NVIDIA GB300: The Core of AI Compute Power and Market Dominance
NVIDIA's Blackwell platform, particularly the GB300/B300 series, is rapidly becoming central to the AI compute power landscape. This chip series entered mass production in Q4 2025 and is projected to account for over 70% of high-end AI GPU shipments in 2026, further solidifying NVIDIA's leadership in the AI accelerator market. NVIDIA has confirmed that Blackwell chips are produced at TSMC's Phoenix fab, with supercomputer manufacturing bases established in Texas through Foxconn's Houston facility and Wistron's Fort Worth facility. Wistron has activated its first manufacturing base in Fort Worth, USA, assembling the NVIDIA GB300 Grace Blackwell Ultra Superchip. Major OEMs such as Dell, HPE, Lenovo, and Supermicro are expected to begin shipping products based on the Vera Rubin platform (the next generation of GB300) in Fall 2026. NVIDIA has also adjusted its architectural update cycle to annually to meet rapidly iterating market demands.
Surging AI Data Center Power Demand: New Challenges Emerge
Between 2025 and 2026, the electricity demand of AI data centers has surged from a marginal share of total data center load to a primary growth driver. Global data center electricity consumption is projected to reach 565 TWh in 2026, a 26% increase from 447 TWh in 2025. Of this, AI-optimized servers are expected to account for 31% of total data center electricity consumption in 2026 and may surpass traditional servers' consumption in 2027. In the US, for example, data center IT power demand is projected to increase from 9.19 GW in 2025 to 78.57 GW in 2029, a 755% increase in four years. In 2026 alone, demand is expected to double to 17.96 GW. By 2030, AI data center power consumption could account for 8% to 12% of total US electricity consumption.
Power: The New Bottleneck for AI Infrastructure Expansion
The exponential demand for power from AI data centers is placing immense pressure on grid supply. Grid interconnection delays, sometimes exceeding three years, have become a major business risk for AI infrastructure expansion. Concurrently, wholesale power costs near US data centers have risen by 267%, further confirming the severe reality that grid supply cannot meet current demand. Several analytical agencies point out that power has become the primary limiting factor for AI infrastructure expansion:
- Enki.AI (January 2026): The primary constraint on AI infrastructure expansion is no longer capital or technology, but the inability of public grids to provide sufficient reliable power.
- Morgan Stanley (September 2026): AI data centers are consuming power so rapidly that chip supply is no longer the limiting factor for industry growth.
- Gartner (June 2026): AI capacity is now limited by power availability, making data center power security a new battlefield in global AI competition.
- PwC (September 2026): Power will be the decisive factor determining the flow of AI infrastructure investment.
- Omdia (April 2026): The reality of 2026 is severe shortages of power, copper, and critical gases, with scarcity becoming the most profitable product.
NVIDIA's flagship rack-scale AI product, the GB300 NVL72, which includes 72 Blackwell Ultra GPUs and 36 Grace CPUs, has a nominal power consumption of 132 to 142 kW, with peaks approaching 155 kW. In contrast, a traditional enterprise rack averages about 9 kW. The GB300 NVL72 rack weighs 1,580 kg and can dissipate approximately 90% of its heat through liquid cooling, highlighting the stringent requirements of high-density AI compute power on cooling and power infrastructure.
Industry Chain Layout and Response Strategies
Facing the power bottleneck, all parties in the AI industry chain are actively adjusting their strategies:
Power Infrastructure Investment and Independent Power Supply
The industry is shifting from sole reliance on grid power to direct investment in dedicated, on-site power generation solutions. Major companies like Microsoft and Google are securing power independence by directly partnering with energy producers and deploying on-site generation. Meta has signed agreements for over 6 GW of nuclear power to supply its upcoming data centers. By 2030, 30% of data centers are expected to use on-site power generation, up from 13% in early 2024. Institutions like MIT are also researching low-carbon or zero-carbon energy solutions, grid management, and power market policies.
Supply Chain Resilience and Diversification
The bottleneck in the AI chip supply chain has shifted from silicon manufacturing to advanced packaging (e.g., 3D stacking, CoWoS) and High Bandwidth Memory (HBM). The HBM market, dominated by SK Hynix, Micron, and Samsung, has all its 2026 capacity pre-allocated, with gross margins as high as 60-70%. Supply chain concentration and geopolitical friction are driving companies toward diversification and resilience strategies, such as building new manufacturing and packaging facilities in the US and Europe through government incentives (e.g., the CHIPS Act). Besides chips, eight other critical materials are facing shortages: helium, T-glass substrates, ABF substrate films, power management ICs, high-speed networking and optical components, liquid cooling, and power transformers.
Data Center Design and Operations Optimization
Data center operators are implementing more innovative cooling systems (e.g., liquid cooling), integrating renewable energy, and adopting on-site power solutions to improve energy efficiency. DCIM (Data Center Infrastructure Management) software is crucial for planning, deploying, and managing high-density AI infrastructure. NVIDIA also offers the DGX SuperPOD reference architecture as a turnkey data center-scale product for AI factories, including data center design guidance and OT integration of the NVIDIA Mission Control software stack.
Challenges and Risks
Despite active responses from all parties in the industry chain, the rapid expansion of AI infrastructure still faces numerous challenges. Grid interconnection delays, land availability, permitting approvals, and skilled labor shortages are severe challenges for data center development. Public opposition to data centers' impact on water resources, electricity costs, and community quality of life is growing. Furthermore, Bain & Co. predicts that by 2030, AI companies will need $2 trillion in annual revenue to support compute capacity demands, but revenue may fall short by $800 billion, potentially leading to economic unsustainability. Despite improved chip efficiency, cheaper token generation will drive higher overall usage, leading to continued growth in electricity consumption.
