What this post is about
AMD makes computer chips. Two years ago, most people knew them for the chips inside gaming PCs and regular business computers. In the last 12 months, they signed deals worth more than $120 billion to make AI chips for the world’s largest AI companies. This post explains what changed, why it matters, and what to watch next. It is written for readers who do not work in finance or technology.
A chip, by the way, is a small flat piece of silicon that does math. A modern AI chip is about the size of a deck of cards but does more math in one second than a desktop computer does in a week. The fancy ones are called GPUs (short for graphics processing unit — they were first designed for video games, then it turned out the same kind of math runs AI). AMD and Nvidia make the most advanced GPUs in the world.
TL;DR (the short version)
- AMD just had its best three months ever. Sales hit $10.3 billion in the first quarter of 2026 — a quarter is the three-month chunk companies use to report results. Sales of chips for big AI computers grew 57% to a record $5.8 billion in that one quarter.
- AMD signed a huge AI chip deal with OpenAI — the company that makes ChatGPT — in October 2025. The deal could be worth up to $90 billion in chip sales.
- AMD signed a similar deal with Meta, the parent company of Facebook and Instagram, in February 2026. That deal is worth around $60 billion.
- AMD’s CEO Lisa Su now says the worldwide market for AI computer parts will reach $1 trillion by 2030. That is double the figure she gave investors a year earlier.
- AMD bought a company called ZT Systems for about $5 billion so it could build whole AI computers, not just chips. Then it sold the factory part for $3 billion and kept the engineers.
- A new AMD product called Helios — a finished AI computer the size of two refrigerators — ships in the second half of 2026.
About the Author
David Berkowitz runs the ValueAligned Portfolio (VAP) at VAP Wealth Advisors. He also publishes free investing videos and articles under the brand Berk on Value (@BerkonValue). He has spent his career studying how companies make money and how investors should think about them. Nothing in this post is investment advice. It is a research note for people who like to understand what they own.
What does “$120 billion pivot” actually mean?
A pivot is a planned change of direction — when a company decides to do something meaningfully different from what it did before. The $120 billion is the size of the change.
AMD signed two giant contracts. The first was with OpenAI in October 2025. That deal could be worth up to $90 billion in chip sales over several years. The second was with Meta in February 2026. That deal is around $60 billion over five years. Add them up and you cross $120 billion. That is more than four times AMD’s total sales last year.
That money is not in the bank. It is a promise. OpenAI and Meta have agreed to buy a lot of AMD chips over time, but they have to actually buy them for the cash to show up. Most of the revenue (the word companies use for “money coming in from sales”) will land in 2026, 2027, and beyond.
But here is the key point. Two years ago AMD had no AI deals like this. Today it has two of the largest AI customers in the world locked in for years. That kind of contract did not exist for any chip company except Nvidia until late 2025.
What is a “gigawatt” deal, and why is it the unit?
Each AMD deal is described as “6 gigawatts.” That is a measurement of electricity, not money. AI chips eat huge amounts of power, so the customer says how much electricity its data center will use, and the chip company figures out how many chips that takes. A data center, in case the term is new, is a giant building full of computers — picture a Costco-sized warehouse stacked floor to ceiling with whirring machines.
A gigawatt is one billion watts of power. A regular light bulb uses about 60 watts. The average American home uses around 1,200 watts at any given moment, so one gigawatt is enough electricity to run roughly 750,000 homes at the same time. Six gigawatts is the power use of a small country.
Each AI chip needs about a thousand watts when it is running flat out. Six gigawatts works out to roughly six million chips per customer over the life of the deal. That is the math.
Why does the OpenAI deal matter more than the dollar sign?
OpenAI is the company that built ChatGPT. Until October 2025 it bought almost all its AI chips from Nvidia. Then it signed a deal with AMD and called AMD a “core strategic compute partner”—corporate-speak for “we are now relying on AMD just as much as we rely on Nvidia.”
Here is the strange part. AMD gave OpenAI a coupon to buy 160 million AMD shares for one cent each. Yes, one cent. That kind of coupon is called a warrant — basically a ticket that lets you buy stock at a set low price by a future date, provided certain conditions are met. The warrant is worth real money only if OpenAI actually buys the AMD chips and AMD’s stock price rises. If OpenAI does not buy the chips, the coupon expires worthless. If OpenAI does buy them and AMD does well, OpenAI ends up owning up to 10% of AMD almost for free.
That setup means OpenAI makes money in two ways from being an AMD customer. It gets the chips and a chunk of AMD’s stock-price upside. That is why the deal is bigger than just $90 billion in chip sales. It has linked the two companies together for years.
The first chips ship in the second half of 2026. The chip is called the MI450. It carries 432 gigabytes of very fast memory built right onto the package — engineers call it HBM4, short for High Bandwidth Memory version 4, and it lets the chip move huge amounts of data quickly. The MI450 is built at TSMC, the Taiwanese chipmaker that makes most of the world’s most advanced chips, using what they call a “two-nanometer” manufacturing process. Nanometers refer to the size of features on the chip. Smaller numbers mean newer technology and better performance. Two nanometers is the most advanced anywhere in early 2026.
What is the Meta deal really buying?
Meta is the company that owns Facebook, Instagram, and WhatsApp. Its AI work powers everything from photo recognition to the chatbots inside its apps.
In February 2026, Meta agreed to buy six gigawatts of AMD chips over five years. The deal is about $60 billion. Meta also got a coupon for AMD stock just like OpenAI did.
Two things make the Meta deal special. First, AMD is making a custom chip just for Meta — a slightly different version of the MI450 designed for Meta’s specific software. Custom chips are expensive to design, so companies only do them for very large customers. Second, Meta is the first major customer for AMD’s next big regular server chip, code-named Venice. Server chips do the everyday work of a data center — running databases, serving web pages, that kind of thing — while AI chips do the AI math. Meta is buying both kinds of chips from AMD for the same data centers. AMD’s server chip family is called EPYC, and Venice is the upcoming sixth-generation version.
That tells you Meta wants AMD to be its main supplier, not a backup.
Why did AMD buy ZT Systems and immediately sell half of it?
This is the part most articles got wrong, so it is worth slowing down.
In 2024, AMD realized something. Selling chips one at a time was no longer good enough. The biggest AI customers wanted a finished computer, not a box of parts. They wanted a tall metal cabinet — what the industry calls a rack — filled with the right chips, cables, software, and cooling, already built and tested. Each rack stands roughly the height of a refrigerator and holds dozens of AI chips wired together. A finished AI rack costs millions of dollars, and the companies that design and build them have very rare skills.
ZT Systems was one of those companies. It built racks for the world’s largest AI customers. AMD bought ZT for about $4.9 billion in March 2025 — an acquisition, as you call it, when one company buys another. That gave AMD the engineers who knew how to design AI racks at scale.
But AMD did not want to run the factory side. Running factories ties up cash and adds risk. So AMD turned around and sold the factory part to a company called Sanmina for $3 billion — that kind of partial sale is called a divestiture. The deal closed in October 2025.
The math: AMD paid about $1.9 billion net of the sale and kept the people. Sanmina got the factory. AMD now has the design skill it needed without the factory headache.
What is Helios, and why does it matter?
Helios is the name of AMD’s first finished AI rack. Think of it as AMD’s answer to Nvidia’s pre-built AI computer.
Each Helios rack holds 72 of AMD’s biggest AI chips, all wired together to act like one giant brain. It includes the regular Venice server chips, the AI chips, the network cards (the parts that connect the rack to the rest of the data center), and the cooling. A customer can plug it in and start training AI models the same day.
Helios is built to a shared industry blueprint called Open Compute — an industry group that Facebook started in 2011 to publish open hardware designs that any company can copy. Meta likes Open Compute because it means Meta is not locked into one vendor’s rack design. AMD likes it because it gives customers a standard product they can deploy without re-engineering their data centers.
HPE — Hewlett Packard Enterprise, one of the biggest sellers of business computers — became the first big company to commit to selling Helios to its customers. Helios ships at scale in the second half of 2026.
Is AMD’s software finally good enough?
Almost. Not quite.
Every AI chip needs software to function. Without software, a chip is just an expensive paperweight. NVIDIA’s software, CUDA, has been the industry standard for over a decade. AMD’s competing software is called ROCm. For years, CUDA was much better, and big customers stuck with Nvidia partly because their AI engineers already knew how to write CUDA code. Switching to AMD meant rewriting code — what programmers call porting —, and that takes time and money.
AMD released ROCm version 7.2 in late 2025. It now supports the kinds of AI math the newest models need (engineers call them FP4 and FP8 — short for floating-point math using 4 bits and 8 bits, where smaller numbers mean faster but slightly less precise calculations, which turns out to be fine for most AI work). It runs much faster on AMD chips and works with the popular AI tools that engineers actually use day to day.
The honest answer: ROCm has caught up enough that big companies will pay AMD’s engineers to make the change. Eight of the ten biggest AI companies now run live workloads on AMD chips. Two years ago, that number was three.
But ROCm is still behind for small teams that want chips to “just work” without rewriting their code. That gap matters for AMD’s long-term position with universities, startups, and individual researchers.
The chip release calendar in plain English
AMD has three generations of AI chips happening at the same time. Each generation has a code name that roughly indicates how new it is — higher numbers mean newer.
Right now, late 2025 through 2026: the MI350 family. This is the chip that drove AMD’s record sales last year. Oracle, Meta, Microsoft, Cohere, and TensorWave (a smaller cloud company that specializes in AMD-based AI) all run it. The MI350 uses TSMC’s 3-nanometer manufacturing process.
Coming in the second half of 2026: the MI400 family. This includes several variants — MI430X, MI440X, MI455X, and MI450 — each tuned for slightly different jobs. The MI450 is the version going to OpenAI and Meta. The MI400 family uses TSMC’s newer two-nanometer manufacturing — the most advanced chip-making process in the world right now.
Coming in 2027: the MI500 family. This will use even newer technology and faster memory. AMD claims it will be 1,000 times more powerful than the MI300 family from 2024 — though that figure includes everything (chip, rack, and software), not just the chip.
The big shift: AMD now puts out a major new AI chip every year. That matches Nvidia’s pace. Two years ago AMD’s release schedule was much slower. The architecture name underneath all these chips is CDNA — think of it like the engine type in a car, the basic design that all the chips share. The newest chips use CDNA 5, and the 2027 chips will use CDNA 6.
How big is this market really?
Lisa Su, AMD’s CEO, said at her November 2025 investor meeting that the worldwide market for AI computer parts will reach $1 trillion by 2030. She had said $500 billion by 2028, the year before, so she doubled her number in twelve months. (An investor meeting, sometimes called an investor day or analyst day, is a big company event where executives walk Wall Street through their long-term plans.)
That figure is what economists call a “total addressable market” (TAM) — the total amount of money that all customers worldwide could spend on a product over a year. It includes everything: chips, networking gear, full racks, and the optical fiber and power gear that goes with them. One trillion dollars is bigger than the entire mobile phone industry today.
Su thinks AMD’s total sales will grow about 35% per year for the next three to five years, and the data center part of AMD will grow about 80% per year. She projects AMD will earn “tens of billions of dollars” in AI sales by 2027.
These are aggressive numbers. They depend on AMD shipping chips on time, TSMC making them on time, and the customers actually using them. Any of those can slip.
What did the latest financial report show?
AMD reported earnings — meaning profit, the money left after expenses — for the first three months of 2026 on May 5. The numbers:
- Total sales: $10.3 billion (up 38% from a year earlier).
- AI and data center sales alone: $5.8 billion (up 57%).
- Profit: $1.4 billion.
- Forecast for the next three months: $11.2 billion in sales (up another 9%).
The forecast — sometimes called guidance — is the company’s prediction for what the next quarter will look like. Wall Street pays very close attention to it because it tells investors what to expect. A weak forecast, even alongside good current results, often hurts the stock.
In this case the stock went the other way. AMD shares rose 16% the day after the report. AMD now trades around $418. (A share, or stock, is a small piece of a company you can own. AMD trades on the Nasdaq stock exchange under the symbol AMD.)
Lisa Su told investors AMD has “strong and increasing confidence” in reaching tens of billions of dollars in AI sales next year.
Important caveat: these record numbers are mostly from the older MI300 and MI350 chips. The big OpenAI and Meta deliveries do not begin until late 2026. The numbers from those deals will start showing up in the third and fourth quarters of 2026 (July through December) and grow from there.
What could go wrong?
Five real risks, in order of how likely they are.
The chips ship late. TSMC’s two-nanometer manufacturing is brand new, and new means risky. If yields are bad — yield is the percentage of chips coming out of a factory that actually work, and bad yield means lots of chips are broken and have to be thrown away — AMD might not have enough working chips for OpenAI and Meta on time. A six-month delay would push first shipments from late 2026 to mid-2027 and reset the financial story.
OpenAI or Meta buy fewer chips than promised. The contracts say “up to” six gigawatts. They are not “must buy” agreements (which industry calls “take-or-pay” deals, where the buyer pays even if they do not take delivery). If AI demand cools, or if either company finds they need less computing power than they thought, the back end of the contracts may not ship. Investors who model these as guaranteed are making a mistake.
AMD’s software still trips up new users. AMD has the top ten AI customers. The wider world of AI researchers — at universities, at startups, at small companies — still picks Nvidia by default. If AMD does not fix that, it stays a “second supplier” rather than the default choice.
Nvidia hits back. Nvidia is not standing still. Their next chip, code-named Rubin, launches around the same time as the MI450. If Rubin is meaningfully better, customers may slow their AMD orders. AMD’s deals are gigawatts of capacity, but capacity gets bought one quarter at a time.
The stock coupons cost shareholders. The OpenAI and Meta warrants combined could turn into 320 million new AMD shares — about 20% of the company. If both customers hit their targets, current shareholders end up owning less. The technical word is dilution: when a company creates new shares, every existing share owns a smaller piece of the company, like cutting a pizza into more slices. Dilution is the cost of locking in the demand. It is real, and it shows up in the numbers when it happens.
How does AMD compare to Nvidia today?
| Item | AMD (Q1 2026) | Nvidia (most recent quarter) |
| Sales from AI/data center | $5.8 billion | About $32 billion |
| Growth rate | +57% per year | About +60% per year |
| Big AI customer deals | OpenAI 6GW + Meta 6GW | Many large but undisclosed |
| Newest chip on sale | MI355X (2025) | Blackwell B200 (2025) |
| Next chip | MI450 (late 2026) | Rubin (late 2026) |
| AI software | ROCm (open) | CUDA (proprietary) |
| Pre-built AI computer | Helios (open standard) | NVL72, DGX (proprietary) |
The simple read: Nvidia is still about five times bigger than AMD in AI today. But AMD is now growing about as fast in percentage terms. The gap closes only if the MI450 ships on time and ROCm keeps closing the software gap.
What should you watch in the next year?
Three checkpoints.
The third quarter (July–September 2026) earnings report. This will be the first time the new MI450 chips show up in AMD’s revenue numbers. If AI sales jump well past $7 billion in that quarter, the story is on track. If they stall, the OpenAI ramp is running late.
The first OpenAI and Meta data centers going live. Both customers’ first one-gigawatt deployments are scheduled for late 2026. If those open on time, with AMD chips inside, the contracts move from promises to reality.
ROCm software updates. Watch how often AMD ships ROCm updates and what they add. If the third-party support keeps getting wider — meaning more popular AI tools start working on AMD by default — the long-term competitive position improves. If updates slow down, that is a warning sign.
The AMD that existed in 2023 was a chip company with a side bet in AI. The AMD that exists in May 2026 is a primary chip supplier to the two most important AI companies in the world, with a finished AI computer (Helios) shipping later this year. That is a different company.
Whether the stock price already reflects all of that — or has more room to run — is a question every investor has to answer. The data is on the page. The decision is yours.
Endnotes
- AMD first quarter 2026 results — AMD investor relations. https://ir.amd.com/news-events/press-releases/detail/1284/amd-reports-first-quarter-2026-financial-results
- AMD–OpenAI 6 gigawatt deal announcement — AMD newsroom. https://www.amd.com/en/newsroom/press-releases/2025-10-6-amd-and-openai-announce-strategic-partnership-to-d.html
- AMD–OpenAI deal full terms with warrants — AMD investor relations. https://ir.amd.com/news-events/press-releases/detail/1260/amd-and-openai-announce-strategic-partnership-to-deploy-6-gigawatts-of-amd-gpus
- AMD–Meta deal coverage including warrant structure — Tom’s Hardware. https://www.tomshardware.com/tech-industry/artificial-intelligence/amd-meta-100-billion-deal
- Lisa Su $1 trillion AI market by 2030 — CNBC. https://www.cnbc.com/video/2025/11/11/lisa-su-announces-more-than-1t-forecast-for-amds-total-addressable-market.html
- ZT Systems acquisition close — AMD newsroom. https://www.amd.com/en/newsroom/press-releases/2025-3-31-amd-completes-acquisition-of-zt-systems.html
- ZT Systems integration into Data Center Solutions — AMD investor relations. https://ir.amd.com/news-events/press-releases/detail/1240/amd-completes-acquisition-of-zt-systems
- Sanmina $3 billion ZT factory purchase — TechCrunch. https://techcrunch.com/2025/05/19/amd-strikes-a-deal-to-sell-zt-systems-server-manufacturing-business-for-3b/
- Helios rack-scale platform on Open Compute Project standard — AMD newsroom. https://www.amd.com/en/newsroom/press-releases/2025-10-14-amd-showcases-helios-rack-scale-platform-built-o.html
- CES 2026 MI400 lineup and Helios rack details — Tom’s Hardware. https://www.tomshardware.com/tech-industry/artificial-intelligence/amd-touts-instinct-mi430x-mi440x-and-mi455x-ai-accelerators-and-helios-rack-scale-ai-architecture-at-ces-full-mi400-series-family-fulfills-a-broad-range-of-infrastructure-and-customer-requirements
- CES 2026 announcement of Helios and Instinct MI400 — AMD newsroom. https://www.amd.com/en/newsroom/press-releases/2026-1-5-amd-and-its-partners-share-their-vision-for-ai-ev.html
- MI500 1000x performance claim — Data Center Dynamics. https://www.datacenterdynamics.com/en/news/amd-unveils-full-mi400-product-lineup-claims-mi500-chips-will-deliver-1000x-increase-in-ai-performance/
- ROCm 7.2 software release — AMD ROCm Blogs. https://rocm.blogs.amd.com/software-tools-optimization/rocm7.2/README.html
- MI350 series shipping in 2025 — AMD blog. https://www.amd.com/en/blogs/2025/amd-instinct-mi350-series-and-beyond-accelerating-the-future-of-ai-and-hpc.html
- AMD record full-year 2025 results — AMD investor relations. https://ir.amd.com/news-events/press-releases/detail/1276/amd-reports-fourth-quarter-and-full-year-2025-financial-results
- AMD–OpenAI tens-of-billions value — TechCrunch. https://techcrunch.com/2025/10/06/amd-to-supply-6gw-of-compute-capacity-to-openai-in-chip-deal-worth-tens-of-billions/
- AMD–Meta $60 billion deal context — TECHi. https://www.techi.com/amd-meta-60b-deal-nvidia-ai-monopoly/
- Meta as lead Venice EPYC customer — Hardware Busters. https://hwbusters.com/news/amd-lands-6-gigawatt-meta-deal-targets-telecom-with-new-epyc/
- AMD–Meta expanded structure with 160M-share warrant — The Markets Daily. https://www.themarketsdaily.com/2026/02/27/advanced-micro-devices-expands-meta-ai-deal-6gw-instinct-gpu-plan-custom-mi450-and-160m-share-warrant.html
- AMD November 2025 Financial Analyst Day — Financial Content / StockStory. https://markets.financialcontent.com/stocks/article/stockstory-2025-11-12-why-is-amd-amd-stock-rocketing-higher-today
- HPE first OEM Helios commitment — HPE newsroom. https://www.hpe.com/us/en/newsroom/press-release/2025/12/hpe-accelerates-ai-deployments-with-first-amd-helios-ai-rack-scale-architecture-with-open-scale-up-networking-built-with-broadcom.html
- Q1 FY2026 stock reaction — TIKR. https://www.tikr.com/blog/amd-stock-beats-q1-estimates-as-data-center-revenue-hits-record-5-8b
- AMD Q1 FY2026 print and Lisa Su commentary — CNBC. https://www.cnbc.com/2026/05/05/amd-q1-2026-earnings-report.html
- Average U.S. home electricity use — U.S. Energy Information Administration. https://www.eia.gov/energyexplained/units-and-calculators/
