Who Buys AMD AI Chips? The Customers and Concentration Risk

A chipmaker's sales look like one number on an earnings report. At the top end of AI hardware, though, that number rests on a short list of very large buyers. Each of them signs, delays or spreads its orders on its own timetable, and any one of them can change the chipmaker's year.

Headlines report the biggest deals in gigawatts, and some call them circular. They rarely say who is paying today, how much of a deal is firm, or how much of the buyer's total spending the supplier really wins.

For any AI-chip stock, the customer list works like a risk map. A broad list spreads the revenue across many buyers, while dependence on a few names concentrates it, so one delay or one smaller order can move the whole business.

Advanced Micro Devices (AMD) shows the pattern clearly. In about two years its AI chips have gone from running services inside Microsoft's cloud to agreements measured in gigawatts, signed directly with AI labs. Those buyers sit at the center of AMD's AI and data-center business, and on BloFin, AMD's share price trades as the AMDUSDT Perpetual.

Who buys AMD's AI chips?

AMD's AI chips are bought by a small group of AI labs, cloud providers and state-backed buyers. On August 4, 2026, AMD named Anthropic, Cirrascale, HUMAIN, Meta, Microsoft, OpenAI, Oracle, TensorWave and Vultr among the companies deploying its Helios racks (source: AMD Q2 2026 results). No customer reached 10% of AMD's revenue in 2025 (source: AMD 10-K, 2025).

AMD's AI chips are its Instinct GPUs. A graphics processing unit (GPU) runs thousands of small calculations at once, and that is the kind of work that training and running AI models needs. Helios is AMD's complete AI server rack, built to hold its newest Instinct GPUs. AMD's own list of Helios buyers ends with the words "and others", so it names a sample of the buyers.

Most lists of AMD customers online mix these AI buyers with the many companies that use AMD's PC or server processors. The AI buyers are a narrower group of AI labs, hyperscalers (the giant public clouds), neoclouds (newer clouds built to rent out GPUs) and sovereign buyers (state-backed projects), and each has made its own dated public move with AMD:

Each of these buyers has a different reason to buy, and a different way of paying. Those differences decide whose name ends up on AMD's invoice, and how much weight any single buyer carries.

Hyperscalers, AI labs, neoclouds and sovereign buyers

AMD's AI customers fall into five groups. Hyperscalers are giant public clouds such as Microsoft Azure. AI labs such as OpenAI build the models, neoclouds such as TensorWave rent out GPUs, and sovereign buyers are state-backed projects such as Saudi Arabia's HUMAIN. Server makers buy AMD chips to build into systems they sell (source: AMD 10-K, 2025).

A hyperscaler is a company that rents computing power from its own huge data centers to thousands of customers at once. AMD's annual report says large public cloud providers and hyperscale private data centers buy a substantial portion of its data-center products, directly and indirectly (source: AMD 10-K, 2025).

An AI lab is a company that builds AI models, such as OpenAI's GPT models or Anthropic's Claude. A neocloud is a newer cloud company built mainly to rent out GPUs for AI work. A sovereign buyer is a government, or a company a state owns, buying AI computing so that it stays inside the country. HUMAIN's chief executive said demand for sovereign, high-performance AI computing keeps growing (source: AMD, Cisco and HUMAIN).

The labs often pay through someone else. Lisa Su, who runs AMD, said on the company's earnings call of August 4, 2026 that OpenAI, Anthropic and Meta "will be consuming through a number of CSPs, both hyperscalers and others" (source: AMD Q2 2026 earnings call). A CSP, or cloud service provider, is any company that sells computing as a service, whether it is a hyperscaler or a neocloud. OpenAI's agreement lets it buy directly, through its affiliates or indirectly through authorized third-party purchasers (source: AMD 8-K, October 2025). So the path from agreement to sale can run like this:

Say you read that an AI lab has signed for a gigawatt of AMD chips. A cloud provider may buy those racks, run them and rent the computing to the lab, so the sale can reach AMD's books under the cloud's name.

Beyond the gigawatt deals sits a wider base of buyers. Su said on the same call that many other customers are interested in Helios at a more regular scale than the gigawatt deals (source: AMD Q2 2026 earnings call). AMD's annual report lists server makers among its data-center customers, and they build its chips into finished machines that they sell (source: AMD 10-K, 2025).

Most of the big names are future buyers of AMD's next chip, the MI450. The chips earning AMD money today are older models, and they run in a different set of places.

Who runs AMD Instinct GPUs today?

AMD Instinct GPUs run today in cloud data centers, AI labs, AI start-ups, national labs and a sovereign AI project, led by the MI350 Series. AMD said its second-quarter 2026 Instinct sales were more than double those of a year before, ahead of the first Helios shipments (source: AMD Q2 2026 earnings call).

Instinct is AMD's family of data-center GPUs. The installed base is the set of those chips already running in customers' data centers. That base earns AMD revenue now, while the gigawatt agreements describe what buyers may take later.

AMD reports its GPUs inside its Data Center segment, which also sells its EPYC server processors. That segment had revenue of $6.7 billion in the quarter ending June 27, 2026, up 107% on the same quarter of 2025. AMD credited the rise mainly to demand for EPYC processors and MI350 Series GPUs (source: AMD 10-Q, Q2 2026).

The installed base is older than the headline deals. Microsoft's Azure cloud made MI300X instances generally available on May 21, 2024, and AMD said at the time that MI300X chips were powering Azure OpenAI's GPT-3.5 and GPT-4 services (source: AMD, May 2024). That was more than a year before OpenAI signed with AMD directly.

Oracle's cloud launched MI300X instances in 2024 and later made MI355X GPUs generally available (source: AMD and Oracle). HUMAIN was delivering production AI computing on MI355X GPUs in Saudi Arabia by August 31, 2026 (source: AMD, Cisco and HUMAIN). TensorWave, a neocloud that calls itself the emerging leader in AMD-powered AI infrastructure, said in May 2025 that it was deploying more than 8,000 MI325X GPUs for one training cluster (source: TensorWave).

The bigger numbers in the headlines are commitments for later. Those commitments come in different units, and only some of each one is binding.

What each AMD AI customer has committed to buy

OpenAI's October 2025 agreement and Meta's of February 2026 each cover up to 6 gigawatts of AMD GPUs that the customer intends to deploy (source: AMD 10-Q, Q2 2026). Anthropic plans up to 2 gigawatts from the first half of 2027, Oracle 50,000 MI450 GPUs from the third quarter of 2026, and Microsoft has published no size.

A commitment here means what a buyer has said it will take, and the words around each number matter. "Up to" sets a ceiling, and a buyer can stop short of it. OpenAI's purchase of its first gigawatt of MI450 Series GPUs is binding (source: AMD 8-K, October 2025). Meta's binding purchase covers the first gigawatt equivalent of Instinct GPUs (source: AMD 8-K, February 2026). Anything past that first gigawatt, up to the 6-gigawatt ceiling, is what AMD's filing says the customer intends to deploy.

The buyers also count in different units. One gigawatt equals a billion watts, and AI deals use that unit of electrical power to size a data center by what its chips draw. A megawatt is a thousandth of a gigawatt. Oracle counts in GPUs and Microsoft gives no count, so the commitments cannot simply be added together.

Anthropic plans up to 2 gigawatts of MI450 GPUs in Helios racks, with the first gigawatt starting in the first half of 2027 (source: AMD and Anthropic). Anthropic's agreement also includes a promised equity investment by AMD. Oracle announced an initial 50,000 MI450 GPUs, to start in the third quarter of 2026 and expand in 2027 and beyond (source: AMD and Oracle). AMD has not said whether those GPUs count toward OpenAI's gigawatts.

Microsoft said on July 20, 2026 that it is bringing Helios to Azure for AI inference, and it gave no quantity (source: Microsoft). Inference is the work of running a trained model to answer requests. AMD, Cisco and HUMAIN plan up to 250 megawatts through their joint venture from 2027, and up to 1 gigawatt by 2030 (source: AMD, Cisco and HUMAIN). Oak Ridge National Laboratory expects to take delivery of its Discovery supercomputer in 2028 (source: AMD and Oak Ridge National Laboratory).

Set beside AMD's own filings, these commitments sharpen a simple question: how much does AMD lean on a few buyers?

AMD customer concentration in the 10-K

AMD's annual report shows no customer at 10% or more of revenue in either 2025 or 2024. In 2023, one customer in its Client and Gaming segment took 18%. At the end of 2025, one unnamed customer accounted for about 11% of the money customers owed AMD (source: AMD 10-K, 2025).

Customer concentration measures how far a company's sales rest on a handful of buyers. AMD's 10-K, its annual report to the US Securities and Exchange Commission, gives three lines on it, and each one says something different.

The first line is the revenue test. No customer reached 10% of AMD's revenue in fiscal 2025 or 2024 (source: AMD 10-K, 2025). Say you want the most any one buyer could have paid AMD in 2025. AMD's revenue was $34,639 million that year, so 10% is about $3.46 billion, and every customer paid less than that.

The second line looks back to 2023, when one Client and Gaming customer took 18% of AMD's revenue. That segment sells PC and gaming chips, so the buyer came from that side of AMD, and the 10-K leaves it unnamed.

The third line covers receivables, the money customers owe for chips already delivered. One customer accounted for about 11% of AMD's receivables on December 27, 2025, and another customer for 24% on December 28, 2024 (source: AMD 10-K, 2025). Those are two different buyers. AMD said it saw no significant credit risk in either balance, based on past collections and a review of their current credit quality.

The revenue mix has shifted since that annual test. Data Center brought in $6,718 million of AMD's $11,536 million of revenue in the second quarter of 2026, about 58%. A year earlier the segment was $3,240 million of $7,685 million, about 42% (source: AMD 10-Q, Q2 2026).

The 2025 test also describes the year before any MI450 shipments. The gigawatt plans are ceilings, while AMD's own purchase commitments are firm. AMD had $30,276 million of unconditional commitments at June 27, 2026, mostly for wafers, substrates and components from suppliers, plus cloud computing capacity (source: AMD 10-Q, Q2 2026). Its risk factors say AMD depends "on a small number of customers for a substantial portion" of its business and expects that to continue (source: AMD 10-K, 2025). That dependence is a central part of the bear case on AMD.

What AMD's filings leave out about its AI buyers

AMD's quarterly filing gives no separate revenue line for Instinct GPUs. The 10-Q reports Data Center as one segment, so the GPU share and each buyer's share stay inside that figure, including the share of sovereign and national-lab buyers such as HUMAIN and Oak Ridge National Laboratory (source: AMD 10-Q, Q2 2026). The 10-Q carries no customer percentage either, because the 10% test is an annual disclosure. The 2023 customer and the two receivables customers are unnamed too.

Some commitments appear in the filings without a name. After June 27, 2026, AMD committed to investments of up to $5.0 billion, with contingencies attached, expected to run through fiscal 2028, and the 10-Q leaves the recipient unnamed (source: AMD 10-Q, Q2 2026).

The next concentration line will come with AMD's 10-K for fiscal 2026. That report will cover the third quarter of 2026, when Helios shipments were due to begin, so it will be the first annual test that can reflect the new racks. Until then, a new customer headline is best read against five questions:

1. Is it a binding purchase or an "up to" ceiling?

2. What unit is it in: gigawatts, GPUs, megawatts, or no size at all?

3. When can the first chips ship, and does the buyer have the site and power ready?

4. Who pays AMD: the lab, a cloud acting for it, or a joint venture?

5. Does AMD give anything back, such as shares, an investment or a capacity commitment?

Concentration measures how much AMD depends on its buyers. The mirror question is how much of each buyer's spending AMD actually wins.

AMD's share of each buyer's AI chips

AMD is one of several chip suppliers to each of its largest AI buyers. OpenAI also buys from Nvidia and is designing its own chip with Broadcom, while Anthropic trains Claude on Amazon's Trainium chips and Google's tensor processing units (TPUs). On March 31, 2026, OpenAI named AMD as one of five chip sources (source: OpenAI, March 31, 2026).

A letter of intent is a signed statement that two companies plan to do a deal, made before any final contract. Nvidia and OpenAI announced one on September 22, 2025, for at least 10 gigawatts of Nvidia systems (source: Nvidia and OpenAI). On February 27, 2026, OpenAI announced $30 billion of new investment from Nvidia. It also announced 3 gigawatts of inference and 2 gigawatts of training capacity on Nvidia's Vera Rubin systems (source: OpenAI, February 27, 2026).

OpenAI's own chip is a joint project. OpenAI designs the accelerators, and Broadcom deploys the racks, targeted to start in the second half of 2026 and finish by the end of 2029 (source: OpenAI and Broadcom). Inside OpenAI's plans, AMD competes with Broadcom as well as with Nvidia.

Anthropic has spread its orders the same way. On April 6, 2026, it said it trains and runs Claude on Amazon Web Services (AWS) Trainium, Google TPUs and Nvidia GPUs. The same day it announced multiple gigawatts of TPU capacity from Google and Broadcom, starting in 2027 (source: Anthropic, April 6, 2026). Two weeks later it secured up to 5 gigawatts with Amazon, including Trainium2 and Trainium3 capacity (source: Anthropic, April 20, 2026). Its AMD plan followed in July.

A sovereign buyer spreads its orders too. HUMAIN is targeting 200 megawatts of Qualcomm AI200 and AI250 racks starting in 2026 (source: Qualcomm and HUMAIN). That sits beside its AMD plan of up to 250 megawatts from 2027. Meta splits its orders as well, and Meta's other chip deals include Nvidia and Broadcom.

Each of those rival suppliers announced its own numbers, on its own date, and set side by side they show AMD's slice of each buyer:

The largest buyers also have their own reasons to keep more than one supplier in the running, which is why Nvidia's largest customers want AMD to succeed. A supplier that is one of several has a reason to sweeten its terms, and that is where the circular financing debate starts.

AMD and the AI circular financing debate

Circular financing is when a supplier puts money or shares into a customer that then buys the supplier's products. AMD gave OpenAI and Meta warrants covering up to 160 million AMD shares apiece, tied to GPU purchases. After June 27, 2026, it also committed to investments worth up to $5.0 billion (source: AMD 10-Q, Q2 2026).

A warrant gives its holder the right to buy newly issued shares at a fixed price, which is $0.01 each for these two. Each warrant vests in stages, called tranches, and OpenAI's depend on its GPU purchase milestones, on AMD stock price targets and on stock performance thresholds (source: AMD 8-K, October 2025). Meta's tranches depend on its purchase milestones and on AMD stock price thresholds (source: AMD 8-K, February 2026). As of June 27, 2026, none of the tranches had vested, and the warrants left AMD's results for the first half of 2026 untouched (source: AMD 10-Q, Q2 2026). The warrant arithmetic shows what they could do to AMD's share count.

An equity investment is money put into a company in return for a stake in it. AMD's July 2026 release says AMD has promised to put up to $5 billion into Anthropic's equity at a later date (source: AMD and Anthropic). AMD also works with Cisco and HUMAIN through a joint venture, a business the partners set up together, which will deploy HUMAIN's next AMD capacity (source: AMD, Cisco and HUMAIN). HUMAIN already sells computing on its current AMD systems to others as GPU-as-a-service (source: AMD, Cisco and HUMAIN).

Some of the money runs the other way. AMD said it will broadly adopt Anthropic's Claude across its engineering and product teams, which makes AMD a customer of its own customer (source: AMD and Anthropic).

A delivered GPU is still a real sale when the buyer pays for it. What these deals change is the spread of AMD's risk, since more of AMD's revenue, and some of its capital, is tied to the same few buyers.

Neocloud financing: AMD Ventures and cloud capacity

AMD Ventures is AMD's venture arm, the unit that invests in young companies. In May 2025 it co-led, with Magnetar, a $100 million Series A round for TensorWave, a Series A being an early funding round (source: TensorWave). TensorWave said then that it was on track to end 2025 with a revenue run rate, its recent revenue scaled up to a full year, of more than $100 million. The release gave the round's total, and it left AMD Ventures' own share unstated.

A backstop is a promise to step in and pay if something falls short. In June 2025, SemiAnalysis, an industry research firm, reported that AMD had offered cloud providers a deal on capacity. In return for buying more AMD GPUs, the clouds could rent a large part of that capacity back to AMD on long-term contracts for AMD's own software work (source: SemiAnalysis). SemiAnalysis named AWS, Oracle Cloud Infrastructure, DigitalOcean, Vultr, TensorWave and Crusoe. It also reported that for some neoclouds, AMD would rent the capacity itself if the cloud could not sell it.

That is an outside report, and AMD's own filings say less. AMD's 10-Q lists "multi-year cloud compute capacity arrangements with cloud service providers" among its commitments (source: AMD 10-Q, Q2 2026). They sit inside the $30,276 million of unconditional commitments at June 27, 2026, and the filing names no cloud and gives them no separate size.

Put together, money can run in a loop at the neocloud level as well. AMD's venture arm helps fund a cloud, the cloud buys AMD GPUs, and AMD holds multi-year commitments to buy computing capacity from cloud providers. The filings describe that last step only in general terms, so the size of any loop stays unknown.

Can AMD's AI customers pay for what they signed?

Whether AMD's biggest buyers can pay is a separate question from circularity. OpenAI said on March 31, 2026 that it had closed $122 billion of committed funding and was earning $2 billion of revenue a month (source: OpenAI, March 31, 2026). Anthropic put its run-rate revenue at more than $30 billion in April 2026 (source: Anthropic, April 20, 2026).

Counterparty credit is the risk that the other side of a contract cannot pay. For AMD, it asks whether a buyer that signed for gigawatts will have the money when the racks arrive. AMD's 10-K says the risk of customers defaulting or paying late is "heightened" because it expects a small number of customers to keep providing a substantial part of its revenue (source: AMD 10-K, 2025).

OpenAI's March round valued it at $852 billion after the new money. OpenAI also expanded its revolving credit facility, a loan it can draw, repay and draw again, to about $4.7 billion (source: OpenAI, March 31, 2026). Reported internal projections have OpenAI losing money in 2026, and OpenAI's own finances show how large those projected losses are.

A credit rating is an agency's grade of how likely a borrower is to repay. In August 2026 Nvidia agreed to guarantees capped at $105 billion on leases taken by an OpenAI affiliate in Pike County, Ohio. The leases cover about 4.25 gigawatts of power for the computing equipment itself. Those guarantees end on events that include OpenAI "achieving a satisfactory credit rating" (source: Nvidia 10-Q, July 2026).

Anthropic said its run-rate revenue had passed $30 billion by April 20, 2026, up from about $9 billion at the end of 2025. Amazon was investing $5 billion at that point, with up to $20 billion more possible later (source: Anthropic, April 20, 2026). Anthropic has also spent much of 2026 in a legal fight with the Pentagon. On August 27, 2026, a federal judge ruled that the Pentagon could not designate Anthropic a supply chain risk, a label it had tried to apply earlier in the year (source: NPR). The Pentagon had applied that label under two different laws, and on September 25, 2026 a federal appeals court in Washington upheld the other designation in a 2-1 decision (source: Reuters via Military Times). Anthropic says the designation has cost it billions of dollars in lost business, and it said it was considering a review by the full appeals court (source: Reuters via Military Times).

Before you count a new gigawatt agreement with an AI lab as future AMD revenue, check the lab's latest funding round, its reported revenue and any credit backing from others.

Customer data-center readiness

A signed gigawatt reaches AMD's revenue only when a data center can take the racks. That needs land, grid power and a finished building, which the industry calls a shell. During the August 4, 2026 call, a Goldman Sachs analyst asked whether AMD's customers had the land, power and buildings ready to accept its products (source: AMD Q2 2026 earnings call).

Su answered that she saw nothing that would give AMD pause about hitting its targets. She added that the range of outcomes "will depend on the ability to bring on more capacity in a timely fashion" (source: AMD Q2 2026 earnings call). Earlier in the call she said AMD was timing its deliveries to when the larger data center build-outs are ready, working with each customer on its data center plans.

The dates show how readiness sets the timing. On August 4, 2026, AMD said Helios was in production, with first shipments due later in the third quarter and a ramp through the fourth quarter into 2027 (source: AMD Q2 2026 earnings call). On September 8, 2026, Jean Hu, AMD's chief financial officer, said the third quarter would be the very beginning of production shipments. She expected revenue in that quarter and a much larger step up in the fourth (source: AMD at Citi's technology conference). Capacity from the HUMAIN joint venture is expected to start coming online in the second half of 2027 (source: AMD, Cisco and HUMAIN). Anthropic's first AMD gigawatt begins in the first half of 2027, and Nvidia's OpenAI guarantees cover a land, power and shell buildout in Ohio.

A buyer's delay at its own site can push AMD's revenue into a later quarter even when the order stands. The chips' own production schedule is the other limit on timing, set by AMD's MI450 roadmap. For a reader of customer news, a building finished late matters as much as a contract signed early.

Customer delivery news reaches traders through AMD's share price, which the AMDUSDT Perpetual follows, and the contract's live price is on the AMDUSDT Perpetual page.

AMD's customers and the AMDUSDT Perpetual

The AMDUSDT Perpetual is BloFin's perpetual contract on AMD's share price. It settles in USDT, a stablecoin pegged to the US dollar, and offers leverage of up to 20x (source: BloFin AMDUSDT contract information). BloFin also lists perpetuals on Microsoft, Oracle and Meta, three of AMD's named AI customers, so one buyer's news can move two contracts on one venue.

A perpetual never expires, so you can keep one open for as long as your margin holds out. BloFin listed the contract on May 6, 2026, with its trading hours given as 24/7 (source: BloFin listing notice). Stock futures on BloFin follow the share price while conveying no ownership interest, so holding AMDUSDT brings no AMD shares and no vote (source: BloFin stock futures notice).

A delayed data center, a smaller order or a new gigawatt agreement can all move AMD stock. That news can break at any hour. Outside regular weekday trading hours, and on weekends and market holidays, BloFin's stock futures may accept only orders that shrink or close a position, known as reduce-only orders (source: BloFin stock futures notice).

Say you hold an AMDUSDT long on a Saturday when a large AI lab announces a delay at one of its data centers. Until regular weekday trading hours resume, BloFin may limit you to reduce-only orders, so you could cut the position while any addition waits for Monday. Maintenance margin is the least margin that keeps a position open. If AMD's price moves far enough against you first, BloFin can partly reduce or fully liquidate the position once your margin, after any loss, falls to that level plus closing fees (source: BloFin liquidation rules).

If you expect a customer setback to hit AMD, you can short AMD on BloFin. A perpetual also has funding, a periodic transfer between long and short holders, which you pay or receive for the whole time the position is open. Leverage magnifies both gains and losses, so size any leveraged AMD trade for a sharp move against you on the day a customer reports.

Looking to trade the AMDUSDT Perpetual on news from AMD's customers? To get started, you'll need to first create a BloFin account, fund your account with cryptocurrency, and go to the AMDUSDT Perpetual trading page.

Frequently asked questions

Who is AMD's biggest customer?

AMD names no customer by revenue share, and in its latest annual report no customer reached a tenth of revenue. Lisa Su called OpenAI, Meta and Anthropic AMD's "strategic anchor customers" on AMD's second-quarter earnings call of August 4, 2026 and said all three were "in the process of ramping" (source: AMD Q2 2026 earnings call). She called the third quarter of 2026 the very beginning of that ramp, rising in the fourth quarter and again in the first quarter of 2027.

Does Microsoft use AMD AI chips?

Microsoft uses them on Azure. On July 20, 2026, it said AMD's Helios platform will power ND MI455X v7 virtual machines, built for the reasoning, search and agent workloads that modern AI services run (source: Microsoft). AMD's EPYC processors will power two more Azure families, HDv2 for data processing and HXv2 for chip-design software. In AMD's August 4, 2026 call, Su said Microsoft will deploy Helios at scale for frontier model inferencing across Microsoft, its AI customers and Azure AI services (source: AMD Q2 2026 earnings call).

How many AMD GPUs is Oracle buying?

Oracle's announced initial deployment is 50,000 AMD Instinct MI450 Series GPUs, set to start in the third quarter of 2026 and expand in 2027 and beyond. When the plan was announced on October 14, 2025, AMD said Oracle would be the first hyperscaler to offer a publicly available AI supercluster built on MI450 GPUs (source: AMD and Oracle). AMD said at the time that Oracle's MI355X supercluster can scale to 131,072 GPUs. AMD's announcement stated the MI450 order in GPUs and gave it no dollar value.

Which governments buy AMD AI chips?

Saudi Arabia and the United States are two named cases. HUMAIN, which AMD calls "a PIF company", after Saudi Arabia's Public Investment Fund, runs AMD GPUs and offers the computing as GPU-as-a-service (source: AMD, Cisco and HUMAIN). In the US, AMD said the Lux AI supercomputer at Oak Ridge National Laboratory would be the first dedicated US AI factory for science. Once fully deployed, Lux and Discovery will represent a combined $1 billion investment of public and private funding, and Discovery's user operations are expected in 2029 (source: AMD and Oak Ridge National Laboratory).

Does AMD rent computing capacity from cloud providers?

AMD has multi-year commitments to buy computing capacity from cloud service providers, listed in its quarterly report for the three months ended June 27, 2026. The providers can reduce, end or sell that capacity to others, and AMD's commitment then shrinks. AMD expects to use the capacity in its own operations or assign it to third parties. The commitments sit under noncancellable purchase orders and contracts, and canceling them would bring significant penalties (source: AMD 10-Q, Q2 2026).

Researched and written by the BloFin Academy editorial team with AI-assisted drafting. All facts independently verified. Primary sources include AMD's Form 10-K for fiscal 2025, its Form 10-Q for the second quarter of 2026, its Forms 8-K of October 2025 and February 2026, and its second-quarter 2026 results and earnings call. They also include a September 2026 Citi conference transcript, announcements by AMD, Oracle, HUMAIN, Microsoft, OpenAI, Broadcom, Nvidia, Anthropic, Qualcomm and TensorWave, and Nvidia's Form 10-Q for the quarter ended July 26, 2026. The rest are reporting by SemiAnalysis, NPR and Reuters, and BloFin's AMDUSDT contract information, listing notice, stock futures notice and liquidation rules, current as of September 28, 2026.

Nothing in this article constitutes financial advice, and nothing in it is a recommendation to buy, sell or hold AMD in any form. Most of AMD's announced AI demand rests on a handful of buyers whose commitments are largely "up to" ceilings, so a delay, a smaller order or a funding problem at one of them can move AMD's price sharply. That news can land outside regular weekday trading hours, on a weekend or on a market holiday, when BloFin may accept only orders that reduce or close an AMDUSDT Perpetual position. A leveraged position that moves against you then can fall to its maintenance margin and be partly or fully liquidated while you are only able to cut it. Past performance does not indicate future results. Do your own research and consider your risk tolerance before you trade on BloFin.