AMD received a reiterated buy rating from Goldman Sachs following its San Francisco event on July 23, with a 12-month price target of $640. The core rationale is that AMD’s growth narrative in AI infrastructure is strengthened by key partnerships with Anthropic and Microsoft, along with the forecast of a $2 trillion total computing market by 2030.
This is not merely a GPU specification release. AMD is attempting to shift the AI competition from individual chips to an end-to-end delivery model encompassing entire racks, CPUs, networks, DPUs, software tools, and large-scale customer deployments. For investors, the key question has shifted from “Does AMD have a stronger GPU?” to “Can it secure sufficiently large AI cluster orders and deliver them on time?”
The report's target price of $640 is based on an estimated normalized EPS of $20 and a P/E ratio of 32x. At the current stock price of approximately $539.69, this implies a potential upside of about 18.6%; the company's market capitalization is approximately $890.5 billion.
The most direct order signal comes from Anthropic.
AMD has entered into a strategic partnership with Anthropic, under which Anthropic will deploy a total of 2 GW of Instinct M1450 GPUs in its Helios rack-scale AI systems. The first 1 GW is expected to begin deployment in the first half of 2027. AMD also plans to provide Anthropic with up to $5 billion in strategic equity investment, and the two companies will collaborate to optimize Claude model performance on AMD GPUs and accelerate ROCm software development.
The significance of this partnership goes beyond chip sales. For AI chip suppliers, the adoption by leading model companies determines ecosystem credibility. If Anthropic consistently migrates or scales its Claude workloads to AMD’s platform, it will help AMD demonstrate that its GPUs, rack systems, and software stack can meet the demands of large-scale training and inference.
Microsoft offers an alternative implementation path. Following the expansion of their Azure partnership, Microsoft plans to begin receiving Helios racks, Venice CPUs, networking equipment, and software starting in the second half of 2026 for frontier model inference, Microsoft’s own AI services, and customer applications. Microsoft will also launch two new virtual machines based on the next-generation 2-nanometer Venice CPU and expand the deployment of Pensando DPUs in its networking services.
This means AMD aims to sell GPU, CPU, DPU, and networking equipment together in cloud provider scenarios, rather than merely positioning itself as a supplier of accelerator cards. For Azure, an AMD platform that delivers sufficient performance and supply flexibility could also help reduce reliance on a single supply chain.
The most striking figure in the report is AMD's upward revision of the total computing market size forecast for 2030 to $2 trillion.
Among these, the TAM for data center AI accelerators has been raised from $200 billion to $1.4 trillion, representing a 40% CAGR; the TAM for server CPUs has been increased from $26 billion to $220 billion, reflecting a 50% CAGR. AMD also expects to capture a 50% market share in the data center CPU market by 2030.
Behind this assumption is agent-based AI. Unlike single-question-and-answer interactions, agent-based AI requires calling tools, planning tasks, reading context, executing multi-step workflows, and handling more complex reasoning and orchestration requests. The computational demand is not limited to GPUs—CPUs also bear the load of scheduling, data preprocessing, system services, and orchestrating multi-agent workflows.
This is also the new story AMD wants to tell: the expansion of AI infrastructure not only increases demand for accelerators but also drives demand for server CPUs, networking, and system-level solutions. If the company can bundle EPYC CPUs, Instinct GPUs, Pensando DPUs, and networking equipment into the Helios rack, its revenue potential will be greater than that of selling individual chips.
But this $2 trillion is not an already realized market size; it is a projection based on the large-scale adoption of agentive AI by 2030. It assumes continuous expansion of AI inference deployments by enterprises and cloud providers, as well as AI applications truly transitioning from pilot projects to high-frequency production workloads.
At the product level, Helios is the core vehicle of AMD's event.
The next-generation Helios AI rack platform is built on the CDNA 5 architecture, delivering up to 40 PFLOPS (FP4) and 20 PFLOPS (FP8) peak performance per GPU, equipped with 432GB of HBM4 memory and 23.3 TB/s memory bandwidth. Each rack integrates 75 GPUs connected via UALink over Ethernet, paired with a 96-core EPYC CPU, Salina DPU, and Volcano 800G AI network card.
The focus of these parameters is not on single-point performance, but on AMD’s shift toward rack-scale product architectures. AI clusters are increasingly dependent on system-level design: interconnects between GPUs, memory bandwidth, network throughput, CPU scheduling capabilities, and the software stack all impact final training and inference efficiency.
Helios has entered full-scale production, with shipments scheduled to begin at the end of Q3 and mass production to ramp up in Q4. Microsoft will begin receiving Helios in the second half of 2026, and Anthropic’s initial 1 GW deployment will start in the first half of 2027, meaning substantial customer validation is still ahead.
AMD has also partnered with Cerebras to integrate the Helios system with Cerebras’ wafer-scale engine to create a high-performance AI inference solution, aiming for lower latency, higher energy efficiency, and up to a 5x improvement in tokens per watt. This solution is expected to be launched via Cerebras Cloud in the second half of 2026.
On the software side, the ROCm.ai platform has been officially launched, integrating AI tools such as Cursor, Claude, Codex, and Gemini to provide developers with an AI-driven software development experience. The accompanying Hyperloom optimization layer has optimized over 14,000 models, delivering an average performance improvement of 3.3x compared to ROCm 7.
However, the partnership between ROCm.ai and Cerebras is better suited as a supplement to the Helios ecosystem rather than the main focus of this article. Investors will ultimately care whether customers can reliably use the AMD platform under real workloads, not just whether the list of tools and partnerships is long enough.
Goldman Sachs financial projections show that AMD's revenue is expected to be approximately $50.57 billion in 2026 with an EPS of $6.20; $86.01 billion in 2027 with an EPS of $13.20; and $109 billion in 2028 with an EPS of $17.90.
This set of forecasts reflects high market expectations for AMD’s AI revenue growth, margin expansion, and operational leverage over the next two to three years. The $640 price target accounts not only for current product launches but also for the simultaneous advancement of multiple initiatives, including Helios shipments, Microsoft Azure deployments, the initial rollout of Anthropic’s 1GW capacity, and improvements to the ROCm ecosystem.
The real divergence lies here.
First, the adoption rate of agent-based AI may be slower than expected. If enterprise AI workflows do not expand rapidly, the assumption of a $2 trillion total computing market by 2030 would be revised downward.
Second, there is still a timing lag in large-client GPU deployments. Microsoft’s Helios begins receiving shipments in the second half of 2026, and Anthropic’s first 1 GW is expected to start in the first half of 2027; short-term financial reports cannot yet fully validate the revenue contribution from these partnerships.
Third, competitive pressure will not disappear. NVIDIA still dominates in AI accelerators and software ecosystems; whether ROCm can close the developer experience gap will impact AMD’s substitutability among large-scale customers.
Fourth, the x86 architecture also faces market share risks in enterprise AI scenarios. If customers increasingly adopt custom chips, Arm servers, or other heterogeneous solutions, AMD’s expectations for the CPU TAM and 50% data center CPU market share could be challenged.
This turns AMD’s story into a test of execution: while reports have already raised the market opportunity, customer orders, and target price, whether the stock can continue to absorb these figures depends on whether Helios can ship on schedule, whether Microsoft and Anthropic can expand deployment as planned, and whether AI demand truly supports a $2 trillion market by 2030.
律动 BlockBeats