By Rajesh Dangi
AI has moved from being a workload running inside data centers to becoming the force that defines what a data center is. Power, cooling, networking, storage, file systems, APIs, and application architecture are all being reorganized around the demands of training and inference at scale. In 2026, these pressures have become acute. New York State introduced a moratorium on new data center construction in July, and Texas followed in August by halting new grid connection approvals pending a statewide audit.
OpenAI paused its UK data center project over regulatory and energy cost concerns. Meanwhile, the four largest AI companies have committed 2.4 trillion dollars in future infrastructure spending. The Big Four consulting firms have each staked out distinct positions on this transformation, and their perspectives converge on one insight: AI data centers are not simply a new workload category but a different kind of asset class, one where power procurement, thermal management, regulatory strategy, and capital structuring matter as much as the technology itself.
The Power and Cooling Crunch
The most visible pressure point is power. AI workloads have pushed rack densities from the traditional 5 to 15 kW range to over 140 kW, with next generation platforms targeting as much as 250 kW per rack. The IEA estimates that data centers consumed around 485 TWh of electricity globally in 2025 and projects this to nearly double to 950 TWh by 2030, slightly more than Japan’s annual electricity consumption today.
The strain is already reshaping corporate strategy. In July, OpenAI announced Project Camellia, a 3.2 GW data center in Effingham County, Georgia, representing over 30 billion dollars in investment and enough power for roughly one million households. Microsoft reported adding approximately one gigawatt of AI capacity in a single quarter while reducing GPU deployment times by nearly 50 percent. Meta disclosed 70 billion dollars in future spending commitments, with half allocated to data center leases over 30 years. Alphabet raised its 2026 capital expenditure outlook to 195 to 205 billion dollars.
Cooling has become an equally urgent challenge. Meta is now using closed loop liquid cooling across the majority of its newest AI optimized data centers, circulating a water and glycol coolant that can remain usable for up to a decade. The company also uses reinforcement learning to optimize cooling operations, reducing air cooling fan energy by an average of 20 percent and water consumption by 4 percent. Keppel and Shell have partnered to test immersion cooling fluids in operational data centers in Singapore, with Shell claiming its gas to liquids based fluid can cut energy use by up to 48 percent and boost computing performance by up to 40 percent. Ecolab acquired CoolIT in July to expand its direct liquid cooling platform, targeting 4 billion dollars in high tech business by 2030.
NVIDIA’s Vera Rubin platform, now in full production, incorporates liquid cooling as a design principle. The NVL72 rack contains no cables, fans, or hoses in the chassis, with compute module assembly reduced from hours to one minute. Its 45 degree Celsius liquid cooling inlet enables chiller less dry cooling operation, saving hundreds of gallons of water per megawatt annually for new AI factories.
Networking and the Interconnect Bottleneck
Networking has emerged as another critical bottleneck. As clusters scale to hundreds of thousands of GPUs, GPUs often sit idle waiting for data. NVIDIA’s Spectrum-X Ethernet Photonics, now in production, combines co-packaged optics with switching to deliver 5x better power efficiency and 5x longer AI uptime than networks using traditional transceivers. The technology provides the foundational fabric for million GPU AI factories, with CoreWeave, Lambda, and Oracle Cloud Infrastructure among the first adopters.
Storage and Parallel File Systems
Storage has become the new foundation of AI inference. The challenge is that AI data does not lose its value when a training run ends. It becomes the raw material for the next model, the audit trail for governance, and the baseline for comparison. VDURA and Wasabi announced a technology alliance in August to connect GPU adjacent AI data infrastructure with predictably priced S3 compatible cloud storage. VDURA provides parallel file system storage performance with RDMA data paths and full POSIX workflows, while Wasabi extends the environment with cloud object storage serving as an active archive and long term retention layer.
The operating principle is straightforward: keep AI data close when it is being used and move it freely when it is not. VDURA’s CEO Ken Claffey put it directly: the infrastructure that feeds GPUs is engineered for velocity, and data belongs there while it is doing active work. It should not live there permanently.
APIs and Application Architecture
APIs are evolving into conversational interfaces and agentic control planes. NVIDIA’s Vera Rubin platform was designed specifically for what Jensen Huang calls agentic AI, where one prompt can launch a thousand step journey of reasoning, retrieval, tool use and response generation. The platform delivers 10x agent throughput at scale compared with the previous generation Grace Blackwell platform.
Application architecture is being reorganized around agents and statefulness. The days of monolithic applications serving the same experience to every user are ending. The emerging model treats the application itself as the primary architectural object, with a declarative specification defining the application graph and a behavioural envelope constraining how that graph may behave and evolve.
Operations Move Toward Agent Defined Control
For decades, data centers relied on if then logic, scripts, and human operators watching dashboards. The scale of AI workloads has rendered traditional, human speed management obsolete. Meta’s use of reinforcement learning for cooling optimization is one example. A pilot in the UK with National Grid and NVIDIA found AI software could cut data center energy use during simulated grid strain by more than a third in under a minute.
National Grid Partners’ third annual Utility Innovation Survey found that 74 percent of utility innovation leaders say AI driven data center load growth is impacting grid reliability, and 78 percent are deploying at least one AI application to manage interconnection demand. The role of the site reliability engineer evolves from doing the work to tuning the agents that do the work.
What the Big Four Are Saying
The Big Four consulting firms have each staked out distinct positions on the AI data center transformation, reflecting their different client bases and methodological traditions.
PwC has produced the most sweeping quantitative forecast. Its Global Data Centre Outlook projects cumulative global investment in AI infrastructure will reach 31.6 trillion dollars through 2050, with annual capital expenditure rising from roughly 800 billion dollars in 2026 to 1.8 trillion dollars per year by 2050. Because chips and ICT equipment require upgrades every four to six years, investment will accelerate rather than decline. PwC identifies power as the decisive factor shaping where investment flows. Disrupted chip supply chains could cut cumulative investment by nearly 6 trillion dollars, while a sovereignty driven shift would redistribute rather than reduce global spending.
Deloitte has focused on the practical mechanics of powering the buildout. Its Asia Pacific research warns that data center capacity is growing faster than power generation, creating an energy supply gap. The firm recommends that India leverage its renewable energy base and standardize state level policies to provide predictable round the clock clean power. Deloitte’s survey data shows that 76 percent of respondents see regulatory change as highly impactful, and 90 percent prioritize investment in making infrastructure more intelligent.
EY positions the transformation as a fundamental reclassification of data centers from IT infrastructure to core strategic infrastructure. A typical AI focused data center now consumes electricity equivalent to roughly 100,000 households, and the largest under construction could consume 20 times that amount. EY’s research highlights that 78 percent of CEOs report AI initiatives performing above expectations, and 58 percent expect AI to be a major growth engine over the next two years. EY estimates that global data center investment could reach 7 trillion dollars between 2025 and 2030, with over 2,000 new data centers coming online worldwide.
KPMG has focused on the execution gap. Its research on India’s data center market projects the sector reaching nearly 46 billion dollars by 2033, with AI optimized facilities growing at a 35.1 percent compound annual growth rate. But KPMG argues the sector’s biggest challenge is no longer demand but the complexity of execution. The industry currently depends on fragmented providers across construction, cooling, technology, and operations. KPMG proposes an integrated lifecycle partner model where a single provider oversees the entire chain from land acquisition and power procurement to AI deployment, compliance, and maintenance.
The Bottom Line
AI is not just a workload running in data centers. It is redefining what a data center is. Storage is no longer passive capacity but an active participant in GPU utilization. File systems are no longer just shared directories but the mechanism that prevents expensive accelerators from idling. APIs are no longer rigid interfaces but conversational control planes.
Applications are no longer monoliths serving millions but fleets of agents serving individuals. Operations are no longer human driven but agent defined. Data centers are becoming integrated AI factories where compute, storage, networking, and software are designed as a single system to keep expensive GPUs fed, cooled, and productive. The firms that succeed will be those that treat data center capacity and power as integrated strategic infrastructure rather than an afterthought.