Vorelios, a startup backed by Y Combinator, has launched a physics foundation model that it says predicts aerodynamics in milliseconds, work that a conventional solver takes hours to complete. The company, which lists YC's F26 batch, did not disclose funding.
The launch arrives as Siemens, Ansys and a crop of well-funded startups compete to turn engineering simulation into AI products. Incumbents already ship AI physics tools, and some rivals have raised nine-figure rounds.
What Vorelios says it built
On its website, Vorelios calls the product "a foundation model for engineering physics." It returns full-field predictions such as pressure and skin friction, along with derived loads including lift, drag and moments. Aerodynamics comes first; the company says heat transfer and structural mechanics are next.
The performance claims are the company's own, and nothing in the available research independently verifies them. The site says the model is more than 1,000,000 times faster than high-fidelity computational fluid dynamics (CFD), the numerical method engineers use to simulate airflow, for each design evaluated. It also says the model works through all 2,500 aircraft in its "BlendedNet++" test set in under five seconds. On public aerodynamics benchmarks, Vorelios claims 20 to 30 percent lower error than the prior state of the art for pressure and skin friction.
Founders.io and LinkedIn list Daniel McNeil as chief executive and Krish Kapadia as chief technology officer. LinkedIn puts headcount at 2 to 10, and Founders.io gives a 2026 founding date. YC's profile states the mission this way: "Solving engineering, then using it to solve everything else."
The cost of waiting on a solver
The expense Vorelios is going after is well documented. A technical opinion from the OECD Nuclear Energy Agency said RANS-class CFD (a widely used approach that averages out turbulence) typically takes days on several hundred processor cores, and that setup and meshing times are far from trivial. Sensitivity and uncertainty studies require many runs, which multiplies the bill. A 2025 NEA document put CFD at days or weeks and called uncertainty quantification studies particularly costly.
One example campaign tabulated in a NASA thesis came to roughly 100,000 CPU-hours, or about five days per run.
The market around that bottleneck is large, though analysts disagree on how large. Grand View Research estimates the global computer-aided engineering (CAE) market at $12.9 billion in 2025 and $14.2 billion in 2026, rising to $32.9 billion by 2033, a 12.8 percent compound annual growth rate. Fortune Business Insights puts the 2026 figure at $11.12 billion, and 360iResearch at $14.90 billion. All three project growth.
Incumbents and well-funded rivals
Siemens announced Simcenter PhysicsAI on May 27, 2026, promising near-instant predictions with built-in validation against high-fidelity CFD. It cites speedups "in the order of 1,000×" for design exploration. A Siemens case study with Kinetic Vision reported speedups of 350x to 4,000x over finite element analysis, at 87.5 to 97.5 percent accuracy.
Ansys has shipped SimAI 2026 R1 with new Pro (desktop) and Premium (software-as-a-service) tiers. Cadence completed its roughly €2.7 billion acquisition of Hexagon's Design & Engineering business, which includes MSC Software, on Feb. 23, 2026, and is positioning itself around "Physical AI." NVIDIA has pitched physical AI as models tied to simulation and digital twins, and has partnered with Dassault Systèmes.
The startups are better capitalized than Vorelios:
- PhysicsX raised a $135 million Series B on June 23, 2025, led by Atomico. A $300 million Series C was reported on June 8, 2026, though the research notes that figures vary by source.
- Neural Concept closed a $100 million Series C on Dec. 18, 2025, led by Growth Equity at Goldman Sachs Alternatives. It cited more than 50 enterprise customers, among them GM, GE Vernova, Leonardo, Safran and Renault.
- Vinci claims a $250 million Series B at a $1.5 billion valuation (Oct. 6, 2026). The figure is self-reported.
- Luminary Cloud released SHIFT-Wing, a transonic wing design model, with Otto Aviation in August 2025.
Neural Concept's customer results hint at what buyers actually measure. Its GM pedestrian-safety work, trained on crash simulations from 11 vehicle programs, returned critical metrics in seconds rather than weeks. MAHLE reported 15 percent more efficiency and 4 dB less noise on a radial blower.
Trust is the harder problem
Adoption is real but uneven. A NAFEMS/Teratec survey found 85 percent of organizations testing or deploying AI in simulation workflows. SimScale's State of Engineering AI 2026 report, released March 24, found that simulation and CAE had the highest AI use of the areas it examined, at 19 percent active use.
Speed, in other words, is not what holds buyers back. The NAFEMS/Teratec survey named data quality and credible accuracy, meaning verification, validation and uncertainty quantification, as key obstacles. ASME V&V standards and NASA-STD-7009A govern the credibility of conventional solvers, but consensus protocols for AI surrogates are still emerging. On Oct. 28, 2026, NAFEMS and Rescale announced AIPhysics.org, an independent benchmarking initiative for AI methods in CAE.
That matters for Vorelios. Its accuracy figures rest on public benchmarks and its own test set, and aerospace buyers will likely want validation against trusted solvers or physical experiments. JLR, presenting at NVIDIA GTC in March 2026, described models trained on more than 20,000 simulations and "aligned with CFD and wind tunnel."
What to watch
The research shows no disclosed customers, funding or pricing for Vorelios. Its next test is whether the heat transfer and structures models materialize, and whether independent benchmarks bear out the aerodynamics claims.
The company's own line frames the bet: "A single model that predicts the physics of a new design in milliseconds, where a solver run takes hours."
