Broadcom's AI Chip Revenue Diverges: Surging Sales Amid Lowered Guidance

Broadcom’s Earnings Reveal AI Chip Revenue Divergence: Structural Mismatch Emerges as Industry Enters Deep Inventory Correction Phase

Broadcom’s latest quarterly earnings report posted a record 48% year-on-year revenue growth—superficially stellar, yet concealing turbulent undercurrents. In an unusual move, the company lowered its AI-related revenue guidance for fiscal year 2025, revising expectations downward from $17 billion to $16 billion. This paradoxical signal—“revenue growth without profit growth,” and “high growth alongside lowered expectations”—is no isolated incident. Micron fell another 3.2% pre-market; SanDisk (a joint venture between Western Digital and Kioxia) saw channel inventory pricing drop 8% week-on-week; and SK Hynix’s HBM3 spot prices retreated 22% from their peak. These converging data points confirm a profound structural divergence across the AI chip market: GPU compute layers remain overheated, while supporting layers—including memory, networking, and power management—have sunk into a high-inventory quagmire. The semiconductor industry has quietly passed the peak of conceptual hype and entered a rigorous calibration phase defined by real-world delivery validation.

AI Server Demand Exhibits “Top-Heavy” Structural Mismatch

As the global leader in AI server networking chips (e.g., Tomahawk switches, Jericho routers) and premium Wi-Fi 7 SoCs, Broadcom’s performance should serve as a barometer for AI infrastructure health. Yet during its earnings call, CEO Hock Tan explicitly stated: “Customers are slowing their procurement pace for AI training clusters—especially after deploying ultra-large-scale, 10,000-GPU clusters—and their willingness to upgrade accompanying network bandwidth and storage I/O is lower than expected.” This cuts to the core contradiction: While next-generation GPU platforms like NVIDIA’s GB200 NVL72 have entered mass production and delivery, cloud providers (AWS, Azure, GCP) have shifted deployment priorities—from “piling on more GPUs” toward “improving per-GPU utilization” and “reducing total cost of ownership (TCO).” Against this backdrop, orders for high-speed Ethernet switching chips (e.g., Broadcom’s Trident series) and complementary PCIe Gen6 retimers used to interconnect GPUs have softened temporarily. More critically, AI servers’ insatiable demand for memory bandwidth far outstrips their need for storage capacity, causing DDR5/LPDDR5X memory module shipments to surge—while enterprise-grade NVMe SSD inventory turnover days climbed to 142 days (per DRAMeXchange), hitting a three-year high.

Memory Chips Under Collective Pressure: From “Capacity Shortage” to “Channel Logjam”

Micron, Samsung, and SK Hynix have all recently cut capital expenditures—a clear reflection of genuine cooling in end-demand. Notably, this inventory pressure does not stem from consumer electronics weakness but rather from structural substitution driven by AI server architecture evolution. For example, Microsoft Azure’s newly deployed MAIA AI supercomputing cluster adopts a CXL 3.0 memory-pooling architecture, partially migrating traditional SSD storage functions to near-memory computing (NMC) modules—directly diminishing demand for standalone SSDs. Simultaneously, domestic AI server vendors are rapidly adopting Yangtze Memory’s PC300 PCIe 5.0 SSDs, whose cost-performance advantage further squeezes international giants’ mid-tier product lines. Channel data shows enterprise U.2 SSD pricing down 19% from Q1 highs; meanwhile, the price war for consumer NVMe SSDs has spilled over into entry-level PCIe 4.0 models. Memory chips are thus rapidly sliding from a “seller’s market” into a “buyer-driven, deep-negotiation” regime.

Supply Chain Validation Logic Shifts: From “Story-Driven” to “Delivery-Driven”

A turning point in market sentiment is already visible at the micro level. Accelink recently completed its 2024 restricted stock unit (RSU) grant, with vesting conditions explicitly requiring “800G optical module shipments exceeding 500,000 units for two consecutive quarters,” anchoring valuation firmly to real-world capacity ramp-up progress. Jianghai Electronics urgently clarified that it has “no plans to expand MLCC production,” emphasizing its current capacity utilization stands at only 78%—a direct rebuttal to earlier market over-optimism about its entry into the AI server passive component supply chain. This shift implies:

• Equipment vendors face restructuring of order composition: etching and thin-film deposition tool demand is tilting toward HBM interposers and CoWoS packaging, while orders for logic-chip equipment slow;
• Advanced packaging has become the sole area of unambiguous strength: ASE, Siliconware, and Advanced Semiconductor Engineering (ASE) maintained CoWoS capacity utilization above 95% in Q2—but yield bottlenecks constrained actual deliveries to just 68% of planned capacity;
• AI server power management is undergoing value re-evaluation: Following Broadcom’s acquisition of VMware, its integrated AI infrastructure software stack is driving server power systems to evolve from “stable power delivery” toward “power-aware scheduling.” Though TI and Analog Devices’ digital multi-phase VR controller orders rose, cloud providers pushed average selling prices down 12–15%.

Shenzhen Accelerates New Infrastructure Investment: Compute Network Construction May Open New Inventory Outlet

Notably, the Shenzhen Municipal Party Committee’s special meeting explicitly called for “accelerating construction of compute networks and next-generation communication networks,” positioning them as core levers to “stimulate new investment and drive emerging industries.” This is not merely a repetition of China’s “East Data, West Computing” initiative—it signals a foundational reconstruction of AI-era infrastructure: leveraging compute-scheduling networks (e.g., Huawei’s Ascend Intelligent Computing Network) to enable dynamic cross-data-center GPU resource allocation, thereby reducing redundant server configurations at any single site. This strategy could catalyze two opportunities: first, a rebound in rigid demand for ultra-low-latency RDMA networking chips (e.g., Broadcom’s Stellar-X2); second, accelerated retirement of legacy servers, unlocking markets for enterprise SSD refurbishment and tiered reuse. Per IDC, China’s AI server secondary-market transaction volume is projected to reach $830 million in 2024—a 67% YoY increase—potentially serving as a buffer to ease original equipment manufacturer (OEM) inventory pressure.

Conclusion: Navigating the Deep-Water Zone Requires Re-Prioritizing Technology Roadmaps

Broadcom’s earnings serve as a stark warning: the AI chip industry has moved beyond the “all-segments-rise-together” infancy stage and entered a precision division-of-labor era where value realization is gated by technology generation. As cutting-edge technologies—including HBM3, chiplets, and CXL—transition from PowerPoint slides to volume production, the true test lies in whether each layer can simultaneously clear the triple hurdles of yield, cost, and ecosystem maturity. For investors, chasing the “AI+” label is less fruitful than tracking three hard metrics: advanced packaging yield curves; server power efficiency (W/GPU); and actual compute-network scheduling latency (measured in microseconds). Only within these millimeter-scale engineering details can one discern genuinely cyclical-resilient, structural opportunities—because in semiconductors, an industry priced by physical laws, the ultimate bill is always settled in kilowatt-hours, nanoseconds, and functional die.