AMD Holds $476 Into Q2 Print as Data Center Targets $6.5B — 51 Analysts Carry a $576.55 Target With 14 Gigawatts of Disclosed Commitments

Advanced Micro Devices closed Friday at $476.15, down 1.90%, and enters Tuesday's second-quarter report as the single most consequential earnings event in semiconductors this week. The stock sits 18.6% below the all-time high of $584.73 set June 30 and roughly 236% above the 52-week low of $141.60. Across twelve months the shares have climbed 194%, and they remain up more than 100% year to date.

The tape around it has fractured. The Philadelphia Semiconductor Index jumped more than 100% over the past year and then dropped roughly 20% since mid-June, trading at 11,296.79 on Monday, off 0.13%. Processor and graphics names are down 12.7% on average over the past month. AMD has fallen 14.5% over the same stretch, underperforming its own group into the print.

Monday's session offered no resolution. President Trump called off strikes against Iran, sending West Texas Intermediate down 6.21% to $79.41 and lifting the Nasdaq Composite 1.77% to 25,822.62. Cloud infrastructure exploded, with CoreWeave up 14.97% and Oracle up 5.6%. Memory got destroyed in the same hour: Micron fell 3.5% to $794.59 on ChangXin taking DRAM share to 8%, South Korea's KOSPI dropped 4% with Samsung and SK Hynix each down nearly 8%, and the DRAM-focused ETF has now sunk 32% on margin calls.

AMD straddles both camps and that is the problem. Its Instinct accelerators and EPYC server CPUs are the compute the AI cloud names are buying. Its client business sells into a PC market that is getting squeezed by exactly the memory price inflation destroying the Korean and Boise trade.

The late-July damage across the group was extraordinary. Chip stocks shed more than $1 trillion in market capitalization inside a single week. Nvidia lost $238 billion since the prior Friday's close. SK Hynix shed $176 billion, Samsung $173 billion, Micron $113 billion, TSMC $119 billion, and AMD roughly $110 billion. The decline was driven largely by sentiment rather than fundamentals, which is the polite framing for a confidence break in the most crowded trade on the board.

AMD's beta reads 1.75 on one measure and 2.469 on another. In a tape moving this violently, that number does most of the work.

What the Street Actually Expects Tuesday

Consensus calls for adjusted earnings of $1.61 per share on revenue of $11.31 billion when AMD reports after Tuesday's close. Bloomberg's compilation puts EPS at $1.62 on $11.3 billion. Against the year-ago quarter, that represents revenue growth of 47% from $7.6 billion and earnings growth of 235% from $0.48.

The company's own guidance issued on the May 5 call was $11.2 billion in revenue, roughly 46% year-over-year growth, with gross margin at 56%. Ten analysts revised earnings estimates upward following that guide.

AMD has exceeded revenue estimates in each of its past eight quarters and topped earnings expectations in seven of eight. The most recent report delivered a 4.06% revenue surprise and a 7.87% EPS surprise, and the stock moved 23.38% in response.

The segment breakdown is where the report gets decided. Data center is expected to more than double, climbing to $6.5 billion from $3.2 billion in the second quarter of 2025, growth topping 100%. Client is projected at $3 billion, up 20%. Gaming is set to bring in $781 million, down 30%.

That composition tells you the entire investment case in three lines. Data center now represents roughly 57% of expected revenue against 42% a year ago. Client contributes 27%. Gaming has collapsed to under 7% and is shrinking at a 30% annual rate, which at this point functions as a rounding error rather than a business line.

The first quarter set a demanding bar. AMD reported revenue of $10.25 billion, up 37.8% year over year, with EPS of $1.37 against a $1.27 forecast. Data center revenue climbed 57% to $5.8 billion. Gross margin expanded to 55%. Free cash flow tripled to $2.6 billion. Guidance for the current quarter exceeded expectations, and the stock rose 4.02% after hours.

The fourth quarter of 2025, reported in February, delivered $10.27 billion of revenue with data center at $5.4 billion, up 39%.

The mechanical setup is a swing of roughly plus or minus 4.32% on the first trading day after the release, based on options positioning.

Fourteen Gigawatts of Disclosed Commitments

AMD has now announced prospective capacity commitments totaling 14 gigawatts across three customers, and the scale is what separates this cycle from every prior AMD accelerator push.

OpenAI signed a 6-gigawatt agreement in October 2025. Meta announced an equivalent 6-gigawatt commitment in February 2026 covering AMD Instinct GPUs across multiple generations, with initial MI450 shipments starting in the second half of 2026. Anthropic added a 2-gigawatt commitment on July 22 covering MI450 Series GPUs deployed in Helios rack-scale solutions, with the first gigawatt beginning in the first half of 2027.

The Anthropic transaction carries a second leg. AMD committed to a strategic equity investment of up to $5 billion in the company, with no disclosed timeline. A separate multi-year engineering collaboration will use Claude to optimize workloads for Instinct GPUs and accelerate ROCm development, and AMD plans to deploy Claude across its own engineering and product teams. The capacity runs Helios racks with Instinct MI455X GPUs, EPYC Venice CPUs, Pensando networking, and ROCm software, building on Anthropic's existing use of MI355X hardware.

The dollar scale of these commitments is difficult to overstate. AMD executives have said developing one gigawatt of AI computing infrastructure can cost tens of billions of dollars depending on equipment and facilities. Fourteen gigawatts of disclosed commitments therefore represents an addressable pipeline running into the hundreds of billions across the buildout window, of which AMD captures the silicon and increasingly the rack.

Execution risk is explicitly flagged. AMD characterized the deployment schedules, engineering work, and equity investment as forward-looking commitments subject to export rules, tariffs, manufacturing capacity, component availability, product performance, and customer demand. None of these gigawatts is contracted revenue in the accounting sense.

OpenAI has been running GPT-class workloads on Helios racks for three months and expects to deploy at massive scale starting toward the end of this year, accelerating into 2027.

The Warrants Are 320 Million Shares of Dilution

Both hyperscale agreements came attached to warrant structures that shareholders have to price. AMD issued OpenAI a warrant to purchase up to 160 million AMD shares at $0.01 per share, equivalent to roughly 10% of the company, vesting in tranches tied to deployment volume milestones and AMD share-price thresholds through October 2030. Meta received an identical arrangement in February: up to 160 million shares at $0.01, vesting against shipment milestones and share-price levels.

Combined, the two warrants represent potential dilution of up to 320 million shares issued at effectively zero cost.

The economic logic is coherent. AMD is paying for guaranteed volume at scale in a market where Nvidia holds the incumbent position and where customer concentration in accelerators is extreme. Locking OpenAI and Meta into multi-generation commitments converts AMD from a challenger hoping for design wins into a second source with contractual visibility.

The shareholder cost is real and it is nonlinear. The warrants vest against share-price thresholds, which means the dilution triggers precisely when the stock works. Investors buying AMD at $476 are buying an equity whose share count expands as the thesis succeeds, and the vesting schedule extends to October 2030.

That structure sits underneath every valuation model on the Street and it is the cleanest explanation for the dispersion in analyst targets. A model that treats the warrants as fully vesting produces a materially different per-share number than one that discounts them by probability.

Anthropic's arrangement inverts the direction: AMD invests up to $5 billion into the customer rather than issuing equity to it. That structure costs cash rather than shares and it aligns the two companies on ROCm development, which is the software gap that has kept AMD accelerators from taking share against CUDA for a decade.

Anthropic runs a deliberately diversified compute stack. It has secured up to 5 gigawatts from Amazon, uses more than one million Trainium2 processors with Trainium3 deployments planned during 2026, and has multiple gigawatts of next-generation TPU capacity from Google and Broadcom expected from 2027. AMD is one supplier among several.

Helios Economics at $5.25 Million a Rack

The commercial development that reframes AMD's model is that it stopped selling accelerators and started selling racks. Helios rack-scale pricing has been confirmed at $5 million to $5.5 million per rack, and that shift moves AMD from component vendor to systems vendor with the margin structure that implies.

The arithmetic on the disclosed pipeline is substantial. At a $5.25 million average per rack, converting even 10% of the combined 12-gigawatt Meta and OpenAI commitment produces roughly 120 megawatts of rack deployments. At approximately 20 to 30 racks per megawatt, that alone generates a revenue pipeline running into the tens of billions across the 2026 to 2028 delivery window. Add Anthropic's 2 gigawatts and the number expands further.

For the past decade AMD sold accelerators and let customers handle integration. Helios changes that, bundling Instinct MI400 series GPUs, sixth-generation EPYC Venice CPUs built on TSMC's 2-nanometer node, Pensando networking, and ROCm into one integrated platform. AMD started sampling MI450 GPUs with lead customers in May and has said Helios production shipments remain on track for the second half of 2026.

Performance claims are aggressive and unverified. AMD projects up to 10 times more performance on frontier-model workloads from the MI400 series against the MI300X, a comparison spanning roughly three hardware generations. Those figures come from AMD's own specifications and await independent MLPerf benchmarks on shipping hardware.

The customer validation has been more persuasive than the spec sheet. Anthropic's chief compute officer described connecting a prior-generation MI355X rack to Claude, instructing the model to bring the machine up, and returning after a weekend to a performance curve that climbed continuously. That anecdote matters because it addresses the ROCm friction that has been AMD's structural disadvantage against CUDA, not the raw FLOPS.

Sam Altman appeared on stage with Lisa Su at the Advancing AI 2026 keynote at San Francisco's Moscone Center on July 22, which is the visible signal of how deep the OpenAI relationship runs.

The CPU Story Nobody Was Modeling

The quieter driver behind AMD's 194% twelve-month run is not the GPU line. It is that agentic AI turned out to be CPU-intensive.

AI agents rely on central processing units to execute tasks for users, whether preparing spreadsheets or sending emails. Agentic systems require CPUs to coordinate multi-step reasoning, manage memory state, and orchestrate workloads across GPU clusters. That makes the EPYC server line far more important to hyperscalers like Microsoft and the AI cloud complex than the pre-agentic modeling assumed.

UBS estimates the server CPU total addressable market could expand from roughly $30 billion in 2025 to $170 billion by 2030, a 5.7x increase across five years, with AMD positioned as a primary beneficiary. AMD has been taking share from Intel in server CPU for six consecutive years, and a TAM expanding at that rate against a rising share position compounds twice.

TSMC's earnings call flagged the same phenomenon from the foundry side. Chairman C.C. Wei cited agentic AI reviving CPU demand as one of the drivers behind raising full-year 2026 revenue growth guidance to slightly above 40% and lifting capital expenditure to $60 billion to $64 billion. That is independent confirmation of the CPU story from the company that fabricates AMD's silicon.

Zen 6 EPYC Venice on the 2-nanometer node is the product carrying it, launched commercially at the Advancing AI conference and going into Helios racks alongside the MI400 accelerators.

The high-margin server CPU segment is what diversifies AMD's growth beyond a single accelerator cycle. If Instinct disappoints and EPYC holds, the earnings power survives. If both work, the $576.55 consensus target is conservative.

Data center revenue of $6.5 billion in the reported quarter would combine both lines, and AMD does not break out the GPU and CPU split cleanly. That opacity is why the market wants clearer AI GPU revenue figures, customer deployment scales, and next-quarter guidance rather than another round of demand adjectives.

The PC Segment Is Walking Into a Memory Squeeze

The client business expected at $3 billion this quarter faces a problem the AI narrative has been drowning out. The PC industry is heading into a downturn as a direct result of increased memory prices, which lands on both AMD and Intel PC chip revenue over the coming months.

The mechanism is simple. DRAM and NAND pricing has spiked as memory capacity redirected toward high-bandwidth memory for AI accelerators, and every notebook and desktop carries memory as a fixed bill-of-materials cost. When DRAM contract prices rise, OEMs either compress their own margins, raise system prices into a soft consumer market, or cut build plans. Historically they cut build plans.

The memory tape confirms the pressure is real rather than theoretical. Micron fell 3.5% to $794.59 Monday. Samsung and SK Hynix each dropped nearly 8% in Seoul as the KOSPI shed 4% and completed its worst month since 2008. The DRAM ETF has sunk 32%, with margin calls forcing mechanical selling.

The offset for AMD is that client revenue is still guided up 20% year over year, so any memory-driven weakness shows up in guidance for the September and December quarters rather than in Tuesday's reported numbers. That makes the forward commentary the risk, not the print.

Gaming provides no cushion. At an expected $781 million and declining 30% annually, the segment has become immaterial to the earnings picture. Console cycle dynamics and discrete GPU competition have compressed it to under 7% of revenue.

The composition risk is straightforward: if data center delivers $6.5 billion and client guides down on memory costs, AMD reports a strong quarter and trades lower. That is precisely what happened to TSMC on July 16, when record results and raised guidance produced an 8% two-session decline because the capex number spooked the market.

AMD's own history contains the same pattern. In the year-ago quarter the stock declined after hours because data center revenue only met expectations rather than beating them decisively.