Advanced Micro Devices

AMD designs processors and has them made by TSMC in Taiwan, which manufactures all of its advanced chips. It sells three kinds of thing. Data-centre chips: EPYC server processors, the Instinct accelerators used to train and run AI models, and networking parts. Client and gaming chips: Ryzen processors for PCs and laptops, Radeon graphics cards, and the custom chips inside the PlayStation 5 and Xbox. And embedded chips, mostly from Xilinx, which it bought in 2022, for cars, factories, networks and aircraft.

It is one of the great turnarounds in American business. When Lisa Su became chief executive in October 2014, AMD was losing money, its server business had almost disappeared and its shares traded for a few dollars. Under Su it built a better processor than Intel's, took a third of the PC and server market from a company that once held almost all of it, bought Xilinx, and entered the AI accelerator market as the only real alternative to NVIDIA's general-purpose chips. In 2025 it took in $34.6bn; this year, on its own guidance, it will take in about $52bn.

In the third quarter of 2026 the data centre was 64% of revenue, up from 48% in 2025. AMD is no longer mainly a PC company. Its fortunes now ride on the same thing as NVIDIA's: how much the world's largest companies and AI laboratories spend on computing, and how they finance it.

Few large companies have had a year like AMD's shares. In February the first-quarter guidance disappointed and the stock closed at its low, $190.95, on 3 March. Then came a deal with Meta for up to six gigawatts of AI chips, a quarter in May that beat expectations by enough to lift the shares 18.6% in a day, and a summer in which every AI laboratory seemed to want a second supplier. On 6 October the shares closed at $649.42.

Then the AI trade turned. Worries about how the build-out would be financed, then the midterm elections and promises of data-centre moratoriums in several states, took the semiconductor sector down; on 5 November AMD fell 17.3% in a day, to $404.40, the day after reporting a quarter that beat every estimate. On 9 November reports of an investigation at NVIDIA sent investors looking for alternatives: AMD rose 13.3% that day. NVIDIA's chief executive resigned on 11 November, and by 12 November AMD had closed at $585.30, 45% above its low of a week earlier. On 13 November it closed at $534.06, −8.8% on the day; we found no company news to explain the move.

Why this matters to an owner. A share price that moves 17% in a day on news that does not change the company's earnings is telling you the market has no firm idea of what the business is worth. That is not a reason to avoid it. It is a reason to be very clear, before buying, about what you think it is worth, because the market will not tell you.

What the 13 November fall says. The day had a pattern. The large semiconductor index fund (SOXX) fell 3.6%, and Broadcom, the other chipmaker bid up on NVIDIA's leadership news, fell 4.8%. NVIDIA itself fell only 2.0%, and the Nasdaq 100 tracker 1.2%. AMD fell −8.8%, more than twice the sector fund, on a day with no filing, no change to guidance and no customer announcement from the company. In the five sessions before, AMD had added 45%, about $180 a share; Friday took back more than a quarter of that. The money that ran into AMD as "the alternative to NVIDIA" began to run back out two trading days after the resignation, and nothing in AMD's order book had changed in either direction.

An owner can draw two lessons from that. First, in a week like this one a large part of AMD's price is a bet about NVIDIA, not about AMD. The warrants, the leases and the gigawatt contracts examined below are the facts about AMD, and none of them moved on Friday. Second, swings like these help a patient buyer only if that buyer has a number written down before the swing arrives. Ours is in the valuation chapter, and at $534.06 the price is still well above it.

The third quarter, reported on 3 November, was excellent. Revenue was $13.38bn, up 45%; the data centre $8.55bn, up 97%, as volume shipments of the new MI450 accelerators began and the new "Venice" server processors started shipping. The adjusted gross margin was 56%. For the fourth quarter AMD guided $17.0bn, plus or minus 3%, up about 66%, with the data centre up about half again in one quarter. AMD has beaten the midpoint of its own guidance in each of the last four quarters.

2027. On the call, according to a third-party transcript, Su gave "markers" rather than guidance: revenue up more than 80% next year, server revenue doubling, and the data centre above 80% of revenue. These figures do not appear in the release, and we treat them as management's ambition rather than a forecast. Analysts have taken them at face value: their 2027 revenue consensus is about $96bn.

Two earnings per share. AMD earned $1.46 a share under accounting rules and $2.09 on its own adjusted measure. The difference was $1.04bn of after-tax items: $523m of stock-based pay, $534m of amortisation of the Xilinx and other acquisitions, and a $1.1bn non-cash charge from revaluing the warrants given to OpenAI and Meta, offset by a large tax benefit. We count stock pay as a real cost: it is paid in shares instead of cash, and the share count shows it. Amortisation of past acquisitions is a genuine non-cash item. The warrant charge is real too, in a way we explain in Part VI.

Where the strain shows. Free cash flow in the quarter was only $554m, down from $1.56bn, because inventory rose to $11.5bn and AMD prepaid suppliers to secure wafers, memory and packaging. Inventory now stands at about 177 days of cost of sales, against 159 a year earlier. Su said on the call that "demand is higher than supply". That is the right problem to have, as long as the demand is as solid as the commitments AMD is making to meet it.

Two businesses inside the data centre. AMD does not say how much of its data-centre revenue comes from server processors and how much from AI accelerators. It said on the call that accelerators will exceed server CPUs in the fourth quarter, which on our arithmetic puts accelerator revenue above $6bn a quarter, a run-rate of more than $25bn a year. That matters for the margin: EPYC processors, sold into a market AMD has been winning for years, carry higher margins than GPUs sold against NVIDIA. The fourth-quarter guidance shows the mix already working against the gross margin, which is guided down from 56% to about 55%.

AMD has two businesses with very different moats. In processors it is half of a duopoly. Every PC and almost every server in the world runs on the x86 instruction set, and only Intel and AMD have the licences, the designs and the decades of compatibility to make competitive x86 chips. AMD's share has risen for eight years: by Mercury Research's count about 35% of desktop units and about 35% of server units in the second quarter of 2026, and about 46% when only EPYC is set against Intel's Xeon. Arm-based chips from Apple, Amazon and others are the long-term threat; inside x86, AMD's position is strong.

In AI accelerators AMD is the challenger. NVIDIA's advantage is not the chip but everything around it: CUDA, the software almost every AI developer writes for, the networking that links thousands of chips, and complete racks. AMD's answer is ROCm, its own software, now at version 10, and Helios, a 72-GPU rack built with partners. Helios racks are running at Anthropic, Meta, Microsoft, OpenAI and Oracle. But AMD's 10-Q is careful to say that it does not manufacture or sell the complete rack, and its own releases describe ROCm's progress, not parity. AMD's gross margin, 56%, is 19 points below NVIDIA's 75%. That is the price of being second.

Why customers want AMD. The strongest argument for AMD is not that it beats NVIDIA but that the largest buyers of AI chips do not want to depend on one supplier. This week showed why: a company whose fortunes were tied to one leader lost that leader overnight. Being the credible second source in a market of several hundred billion dollars a year can be an excellent business. It is not a moat in Buffett's sense, because the customer, not AMD, holds the power: the same laboratories that buy MI450s also run Google's TPUs, Amazon's Trainium and NVIDIA's racks, and design their own chips.

Three lessons from the events of 8–12 November apply directly to AMD. First, key people. NVIDIA's strategy had been one person's for thirty-three years, and the market took about $490bn off its value in the week that person left. AMD is also identified with one chief executive, who has run it for twelve years and is chair of the board. AMD's bench is deep (Papermaster, Norrod and Hu have run their businesses for years), but owners should note the parallel.

Second, financing. About 72% of NVIDIA's fall from its October high came before the scandal, from worries about how AI data centres would be paid for and whether politics would slow them. AMD fell further in the same weeks (−38% from its high to 5 November against NVIDIA's −24% to the same day), because the second supplier is more exposed to a pause than the first. The same forces that hurt NVIDIA in October will hurt AMD more if they return.

Third, the substitute is not the winner. When investors sold NVIDIA they bought AMD and Broadcom at the same time. That is a reminder that AMD's AI business is valued, in part, as an alternative to NVIDIA rather than for its own earnings. Alternatives are valuable while the leader is in trouble. Customers, however, buy on performance, software and supply, and none of those changed this week.

The warrants. In October 2025 AMD agreed to supply OpenAI with up to six gigawatts of AI chips over several generations, starting with one gigawatt of MI450 in the second half of 2026; its CFO spoke of "tens of billions of dollars in revenue". In February 2026 it signed a similar deal with Meta. In each case AMD gave the customer a warrant for 160 million AMD shares at one cent. Together that is 320 million shares, 19.5% of those outstanding. They vest in tranches as the customer buys gigawatts, as AMD's share price passes hurdles that rise to $600, and as the customer meets technical and commercial conditions. The detailed schedules are redacted in the filings. None had vested by 26 September, though the 10-Q says "most market conditions had been met".

How to think about them. In plain terms, AMD is paying for orders with its own shares. If the deals deliver what they promise, the customers will have bought tens of billions of dollars of chips and received, at a $600 share price, up to $190bn of AMD stock between them. The accounting tries to capture this. The fair value of the warrants at the time of the deals, $4.5bn still to be recognised, will be deducted from future revenue as the chips are sold; the change in their value since is booked each quarter as a non-cash charge outside operating profit, $1.1bn in the third quarter, and management expects it to "likely grow". The adjusted figures leave both out. An owner should not: if the warrants vest, today's shareholders own 16% less of the company, and every per-share figure in the forecasts shrinks with them.

The investments and guarantees. AMD has also bought into its customers. It is committed to invest up to $5bn in Anthropic, which has agreed to buy up to two gigawatts of MI450s; it put $1bn into Meta's share offering in October. And it has taken on obligations to make the data centres happen: $13.3bn of data-centre leases not yet started, which it intends to pass to partners; $8.5bn of guarantees of partners' leases and debt; and $6.2bn of cloud capacity it must buy if it cannot place it with others. Nothing has been paid under the guarantees and no partner is in default. On top of that, purchase commitments to suppliers rose from $12.2bn at the end of 2025 to $48.8bn, mostly wafers and memory. Against all of this AMD has about $8.7bn of net cash.

The pattern is familiar. We wrote in September that NVIDIA was financing part of its own demand. AMD is doing the same thing from a weaker position: as the second supplier it has had to offer more, in equity rather than in cash, to win the orders. None of this is hidden; it is in the 10-Q. But the revenue in AMD's forecasts now depends, to a degree that is hard to measure, on the continued ability of a few AI laboratories and their partners to raise money.

As an operator, Su has few equals. AMD has delivered a new generation of processors every year for nearly a decade, entered and grown in AI accelerators against the most dominant competitor in technology, and beaten its own guidance four quarters running. The board and shareholders agree: Su was re-elected with 93% of the vote and pay was approved with 92%.

As a capital allocator, the record is written in shares. Xilinx was bought in 2022 for $48.8bn, mostly by issuing 429 million new shares. ZT Systems was bought for $4.4bn in 2025 and its manufacturing arm sold on for $2.4bn plus earn-outs. Two more deals are pending: World Labs, about $8.2bn, all in shares, and Taalas, about $2.9bn. Then the warrants, worth up to 320 million shares. Stock-based pay runs at about $2bn a year. Against that, buybacks have been small: AMD bought 1.1 million shares in the first nine months of 2026, while it spent $1.8bn withholding shares for employees' taxes. Since 2021 the weighted share count has risen 35%.

The $600 coincidence. The final hurdle on the customers' warrants is $600 a share. So is the top hurdle on the chief executive's new five-year award. We do not suggest anything improper; aligning a chief executive with shareholders is what an award should do. But owners should know that the same price triggers a large transfer of shares to customers and the largest payout to management.

On the balance sheet AMD is strong: $15.8bn of cash and short-term investments against $7.1bn of debt, none of it due before 2028, and $10.5bn of undrawn credit and commercial paper. Off the balance sheet the picture is the one in Part VI: about $77bn of purchase commitments, leases, guarantees and capacity, plus up to $5bn of investment commitments, against roughly $68bn of shareholders' equity, much of which is Xilinx goodwill. Both pictures are true. The first tells you AMD can pay its bills; the second tells you how much of its future it has committed to one market.

AMD pays no dividend and has said nothing to suggest it will. That is the right choice for a company growing this fast. The question for a shareholder is instead whether each share is becoming more valuable, and here the record is mixed: the business has grown enormously, but the number of shares it is divided among has grown by about a third in five years and may grow by a fifth again. The buyback programme, $9.2bn of authority left, has mostly offset stock pay rather than reduced the count.

The model. We estimate revenue in 2030, apply an owner-earnings margin after stock pay and tax, divide by a share count that includes the warrants that would vest in each case, put a multiple on the result, discount 4.1 years at 10% and add net cash. The forecasts on the screens assume AI spending compounds without a pause; we use them only for the third case.

The middle case moves to about $270 at 18 times and $385 at 26 times; to $265 if 2030 revenue is $130bn and $410 if it is $200bn; to $290 at a 22% margin and $365 at 28%. The range across the three cases runs from about a fifth of today's price to above it. That is what "too hard" looks like.

What is priced in? At $534.06, on our middle margin, share count and multiple, AMD needs revenue of about $260bn in 2030: about 50% a year from this year's $52bn, for four years. AMD's own long-term target, set a year ago, was more than 35% a year from 2025. The price is not paying for AMD's plan. It is paying for the plan being beaten by a wide margin for five years in a row, in a market where every customer is also building its own chips.

A number for readers inside the circle. We do not think AMD can be valued with the confidence we ask for: the range is too wide and depends on things nobody can know. For a reader who disagrees and wants a price, ours would be about $230, 30% below the middle case, about 13 times the 2027 adjusted consensus. At that price a pause in AI spending would be paid for rather than suffered.