Computational fluid dynamics has been the unloved corner of engineering software for thirty years. A few thousand mechanical engineering shops, a handful of insanely expensive Ansys and Siemens licenses, weeks-long simulation runs on HPC clusters, output that's 80% accurate and rendered in software that looks like it shipped on a CD-ROM in 1998. Now data centers are about to consume 12% of US electricity by 2028 and the people running them are stuck doing thermal layout in spreadsheets. Inviscid AI (YC W2026) thinks physics-informed neural networks are the answer, and they might be right.
The bet
Inviscid AI is building a CFD platform that runs in real time, accepts live IoT sensor data as boundary conditions, and continuously optimizes HVAC, airflow, and energy in buildings and data centers. The technical bet is on physics-informed neural networks, the architecture where the loss function includes the residual of the governing PDE itself, in this case the Navier-Stokes equations plus heat transport. PINNs are not new in research, but moving them into a production simulator that data center operators can rely on is genuinely hard. If they pull it off, they will be selling a product that is 240x to 1000x faster than what every Tier 4 facility in the world buys today, at maybe 5% of the cost.
The story Inviscid AI tells is real numbers backed up by their published case studies: 40% improvement in air circulation through optimized vent placement, 30% reduction in stagnant zones, simulations completed 240x faster than traditional CFD on the same geometry. None of those are imaginary. They are what a well-trained PINN on a moderately constrained domain delivers. The engineering question is whether they can hold those numbers across the messy geometries of real production data halls, retrofit campuses, and mixed-use buildings.
Why the timing is right (and why you should care)
Three things converged in the last 18 months to make this company possible. First, GPU compute became cheap enough that training a domain-specific PINN for a single building footprint costs hundreds of dollars instead of hundreds of thousands. Second, IoT instrumentation in commercial buildings finally crossed the threshold where you can pull live temperature, pressure, and airflow telemetry off a BACnet stack without writing custom drivers for every site. Third, the AI compute boom created a ferocious demand for cooling efficiency in data centers, where every kilowatt saved in HVAC is a kilowatt redirected to revenue-generating compute.
Hyperscalers are already burning the candle here. Google has published on its internal use of reinforcement learning for cooling optimization. Microsoft has invested in liquid immersion cooling. Meta has talked about chiller plant automation. None of them are going to buy from Inviscid. But there are roughly 8,000 colocation data centers in the world that are NOT hyperscalers, plus a long tail of enterprise on-prem facilities, plus every commercial building larger than 100,000 square feet, and almost none of those have a meaningful CFD-driven optimization layer today. That's the addressable market.
What they actually ship
Based on their public materials, the platform looks like three layers stacked together.
The simulation engine. A PINN trained on the building's geometry that learns to satisfy the Navier-Stokes residual plus thermal energy conservation, given boundary conditions from sensors. Once trained, inference is forward-pass-only and runs on a single GPU in seconds rather than the multi-hour solve traditional CFD requires. The training is the expensive part; the deployment is fast.
The sensor fusion layer. Live data from BACnet, Modbus, or OPC UA stacks feeds boundary conditions into the simulator. This is also where the digital twin part lives: a continuously-updated model of the actual physical state of the facility, with simulated counterfactuals on top. "If we close vent 14 by 30%, what happens to row 7?" becomes a question you can answer in 30 seconds instead of a week.
The control layer. A recommendation engine that proposes set-point changes (vent positions, chiller staging, fan speeds) and either pushes them autonomously through a building management system or hands them to an operator for approval. This is the part that turns a simulation toy into an operational tool, and it's also the part where they will accumulate the most lock-in.
Difficulty score, by stack
If you wanted to clone this, here is what you are signing up for.
ML/AI: 9/10. PINNs are an active research field, not a settled engineering practice. Getting a Navier-Stokes PINN to converge stably on a non-trivial 3D geometry with realistic Reynolds numbers is genuinely hard. Most academic papers in this area work on toy 2D domains. Production-grade requires custom loss balancing, adaptive sampling, mesh-free coordinate encodings, often Fourier features, and a lot of empirical taste. There are maybe 200 people in the world who have shipped this kind of model into production. Hiring is brutal.
Data: 7/10. Building geometry comes from BIM files (IFC, Revit) which are a tar pit of vendor-specific encodings. Sensor data comes from BMS protocols that range from clean (BACnet/IP) to nightmares (legacy serial Modbus over RS-485 with site-specific addressing tables). Once you have it normalized, the data itself is moderate volume; the integration is what eats the engineering quarters.
Backend: 6/10. Standard distributed inference serving, GPU autoscaling, time-series storage for sensor history, geometry storage for building models. Solvable problems with off-the-shelf tools, just a lot of glue.
Frontend: 5/10. 3D visualization of airflow and thermal fields is the only interesting frontend piece. Three.js or Babylon.js carry most of the load. The UX challenge is making CFD output legible to building operators who are not engineers, which is a design problem more than an engineering one.
DevOps: 7/10. Edge deployment matters because some customers will not allow telemetry off-prem. That means shipping inference containers that run on customer hardware, with a control plane back to your cloud for model updates. Air-gapped variants for government and finance customers add complexity. Plus the usual GPU scheduling, model versioning, and reproducibility headaches that ML platforms always have.
Aggregate: about 8/10. This is not a weekend project. It is a 2 to 3 year build for a competent founding team with a deep ML/CFD lead.
The moat, honestly
Two real moats and one fake one.
Real moat one: training data. Every customer building they onboard becomes a labeled dataset for the meta-model that bootstraps faster training on the next building. After 50 facilities they will have a foundation model for HVAC physics that nobody else can replicate without spending two years and several million dollars on customer acquisition. This is the same flywheel that worked for Tesla in self-driving and for Cohere in enterprise RAG.
Real moat two: BMS integration. Once you are inside a customer's building management system, talking BACnet to their Honeywell or Johnson Controls or Siemens stack, ripping you out is a six-month project. Switching costs are real. ServiceNow did not become a $200B company because their UI was good. They became one because their integrations were sticky.
Fake moat: the PINN architecture itself. If Inviscid succeeds, the basic architecture will be cloned by smart competitors within 18 months. There is a wave of CFD-on-AI papers coming out of NVIDIA Modulus, MIT, Stanford, and ETH Zurich. The math is not the moat. The customer wedge and the integration depth are.
What could kill them
Three real risks, in increasing order of how hard they are to fix.
One, traditional CFD vendors (Ansys, Siemens Simcenter) wake up and ship a credible "AI surrogate" mode that captures 80% of the performance benefit while keeping their existing channel. Ansys has the customer relationships and the budget; if they decide to be serious about this, Inviscid is in trouble. The defensible read here is that incumbent CFD vendors have spent two decades building moats around physics rigor and accuracy guarantees, which makes them constitutionally unsuited to ship a "good-enough" approximation.
Two, hyperscalers commodity-ize cooling automation by open-sourcing their internal tooling. This is a low-probability but high-impact risk: Meta has open-sourced Llama, Google has open-sourced TensorFlow, there is a non-zero chance that one of them decides cooling-optimization commoditization is strategic. If Google releases an open-source chiller plant optimizer trained on a billion-hour-corpus of internal telemetry, Inviscid sells niche vertical software for 18 months and then dies.
Three, the regulatory environment for autonomous BMS control hardens. If a model recommends a vent change that creates a hot spot that fries a $4M GPU rack, somebody is getting sued. Insurance markets for AI-driven building automation are not mature. The first big lawsuit in this space will reshape what these systems are allowed to do, and the worst-case outcome is a "human-in-the-loop only" mandate that strips the autonomy out of the product.
Pricing thesis
This is a $30k-$300k ACV product, sold by enterprise reps, with 9 to 18 month sales cycles. The hyperscaler-adjacent colocation operators (Equinix, Digital Realty, CoreSite, QTS) buy at the high end. Mid-market commercial real estate (industrial, healthcare, biotech labs) buys at the low end. Self-serve does not work in this category. There is no $99/mo SaaS tier and there should not be one. The customer who is willing to give an outside vendor write access to their building management system is, by definition, an enterprise customer.
If Inviscid hits 100 customers at a $100k blended ACV in 36 months, they are at $10M ARR. That is the credible base case for a healthy Series A in 2027.
The verdict
Real company, real problem, real technical edge. The risks are real but solvable. The team will live or die on whether they can ship a PINN that converges robustly on actual 3D geometries at customer sites within 18 months, not on whether they can sell it. Selling CFD into data centers in 2026 is the easiest sales motion in commercial real estate technology. Building software that actually works is the hard part.
If you are an engineer and you are not actively shorting Ansys (NASDAQ: ANSS) shares right now, that is a defensible position. If you are a data center operator running a 10MW colo, you are 12 to 24 months from being unable to compete on PUE without something like this in your stack. Inviscid is not the only company that will win this category, but they are well-positioned to be one of the three.
