Kabir Jain's pitch rests on equations that predate the digital age by more than a century. Which is part of the point.
As artificial intelligence pushes data center electricity demand toward what the International Energy Agency projects will be a near-doubling by 2030—from roughly 415 terawatt-hours in 2024 to perhaps 945 TWh—the same AI systems are increasingly being tasked with solving the energy crisis they're creating. It's a paradox that has become something of an industry punchline, except the stakes are no longer academic. Buildings and data centers need to get smarter about energy, and fast. The question is whether the current wave of AI optimization tools can deliver verified savings quickly enough to keep pace with mounting regulatory pressure and power bills that are, in some regions, becoming existential constraints.
Inviscid AI, the San Francisco startup Jain founded in 2025 and brought through Y Combinator's Winter 2026 batch, is entering this market at what might charitably be called a pivotal moment. Or, less charitably, a crowded one. The company's thesis hinges on what Jain describes as "physics-informed models embedding conservation laws and system structure"—a deliberate departure from the black-box machine learning that has dominated the building automation conversation for the past several years. Inviscid fuses computational fluid dynamics with real-time IoT sensor data to build digital twins for buildings and data centers, targeting airflow optimization, HVAC energy reduction, and predictive maintenance.
The company's website makes bold claims: simulations that run "1000x faster" than traditional CFD, HVAC vent optimization "240x faster, 40% better flow." These are vendor-reported figures, not yet backed by third-party measurement and verification. But they signal the performance ambitions of a new generation of physics-first tools, and perhaps more importantly, they reflect a bet that speed—not just accuracy—will be the competitive advantage as compliance deadlines tighten and power constraints bite harder.
Whether a two-person startup can scale that bet into a platform capable of competing with Honeywell and Siemens remains to be seen. But the tailwinds are undeniable.
The Urgency Is New, Even If the Problem Isn't
Buildings account for roughly 30% of global final energy consumption and 26% of energy-related CO₂ emissions, according to the IEA's overview published in 2025. Within that envelope, HVAC systems consume somewhere between 30% and 50% of a commercial building's energy, though precise figures vary by climate zone, building vintage, and how aggressively the operator tunes the system. The 2018 Commercial Buildings Energy Consumption Survey remains the U.S. baseline—a reminder of how slowly definitive data catches up to evolving infrastructure.
Data centers represent the sharper, more urgent edge of the energy challenge. The IEA projected in May 2025 that data center electricity use would climb from about 415 TWh in 2024 toward 945 TWh by 2030, with AI workloads driving much of that surge. Gartner echoed this in November 2025, forecasting a 16% jump in data center electricity demand in 2025 alone and a doubling by 2030, with U.S. data centers expected to consume 7.8% of regional electricity by decade's end. S&P Global Market Intelligence pegged the lower-bound scenario at 860 TWh in 2025 climbing to 1,587 TWh by 2030.
The range depends on deployment pace and efficiency gains, but every forecast points in the same direction. And that direction is up.
Against this backdrop, the market for building management systems and energy management software is expanding rapidly, though estimates diverge—sometimes wildly—on scale and pace. Mordor Intelligence put the Building Management System market at $51.25 billion in 2026, projecting growth to $140.73 billion by 2031, a compound annual growth rate near 22.4%. Other analysts offer more conservative figures. 360iResearch forecast Building Energy Management Systems from $46.10 billion in 2026 to $83.77 billion by 2032, a 10.43% CAGR. Global Market Insights estimated the BEMS market at $14.7 billion in 2024 with a 7.5% CAGR through 2034.
The spread reflects differing definitions of what counts as "management" versus "optimization" and whether services are bundled. But the direction, again, is uniform: building owners and operators are spending more on software and controls, and AI is the headline feature.
Adoption intent surveys reinforce the momentum, if not always the budget follow-through. Honeywell's February 2025 study found that 84% of commercial building decision-makers planned to increase AI use in the following year. JLL's 2025 Facilities Management Report reported that 92% of organizations had either piloted AI tools for commercial real estate or planned to start in 2025, up sharply from 61% in 2024. ABI Research claimed that more than one million buildings had adopted AI energy-efficiency solutions as of 2025.
Yet budget caution persists. JLL noted that only 32% of facilities managers planned to increase software spending in 2025, down from 39% in 2024—a reminder that intent doesn't always convert to immediate budget allocation. Pilot fatigue, perhaps, or ROI uncertainty creeping in.
Three Forces Converging
Three forces are converging to accelerate the shift toward physics-informed, AI-driven building optimization: regulatory mandates that now carry real penalties, technology maturation that has moved PINNs and digital twins from labs to product roadmaps, and the practical realities of a power-constrained world where electrons, not algorithms, are the limiting factor.
The regulatory ratchet is turning. New York City's Local Law 97 required the first emissions reports for calendar year 2024 to be filed by May 1, 2025, with penalties of $268 per ton of CO₂ equivalent for buildings that exceed carbon caps. The first compliance period runs from 2024 to 2029, and the deadlines are now annual. This isn't a future concern—it's happening.
Colorado's HB21-1286 set statewide building performance standards with energy reduction targets of 7% by 2026 and 20% by 2030 against a 2021 baseline. Washington State's Clean Buildings Performance Standard covers buildings 50,000 square feet and larger, with compliance windows running from 2026 to 2028. Seattle's city-level ordinance aims for net-zero emissions by 2050, with phased targets beginning in 2031.
In Europe, the Energy Performance of Buildings Directive recast entered into force on May 28, 2024, with EU member states required to transpose it into national law by May 29, 2026. The directive positions the bloc toward zero-emission buildings and opens the door to future binding measures. The EU also implemented a data center sustainability rating scheme in March 2024, requiring key performance indicators to be reported first by September 15, 2024, and then annually by May 15.
California's SB 253, the Climate Corporate Data Accountability Act, mandates Scope 1 and 2 emissions reporting with the first filings due August 10, 2026, per the California Air Resources Board's updated timeline approved in February 2026.
The compliance burden is real, and it's forcing building operators to shift from periodic tune-ups to continuous optimization. The old playbook of commissioning, forgetting, and re-commissioning every few years no longer works when annual reporting windows and escalating penalties loom.
Technology convergence is making that continuous optimization feasible. Physics-informed neural networks, a hybrid approach that embeds conservation laws and partial differential equations directly into machine learning architectures, have matured from academic curiosities into practical tools. A January 2026 preprint on "NewPINNs" described a framework that delegates physics enforcement to numerical solvers, addressing classic PINN failure modes around shocks and discontinuities—problems that had stymied earlier implementations.
Recent peer-reviewed work through late 2025 explored physics-informed reinforcement learning and model predictive control for HVAC systems, reporting double-digit simulated energy savings. The caveats about safe exploration and real-world validation remain, but the direction of travel is clear.
NVIDIA's Modulus and PhysicsNeMo platforms saw updates across 2025, pushing toward neural operators like Fourier Neural Operators and DeepONet architectures. Omniverse Blueprints integrated assets from Schneider Electric's ETAP and Vertiv for AI factory digital twins, signaling the mainstreaming of physics-informed workflows in industrial settings. Ansys launched TwinAI workshops in early 2025, emphasizing hybrid analytics for near-real-time deployment to IoT platforms. Cadence's Reality DC platform, built on the acquisition of Future Facilities' 6SigmaDCX CFD stack in 2022, positions itself as a data center digital twin standard.
Academic frameworks are proliferating at a pace that suggests commercialization is imminent. BESTOpt and OptAgent, both released in preprints in January 2026, layer agentic AI on top of physics-consistent digital environments for end-to-end building modeling and control. PILLM, published in October 2025, applied physics-informed large language models to HVAC anomaly detection with interpretable rules. A March 2026 preprint explored digital twin-based data center cooling optimization at the Frontier exascale computing site. The research pipeline is dense, and the gap between preprint and product is narrowing faster than it used to.
Power constraints are the unspoken driver. Phaidra's CEO Jim Gao framed it bluntly in an October 2025 interview with GeekWire: "We live in a power-constrained world… the ability for big AI companies to generate revenue is literally limited by electrons available."
That quote has circulated widely in industry circles for a reason. It's true.
Phaidra, which raised over $50 million in a Series B round that month with backing from NVIDIA and others, reports a 25% reduction in cooling energy at data centers using its reinforcement learning platform. The company's pitch centers on orchestration—coordinating power, cooling, and workload to extract maximum compute from constrained infrastructure. A March 2026 trial backed by NVIDIA, U.K. National Grid, and EPRI demonstrated that AI data centers can flex power demand in near real time, supporting grid stability during peak hours.
Microsoft unveiled liquid-cooled chip designs in September 2025 to address AI data center overheating. Vertiv ramped liquid cooling product launches throughout 2025 and into 2026, responding to surging orders from hyperscalers densifying racks with GPU clusters. The message is consistent: efficiency isn't just about cost or carbon anymore. It's about unlocking capacity.
The Inviscid Approach, and the Validation Gap
Inviscid AI's approach sits at the intersection of these forces. The startup claims to accelerate building simulations by a factor of 1,000 compared to traditional CFD methods, enabling what it describes as "neural network simulation that respects physics." The company integrates with building management systems and IoT sensors to create airflow simulations in near real time.
In February 2026, Inviscid announced a memorandum of understanding with the Madhya Pradesh State Electronics Development Corporation to deploy an AI digital twin proof-of-concept at the Madhya Pradesh State Data Center in India. The focus: real-time thermal simulation, hotspot detection, and cooling optimization. The agreement, signed during the India AI Impact Summit 2026, represents the startup's first publicly disclosed government partnership.
Jain articulated the physics-first rationale in a February 2026 LinkedIn post: "The next step-change will come from physics-informed models embedding conservation laws and system structure." The argument hinges on the brittleness of purely data-driven models when confronted with sparse data or regime shifts—situations common in buildings and data centers where sensor coverage is uneven and operating conditions change seasonally or during retrofits. By baking in the Navier-Stokes equations or thermodynamic constraints, the pitch goes, models remain physically plausible even when extrapolating beyond training data.
The vendor-reported metrics are aggressive: 240 times faster HVAC vent optimization with 40% better airflow, 95%+ accuracy, and 24/7 monitoring capabilities. These claims have not yet been validated by third-party measurement and verification aligned with IPMVP protocols or DOE FEMP M&V 5.0 guidance, the industry standards for documenting energy savings.
That caveat matters. A 2024 review of building digital twin use cases in Energy Informatics found that only five of 20 reported quantified benefits, flagging an evidence gap across the sector. Vendor claims and measured savings are two different things, and the industry has learned—sometimes painfully—to distinguish between them.
Trane Technologies offers a counterpoint with more established validation. The company completed its acquisition of BrainBox AI in January 2025 and launched the BrainBox AI Lab in August 2025 to explore agentic AI and physics-informed techniques. BrainBox AI's pre-acquisition case studies reported a 15.8% HVAC energy reduction and $42,000 in annual savings at 45 Broadway in New York City over 11 months, as profiled in Time magazine in December 2024. A multi-site deployment across 120-plus facilities reported 26.1% CO₂ equivalent reduction and 16.7% electricity savings in the first six months, with 1,132 metric tons of CO₂ and $329,000 in savings across 18 months portfolio-wide, according to a 2025 Trane case study.
The numbers come with real customer names and measurement periods, lending credibility that a two-person startup can't yet claim.
Phaidra, the Seattle-based startup now backed by NVIDIA, Collaborative Fund, and others after its October 2025 Series B, approaches the problem through reinforcement learning tuned for data center orchestration. The company's reported 25% cooling energy reduction stems from operational deployments, and CEO Gao has positioned the platform not as a point solution but as a workload-to-electron optimization layer. The broader prize, in Gao's framing, isn't just cooling—it's coordinating compute scheduling, power availability, and thermal limits to maximize throughput in electricity-constrained environments.
Legacy building automation giants are responding, though their integration timelines vary. Honeywell's study finding that 84% of building managers plan AI increases in 2025 reflected not just market appetite but the company's own Forge platform push. Johnson Controls published a Total Economic Impact study in April 2025 on efficiency and cost savings from its OpenBlue platform. Siemens earned recognition as a leader in Verdantix's 2024 Green Quadrant for IoT platforms in building operations. Carrier upgraded its Abound platform with AI-powered features in September 2025. Schneider Electric's EcoStruxure Building Advisor has older measurement and verification results on file with DOE Better Buildings, though some case studies date to 2017 and earlier—a timeline that raises questions about how current the technology stack is.
The competitive landscape reveals a bifurcation. Incumbents have scale, integration with installed hardware, and multi-decade customer relationships. Startups have speed, willingness to deploy bleeding-edge techniques like physics-informed neural networks, and investor appetite for AI-driven climate solutions.
Whether consolidation follows the Trane-BrainBox AI model or startups carve out defensible niches around specific building types or geographies remains an open question. History suggests both outcomes are possible, sometimes simultaneously.
What Happens Next
The trajectory through 2030 appears set: data center electricity demand will roughly double, regulatory compliance windows will narrow, and building operators will face escalating pressure to demonstrate verified savings, not just pilot results. What's less certain is which technology architectures will dominate and how quickly third-party validation catches up to vendor claims.
Physics-informed AI seems positioned to move from research prototype to production standard somewhere between 2027 and 2029. NVIDIA's presence at both the platform layer—Modulus, Omniverse—and as an investor in startups like Phaidra signals that GPU economics favor hybrid approaches that blend neural networks with numerical solvers. Academic output is accelerating. The volume of preprints on physics-informed reinforcement learning, digital twins, and agentic building control in late 2025 and early 2026 suggests a cohort of researchers moving toward commercialization, which historically means products start shipping within 18 to 24 months.
The EU's Energy Performance of Buildings Directive transposition deadline of May 29, 2026, will push continuous optimization and fault detection across European commercial real estate. U.S. state and city building performance standards—Washington's 2026 to 2028 compliance windows, New York City's annual Local Law 97 filings, Colorado's 2026 and 2030 reduction targets—will drive similar behavior domestically. California's SB 253 Scope 1 and 2 emissions reporting, with filings starting in August 2026, adds another compliance lever for companies with large real estate footprints.
The opportunity for startups like Inviscid AI lies in the gap between incumbent product cycles and the urgency of regulatory timelines. Building management system vendors have historically moved on three-to-five-year hardware refresh cycles. Software-first entrants with API integrations can deploy faster—though they face the challenge of proving savings in environments where baseline energy use fluctuates with weather, occupancy, and operational changes that have nothing to do with the software.
DOE FEMP's M&V 5.0 guidance, released in October 2025, and IPMVP protocols offer frameworks for credible measurement, but adoption remains inconsistent. Inviscid's MPSEDC proof-of-concept in India will be an early test of whether the vendor's simulation speed claims translate to verified kilowatt-hour reductions and whether a startup digital twin can coexist with legacy BMS infrastructure without requiring a rip-and-replace that most building operators can't afford.
Data centers represent the most immediate commercial pressure. The power constraint that Phaidra's CEO described isn't hypothetical—hyperscalers are deferring expansion plans in grid-constrained regions and paying premiums for power purchase agreements. Vertiv's rapid expansion in liquid cooling services and NVIDIA's Omniverse DSX blueprints for gigawatt-scale AI factories reflect an industry bracing for density and heat loads that air cooling alone cannot handle.
Physics-informed optimization of airflow, thermal distribution, and workload placement could shave meaningful percentages off cooling energy. And in a power-constrained world, those percentages unlock additional compute capacity. Which means they unlock revenue.
Challenges persist, of course. Academic reviews of PINNs note precision ceilings and failure modes around discontinuities. Real-world HVAC systems have lag, hysteresis, and control hierarchies that don't always map cleanly to simulation. Safe exploration in reinforcement learning—ensuring that an agent doesn't destabilize a chiller or overheat a server room during training—remains an active research problem. The 2024 Energy Informatics review's finding that most digital twin implementations lacked quantified benefits underscores a validation gap that the market hasn't fully closed.
Budget caution also tempers the hype. JLL's report that only 32% of facilities managers planned to increase software spending in 2025, down from 39% in 2024, suggests that pilot fatigue or ROI uncertainty is creeping in. The industry has seen waves of "smart building" promises before. Skepticism is warranted until savings are measured, not modeled.
For investors and operators, the calculus is shifting from "should we pilot AI optimization?" to "which vendors can deliver verified savings at scale, and how fast can we deploy before the next compliance deadline?" Inviscid AI and its cohort are entering a market where timing, perhaps more than any other variable, will determine winners.
The regulatory ratchet is turning. The power constraints are real. The toolchain is maturing. Whether a two-person YC startup can scale physics-informed CFD into a platform that competes with Honeywell and Siemens—or gets acquired by one of them—depends on proving, not just claiming, that simulations 1,000 times faster deliver energy reductions that measurement and verification can confirm.
Kabir Jain's bet on centuries-old
