The irony borders on cosmic.
Artificial intelligence's hunger for electricity is reshaping the global power grid—data centers are projected to consume roughly 415 terawatt-hours worldwide in 2025 and could hit 945 TWh by 2030, according to European Commission projections. Yet the same computational engines driving that surge are now being repurposed to solve the energy crisis they've helped trigger. The twist? These new systems don't rely on machine learning alone. They're wedding it to something far older: the immutable laws of physics.
Buildings and data centers together swallow about 30% of the planet's energy and generate 26% of energy-related carbon emissions, per 2022 data from the International Energy Agency. In U.S. commercial buildings, HVAC systems account for roughly 51% of major fuel consumption, based on the most recent comprehensive federal survey from 2018. For years, the industry has chased marginal gains—better insulation, LED retrofits, demand response nudges. The Department of Energy's Better Buildings Initiative tallied nearly $22 billion in cumulative savings from its partners as of last September, mostly through conventional efficiency measures.
But conventional has stalled out. The Uptime Institute's 2024 survey of global data centers found average power usage effectiveness lingering around 1.56, with industry-wide efficiency improvements essentially flatlined. Meanwhile, regulators are cranking up the pressure. New York's Local Law 97 slaps penalties of $268 per metric ton of carbon dioxide equivalent over the limit. Washington, D.C.'s first building performance standards compliance cycle concluded at the end of 2024, with review expected this year. Boston, Denver, Seattle—all have rolled out phased performance mandates stretching into the 2030s.
Enter what researchers call physics-informed AI: a technical approach that hardwires physical laws directly into machine learning architectures instead of letting models learn patterns from scratch.
The Incumbents and Their Claims
The building energy optimization market isn't exactly virgin territory. Honeywell's Forge platform, Siemens' Building X suite, Schneider Electric's EcoStruxure—all have deployed machine-learning-based building management systems promising 10% to 20% average savings. Trane published a case study with BrainBox AI across more than 120 facilities claiming 26.1% carbon reduction and 16.7% electricity savings at test sites during the first six months, with portfolio-wide cuts of 1,132 metric tons CO2 equivalent and $329,000 over 18 months.
Those numbers are solid, if not revolutionary.
For data centers specifically, early movers like Vigilent reported a 20% PUE reduction at an Evoque facility in Lisle, Illinois, in 2021, cutting the average number of computer room air handlers running from 103 to about 63. Google's DeepMind made headlines back in 2016 with claims of up to 40% cooling energy reduction using reinforcement learning—equivalent to roughly a 15% drop in total PUE overhead.
Yet most of these systems treat buildings as black boxes. They learn correlations between sensor data and outcomes without understanding why air flows the way it does, or how heat actually transfers through materials. That works fine until conditions shift—a sensor fails, occupancy patterns change post-pandemic, equipment degrades. Then the model needs retraining, or starts making suboptimal calls.
The data center surge makes the urgency harder to ignore. The IEA's recent electricity outlook flagged data centers as a key demand driver, with U.S. facilities alone consuming approximately 180 TWh in 2024. Global demand growth is accelerating through the end of the decade, outpacing GDP. Recent research suggests data centers typically burn 30% to 40% of their energy on cooling, a share that's climbed as AI workloads generate denser heat loads in tighter spaces.
The Physics Bottleneck
Computational fluid dynamics—the mathematical modeling of how air, heat, and fluids move through physical spaces—has been around for decades. Engineers use CFD to design aircraft wings, optimize HVAC systems, simulate everything from turbine blades to blood flow. The problem? Traditional solvers are painfully slow for real-time operations. Running a high-fidelity CFD simulation of airflow in a data center or commercial building can consume days or weeks on conventional hardware. Useless for live optimization.
This is where physics-informed neural networks and related techniques create daylight. By encoding physical constraints—conservation of mass, energy, momentum; thermodynamic relationships—directly into neural network architectures or loss functions, these models can approximate complex physics orders of magnitude faster than traditional numerical solvers while keeping reasonable accuracy.
Academic research has exploded over the past few years. A paper from early this year introduced BESTOpt, a modular physics-informed machine learning framework for building modeling, control, and optimization. Another recent preprint candidly examined "lessons learned from field demos" of model predictive control and reinforcement learning for HVAC, discussing real-world deployment costs and the chronic problem of data scarcity. NVIDIA's Modulus framework, continuously updated since 2021, now offers pre-trained physics models with fine-tuning recipes for aerodynamics and other applications. One study last year demonstrated 10x to 100x speed-ups versus classical solvers across fluid flow benchmarks.
The bridge from research to production, though, remains perilously narrow.
The New Entrants
Inviscid AI emerged from Y Combinator's Winter 2025 batch with a straightforward pitch: real-time building simulations that marry live IoT sensor data with GPU-accelerated computational fluid dynamics. Founded in 2025 by Kabir Jain and Ziming Qiu, the San Francisco two-person team claims their approach runs "approximately 1000x faster than traditional solvers" with "95%+ accuracy," according to their YC profile and company website.
The startup's positioning is aggressive. They promise to "transform weeks of CFD into seconds," enabling what they call "live digital twins" that can autonomously optimize buildings and data centers in real time. The company claims 15% to 30% energy reduction potential, with early results including "2x cooling efficiency gains in a data center" mentioned in a LinkedIn post from the seventh week of their YC batch. They've signed a memorandum of understanding with Madhya Pradesh State Data Center in India to deploy a live digital twin at a government facility.
Treat those numbers with the appropriate skepticism—no third-party case studies or rigorous measurement-and-verification reports are publicly available yet. Still, the claims align with what established players have achieved over longer timelines using less sophisticated methods.
Inviscid isn't alone. PassiveLogic, which won a 2024 AI Breakthrough Award, deploys physics-based digital twins through its "Hive" and "Autonomy Studio" ecosystem for autonomous building control. The company markets its approach as distinct from black-box machine learning, emphasizing that physical constraints prevent unsafe control actions—a pitch aimed squarely at facility managers who've watched too many YouTube compilations of robots going haywire.
For data center cooling specifically, Phaidra has gained traction with autonomous AI agents. The company announced a non-binding memorandum of understanding with UAE-based Khazna data centers in February for pilot deployments, and reported NVIDIA tests showing roughly 80% reduction in thermal spikes for liquid cooling systems. Vigilent continues expanding its AI-based cooling control for computer room equipment.
On the incumbent side, Schneider Electric ran field trials in Canada from mid-November to late January on "edge AI" room controllers that achieved an average 5% energy reduction versus non-AI controllers, with reductions reaching 15% under specific conditions, according to a trade publication report from last March. The controllers process control logic locally rather than relying solely on cloud-based optimization—potentially addressing latency and resilience concerns that keep some operators up at night.
When Penalties Concentrate the Mind
Perhaps nothing focuses attention like structured fines.
New York City's Local Law 97, with its first compliance period running 2024 through 2029, imposes $268 per metric ton of carbon dioxide equivalent for emissions over established limits. For a 100,000-square-foot office building exceeding its cap by 200 metric tons annually, that's $53,600 in penalties. Every year. Multiply across a portfolio and the numbers become existential.
California's SB 253, requiring Scope 1 and 2 emissions reporting with first filings that came due last August, applies to thousands of large companies operating in the state. (SB 261, which would have required climate-related financial risk reporting, was stayed by the Ninth Circuit last November, though the larger disclosure trend persists despite legal uncertainty.)
Washington D.C.'s first Building Energy Performance Standards cycle concluded at the end of 2024, with review expected this year. Boston's BERDO 2.0 emissions standards kick in by building size through 2025 and 2030. Denver's Energize Denver targets phase in for buildings over 25,000 square feet. Seattle's building performance standards, signed in late 2023, set first compliance dates from 2031 through 2035 depending on size.
Internationally, the European Union adopted a data center sustainability rating scheme that requires operators to report key performance indicators annually by May 15. The Commission is expected to unveil a broader data center energy efficiency package sometime this spring. These frameworks create transparent benchmarking that—whether operators like it or not—will expose underperformers.
JLL's recent policy tracker notes that city-level regulations are "driving building transformation," forcing investors and occupiers to rethink asset strategies. You can either retrofit and optimize proactively, or pay penalties and watch valuations erode. The business case for AI-driven optimization suddenly looks considerably less speculative when the alternative is structured bleeding.
What Actually Works
The most mature applications remain HVAC sequencing and set-point optimization, where sensor-rich environments and clear feedback loops enable robust machine learning. The Department of Energy's Smart Energy Analytics Campaign found energy management information systems paired with fault detection delivering median savings of 4% to 9% within two years, with some portfolio cohorts achieving higher. Commissioning literature shows 3% to 16% median savings with typical paybacks around two years.
At the high end, a paper published last year in the journal Energy described a "generic framework for integrating AI into building automation systems" that reported a chiller sequencing AI deployment saving approximately 3,945 kilowatt-hours per day. One case study hardly constitutes universal proof, but it suggests the upper bound of what's possible under favorable conditions.
75F, an IoT-based building management provider targeting small-to-medium portfolios, cites evaluation work claiming "up to 31% total building energy savings" across 14 building types in its marketing materials. (Direct verification of that figure proved elusive during reporting.) Their hospitality and retail case studies show HVAC energy reductions varying significantly by site characteristics—which is perhaps the most honest thing anyone in this space has admitted.
For data centers, the frontier is shifting toward liquid cooling optimization as density increases with AI workloads. A recent preprint on digital twin models for Frontier supercomputer infrastructure suggested 20% to 30% energy savings potential under constrained optimization of hot-water cooling loops. Whether those gains translate to typical enterprise or colocation facilities is an open question, but it signals where hyperscale operators are placing bets.
Inviscid AI's approach—accelerating CFD to near-real-time with physics-informed neural operators—targets a different bottleneck: design iteration and live airflow/thermal optimization rather than just reactive control. Their case examples include HVAC vent optimization claimed to be "240x faster" with "40% better flow," as well as non-building use cases like coastal infrastructure stress analysis and storm surge forecasting. The generality of the platform is notable. It also raises questions about specialization versus breadth—whether trying to solve everything means solving nothing particularly well.
PassiveLogic's emphasis on physics-based digital twins that "understand" building systems addresses a risk that's kept many facility managers skeptical: black-box ML models can issue control commands that violate safety constraints or damage equipment if not properly constrained. Encoding thermodynamic limits and equipment operating ranges directly into the control logic—whether through physics-informed training or explicit guardrails—reduces operational risk. It's the difference between a model that "learns" not to melt a chiller by trial and error versus one that knows from first principles that certain commands are thermodynamically impossible.
Still, real-world hurdles persist. A recent preprint examining field deployments of advanced HVAC control candidly discussed deployment and maintenance costs, data scarcity when sensors are sparse or unreliable, and model drift when building usage patterns shift. Cybersecurity of operational technology networks remains a concern anytime cloud-based AI interacts with legacy building management infrastructure. Integration complexity is real, and measurement-and-verification rigor is often underwhelming in vendor case studies. Most companies report savings against modeled baselines rather than metered pre-and-post comparisons, which leaves considerable room for creative accounting.
The Next Act
The convergence of regulatory pressure, energy cost volatility, and technical maturation suggests this isn't hype. It's infrastructure necessity—or at least, that's the pitch.
Market research firms project the global smart buildings market growing from roughly $142 billion in 2025 to somewhere between $548 billion and $554 billion by the early 2030s, though methodology varies wildly across sources. Digital twins for buildings specifically are pegged by some analysts to balloon from around $2 billion last year to over $26 billion by 2033.
Take those projections with industrial-grade skepticism. But the directional signal is hard to argue with: building owners and data center operators face a multi-decade capital cycle to hit net-zero targets, and software-driven optimization is orders of magnitude cheaper than wholesale equipment replacement.
The next year or so will likely separate science projects from scalable platforms. Startups like Inviscid AI need to move beyond founder-reported claims to third-party-verified case studies with rigorous measurement and verification. Incumbents like Siemens, Schneider, and Honeywell will continue integrating physics-informed techniques into existing platforms, leveraging installed base and channel relationships that took decades to build. Data center operators, facing both capacity constraints and efficiency mandates, may prove the most aggressive early adopters—particularly for liquid cooling optimization as GPU clusters proliferate.
For proptech investors, the diligence questions are sharpening. Does the platform embed physical constraints, or just learn correlations? Can it handle sensor failures and regime shifts without catastrophic performance degradation? What's the integration path for legacy building management systems, and who bears the cybersecurity risk? Most importantly, what's the measurement framework—are savings calculated against a modeled baseline, or metered pre-and-post with independent verification?
FERC Order 2222, which enables distributed energy resource aggregation in wholesale markets with implementation advancing through this year and next, opens additional revenue streams for buildings capable of grid-interactive flexibility. Pairing physics-informed optimization with demand response and time-of-use arbitrage could unlock stacked value that justifies higher upfront investment. Maybe.
The phrase "AI for buildings" has been overused to the point of meaninglessness—right up there with "synergy" and "paradigm shift" in the corporate jargon hall of shame. What's emerging now feels different, though perhaps that's wishful thinking. The physics doesn't lie. Air moves, heat transfers, energy dissipates according to laws that don't change with the weather or the market or the latest earnings call. The question is whether startups can operationalize that reality fast enough to matter, or whether the window closes as incumbents absorb the best ideas and lock in distribution.
Given the stakes—trillions in potentially stranded assets if buildings can't hit net-zero, grid instability if data centers keep doubling consumption—it's a question the industry can't afford to get wrong. The irony would be if AI, having created the energy crisis, can't quite solve it either.
