Micron Technology reports earnings amid speculation that $100 billion long-term AI infrastructure contracts could drive memory and HBM demand

Meta's Muse and OpenAI's Astra are driving a fundamental shift in AI applications toward sustained, multi-step execution and collaborative agent workflows. This expansion is forcing a broad reassessment of the CPU, memory, and storage requirements underpinning these systems, with significant implications for the entire AI infrastructure supply chain.

The market has responded by elevating growth expectations for leading computing-power suppliers—SK Hynix, Samsung, Micron, NVIDIA, AMD, and Taiwan Semiconductor among them. As a result, Micron's earnings report, scheduled for September 30, has become a critical barometer for both the memory supercycle trajectory and the scale of AI infrastructure investment ahead.

Micron reported third-quarter revenue of $41.456 billion, up roughly 346% year-over-year, with adjusted earnings per share of $25.11. For the fourth quarter, the company has guided for revenue of $50 billion, plus or minus $1 billion, with a gross margin of approximately 86% and adjusted EPS of $31, plus or minus $1. More significantly, Micron disclosed that among its 16 strategic customer agreements, 14 are priced at the contract-minimum rate, generating cumulative revenues of approximately $100 billion over the remaining contract terms. This provides exceptional visibility into future revenue and profitability.

The strength of this guidance, aligned with Wall Street consensus, signals that the market's focus has shifted beyond the current quarter's potential upside to whether prices, shipment volumes, and long-term orders in fiscal 2027 can sustain further earnings-estimate revisions. UBS analysts have raised their 2027 operating cash-flow forecasts for global hyperscale cloud providers by 30%, while Morgan Stanley projects that the combined capital expenditures of North America's four major superclouds and AI vendors will surge from $917 billion in 2026 to $1.47 trillion in 2027 and $1.64 trillion in 2028. According to S&P Global's September 22 consensus, Meta, Alphabet, Amazon, and Microsoft are expected to spend a combined $2.8 trillion on capital expenditure between 2026 and 2028. Wall Street has correspondingly raised Micron's fiscal 2027 EPS forecast from approximately $90 in March to about $156.

Why do CPUs and memory remain indispensable even as AI becomes more capable at execution tasks?

The profound shift lies in the ability for a single user command to initiate continuously running workflows. Agent tasks—research, programming, or enterprise operations—typically involve planning, information retrieval, web browsing, file reading, code execution, result verification, and error correction. These steps frequently invoke models multiple times, generating new contextual data at each iteration. Meta's Muse operates within a secure cloud-based virtual machine with its own independent browser, allowing tasks to persist even after closing the application. OpenAI emphasizes Astra's enhanced capabilities in computer operations, browser usage, and cross-software professional workflows.

As such tasks scale from small groups of power users to enterprise-wide processes, infrastructure requirements are driven simultaneously by user-base size, task frequency, task duration, and peak concurrency. This explains the market's renewed focus on data-center CPUs based on AMD, Intel, and Arm architectures. While GPUs and ASICs handle large-scale neural-network computations, CPUs manage virtual machines, browsers, code sandboxes, task scheduling, data preprocessing, and tool execution. Efficient task completion depends on smooth coordination between model inference and actual execution; any delays in the CPU, database, or input/output stages reduce utilization of costly accelerators. AMD's PACE technology explicitly adopts a collaborative approach, with CPUs orchestrating and GPUs executing heavy-duty inference. Arm similarly identifies agent sandboxes and tool execution as key cloud-CPU workloads, providing concrete engineering rationale for expanding CPU demand and positioning server DRAM, storage, and networking as critical links in the same demand chain.

From a large-model inference perspective, the prefill stage processes input context while the decoding stage continuously generates output; model weights and the key-value cache consume memory capacity and bandwidth. Longer contexts, higher concurrency, and more parallel tasks increase the volume of state the system must maintain simultaneously. This creates a clear division of storage responsibilities: HBM supports high-bandwidth access for accelerators, server DRAM handles CPU workloads and portions of the cache, and enterprise-grade SSDs store files and application states, also hosting portions of inference context in tiered-caching systems. NVIDIA's CMX solution integrates an Ethernet-connected flash layer between GPU memory and shared storage, enabling context to be stored, migrated, and reused across nodes. Storage capacity, access speed, and network bandwidth collectively influence inference-service throughput.

The scale of these demands is already evident in production systems. OpenAI disclosed that its Habitat online storage platform processes over 70 million requests per second and manages data exceeding 500 petabytes, handling routing, compression, encryption, and checksum verification across multiple products. As more agents move into production environments—adding execution nodes, database accesses, and inter-node data exchanges—demands on servers, memory, enterprise-grade SSDs, and high-speed interconnects will intensify further. Deployed capacity is expected to expand from 35 gigawatts in 2025 to 145 gigawatts by 2028, driving procurement needs that extend to optical interconnects, power distribution, and cooling systems.