AMD AI and Data Center Story Explained: What Drives It Now

Over the last three years the companies building artificial intelligence have become some of the largest buyers of chips in the world, and the way the market describes a chipmaker has changed with them. A company that once made processors for gaming rigs and laptops, and spent a decade as the challenger closing a gap with Intel, can now be introduced in five words: it is an AI company.

For AMD, that sentence hides three claims: one about products, one about who is buying them, and a third about time that nobody says out loud. The products ship today and the buyers have signed, but the deals run to the end of the decade while the share price reprices every trading day, so the claim about time is the one still open.

What changed is not only what AMD makes but how it is sold. Most of what AMD sold used to be parts: processors and graphics chips that went into someone else's machine. It now signs agreements measured in gigawatts of electricity, with a handful of named buyers, running years into the future. It also designs whole racks, the tall frames that hold a data center's servers and networking gear stacked together, and customers build those racks from its chips.

Electricity is the reason the unit changed. A site can only draw the power its grid connection and its cooling allow, and that ceiling is fixed long before anyone chooses a chip. So a buyer working to that ceiling asks how much work a rack returns for each megawatt it can spare.

What is AMD's AI data-center business?

AMD's AI data-center business is the Data Center segment. It sells Instinct accelerator chips for training and running AI models, EPYC server processors and networking parts, with AMD's own ROCm software layer for programming its graphics processing units (GPUs). In the quarter ended June 27, 2026, it was AMD's largest segment by both revenue and profit.

An accelerator is a chip built to do one narrow kind of math very fast, and AI models are made of that math. The chip sits next to a general-purpose processor, the central processing unit (CPU), which runs everything else, and that pairing is why AMD sells GPUs and server CPUs into the same rack.

Those figures come from AMD's quarterly filing (source: AMD Form 10-Q, Q2 2026). The results release put the segment's growth at 107% from a year earlier (source: AMD second-quarter 2026 results).

A year earlier the same segment brought in $3.2 billion out of $7.7 billion. So it was 42% of AMD then and 58% now. The share moved because that one segment roughly doubled while the rest of the company grew slowly.

The profit line is the more startling one, and it has a dull explanation. Data Center made $2.1 billion of operating income in a quarter when the whole company made $2.0 billion. AMD reports a further $1.1 billion of costs outside its three segments, so segment profits are struck before those costs come out. What AMD stock is walks through the full split.

Read those two facts together and the shape of the company is clear. One segment now brings in most of AMD's revenue and more than all of its reported operating profit. AMD's share price is, to a large degree, a price on that segment's future.

How AMD got here: From MI300 to Helios racks

AMD's data-center products grew out of graphics. The Instinct line began as GPUs and became chips built for AI work, and AMD's server processors are sold under the EPYC name. It now supplies both, with networking and software, for a rack design called Helios that customers and their manufacturers build (source: AMD Form 10-Q, Q2 2026).

The annual report lists the Instinct family as the MI200, MI300, MI325 and MI350 series, all built on an architecture AMD calls CDNA. It describes those parts as built for AI training, AI inference and exascale-class scientific computing (source: AMD Form 10-K, fiscal 2025). Graphics chips turned out to be good at the math AI needs, and AMD built a separate line for that work.

AMD's annual report names faster growth in the Data Center segment as a key priority for 2025. It says demand for its AI accelerators was strong as large hyperscale customers, original equipment manufacturers (OEMs) and original design manufacturers (ODMs) deployed its Instinct MI350X Series GPUs (source: AMD Form 10-K, fiscal 2025). A hyperscaler runs data centers at the scale of a Microsoft or a Meta. An OEM builds and brands the finished server, while an ODM builds it for someone else to brand. Add the AI labs that now sign with AMD directly and a few specialist cloud providers, and that is the customer base for a chip like this. There are not many of them.

The filing then describes a schedule. AMD says it moved its Instinct GPUs onto a yearly release cycle, beginning with the MI350 Series in 2025, and it announced the next-generation MI355X for large-scale AI deployments. A yearly release schedule is a commercial promise as much as an engineering one. A customer planning a data center wants to know what will be on sale when the building opens. Each generation on AMD's GPU roadmap has its own launch window, and a customer plans the build against those dates.

EPYC, ROCm and the turn to rack designs

Two things sit either side of the accelerators, and the first is the server processor line: AMD launched its 5th Gen EPYC server processors in 2025, and those are the general-purpose chips that run the rest of the machine. The second is software for the GPUs, where AMD says it keeps investing in its ROCm platform and open ecosystem, adding features for AI training and inference (source: AMD Form 10-K, fiscal 2025). ROCm is the layer a developer writes against, and it decides whether a team can move an existing model onto AMD hardware in an afternoon or a quarter. Fast chips that are awkward to program get bought slowly, which is why the software line in a chip company's filing is worth as much attention as the product line.

The turn to racks is where the product story ends up. In 2025 the company previewed Helios, a rack-scale platform that puts all of its data-center products in one rack: processors, accelerators and the networking between them. AMD supplies the chips and licenses the design. Customers and their manufacturing partners build the finished rack. Designing the rack means AMD competes on what the whole rack delivers for the electricity it draws. That is why its deals are now measured in gigawatts, and its claims in output per rack.

What a gigawatt is worth, and when it arrives

A gigawatt is a billion watts of electrical power, the amount a site draws at any moment. AMD's largest AI agreements are written in gigawatts, so a buyer commits to power rather than to a number of chips. Three buyers have announced about fourteen gigawatts, and AMD puts the revenue from each gigawatt in double-digit billions.

That is worth pausing on, because it changes what a deal means. A gigawatt says nothing directly about how many chips were bought. It describes how much power the buyer's sites can draw. Two buyers with the same gigawatt figure could take different numbers of racks, in different years, at different prices. So a gigawatt is a budget for electricity, and the revenue attached to it is a separate question.

AMD has given one public answer to that question. Asked on the second-quarter call, chief executive Lisa Su said: "In terms of the revenue per gigawatt, I think we've said double-digit billions." She added that AMD was still in that range (source: AMD Q2 2026 earnings call, August 4, 2026). Double-digit billions means somewhere between $10 billion and $99 billion. That is a wide range, and it is the only per-gigawatt figure AMD has put its name to.

Now the arithmetic, which is simple and needs its caveats kept close. Fourteen gigawatts have been publicly announced, and at the very bottom of that range it is about $140 billion of revenue. Set that beside a segment booking $6.7 billion a quarter and it comes to roughly five years of the current run rate, spread across a ramp that reaches into 2031.

A large number is not the same thing as a large share price, and this is where the comparison gets interesting. Writing on September 12, 2026, with AMD's market value near $870 billion, one analysis worked backwards from the price. Suppose the market eventually pays a mature 25 times earnings, it argued, down from a forward multiple of about 33 at the time. AMD would then need about $35 billion of annual profit just for its current price to hold (source: The Motley Fool, September 12, 2026). Turning profit into revenue takes one more figure, and the same piece supplies it. AMD kept about 24 cents of every revenue dollar as adjusted profit in the second quarter, so $35 billion of profit would need roughly $145 billion of annual revenue (source: The Motley Fool, September 12, 2026).

Put that beside the gigawatts. If AMD's range means revenue per gigawatt delivered, not per year, then on those September 12 assumptions the $140 billion from the announced gigawatts is about one year of the revenue the price would need, spread across five or six. The same analysis reaches the point another way: on its reading, the price assumes AMD books about 7% of a $2 trillion market by 2030 (source: The Motley Fool, September 12, 2026). Either route says the same thing. The announced deals are big enough to matter and too small, by themselves, to explain what the shares cost.

Where the fourteen gigawatts come from

Three announced agreements carry the capacity, and each is a ceiling rather than a schedule.

The OpenAI agreement began the sequence in October 2025. The annual report records a product purchase agreement for OpenAI to deploy 6 gigawatts of AMD GPUs, and the filing made at signing sets out the vesting (source: AMD Form 8-K, October 6, 2025). The Meta agreement repeated that structure four months later, in nearly identical language (source: AMD Form 8-K, February 24, 2026, Item 1.01).

The capacity figure and the build schedule for Meta come from the announcement AMD published the same day, which expects shipments for the first gigawatt to begin in the second half of 2026 (source: AMD and Meta announcement, February 24, 2026). The Anthropic agreement came in July and runs the other way on equity. It carries no warrant. AMD has committed to a strategic equity investment of up to $5 billion in Anthropic at a future date, and the first gigawatt is due to deploy from the first half of 2027 (source: AMD newsroom, July 22, 2026).

Which of the fourteen gigawatts AMD's customers have committed to buy

Now attach the caveats, because they do most of the work. Twelve of the fourteen gigawatts are ceilings. AMD's quarterly filing puts it plainly for the two largest deals, saying each customer intends to deploy up to six gigawatts (source: AMD Form 10-Q, Q2 2026). The Anthropic capacity was announced the same way. The exception is the first gigawatt of the OpenAI and Meta agreements: each buyer has signed a binding commitment for that much, which makes two committed gigawatts out of fourteen announced. AMD has disclosed no binding commitment for Anthropic's first gigawatt.

The per-gigawatt figure has its own limits. AMD has not said whether its per-gigawatt range is what a gigawatt of capacity brings in over its life or in a year, and it has never split the figure by customer or by year, so treat it as an order of magnitude. And the deliveries are slow. The first Meta gigawatt was due to start shipping in the second half of 2026, and Anthropic's first is not due until the first half of 2027. So the model that fits is a ramp measured in years with a ceiling on top, in which any single quarter is a thin slice.

AMD's 2030 market figures, and what each one counts

AMD's managers used several 2030 market figures between July 23 and September 8, 2026, and they are easy to quote interchangeably. They describe three scopes: the data center AI accelerator market, the server processor market, and AMD's own total addressable market, meaning the yearly spending on every kind of chip AMD sells.

All three came up on the August 4 call, and Su gave each a growth rate as well as a destination. The accelerator market is growing "more than 45% annually to approximately $1.4 trillion by 2030". The server CPU market is growing "more than 50% annually to approximately $220 billion by 2030". The wider picture is "the overall market for high performance in AI computing growing approximately 40% annually over the next several years, approaching $2 trillion by 2030". She finished that sentence by saying AMD expects to grow well above the market (source: AMD Q2 2026 earnings call, August 4, 2026).

AMD's own total addressable market is where the figures get confused. AMD's release for its July 23 Advancing AI event put it at about $2 trillion in 2030, across data center, PCs, edge and embedded processors (source: AMD, July 23, 2026). At Citi's conference on September 8, AMD's chief financial officer said the company had updated its market opportunities at its July event "to now $2 trillion in 2030" (source: Stock Analysis, AMD Citi conference transcript). The Motley Fool reported the same remarks as $2 trillion to $3 trillion (source: The Motley Fool, September 12, 2026). The transcript records only the $2 trillion figure. Add the $1.4 trillion accelerator market to it and you count that market twice, because the $2 trillion already contains it.

The figure Su gave, approaching $2 trillion, and the $2 trillion Hu repeated are the same market under two names. Asked on the August 4 call how fast AMD could grow against that market, Su described it as spanning server CPUs, data center AI, embedded and PCs, the same areas AMD's July release lists, and put its growth above 40% a year.

When an AMD market figure appears in a headline, ask three things first: who said it, on what date, and what does it count. Each figure survives that test on its own, and adding the accelerator or the server figure to the $2 trillion counts it twice.

Try it on the most quoted of them. The $1.4 trillion is the data center AI accelerator market, and it says nothing about AMD's share of that market, so it is not a figure about AMD's revenue at all.

What AMD says its own plan is

Two more statements from the Q2 call connect the size of those markets to AMD's own accounts, and the first is about the segment itself. In her prepared remarks Su said AMD now expects server processor revenue to grow more than 80% from a year earlier in the second half of 2026 and more than 70% in 2027. She added that more and larger Helios and MI450 deployments put data-center AI on course for growth that speeds up in 2027. Taken together, AMD expects Data Center segment revenue to more than double in 2027 (source: AMD Q2 2026 earnings call, August 4, 2026). The second is about the company. Su said AMD is tracking materially ahead of the long-term financial model it shared at its Financial Analyst Day in November 2025, and she named the two targets it is ahead of. AMD now expects revenue to grow substantially faster than its earlier target of more than 35% a year, and to significantly exceed its $20 annual EPS target within its strategic timeframe. EPS is earnings per share: the profit divided by the shares outstanding. That is management describing its own plan, and a plan about earnings only tells you what the shares are worth once you settle what multiple to pay for those earnings.

Inference, training, and why AMD talks about tokens per dollar

Two jobs happen on this hardware, and their economics differ. Training is the one-off, very expensive job of building a model from data. Inference is running the finished model to answer a request, which happens every time someone asks a chatbot a question. Training is a project. Inference is a utility bill, and it keeps arriving.

That difference explains a change in vocabulary: the industry used to sell raw speed, and now it sells cost per unit of output. A token is the unit a model reads and writes, roughly a fragment of a word, so tokens per dollar measures how much usable output a rack produces for each dollar spent running it. Once a model serves millions of requests, the operator stops asking how fast it trains and starts asking what each answer costs.

AMD has picked its ground there. On the second-quarter call Su made a bounded claim: "Across a broad range of inferencing workloads, Helios delivers up to 15% more throughput at the same rack power and up to 30% more tokens per dollar than the competition" (source: AMD Q2 2026 earnings call, August 4, 2026). Three details in that sentence separate a benchmark from a slogan. It is scoped to inference, and both figures are "up to", which is a best case. And the throughput comparison holds rack power fixed, which is the comparison that counts when electricity is the thing you have run out of.

Say you are choosing between two racks and your site can spare one megawatt. The question is no longer which chip is quicker. It is how many answers each rack returns for that fixed megawatt, and what each of those answers costs you. On AMD's own figures, a Helios rack returns up to 15% more work for the same power. That is a claim a customer can measure on their own hardware. It is still a vendor claim, made by the vendor, on the vendor's earnings call, and it should be read as one.

AMD can make a rack-level claim at all only because it designs the Helios rack. Output at a given power draw depends on processors, accelerators, networking and cooling working together, so only the designer of the whole rack can speak for it. Whether a customer reaches those numbers then depends on the software, and ROCm against Nvidia's CUDA is the contest that decides it.

AMD Instinct revenue: What AMD reports, and what it leaves out

AMD reports one Data Center number. Search the segment note in its annual report for a separate Instinct figure, or a GPU line, or any split between accelerators and server processors, and there is nothing there. The filing gives the segment total and stops (source: AMD Form 10-K, fiscal 2025).

So anyone quoting you a precise figure for AMD's AI accelerator revenue is estimating, however confident the chart looks. Those estimates start from the segment total and divide it between GPUs and CPUs using channel checks, supply-chain reports and assumptions about pricing. They can be reasonable, and they can be badly wrong. You have no way to grade them, because the one company that could settle the question reports a single line.

That leaves three things you can measure, and they are enough to hold a view with.

The first is the segment total and its year-over-year change, which is audited. It covers accelerators, server processors and networking together, so the mix is blurred, but the number itself cannot be argued with.

The second is the guide AMD publishes for the coming quarter. For the third quarter of 2026 that was "approximately $13 billion, plus or minus $300 million" for the company as a whole (source: AMD second-quarter 2026 results). It is AMD's own near-term estimate, and it gets read harder than the result itself, as what moves AMD stock on any day explains.

The third is the shipment language in each announcement. Meta's first gigawatt was due to begin shipping in the second half of 2026. A commitment that has started turning into shipments carries evidence that a signature on its own does not.

Watching those three is less satisfying than one accelerator revenue figure. It has one large advantage: every one of them comes from AMD, not from someone modeling AMD.

Risks to AMD's data-center growth

The case for AMD's data-center business rests on four supports. They are a customer base built on three announced gigawatt-scale buyers, a horizon running to 2030 and 2031, a software layer competing with Nvidia's CUDA, and export rules AMD cannot schedule. Each has its own way of failing.

The customer base is the first support and the most obvious. Who buys AMD's AI chips matters as much as how many it sells. Three announced agreements account for all of the announced capacity, and the buyers at this scale are a small group: hyperscalers, the server makers that build for them, and the companies training frontier models. Concentration cuts both ways. One signature adds gigawatts, and one delay removes them, with no broad base of small buyers to cushion the difference.

The horizon is the second, and it catches people who trade on quarterly results. The two warrants attached to these agreements expire in October 2030 and February 2031. So a quarter that lands short tells you a great deal about the share price that week and very little about whether fourteen gigawatts eventually ship. The reverse trap costs just as much: a quarter that beats expectations does not pull 2031 forward.

The software ecosystem is the third, and it is the support AMD controls least. ROCm competes with Nvidia's CUDA, the programming layer a generation of AI researchers learned first. A customer choosing hardware is also choosing how much of its existing code it will have to rewrite. AMD keeps naming software investment in its filings for exactly that reason.

Export rules are the fourth, and AMD's own annual report says that "an inability to obtain the required export licenses, or to predict when or pursuant to which conditions they will be granted, increases the difficulties of forecasting shipments" (source: AMD Form 10-K, fiscal 2025). That is a company telling you a policy decision can reprice its forecast without warning. An export-control decision has already cost AMD about $440 million once, with no such charge in 2024 or 2023 (source: AMD Form 10-K, fiscal 2025). That one charge is what makes China and export-control risk concrete.

The risk outside AMD's control

One risk sits with the buyers, and it is easy to miss because AMD cannot report on it. The capacity in those agreements gets built only while the customers keep spending, and their capital plans are theirs to publish. The contracts make that visible: the Meta and OpenAI warrants vest in full only when six gigawatts have been bought, so AMD's upside depends on purchases nobody has made yet. Two further conditions sit on top of that in both filings. Vesting also depends on AMD's own share price reaching levels set in the contract. And each tranche that does vest is "subject to the fulfillment of certain other technical and commercial conditions" before it can be exercised (source: AMD Form 8-K, October 6, 2025). So a milestone that has been hit does not automatically become stock. That structure shows how both sides saw the risk when they signed. On the two warrant deals the buyer's equity vests only as it buys capacity. On the Anthropic agreement the equity moves the other way: AMD has committed to invest up to $5 billion in the customer, at a future date the announcement does not fix.

Before you read a weak quarter as the thesis breaking, check which of the four supports it touched. A soft guide with the agreements intact is a timing story. A customer stepping back from committed capacity is a different event, and so is a license decision that removes a market. The four supports fail in different ways, so name the one a quarter touched before you judge it.

Trading the AMD data-center thesis on a perpetual

BloFin lists the AMDUSDT Perpetual, which follows an index of AMD's share price rather than the listed share itself, and conveys no ownership of it. Its contract value is 0.01 AMD, its tick size is one cent, and it offers up to 20x leverage with settlement in USDT. It has been live on the venue since May 6, 2026.

There is a mismatch here worth taking seriously. The thesis is dated 2030 and 2031, and the contract settles funding every eight hours, three times a day.

Funding is a payment passed between the long and short sides: when the rate is positive, longs pay shorts, and when it is negative, shorts pay longs, so the sign tells you which side is crowded. The AMD contract settles it at 00:00, 08:00 and 16:00 UTC. Across the thirteen settlements from 16:00 UTC on September 19 to 16:00 UTC on September 23, 2026, the rate ran from negative 0.0120% to positive 0.0619% and averaged positive 0.014808% (source: BloFin, AMDUSDT funding data, which labels each settlement in your own timezone rather than UTC). The settlements at 00:00 and 08:00 UTC on September 23 each paid longs 0.0120%, and the one at 16:00 came back to positive 0.0008%.

Say you hold a $5,000 long. The window's average rate was positive, so at that average you would pay about $0.74 a settlement, or $2.22 a day. Carried at the same average for a year, which is 1,095 settlements, that comes to roughly $811, or about 16.2% of a $5,000 position. The venue's cap and floor on this contract are plus and minus 3% a settlement, far outside anything in that window. What matters is the order of magnitude: carrying a multi-year view on an eight-hour contract is not a rounding error. A rate that stays one side of zero for weeks costs far more than any single settlement. That accumulation is the whole of funding on the AMD perpetual.

The trading windows matter for a specific reason: some of the news that moves this thesis arrives outside US market hours, and quarterly results reliably do, because AMD publishes them after the close. Other news lands inside the session, as the chief financial officer's conference remark did on September 8, 2026. Outside regular weekday hours, and on weekends and market holidays, BloFin may limit stock futures trading to reduce-only orders, which let you close a position but not open one (source: BloFin stock futures trading update, April 21, 2026). So the price can move on that news before you can open anything new. A leveraged position carried through one of those windows meets the new price at once. That is how an overnight or weekend gap liquidates what a gradual move would not.

If you would rather hand the position to a trader with a record, check the instrument list first: on September 25, 2026, AMDUSDT was missing from the markets BloFin's copy trading covers.

A higher leverage setting narrows the price move a position survives, so how much leverage to take decides how much room the thesis has.

Looking to trade the AMD data-center story on the AMDUSDT Perpetual? To get started, you'll need to first create a BloFin account, fund your account with cryptocurrency, and navigate to the AMDUSDT Perpetual trading page.

Frequently asked questions

Does AMD do data centers?

Yes, and by revenue it is now AMD's largest business. AMD reports its data-center products as a single segment, covering accelerators, server processors and graphics processors together with networking parts, so its data-center disclosure describes a bundle. AMD expects the first Helios-based systems, built by customers and their manufacturing partners from AMD's chips, to start shipping in the second half of 2026 (source: AMD Form 10-Q, Q2 2026).

How much of AMD's revenue is data center?

58% in the quarter ended June 27, 2026, from $6.7 billion of $11.5 billion in total revenue, against 42% a year earlier. What that leaves is worth naming, because AMD has two other segments and neither is small: in the same quarter Client and Gaming brought in $3.8 billion and Embedded $977 million. Over the six months to June 27, 2026, Data Center revenue was $12.5 billion, up 81% from a year earlier (source: AMD Form 10-Q, Q2 2026).

Which AMD products go into a Helios rack?

Four AMD product lines go into it. On the second-quarter call AMD described Helios as its rack-scale AI platform combining EPYC Venice processors, MI450-Series GPUs, Pensando networking and ROCm software (source: AMD Q2 2026 earnings call, August 4, 2026). MI450 is the accelerator the OpenAI and Meta first gigawatts are built on, and Pensando is the networking business AMD agreed to buy for about $1.9 billion in April 2022 (source: AMD, April 4, 2022). Helios is the rack at the center of AMD's MI450 roadmap, and the customer, working with its manufacturing partners, builds the finished rack from those parts.

How much revenue does a gigawatt of AMD GPUs bring?

AMD's chief executive puts it in the double-digit billions of dollars of revenue per gigawatt, anywhere from $10 billion to $99 billion. She gave the range on the August 4, 2026 call when an analyst asked about AMD's monetization per gigawatt, so it describes what AMD earns from capacity its customers deploy, not the cost of a data center. She also set out when that revenue starts: the third quarter of 2026 is "the very beginning of the ramp", with a step up in the fourth quarter and another in the first quarter of 2027 (source: AMD Q2 2026 earnings call, August 4, 2026).

Does AMD sell more GPUs or CPUs to data centers?

AMD's filings do not answer this, and the question is narrower than the segment. Beyond the accelerators, server processors and GPUs the annual report's segment definition opens with, the same segment also covers accelerated processing units, data processing units, AI network interface cards, field programmable gate arrays and adaptive system-on-chip products (source: AMD Form 10-K, fiscal 2025). Eight categories in all, so a GPU-against-CPU split would leave most of the list unaccounted for. One Data Center total is reported and no split of any kind is given, which makes every precise figure you meet somebody's estimate.

Is AMD's AI business bigger than Nvidia's?

Nvidia's is far larger. Nvidia reported $89.0 billion of Data Center revenue for its quarter ended July 26, 2026 (source: NVIDIA second-quarter fiscal 2027 results). That is about 13 times the $6.7 billion AMD's Data Center segment brought in during its quarter ended June 27, 2026, a month earlier (source: AMD Form 10-Q, Q2 2026). AMD's own annual report says Nvidia uses its position in data-center GPUs, its financial resources and its proprietary software ecosystem to promote its systems and influence AMD's customers (source: AMD Form 10-K, fiscal 2025). The two segments bundle different products, so the ratio is a rough measure, and AMD against Nvidia sets out what can be compared directly.

Researched and written by the BloFin Academy editorial team with AI-assisted drafting. All facts independently verified. Primary sources include AMD's second-quarter 2026 results release, its Form 10-Q for the quarter ended June 27, 2026, its Form 10-K for fiscal 2025 and its Forms 8-K of October 6, 2025 and February 24, 2026 filed with the SEC, the transcript of AMD's second-quarter 2026 earnings call, the Stock Analysis transcript of AMD's Citi conference session, AMD's announcements of April 4, 2022 and February 24, July 22 and July 23, 2026, NVIDIA's second-quarter fiscal 2027 results, The Motley Fool's reporting of September 12, 2026, and BloFin's stock-futures trading notice and AMDUSDT funding data, current as of September 25, 2026.

Nothing in this article constitutes financial advice, and nothing in it is a recommendation to buy, sell or hold AMD in any form. Management's estimates of future markets and of revenue per gigawatt are reported here as the company's own statements, and none of the announced capacity is guaranteed to ship on the schedule given. Carrying a view that runs to 2030 on the AMDUSDT Perpetual means paying or receiving funding every eight hours for the whole period, and a leveraged position can be liquidated by a price move far smaller than the leverage multiple suggests. Past performance does not indicate future results. Do your own research and consider your risk tolerance before you trade on BloFin.