Advanced Micro Devices (AMD): Gaining Share in AI Compute

Advanced Micro Devices (AMD): Gaining Share in AI Compute

Data and estimates are as of the July 31, 2026 close. AMD is scheduled to report fiscal second-quarter 2026 results on August 4, 2026, so the analysis below excludes that pending report.

Executive view

AMD has moved from being a credible alternative to NVIDIA in AI accelerators to becoming a meaningful second source for hyperscale AI infrastructure. The company’s MI300X, MI325X, and MI350-series products have established a credible hardware roadmap, while ROCm has improved materially in model support, inference performance, and multi-node scaling.

The investment case is increasingly dependent on three variables:

1. MI350 adoption and MI450/Helios execution
2. ROCm’s ability to reduce the CUDA migration penalty
3. Whether AI infrastructure demand produces a second wave of EPYC server CPU growth

At the July 31, 2026 close, AMD traded at approximately $476.15 per share, representing an equity value of roughly $0.8 trillion. AMD stock forecast and valuation data (stockanalysis.com)

The current Street framework is already aggressive: approximately $49.7 billion of 2026 revenue and $7.45 of adjusted EPS, followed by roughly $79.7 billion of 2027 revenue and $13.82 of adjusted EPS. AMD consensus estimates (stockanalysis.com)

My conclusion is 65/100 on valuation versus upside. AMD has substantial long-term potential, but the stock price already discounts a successful MI450 ramp, continued CPU share gains, and meaningful ROCm progress. The risk-reward is attractive on execution-driven pullbacks, but less compelling as an unqualified momentum purchase.

Mid-2026 price and current expectations

Because July 31, 2026 was the last completed trading day before the stated current date of August 1, I use the $476.15 close as the mid-2026 reference price. AMD’s upcoming August 4 earnings report could materially change the near-term valuation framework.

The arithmetic is important. AMD’s valuation looks extremely high against 2026 earnings, but much more reasonable against 2027 earnings. At $476.15, the stock is valued at approximately 34.5 times current 2027 consensus EPS. That is still a premium to most semiconductor companies, but it reflects the possibility that consensus estimates do not yet fully capture large-scale MI450 deployments.

The average analyst target of approximately $576.55 implies roughly 21% upside, but that target is not a substitute for a valuation model. It indicates that the Street is already assuming strong execution. AMD analyst forecast data (stockanalysis.com)

AMD’s AI product roadmap

MI300X: the foundation

The MI300X, introduced in 2023, was AMD’s first serious hyperscale generative-AI accelerator. It uses CDNA 3 architecture and includes:

• 192GB of HBM3 memory
• Approximately 5.3TB/s of memory bandwidth
• 304 compute units
• Eight-GPU platform configurations

Its most important contribution was commercial rather than merely technical. MI300X established that major cloud providers and AI developers were willing to deploy AMD accelerators for production workloads. It also gave AMD and its customers a real-world platform on which to improve ROCm, distributed training, inference serving, and cluster management.

AMD’s relationship with OpenAI began with MI300X and continued through MI350X, according to AMD’s strategic partnership announcement. AMD MI300-series product information AMD–OpenAI strategic partnership (amd.com)

MI325X: a high-memory bridge product

The MI325X is best understood as a high-memory refresh of the MI300 platform rather than a completely new architectural generation. Its key specifications include:

• 256GB of HBM3E
• Approximately 6TB/s of memory bandwidth
• 304 compute units
• Drop-in compatibility with the MI300X platform

That compatibility matters. Customers can upgrade accelerator capacity without redesigning an entire server platform, reducing deployment friction and shortening qualification cycles.

MI325X is especially relevant for large-model inference and memory-intensive workloads. More memory can reduce the number of GPUs required for a model, potentially lowering networking overhead and improving total cost of ownership. AMD MI325X and MI300 product page (amd.com)

MI350X and MI355X: the real inflection point

The MI350 series, based on AMD’s fourth-generation CDNA architecture, is the more important product for AMD’s share-gain thesis.

The MI350X and MI355X offer:

• Up to 288GB of HBM3E memory
• Up to 8TB/s of memory bandwidth
• Support for FP4, FP6, and FP8-style low-precision workloads
• Up to approximately 10.1 petaflops of MXFP4/MXFP6 performance
• Eight-GPU platforms with approximately 2.3TB of total HBM3E memory

The MI350P PCIe product expands the addressable market beyond hyperscale customers into enterprise deployments and existing server infrastructure. AMD MI350-series specifications (amd.com)

The MI350’s strategic advantage is not that AMD has suddenly overtaken NVIDIA across every workload. Rather, the product gives AMD a credible offering in several economically important areas:

• Memory-heavy inference
• Open-model deployment
• Enterprise AI
• High-performance computing
• Cost-sensitive hyperscale workloads
• Large-model serving where GPU memory capacity is a constraint

AMD’s 2026 MLPerf results provide evidence of improving competitiveness. In MLPerf Inference 6.0, AMD reported that MI355X delivered approximately 3.1 times the throughput of prior MI325X results on a Llama 2 70B workload. It also reported near-parity or better-than-parity results against NVIDIA B200 in selected inference configurations. AMD MLPerf Inference 6.0 results (amd.com)

For training, AMD reported MI355X results within approximately 5% to 6% of NVIDIA B200 on two selected LLM workloads, along with a 512-GPU multi-node submission. These are meaningful developments, although benchmark results should not be treated as universal production performance. AMD MLPerf Training 6.0 results MLCommons MLPerf Training results (amd.com)

MI450 and Helios: what the stock is really pricing

Although the immediate product discussion centers on MI300, MI325, and MI350, AMD’s valuation increasingly depends on MI450 and Helios.

AMD expects the Helios rack-scale platform, powered by MI450 GPUs, to begin ramping in the third quarter of 2026. The platform combines:

• Instinct GPUs
• EPYC server CPUs
• Pensando networking
• ROCm software
• Rack-level system integration

AMD has announced multi-year agreements for up to 6 gigawatts of GPUs with OpenAI and up to 6 gigawatts with Meta. Initial one-gigawatt deployments for each partnership are expected to begin in the second half of 2026. AMD Financial Analyst Day strategy AMD–OpenAI agreement AMD–Meta agreement (ir.amd.com)

The agreements are strategically important, but they should not be treated as fully secured revenue. Actual revenue depends on:

• Product qualification
• Supply availability
• Deployment schedules
• Customer acceptance
• Performance milestones
• Revenue-recognition timing

The agreements also include performance-based warrants of up to 160 million shares each for OpenAI and Meta. If fully vested, the two warrants together could represent approximately 20% of AMD’s current share count, before considering any buybacks. This is a significant potential dilution factor. (ir.amd.com)

ROCm: materially better, but not yet CUDA-equivalent

Evidence of real progress

ROCm is AMD’s open-source GPU software stack, covering compilers, runtimes, libraries, profilers, communication tools, and AI frameworks. The stack is built around HIP, which provides a path for adapting CUDA-oriented code to AMD GPUs.

ROCm’s maturity has improved in several measurable ways:

• ROCm 7 added MI350 support.
• It added functional support for FP4, FP6, and FP8-related data types.
• It expanded PyTorch support.
• It added GPU partitioning and virtualization improvements.
• Current documentation covers PyTorch, JAX, vLLM, SGLang, Kubernetes, distributed inference, and multi-node deployment.
• AMD reported a tenfold year-over-year increase in ROCm downloads in 2025. ROCm 7 release notes ROCm ecosystem documentation AMD 2025 annual report (rocm.docs.amd.com)

The most encouraging evidence is that AMD is moving beyond single-GPU demonstrations. Its MLPerf Training 6.0 submission included multi-node training, Primus training workflows, and partner participation across cloud and server vendors. AMD said partner results were within approximately 6% of its own submissions across major LLM workloads. That suggests the performance is becoming more reproducible outside AMD’s internal laboratory environment. (amd.com)

The remaining software discount

ROCm is now credible for selected production workloads, but credible is not the same as ecosystem parity.

The key investment risk is the software migration tax. A customer may accept a lower hardware price or better memory configuration, but only if the engineering work required to port, debug, tune, and maintain applications is manageable.

AMD still faces several disadvantages:

• Less installed-base familiarity than CUDA
• More dependence on version-specific tuning
• Fewer years of accumulated third-party optimization
• More uncertainty around new model bring-up
• Greater dependence on cloud providers and systems partners for deployment support

ROCm’s own release notes continue to list compatibility and known-issue items, which is normal for a rapidly evolving platform but reinforces the view that the stack is still maturing. ROCm 7 release notes (rocm.docs.amd.com)

My assessment is that ROCm has moved from “too immature for broad deployment” to “production-capable for customers willing to optimize.” That is a major improvement, but it does not justify assigning AMD the same ecosystem premium as NVIDIA.

AI-related server demand should lift EPYC

AI infrastructure creates demand for more than accelerators. Every large AI cluster also requires:

• Host CPUs
• Memory and storage controllers
• Networking and orchestration
• Data preprocessing
• Cluster management
• CPU-side inference and agent workloads

AMD’s Data Center segment generated $16.6 billion in 2025 revenue, with operating income of approximately $3.6 billion. In the first quarter of 2026, Data Center revenue reached approximately $5.8 billion, up 57% year over year, driven by EPYC processors and Instinct GPUs. AMD 2025 annual results AMD Q1 2026 results (ir.amd.com)

The CPU opportunity is strategically important because EPYC can benefit even when AMD does not win every accelerator socket. A cloud provider may use NVIDIA GPUs but still choose EPYC CPUs based on:

• Performance per watt
• Core density
• Memory bandwidth
• Total cost of ownership
• Existing EPYC fleet compatibility

The AI-related CPU uplift is not separately disclosed, so the following estimates are scenario assumptions rather than company guidance:

• Bear case: $1 billion to $1.5 billion of incremental 2027 CPU revenue from AI infrastructure
• Base case: $3 billion to $3.5 billion
• Bull case: approximately $5 billion

The Meta agreement is particularly important because it explicitly includes sixth-generation EPYC “Venice” CPUs alongside MI450-based accelerators and Helios racks. AMD–Meta partnership announcement (ir.amd.com)

Data-center GPU share and earnings scenarios

AMD does not disclose a precise share of the merchant AI accelerator market. The following framework uses a modeled accelerator revenue pool of approximately $175 billion in 2026 and $235 billion in 2027. These are analytical assumptions, not reported market data. The share figures refer to merchant data-center AI accelerator revenue, excluding internally designed cloud ASICs and general-purpose CPUs.

Bear case

The bear case assumes:

• MI350 performs well technically but does not produce broad customer conversion.
• MI450 and Helios deployments are delayed.
• Customers continue to favor NVIDIA for training and use custom ASICs for inference.
• ROCm remains a material engineering burden.
• AMD’s gross margin remains in the mid-50% range.
• The OpenAI and Meta agreements generate less 2026–2027 revenue than expected.

At a 25-times adjusted EPS multiple, the bear case produces a value near $230, or more than 50% below the current price.

Base case

The base case assumes:

• MI350 becomes a credible second-source platform.
• AMD reaches roughly 11% of the modeled merchant accelerator market by 2027.
• Inference grows faster than training as model serving, agentic AI, and enterprise deployment expand.
• MI450 deployments begin in the second half of 2026 but contribute more meaningfully in 2027.
• EPYC continues gaining share in AI and general-purpose servers.
• Non-GAAP gross margin rises from the mid-50% range toward approximately 58%.

The base case implies approximately 10% upside to current 2027 revenue and EPS consensus:

• 2027 revenue: $88 billion versus $79.7 billion consensus
• 2027 adjusted EPS: $15.20 versus $13.82 consensus

Applying a modest growth premium of approximately 37 times adjusted EPS produces a value range of $560 to $580.

Bull case

The bull case requires more than strong MI350 sales. It assumes:

• Initial OpenAI and Meta gigawatt deployments are delivered on schedule.
• Helios becomes a full-system alternative rather than merely a GPU platform.
• ROCm achieves reliable day-zero support for more frontier models.
• AMD wins additional hyperscale and sovereign AI deployments.
• Inference becomes the dominant portion of AMD’s AI revenue.
• GPU and EPYC mix drives non-GAAP gross margin toward 60%.

Under that outcome, AMD could approach $100 billion of revenue in 2027 and generate approximately $19 of adjusted EPS. A 42-times multiple would imply a value near $790, or approximately 65% upside.

Valuation versus NVIDIA, Broadcom, and semiconductor peers

NVIDIA’s forecast revenue and earnings remain much larger than AMD’s, with approximately $393.6 billion of forecast revenue and $8.99 of EPS in the relevant forward fiscal year. NVIDIA’s forward multiple is lower than AMD’s despite stronger margins and a more entrenched software ecosystem. NVIDIA forecast data (stockanalysis.com)

Broadcom’s forward valuation is also lower than AMD’s on a next-year earnings basis. Broadcom’s forecast includes a combination of custom AI accelerators, networking, and infrastructure software, producing a more diversified earnings profile. Broadcom forecast data Broadcom valuation ratios (stockanalysis.com)

AMD’s premium can be justified only by higher earnings growth from a smaller base. The stock should not receive a full NVIDIA-like multiple because:

• ROCm remains less sticky than CUDA.
• AMD’s gross margin is materially lower.
• Its AI product revenue is less established.
• Helios and MI450 still require large-scale execution.
• Potential customer warrants create dilution risk.

At current consensus, AMD’s implied 2027 P/E is approximately 34.5 times. That is above a normalized semiconductor reference of about 30 times and well above NVIDIA or Broadcom. The market is therefore already pricing in meaningful estimate growth.

Catalysts

1. August 4 earnings and guidance

The most immediate catalyst is AMD’s second-quarter report and forward guidance. AMD previously guided to approximately $11.2 billion of Q2 revenue, plus or minus $300 million, with approximately 56% non-GAAP gross margin. AMD Q1 2026 results and Q2 outlook (ir.amd.com)

The most important items will be:

• Data Center revenue
• MI350 demand commentary
• MI450 supply and customer qualification
• Gross-margin trajectory
• EPYC server demand
• 2026 revenue guidance

2. Design wins converting into shipments

The market needs evidence that OpenAI, Meta, Oracle, Microsoft, and other customers are moving from announced partnerships to production deployments.

3. ROCm performance outside AMD’s own benchmarks

Third-party replication, customer case studies, and sustained performance on vLLM, SGLang, PyTorch, and newer frontier models would be more valuable than another theoretical peak-performance announcement.

4. EPYC Venice adoption

AI systems that combine AMD GPUs with AMD CPUs and networking could increase the value of each customer relationship and make AMD more competitive at the rack level.

5. Gross-margin expansion

A sustained move toward 57%–60% non-GAAP gross margin would indicate that AMD is monetizing AI infrastructure rather than merely buying share through aggressive pricing.

Key risks

Execution and timing risk

The most important risk is that MI450 and Helios shipments slip, ramp slowly, or require more customer-specific engineering than expected.

Ecosystem stickiness

NVIDIA’s advantage is not only hardware. Its software, tools, libraries, and developer familiarity can make customers reluctant to switch even when AMD offers competitive hardware.

Margin risk

AMD’s AI products require expensive HBM, advanced packaging, networking, and system integration. Rack-scale systems may generate substantial revenue but could carry lower margins than accelerator-only sales.

Custom silicon competition

Cloud providers increasingly design their own accelerators. Broadcom benefits from that trend through custom silicon and networking, while AMD competes for workloads that may otherwise go to internal ASICs.

Dilution

The potential OpenAI and Meta warrants could represent roughly 20% of AMD’s current share count if fully vested. Even if the warrants create strategic value, they reduce per-share earnings growth.

Export controls and China

AMD recorded significant MI308-related inventory and export-control charges in 2025. Similar restrictions could affect product availability, inventory, and regional demand. AMD 2025 annual results (ir.amd.com)

Valuation versus upside score: 65/100

Interpretation: moderately attractive, but execution-dependent.

• AI growth opportunity: 9/10
• Product roadmap: 8/10
• ROCm progress: 7/10
• CPU and EPYC leverage: 8/10
• Current valuation: 5/10
• Execution and ecosystem risk: 5/10

AMD has a credible path to meaningful AI accelerator share gains and could outperform current estimates if MI450, Helios, and ROCm execute well. The company also has an underappreciated advantage in EPYC CPUs, which can benefit from the broader expansion of AI server infrastructure.

However, AMD is no longer priced as a speculative challenger. At roughly 34.5 times 2027 consensus EPS, the market already expects strong execution. My base case is approximately $560–$580, with a bull case near $800 and a bear case around $230.

The stock is therefore best characterized as a high-upside, high-execution semiconductor investment. The most attractive entry points would likely occur when the market temporarily discounts AMD’s software progress or questions near-term MI450 timing—provided customer deployments and gross-margin expansion remain intact.