OpenAI and Meta each in the second half of 2026, and Anthropic's first gigawatt in the first half of 2027 64 65 31. AMD's own growth targets (Data Center AI above 80% a year) assume those ramps land 8.