Global AI Infrastructure Race Heats Up: Capital Expenditures Shift to Physical Compute Hardware
The Global AI Infrastructure Race Heats Up: A Paradigm Shift in Capital Expenditure Is Rewriting the Logic of the Semiconductor and Compute Value Chain
When ByteDance raised its 2024 AI infrastructure capital expenditure by 25% to RMB 200 billion (approx. USD 27.5 billion), and when Apollo Global Management and Blackstone jointly extended a USD 35-billion private credit facility dedicated to AI chips for Broadcom—one of the largest single private-credit transactions ever executed—the global AI industry’s evolutionary coordinates have fundamentally shifted. This is not merely another leap in model parameters; it is a wholesale reconstruction of the compute foundation. It is no longer about lab-based technology validation, but rather the large-scale deployment of gigawatt-level power systems, ten-thousand-GPU clusters, hundred-kilometer optical interconnects, and kiloton-scale liquid-cooling infrastructure. AI investment logic is rapidly pivoting from “algorithm-driven” to “infrastructure-driven.” A global infrastructure arms race—spanning chip design, advanced manufacturing, system integration, and energy support—is now entering a white-hot phase.
A Paradigm Shift in Capex: From “Model Development” to “Scaled Deployment of Hard Infrastructure”
Over the past three years, the dominant narrative of the AI wave has centered on breakthroughs in large language models (LLMs): OpenAI’s GPT series, Google’s Gemini, Meta’s Llama, and China’s vibrant “hundred-models-race” ecosystem. Capital flows were heavily concentrated in algorithmic R&D teams, data acquisition, and small-scale training clusters. Yet starting in Q2 2024, the signal became unmistakable: marginal gains in model capability increasingly depend on compute density and engineering efficiency—not simply scaling up parameter counts. ByteDance’s RMB 200-billion capex plan significantly exceeds its ~RMB 160 billion actual outlay in 2023—and explicitly targets physical-layer infrastructure: building proprietary AI computing centers, taping out custom AI chips, deploying full-stack liquid-cooling systems, and upgrading interconnection architecture for ten-thousand-GPU clusters. According to internal sources cited by the South China Morning Post, the power density of its newly built data centers has reached 40 kW per rack—more than triple that of traditional IDCs—imposing disruptive requirements on power supply, thermal management, and network bandwidth. This marks the transition of AI investment beyond the “technical feasibility” validation stage into the decisive “economic scalability” phase: whoever can orchestrate more effective compute per unit of energy consumption will command pricing power and ecosystem gateways for next-generation AI applications.
Deep Restructuring of the Hardware Demand Chain: Custom GPUs/TPUs, Advanced Packaging, and High-Speed Interconnects Emerge as Critical Bottlenecks
Massive infrastructure build-outs are directly triggering structural demand upstream. Traditional procurement of general-purpose GPUs is no longer sustainable: an NVIDIA H100 GPU consumes up to 700 W; a ten-thousand-GPU cluster consumes over 1.5 TWh annually—equivalent to the annual electricity use of a mid-sized city—and PCIe bandwidth has become a major bottleneck for training efficiency. The market is rapidly shifting toward ASIC specialization, chiplet architectures, and co-packaged optics (CPO). Broadcom’s USD 35-billion financing is primarily earmarked to accelerate volume production of its Tomahawk 5 switching chips and Jericho 3 DPUs—and to fund tape-outs for its next-generation AI inference chip, “Trident.” Built on TSMC’s N3E process node, Trident integrates an 8-nm I/O die with HBM3 stacks, targeting end-to-end communication latency reductions to the nanosecond level. Domestic vendors are moving in parallel: Cambricon’s MLU590 already supports heterogeneous chiplet integration; Huawei’s Ascend 910B achieves a 40% improvement in energy efficiency via synergistic optimization between its proprietary DaVinci architecture and Kunpeng CPUs. Crucially, advanced packaging has become the “invisible valve” governing compute release: ASE and Amkor have expanded CoWoS capacity to 100,000 wafers per month—but lead times remain at 26 weeks. Meanwhile, China’s JCET and Tongfu Microelectronics are accelerating adoption of fan-out RDL technology, yet domestic TSV (through-silicon via) yield for HBM3 remains the single largest bottleneck for localization.
Energy & Thermal Management: Soaring Liquid-Cooling Adoption Amid Mounting Grid Strain
Behind the RMB 200-billion infrastructure investment lies a staggering energy ledger. Industry estimates suggest ByteDance’s new compute clusters may consume over 8 billion kWh annually—equivalent to the yearly electricity demand of a city of one million residents. Air cooling has hit its physical limits at 40 kW/rack, making liquid cooling—especially cold-plate and immersion variants—a hard requirement. Sugon and Inspur reported a 180% year-on-year surge in liquid-cooled server shipments in Q1 2024, yet domestic substitution rates for critical subsystems—including coolant circulation systems, specialized pumps/valves, and high-thermal-conductivity materials—remain below 35%. Even more pressing is grid compatibility: the dense rollout of AI computing centers in Inner Mongolia and Gansu has widened regional peak-to-trough load differentials. In April, industrial electricity consumption in Inner Mongolia grew by 12.7%, markedly outpacing the national average. China’s National Energy Administration has launched pilot programs for “direct green-power supply to AI computing centers,” but bridging the gap between the intermittency of wind/solar generation and the inflexible, constant-load nature of AI workloads urgently requires closed-loop solutions combining energy storage (e.g., CATL’s sodium-ion batteries), intelligent dispatch algorithms, and virtual power plants (VPPs)—sparking a fundamental re-rating of emerging energy-tech companies.
Global Supply Chain Realignment: Structural Opportunities in Equipment, IDMs, and Optical Modules
The capex wave is exerting dual pressure on the global semiconductor supply chain. On one hand, it forces equipment vendors to accelerate innovation: ASML’s High-NA EUV lithography tools have order backlogs extending to 2026; Lam Research’s etch-equipment revenue surged 42% YoY in Q1 2024, driven by both advanced packaging and memory-chip expansion. On the other, IDM (Integrated Device Manufacturer) models are undergoing value re-assessment: Under Intel’s IDM 2.0 strategy, its foundry division has secured orders from Broadcom and Qualcomm; SMIC is fast-tracking its N+1/N+2 process platforms to meet domestic AI chip tape-out demand. The optical module segment is experiencing especially explosive growth: the unit price of 800G DR8 modules is nearly three times that of 400G modules, propelling YOY revenue growth of 110% and 95% for Innolight and Eoptolink in Q1 2024. CPO technology will further accelerate domestic substitution of silicon photonics chips and lasers. Notably, data-center REITs—such as U.S.-based EQIX and China’s Penghua Shenzhen Energy REIT—are emerging as new investor favorites: their underlying assets are shifting from “space leasing” to “compute-service contracts,” with rental income tightly linked to AI cluster rack-up rates—and valuation anchors evolving from P/FFO metrics toward “annual cash flow per watt of compute.”
Conclusion: The Infrastructure Race Is, Fundamentally, a Contest for National Compute Sovereignty
ByteDance’s RMB 200 billion and Broadcom’s USD 35 billion may appear as corporate initiatives—but they are, in essence, microcosmic projections of national strategies for sovereign control over AI infrastructure. As AI becomes the core productive force in great-power competition, compute infrastructure is equivalent to the “national highway network” and “ultra-high-voltage power grid” of the digital era. It is no longer solely about commercial efficiency; it implicates data security, industrial standard-setting authority, and agency in technological generational leaps. This race has no finish line—only perpetual dynamic equilibrium: between Moore’s Law driving technical iteration and physics imposing energy constraints; between global supply-chain efficiency and geopolitical resilience. Investors who continue viewing semiconductors through the lens of consumer-electronics cycles risk missing the deepest structural opportunity of this cycle. The true winners will be “hard-tech integrators”—those uniquely capable of mastering the physical limits of chips, the complexity of energy systems, and the tensions within global supply chains.