AI Compute Power Boom: Dual Narrative of Surging Profits and Capital Expenditure Risks
Dual-Theme Resonance: AI Compute Power Infrastructure’s Profit Surge and Capital Expenditure Risks Are Widening the Valuation Rift in the Semiconductor Industry
The National Bureau of Statistics’ latest data reveals that China’s electronics industry posted a staggering 103.9% year-on-year profit growth for January–May 2026—contributing 43.1% to the overall profit growth of all designated-size industrial enterprises. This figure far exceeds both market consensus expectations and the equipment manufacturing sector’s aggregate profit growth of just 14.1%. Behind these numbers lies a global AI compute infrastructure build-out entering an “ultra-accelerated gear-shifting phase”: from NVIDIA’s mass production and delivery of the H200 GPU, to TSMC’s CoWoS advanced packaging capacity remaining fully booked for three consecutive quarters, to Micron and Samsung ramping up HBM3e production—this AI compute value chain—centered on high-end GPUs, high-bandwidth memory (HBM), advanced packaging (2.5D/3D IC), and high-speed interconnects (CPO, silicon photonics)—is unleashing unprecedented earnings elasticity.
Yet precisely along this same timeline, capital markets have cast a deep shadow: Oracle’s stock plunged 19.4% in a single week—the largest drop since 2001; onsemi fell 23.66% in one day. These declines were not triggered by earnings misses, but rather by structural contradictions exposed in their financial reports: Oracle’s fiscal 2026 capital expenditures (CAPEX) surged 162% year-on-year, while its free cash flow recorded a historic deficit of -$24 billion; onsemi, meanwhile, saw operating efficiency eroded by large-scale M&A integration, resulting in margin pressure and a sequential 18% increase in inventory days. This seemingly paradoxical juxtaposition—robust profits coexisting with mounting CAPEX risks—constitutes the core tension defining today’s AI hardware investment cycle: the dual themes of “profit surge” and “CAPEX risk” are resonating powerfully, rapidly fracturing the semiconductor industry’s value logic and redefining its valuation anchors.
Profit Surge: Full-Chain Prosperity—from Chip Design to Advanced Packaging
The electronics industry’s doubling of profits is not driven by any single link, but rather by AI compute demand generating a “cascading amplification effect” across multiple technology nodes. On the high-end GPU front, orders for NVIDIA’s Blackwell architecture extend through Q2 2027; TSMC’s CoWoS packaging utilization has remained above 115% (including overtime capacity), driving HPC-related orders at OSAT leaders ASE and SPIL to exceed 40% of total revenue. In memory chips, as HBM3e yield improves to 75%, Micron and SK hynix have jointly raised their full-year 2026 HBM shipment guidance by 30%; unit pricing for HBM3e is ~22% higher than HBM3, directly lifting DRAM manufacturers’ gross margins to 41.2% (year-to-date average). Even more telling is the technology spillover effect: DeepSeek’s open-source DSpark inference acceleration framework—developed jointly with Peking University—delivers 60–85% faster per-user generation speed on V4-Flash model deployment. Its performance hinges critically on high-bandwidth memory access and ultra-low-latency chip-to-chip interconnects—thereby sharply increasing demand for CXL 3.0 controllers, silicon photonics transceivers, and advanced substrates (ABF carrier substrates) supporting chiplet architectures. A telling example: a leading AI server OEM’s Q2 2026 procurement list shows ABF substrate prices up 37% year-on-year, with lead times stretched to 26 weeks.
CAPEX Risk: Unsustainable Expansion Is Triggering Financial Alarms
Yet the dazzling income statement cannot obscure the undercurrents roiling the cash flow statement. Oracle has abruptly raised its CAPEX-to-revenue ratio from 12.3% in 2025 to 28.6% in 2026. Of this incremental spend, 63% funds self-built AI data center clusters—including liquid-cooling systems, custom GPU servers, and optical interconnect backbone networks. However, such assets carry a 7-year depreciation horizon, while AI model iteration cycles have compressed to just 9–12 months. This implies: by the time new clusters go live, the model architectures they host may already be obsolete—supplanted by next-generation inference frameworks. onsemi’s case is even more cautionary: following its 2025 acquisition of a European power semiconductor company, integration costs overran by $420 million; moreover, the acquired firm’s 8-inch wafer fab exhibits process incompatibility with onsemi’s 12-inch IDM platform, causing power device yields to fall by 5.8 percentage points in Q1 2026. This “acquisition = impairment” risk is now spreading to equipment vendors: Applied Materials’ Q2 2026 earnings report shows that 35% of its EUV photoresist deposition tool orders come from AI chip customers—but those customers typically budget CAPEX on a “two-year rolling basis.” Should large-model training paradigms shift toward sparsity or Mixture-of-Experts (MoE) architectures, equipment repurchase rates could plummet precipitously.
Divergence Pressure: Valuation Logic Undergoing Structural Reframing Across Sub-Sectors
This dual-theme resonance is actively reshaping the industry’s valuation coordinate system.
- Semiconductor equipment makers face simultaneous “order booms” and “payment defaults”: ASML secured 17 High-NA EUV orders for 2026, yet four customers have requested delayed deliveries and renegotiated payment terms.
- Foundries confront the paradox of “capacity hunger” versus “customer attrition”: SMIC’s advanced-node utilization rate hit 98% in 2026—but AI customers’ prepayment ratio dropped to 32% (from 51% in 2025), shifting instead to “per-die settlement + performance-based wagering,” partially transferring yield risk back to foundries.
- IDMs, like Intel, find their IDM 2.0 strategy challenged in the AI era: foundry services contribute only 11% of total revenue, while yield volatility on in-house AI chips (e.g., Gaudi3) has delayed server deliveries—triggering cloud provider claims.
- AI infrastructure service providers—covering liquid cooling, optical interconnects, and AI computing center operations—are experiencing extreme bifurcation: firms with proprietary optical engine R&D capabilities trade at PE multiples of 42x; pure-play rack-leasing providers now trade below 15x PE—markets are pricing based on “technical depth,” not “scale expansion.”
Conclusion: Navigating the Valuation Rift Requires Identifying “Authentic Cash Flow Generators”
As Oracle burns five years’ worth of future cash flow on AI infrastructure, and as onsemi pays implicit “technology ransoms” for M&A, markets are using stock prices as rulers to measure the widening chasm between corporate “technology narratives” and “financial realities.” True investment opportunities no longer belong to those boasting the most eye-catching CAPEX figures—but to enterprises capable of converting compute demand into verifiable, sustainable, and scalable operating cash flow: For instance, a domestic HBM interface IP core vendor—whose licensing revenue from the DSpark framework has already covered R&D costs by 217%; or ChangXin Memory, recently granted procurement approval by Apple after achieving >92% yield on LPDDR5X—now entering the iPad Pro supply chain. These companies don’t sell stories built on capital expenditure—they deliver profits anchored in technical penetration. The AI compute wave will not recede—but when the tide goes out, financial statements will definitively separate the swimmers from the surfers.