At three in the morning on a Tuesday last month, an operations manager at a mid-sized data center outside Phoenix faced a choice she's made a hundred times before. Server temperatures were climbing into the yellow zone. She could push the chillers harder—guaranteeing another spike in the power bill—or hold steady and pray nothing throttled before sunrise.
She cranked the cooling.
That single decision, multiplied across thousands of facilities worldwide, explains why the global data center cooling market is hurtling toward $100 billion by 2035, up from $26 billion last year. It also explains why a clutch of startups and legacy industrial giants are racing to deploy artificial intelligence that can make those three-a.m. calls obsolete.
The stakes extend far beyond quarterly energy budgets. Buildings consume roughly 30% of global electricity and account for 26% of energy-related CO₂ emissions, according to the International Energy Agency. Data centers alone ate through 415 terawatt-hours in 2024—about 1.5% of worldwide demand—and the IEA expects that figure to more than double to 945 TWh by 2030 as generative AI workloads balloon. In the United States, where data centers used approximately 183 TWh last year, cooling ranks as the second-largest power draw after the servers themselves. Deloitte pegs it at 38 to 40% of a typical facility's energy budget.
The old playbook no longer works. For years, engineers relied on static airflow rules and the occasional computational fluid dynamics simulation to keep server rooms comfortable. That approach held up when racks drew 5 kilowatts and hardware refresh cycles stretched across years. It collapses when AI clusters demand 60 to 120 kilowatts per rack and workloads swing hour by hour. Now a new cohort of companies is fusing physics-informed neural networks, real-time IoT sensor streams, and reinforcement learning to build digital twins that can simulate thermal behavior in seconds rather than days.
The question isn't whether AI can optimize cooling—Google's DeepMind proved that nearly a decade ago, cutting cooling energy by up to 40%. The question is whether the industry can deploy these systems at scale before regulators force their hand.
The Inertia Problem
Step inside most commercial buildings today and you'll find HVAC systems running on schedules written when the lease was signed, perhaps updated once or twice since. The Uptime Institute's 2025 survey delivers a sobering verdict: average data center power usage effectiveness—the ratio of total facility energy to IT equipment energy—has flatlined for six consecutive years, even as rack densities climb. The sustainability metrics that operators promised to improve have "stalled slightly," the report notes, as the AI buildout consumes engineering bandwidth faster than efficiency teams can reclaim it.
Meanwhile, money is pouring into the promise of smarter infrastructure. The smart building software market hit $8.5 billion in 2025, according to research firm Verdantix, and spending on digital twins for buildings is projected to reach $26.2 billion by 2033. Honeywell's February 2025 study found that 84% of commercial building decision-makers plan to increase AI use within the next year, with 55% already deploying AI for energy management. Johnson Controls claims its OpenBlue platform delivers up to 10% energy savings and a 155% three-year return on investment, citing a Forrester Total Economic Impact analysis commissioned by the company.
Yet legacy building management systems remain stubbornly fragmented. Data sits in silos—one vendor for HVAC, another for lighting, a third for security—and semantic tags vary wildly from site to site. The industry is slowly coalescing around standards like BACnet, Brick Schema, and RealEstateCore to enable interoperability, but integration friction and data quality issues persist. Perhaps more revealing, operators still hesitate to hand full control to algorithms. The Uptime survey shows trust in AI peaks for analytics and predictive maintenance; it drops considerably when you ask about direct autonomous equipment control.
Trust, it turns out, has to be earned one watt at a time.
Regulation Tightens the Screws
New York City's Local Law 97 is about to make the trust issue urgent. The statute mandates emissions reporting for calendar year 2024, with filings due May 1, 2025, and fines of $268 per metric ton of CO₂ equivalent over the limit. Those caps tighten further between 2030 and 2034. Washington, D.C., and other jurisdictions have enacted Building Performance Standards with staged compliance cycles that grow more stringent every few years. California's 2025 Energy Code, which took effect January 1, 2026, encourages heat pumps and demand-flexible controls, estimating $5 billion in savings over 30 years if adoption reaches projected levels.
Across the Atlantic, the European Union's revised Energy Efficiency Directive requires an 11.7% consumption reduction by 2030 and introduces mandatory sustainability reporting for data centers. The first key performance indicator reports landed in September 2024, and Brussels plans to unveil a rating scheme with potential minimum performance standards by March 2026. In Ireland—where data centers already consume 22% of national electricity, a figure expected to hit 31% by 2034—regulators now require new facilities to match import capacity with on-site generation or storage. Singapore updated its Green Mark for Data Centre criteria in 2024, emphasizing intelligent systems and carbon reduction.
Policy deadlines, in other words, are no longer theoretical.
Technology Catches Up—Fast
The tools to meet those deadlines are arriving faster than many expected. Physics-informed neural networks can generate computational fluid dynamics surrogates that run thousands of times faster than classical solvers. Recent research reports order-of-magnitude speed-ups with accuracy scores above 95% on benchmark flows. Lawrence Berkeley National Laboratory demonstrated a live digital twin for HVAC, solar, and battery systems in February 2026, enabling real-time testing of energy strategies without disrupting actual operations. Reinforcement learning agents trained on 2,000 hours of production data have delivered 14 to 21% cooling energy savings without safety violations in offline trials, according to published results.
Liquid cooling adoption is inflecting, too. One industry survey found roughly one in three data centers plan to deploy liquid cooling within two years, up from an 11% baseline in Uptime's 2023 count. As rack densities push past 60 kilowatts, air-based systems simply can't keep pace without enormous fan energy. Vendors like Vertiv, Modine, and Schneider Electric are expanding portfolios to capture the shift. The Wall Street Journal recently dubbed cooling equipment "a hot stock market trade"—a headline that would have seemed absurd five years ago.
Add the broader thermal management sector—sensors, controls, liquid cooling systems—and you're looking at a $15 billion market growing toward $33 billion by the early 2030s, according to analyst estimates.
Who's Making It Work
Google's DeepMind team set the benchmark years ago, cutting cooling energy by up to 40% and reducing overall PUE overhead by about 15%. The system later moved to autonomous operation under human supervision. That success inspired a wave of commercial offerings, though not all deliver on the hype.
BrainBox AI, which uses reinforcement learning for HVAC optimization, reports a 15.8% energy reduction and more than $42,000 in annual savings at 45 Broadway in New York City. A pilot across 600 Dollar Tree stores saved approximately 7.98 gigawatt-hours and $1.03 million, prompting expansion to over 2,000 locations, the company says. EkkoSense, which layers AI atop dense sensor meshes, claims up to 30% cooling energy reductions for customers including Virgin Media O2 and Telehouse.
In Spain, a hospital using Siemens' Building X platform cut operational costs by as much as 35% by integrating energy, security, and HVAC under one AI-driven interface, according to a Siemens case study. PassiveLogic raised $74 million in September 2025 to scale what it calls "physical AI," building physics-based digital twins with its Hive controller and Sense Nano sensors. Cadence's Reality Digital Twin platform, which absorbed the former 6SigmaDCX team, claims up to 30% efficiency gains by fusing CFD with live building management system and data center infrastructure management feeds. A healthcare enterprise using Reality DC achieved improved thermal compliance and airflow optimization, the company reports.
Berkeley Lab's model predictive control demonstration for dual-fuel systems—heat pump plus gas furnace—showed a 27% cost reduction and eliminated furnace use over two months in a test building. Academic pilots using what researchers call "reliable protocol" designs report average savings around 13% in commercial buildings, though full deployment costs and rigorous A/B testing remain underreported in many vendor case studies.
One of the newest entrants is Inviscid AI, a Y Combinator Winter 2026 startup founded by Kabir Jain and Ziming Qiu. The company claims its physics-informed neural networks can simulate airflow and thermal behavior roughly 1,000 times faster than traditional CFD, achieving over 95% accuracy on validation benchmarks. Inviscid integrates with building management systems to process thousands of IoT streams in real time, targeting dead zones and wasted cooling capacity. An early HVAC ventilation optimization case showed 40% better flow distribution in 240 times less compute time, according to the founders. They emphasize a "physics-first" approach that marries domain knowledge with neural surrogates, positioning the startup to compete against both established BMS vendors and pure-play machine learning platforms.
Whether Inviscid can scale beyond early pilots remains to be seen, but its pitch resonates with an industry hungry for solutions that respect the laws of thermodynamics.
What Comes Next
The infrastructure is converging, if unevenly. Schneider Electric is preparing a unified EcoStruxure Foresight Operation platform for broader release in the second half of 2026, aiming to tie together energy, power, and building controls in a single pane of glass. Johnson Controls continues to expand OpenBlue with generative AI features and autonomous control modules. Honeywell, Siemens, and Distech Controls are all investing in edge controllers and cloud analytics that can orchestrate portfolios at scale.
Interoperability standards are finally gaining traction. The Brick Schema and RealEstateCore teams are harmonizing their ontologies, and BACnet maintains dominant market share in building automation. As semantic tagging improves—admittedly a slow process—the friction of integrating disparate systems should decline, unlocking portfolio-scale AI deployments that were previously cost-prohibitive.
Physics-informed surrogates will likely move from research labs into production workflows faster than skeptics expect. NVIDIA's Modulus and other frameworks are making it easier to train neural operators for turbulence and transient thermal problems. Expect to see more vendors offering "instant CFD" for design reviews and operational what-if scenarios, compressing iteration cycles from weeks to minutes.
Policy deadlines will force action. New York City building owners must file their first Local Law 97 reports in a matter of weeks. EU data centers face ratings and potential minimum standards next year. California's 2025 code is already live. McKinsey projects data centers could consume 5 to 9% of global electricity by 2050 if current trajectories hold. Without aggressive efficiency measures—liquid cooling, waste heat reuse, AI-driven optimization—that load will strain grids and balloon carbon footprints at a time when utilities can least afford it.
Measurement and verification remain the gating factor. Field studies stress rigorous experimental designs and transparent reporting of integration costs. Operators need proof that savings translate to the bottom line and won't introduce new failure modes. The companies that can deliver auditable, repeatable results with minimal disruption will capture the largest share of a market racing toward $100 billion.
For founders, the message is clear: physics-informed intelligence isn't a nice-to-have. It's the cost of entry in a sector where every watt counts and regulators are watching. That three-a.m. decision in Phoenix? Pretty soon, an algorithm will make it. The only question is whose.
