Who Makes AMD's Chips? The Supply Chain Behind Helios (October 2026)
AMD doesn't manufacture its own chips. TSMC fabricates them, SK hynix and Samsung supply the memory, and Amkor handles packaging and test. Here's who makes AMD chips, and what this month's news means for that chain.
Make Sure I Know It
· 4 min read
AMD's Helios rack-scale AI servers are starting to reach customers this half of the year, and AMD won't have fabricated any of the chips inside them. It designs them. Other companies build them. If you're following AMD this quarter, that's the part of the story worth understanding, because most of this month's chip news is really about those other companies.
The short answer
No, AMD does not make its own chips. It's a fabless company: AMD owns the CPU and GPU designs and outsources the manufacturing.
Who fabricates AMD chips: TSMC builds every one of AMD's leading-edge CPU and GPU dies.
Who supplies the memory: SK hynix supplies the high-bandwidth memory stacked next to Instinct accelerators, and Samsung Electronics is the second source for HBM and regular DRAM.
Who packages and tests them: Amkor Technology provides outsourced advanced packaging and test.
Whose tools make it possible: ASML makes the EUV scanners that print AMD's leading-edge nodes. Applied Materials supplies deposition and etch tools inside those fabs, and Synopsys makes the design software AMD lays its chips out with.
Helios is ramping, and two huge buyers are waiting
Helios puts AMD's Instinct MI455X GPUs and its 6th-gen EPYC "Venice" CPUs into a single rack.
OpenAI is the first big name. In July, AMD said OpenAI expects to start bringing Helios online through multiple deployment partners in the second half of 2026, with deployments picking up through 2027. That's the first phase of the 6-gigawatt deal the two companies announced in October 2025.
Meta Platforms is the second. In February it signed its own agreement with AMD for up to 6 gigawatts of Instinct GPUs. The first deployment uses a custom GPU based on the MI450 architecture, and shipments for the first gigawatt are scheduled to begin in the second half of 2026. Meta is also a lead customer for the Venice CPUs.
Put together, that's up to 12 gigawatts of planned AMD GPU capacity, and all of those GPUs have to come off TSMC's production lines.
TSMC's September numbers show how much demand it's handling
On October 8, TSMC reported September revenue of NT$511.9 billion, up 54.6% from a year earlier. Revenue for the first nine months of 2026 was up 41.1%. TSMC holds its full third-quarter earnings call on October 15.
AI demand is behind most of that growth. AMD, NVIDIA and the cloud companies' custom chips all compete for the same leading-edge wafers, so how much TSMC can build sets a ceiling on how much AMD can ship.
GlobalFoundries will make a packaging part for TSMC in New York
Also on October 8, GlobalFoundries signed a $2 billion, five-year deal to make silicon interposers for TSMC's CoWoS advanced packaging at its fab in Malta, New York. Volume production is expected to ramp in the first half of 2028.
An interposer is the silicon layer that connects a processor to the memory stacks next to it. CoWoS capacity has been one of the main limits on how many AI accelerators the industry can ship, so a new interposer source helps everyone building them. It's also TSMC's first U.S.-based source for the part. Neither company named end customers.
AMD's own calendar
AMD reports third-quarter results on November 3, 2026, after the market closes. On September 28 it announced plans to acquire World Labs, an AI model research company, to support its AI compute roadmap. And on October 8, CEO Lisa Su received the National Medal of Science.
Why AMD's supply chain looks like this
AMD went fabless in 2009, when it spun its factories off into the company that became GlobalFoundries. Since then it has spent its engineering budget on chip design and software, while TSMC pays the tens of billions it costs to build each new fab.
The cost of that choice is dependence on a few suppliers. AMD's leading-edge parts rely on one foundry, HBM from two memory makers, and EUV tools that only ASML builds. When TSMC's leading-edge lines are full, AMD has no other foundry it can turn to at that performance level.
Who buys AMD's chips
The map shows four big buyers:
Microsoft buys EPYC CPUs and Instinct accelerators for Azure.
Meta Platforms runs AMD server CPUs and accelerators in its AI training fleet.
Oracle sells AMD-based compute in Oracle Cloud Infrastructure.
Dell Technologies builds AMD silicon into PowerEdge servers and commercial PCs.
What to watch next
TSMC's call on October 15 should say something about CoWoS capacity and AI demand going into 2027. AMD's results on November 3 could be the first time Helios shows up in the numbers. And if the first OpenAI and Meta deployments ramp on schedule late this year, AMD's orders at TSMC will grow a lot in 2027.
