Haverhill Financial Research · October 12, 2026
Summary: Haverhill price target $605 (12 months)
AMD has become the most credible challenger to Nvidia for AI data center spending: its Data Center sales more than doubled in a year to $6.7 billion in Q2 2026. OpenAI, Meta and Anthropic have signed multi-year deals for up to 14 gigawatts of Instinct GPUs between them, and Oracle has ordered 50,000, with the first large shipments due from the second half of 2026. At $608.10 the stock already prices in most of that ramp, at about 36 times the $17 of 2028 earnings per share in our base case. Our 12-month price target of $605 (through October 2027) sits about 1% below the current price, so we rate the shares Hold.
AMD filings through the Q2 2026 Form 10-Q; Haverhill Financial Research model; price as of Nasdaq close Oct 9, 2026
AI is now the growth engine. Data Center revenue rose 107% to $6.7 billion in Q2 2026, 58% of sales, and AMD's CFO expects Data Center growth to accelerate in the second half.
The orders have names on them. OpenAI (6 GW), Meta (6 GW) and Anthropic (2 GW) signed multi-year Instinct deals, with first shipments from late 2026 into 2027. The cost was warrants for up to 320 million AMD shares.
Our view: the business is in far better shape than a year ago, but at about 36 times our 2028 earnings estimate the price leaves little room for a slow ramp of the new MI450 chip.
Investment thesis in brief
Our thesis: AMD will win a real second-supplier share of AI accelerator spending, but the share price has moved ahead of the earnings that share will produce over the next two years.
What we like. The largest AI buyers want an alternative to Nvidia, and four of them have announced large Instinct deals with AMD since October 2025. Helios, AMD's complete rack-scale AI system, began deploying in Q2 with Anthropic, Meta, Microsoft, OpenAI and Oracle, according to the company. The CPU business keeps gaining as well: Client revenue rose 23% in Q2, and AMD targets more than half of server CPU revenue within three to five years.
What the market may be missing. The deals came at a price. AMD issued OpenAI and Meta warrants for up to 160 million shares each, vesting with shipments and share-price thresholds. Fully vested, that is about 19% of today's 1.66 billion diluted shares (Haverhill calculation). AMD has also guaranteed up to $4.1 billion of partners' data center leases, and after the quarter it committed up to $5.0 billion of equity investments, in line with the stake in Anthropic it announced in July.
What holds us back. At $608.10 AMD is worth about $1.0 trillion. That is about 82 times our 2026 earnings estimate of $7.40 a share and about 51 times our 2027 estimate of $12. The price assumes the gigawatts after the first arrive on schedule.
How does the business actually make money?
AMD designs processors and pays outside foundries, mainly TSMC, to make them. It reports three segments, and in Q2 2026 Data Center was 58% of revenue and about two thirds of segment operating income.
Data Center ($6.7 billion, 58% of revenue). EPYC server processors and Instinct GPUs, the accelerators that train and run AI models, now also sold as full Helios racks. Segment operating margin was 31%. AMD does not split the segment between CPUs and GPUs.
Client and Gaming ($3.8 billion, 33%). Ryzen PC processors, Radeon graphics and custom chips for game consoles. Client sales rose 23% to $3.06 billion while Gaming fell 31% to $779 million. Operating margin was 15%.
Embedded ($977 million, 8%). Programmable and embedded chips for industrial, automotive and communications customers, mostly from the Xilinx purchase. Sales rose 19% at a 40% operating margin.
Who pays for the AI chips. AMD's disclosed Instinct customers are a short list of very large buyers. OpenAI agreed in October 2025 to deploy 6 GW across several generations, starting with 1 GW of MI450 in the second half of 2026; AMD said the deal should bring "tens of billions of dollars" of revenue. Oracle ordered 50,000 MI450 GPUs from calendar Q3 2026. Meta agreed in February 2026 to up to 6 GW, beginning with a custom MI450-based chip. Anthropic signed in July 2026 for up to 2 GW, with the first gigawatt from the first half of 2027.
What do the numbers say?
Revenue grew 50% in a year to a record $11.5 billion in Q2 2026, and non-GAAP operating margin reached 27%. Data Center did most of the work, adding $3.5 billion of the $3.9 billion increase. Q3 guidance is about $13.0 billion, plus or minus $300 million, roughly 41% growth at the midpoint.
Exhibit 1: Data Center revenue has roughly doubled in a year as Instinct shipments ramp
Three things investors should know.
Growth is speeding up. Data Center revenue grew 39% year over year in Q4 2025, 57% in Q1 2026 and 107% in Q2. Q2 2025 was a weak base, with an $800 million charge tied to US export controls on the MI308 chip for China.
Margins have stepped up. Non-GAAP gross margin was 56%, against 43% a year earlier, and is guided to about 56% for Q3. Non-GAAP EPS was $1.66, up from $0.48.
The balance sheet is clean. Free cash flow was $1.56 billion in Q2 and $4.12 billion in the first half. AMD held $13.1 billion of cash and short-term investments against $3.2 billion of debt. Inventory rose to $8.5 billion from $7.9 billion in December, which we read as building for the MI450 ramp.
Competition and moat
AMD's edge is that it sells both leading server CPUs and a complete AI GPU rack, at a time when the biggest buyers are paying to build a second source of supply.
Who it competes with. Nvidia dominates AI accelerators and has the software most AI developers already use; AMD's answer is its open ROCm software, which it extended with ROCm.ai in Q2. Cloud companies' own custom chips compete for the same budgets. In CPUs, AMD competes with Intel's Xeon and with Arm-based chips designed by cloud providers.
Where the moat is real. The GPU contracts span several chip generations, and Meta's first gigawatt is a chip customized for its own workloads, which makes switching costly. Meta is also a lead customer for the 6th Gen EPYC launched in Q2. AMD does not publish its current server CPU share in its results, but at its November 2025 analyst day it set targets of more than 50% of server CPU revenue and more than 40% of client CPU revenue within three to five years.
Where it is thin. The warrants show AMD had to pay to win its two largest GPU customers. Nvidia's software lead remains wide, and AMD depends on TSMC for manufacturing and on suppliers such as Samsung for high-bandwidth memory.
What is it worth?
We value AMD at $605 per share over the next 12 months (to October 2027), a probability-weighted blend of three scenarios. That is about 1% below today's price, which supports a Hold.
Our method. AMD is profitable and growing fast, so we value it on earnings. In October 2027 investors will be looking at 2028, so we forecast 2028 revenue by segment, apply a non-GAAP operating margin, take out about 13% tax (our assumption) and divide by 1.72 to 1.82 billion shares, which allows for part of the OpenAI and Meta warrants vesting. We then apply a price-to-earnings multiple and weight the cases 25% bear, 50% base and 25% bull.
Why these numbers. Our base case of about $104 billion of 2028 revenue assumes the OpenAI, Meta, Oracle and Anthropic deployments ship roughly on the schedules AMD has given, in line with AMD's target of more than 35% annual revenue growth. The 34% operating margin sits just below AMD's target of more than 35%, against 27% today. A 36x multiple assumes growth above 30% into 2029.
What the price implies. At $608.10 the market is paying about 36 times our base-case 2028 EPS of $17, so it already assumes the base case. Each turn of the multiple is worth about $17 a share, and each extra dollar of 2028 EPS about $36 at 36x.
What are the real risks?
The MI450 ramp slips. Oracle, OpenAI and Meta deployments are all due to start in the second half of 2026. A two-quarter delay would push several billion dollars of revenue out of 2027 (our estimate).
A few buyers matter too much. Much of the planned GPU growth rests on four customers, two of them private AI labs rather than cash-rich cloud companies. A pause at any one would show up quickly.
Dilution grows with success. The warrants for up to 320 million shares vest as shipments scale and as the share price rises, so dilution is heaviest in exactly the scenarios where the stock does well.
Nvidia and custom chips fight on price. A price cut by Nvidia, or a buyer shifting volume to its own chip, would pressure the 56% gross margin.
Export rules and Taiwan. US controls already cost AMD $800 million on the MI308 in 2025, and its leading chips are made in Taiwan.
New financial commitments. Up to $4.1 billion of lease guarantees and $5.0 billion of investment commitments tie more of AMD's balance sheet to its customers' success.
What would change our mind?
Our Hold rests on one idea: the AI ramp is real, but the price already reflects our base case.
Signals that would move us to Buy
Data Center revenue above $10 billion in a single quarter by Q2 2027.
Non-GAAP operating margin above 32% in any quarter of 2027.
A new multi-gigawatt customer signed without a warrant attached.
The price falls below about $525 with no change in our earnings view.
Signals that would move us to Sell
Q4 2026 guidance implying flat Data Center revenue from Q3.
A delay to any first-gigawatt deployment announced by AMD or a customer.
Non-GAAP gross margin below 53% in any quarter.
The price runs above about $700 with no change in our earnings view.
Catalysts to watch
November 3, 2026: Q3 2026 results, after the market close (AMD announcement).
Second half of 2026: first gigawatt of MI450 shipments for OpenAI and Meta; Oracle's 50,000-GPU deployment from calendar Q3.
Early 2027 (Haverhill estimate): Q4 2026 results and AMD's first look at 2027.
First half of 2027: Anthropic's first gigawatt of MI450 in Helios racks.
2027: planned launch of the MI500 Series, the next Instinct generation.
Sources and disclosures
Every company figure in this report comes from Advanced Micro Devices's own filings and releases, linked below. The analysis, model and price target are Haverhill Financial Research's own work.
Disclosures. Haverhill Financial Research is a publisher of general financial commentary. This report is not personalized investment advice and does not consider any reader's financial situation; readers should do their own research. Price targets are estimates of value, not guarantees. Haverhill ratings follow our published bands: Buy when our 12-month target is at least 15% above the price, Sell when it is at least 15% below, Hold in between. The author of this post does not hold a position in AMD. Data as of October 9, 2026.
