Godela Seed

Godela Deep Dive: A YC X25 Physics-AI Simulation Startup Whose Ex-Apple, Google and Intel Team Uses Physics-Constrained Surrogate Models to Cut Simulation from Days to Seconds, at Seed Stage with Pilot-Only Access

A YC X25 physics-AI simulation startup founded in 2025 in San Francisco by ex-Apple, Google and Intel engineers; it fits physics-constrained surrogate models to a customer's own simulation data and returns solver-level accuracy in seconds, backed by Y Combinator and six named seed investors at the Seed stage, with pilot-gated access and no disclosed priced round or revenue.

Executive Summary

Godela is a physical AI startup founded in 2025 in San Francisco and part of Y Combinator's Spring 2025 batch, X25. Founders are Cinnamon Sipper as CEO, who worked on laptops and iPads at Apple and on AI products at Google, and Abhijit Pranav Pamarty as CTO, who researched physics-driven optimization at Harvard, generative models for fluid mechanics at Stanford and surrogate models for semiconductor manufacturing at Intel.

The product is positioned as an AI physics engine replacing simulations and prototypes. It runs a three-step loop: Train, uploading simulation data to fit a physics-constrained model; Explore, asking what-if questions in natural language such as raising airflow in one zone and getting a prediction instead of queuing a solver run; and Optimize, searching for better configurations. CAD files and experimental results can be uploaded alongside simulation data, covering five domains in data center thermals, electronics cooling, aerodynamics, electromagnetics and structures plus a build-your-own path for other physical systems.

The published demonstration is a rack cooling study with 2,400 simulations and 12 parameters answered in 0.8 seconds. One investor note claims the platform runs 4,500 times faster than legacy tools.

On funding, a seed round closed on June 3, 2025 with Y Combinator, Aera VC, Imagination Capital, MGV.VC, Swell Partners, CLAI Ventures and Network VC, seven investors in total, with the amount undisclosed. A third-party database estimates annual revenue between two and four million dollars, an estimate rather than a disclosure. Headcount figures differ across sources, and the official figure has moved: the YC company page, accessed November 2, 2026, lists ten, two third-party directories list ten, one lists four with an October 2026 update date and one lists a range of two to ten.

The business model is pilot-based, with no public pricing, self-serve trial or documentation, and access through a booking request, indicating enterprise sales and custom delivery rather than a productized SaaS.

In physical AI terms Godela represents physics-constrained surrogates trained on customer data, which differs from Fainite's limited-data route since it explicitly requires customer simulation data. That means it cannot serve scenarios without historical simulation data while it better fits a customer's specific geometry and conditions.

One framing point matters. This is an early company with a YC halo, polished demos and missing validation. Its strengths are the team background across Apple, Google, Intel, Stanford and Harvard, YC resources with seven investors and a clear three-step product logic. Its weaknesses are the absence of published accuracy validation against commercial solvers, no customer cases, no pricing or documentation, and an unresolved tension between replacing simulation and becoming its intelligent home.

Company Overview

Team

Sipper combines hardware and AI product experience, having worked on laptops and iPads at Apple and AI products at Google, which suits engineering tools since she understands both hardware workflows and AI interaction design.

Pamarty combines physics, machine learning and industry, having researched physics-driven optimization at Harvard, generative models for fluid mechanics at Stanford and surrogate models for semiconductor manufacturing at Intel. The last is especially relevant since semiconductor process surrogates mapping parameters to yield and morphology are among the most mature industrial surrogate applications, indicating he understands what industrial acceptance requires.

Headcount is ten as listed on the Y Combinator company page, accessed November 2, 2026, alongside the founding date of 2025, the Spring 2025 batch, active status, the San Francisco location and the primary partner. Two third-party directories also list ten, one lists four with an October 2026 update date and one lists a range of two to ten. The official figure is used here and the remainder are recorded as dated variants. The same page carries three open roles, founding engineer for machine learning systems, founding simulation engineer and founding engineer for software engineering, all in San Francisco on founding-level equity bands, so the team is still in an expansion phase.

YC and Investors

The YC X25 batch provides seed capital on standard terms, startup methodology emphasizing fast iteration and product market fit, and an alumni network for early customers and later funding. Among the seven investors, Aera VC, CLAI Ventures and Network VC are early-stage funds while MGV.VC and Swell Partners may provide follow-on.

Internal Tension in Positioning

There is a notable tension. The YC page and website say the product replaces simulations and prototypes, while the founders' launch post says Godela does not replace simulations but becomes their intelligent, adaptive home.

The commercial meanings differ. Replacement is a substitute logic competing with Ansys for budget, while a home is a complement logic coexisting with Ansys as a query layer above it. The former supports a more aggressive narrative and larger valuation imagination, the latter is more pragmatic and easier for customers to accept.

The tension is not fatal, since early companies often speak differently to different audiences, but it shows positioning has not converged.

Technology

What the Three Steps Mean

Train fits a physics-constrained model to uploaded simulation data possibly alongside CAD and experiments, with the technical core being how to learn generalization from limited data of perhaps hundreds to thousands of cases.

Explore takes natural language questions about parameter changes and returns predictions directly, with the core being the chain from language to parameter change to model inference to result explanation, where language understanding is only the interface and the real computation is inference.

Optimize searches for better configurations, exploiting sub-second inference to explore parameter space far more broadly than manual trial and error.

How Physics Constraints Are Implemented

The claim of a physics-constrained model could mean three implementations: physics-informed neural networks with equation residuals as loss terms, physics-enforced architectures with conservation structure, or post-processing consistency correction.

Generalization differs greatly. Soft constraints are easy but do not guarantee compliance at inference, hard constraints generalize better but are hard to train, and post-processing is simple but may damage other model properties. The implementation is undisclosed, a key gap since physics constraints are the core claim distinguishing the product from purely data-driven surrogates.

Reading the Demonstration

The rack cooling case with 2,400 simulations, 12 parameters and 0.8 second answers carries three points of information.

First, 2,400 cases is a moderate training set capable of covering the main variation range of 12 parameters for data center cooling, showing the data requirement is satisfiable unlike methods needing tens of thousands of cases.

Second, 12 parameters is high-dimensional, since ten values each would give a trillion combinations, explaining why a surrogate is needed, with 2,400 samples sparse in that space and the model relying on physical constraints to interpolate.

Third, sub-second inference enables interactive querying and optimization search.

The case is a scripted demo rather than a published benchmark, without comparison to commercial solvers, error distribution or failure conditions, so it proves the system runs rather than that it is trustworthy.

The 4,500 Times Figure

The 4,500 times claim comes from an investor note rather than the company website, with unclear basis: single inference versus single solve, or training plus inference versus many solves.

The former is plausible since a solver taking hours against 0.8 second inference gives a ratio in the thousands, but that comparison ignores training cost, since 2,400 simulations themselves consume substantial time and the first total investment far exceeds a single solve. This is the common feature of all surrogates, front-loaded investment for later speed, and omitting it misleads customers about return on investment.

Technical Shortfalls

First, no published accuracy validation against Fluent, Icepak or Ansys.

Second, undisclosed physics constraint implementation.

Third, unverified cross-domain generalization, since whether five domains share one model or separate ones determines whether the platform's value is technical or merely interface-level.

Fourth, data dependence, since requiring customer simulation data excludes customers without history in new designs, materials or conditions.

Product and Commercial Model

Product Form

The product is a web platform for uploading data, training, natural language querying and optimization search, accessed by pilot request without self-serve signup, pricing pages, documentation or security notes, serving only customers willing to go through a sales process.

Business Model

The model is inferred as project or per-model billing, and without self-serve pricing it resembles enterprise customization rather than subscription SaaS, with high value per account but difficult scaling since each customer needs sales, implementation and support.

Target Customers

Targets are simulation-heavy teams in data center thermal management, electronics cooling, aerodynamics, electromagnetics and structures, with data centers and electronics cooling the most concrete since iterations are frequent and accuracy requirements are relatively forgiving, as temperature trends matter more than absolute values.

Customers and Revenue

No customer names, counts or paid scale are public. The third-party estimate of two to four million dollars in annual revenue is a projection based on funding, team size and industry averages rather than disclosure.

Competitive Landscape

Position in the Field

Godela sits in the second tier of physical AI simulation, early with a clear technical route but unvalidated, alongside Fainite and other European and North American startups, below PhysicsX, Vinci, Neural Concept and Luminary Cloud, and adjacent to corporate AI features from Ansys and Siemens.

Comparison With Fainite

Both are early physical AI simulation companies with different technical routes. Fainite emphasizes limited data and physics enforcement assuming customers have little data, while Godela emphasizes training on customer data and physics constraints assuming customers have simulation history.

That difference determines different customer bases, with Fainite serving new domains and materials lacking history and Godela serving optimization on mature product lines with abundant simulation records.

Relationship With Incumbent CAE

The replacement narrative conflicts directly with incumbent CAE while the home narrative is complementary. Realistically customers will not abandon Ansys for an unvalidated tool but may use Godela as a fast query layer above it, establishing a baseline in Ansys and exploring design space in Godela. That complementary position lands more easily while limiting market size, since customers still buy Ansys.

Relationship With General AI

The differentiation is physics constraints versus general language models, which understand engineering questions but cannot compute physics while physics models compute but cannot grasp complex intent. Combining them with a language model as interface and a physics model as compute is the sensible architecture, and Godela's natural language querying follows it.

Financing and Capital Structure

Funding

The seed round closed June 3, 2025 with seven investors and an undisclosed amount. Under YC standard terms the YC investment is typically 500,000 dollars, and other investors may have brought the total into the one to three million dollar range.

Valuation and Revenue

Valuation is undisclosed. The third-party estimate of two to four million dollars in annual revenue is high for a company founded in 2025 with ten people, implying 200,000 to 400,000 dollars per head, possibly conflating potential contract value or industry averages with actual revenue. This report labels it a third-party estimate rather than fact.

Use of Proceeds

YC funding typically covers 18 to 24 months of operations including team expansion, product development and early sales, enough for a company of this size to grow to 20 to 30 people and complete first customer validation. The three open founding-level roles on the YC page are consistent with that trajectory, and they are also the clearest public signal that the ten-person figure is not static.

DISCREPANCIES

1. Headcount figures, and the official one has moved. The YC page listed three when this report was first compiled and lists ten when accessed again on November 2, 2026. Two third-party directories list ten, one lists four with an October 2026 update date and one lists a range of two to ten. The direction of travel is consistent: the company has grown past the three-person figure, and the four-person entry is a dated snapshot rather than a contradiction. This report uses ten, the current official figure. Source tier: OFFICIAL (YC) crossed with third-party (MEDIA).

2. Conflict between replacing simulation and becoming its home. The YC page and website describe replacing simulations and prototypes while the founders' launch post says the product does not replace simulations but becomes their intelligent home. The commercial meanings differ, and this report presents both without choosing. Source tier: OFFICIAL (company material) internally inconsistent.

3. Source of the 4,500 times speedup. The figure comes from a CLAI Ventures investor note rather than the company website, with unclear basis such as single inference versus single solve or inclusion of training cost. This report labels the source and the definitional issue. Source tier: MEDIA (investor) rather than company framing.

4. Nature of the revenue estimate. A third-party database estimates two to four million dollars in annual revenue without company disclosure, which is high for a 2025-founded company and may be projection error. This report labels it an estimate rather than fact. Source tier: third-party estimate, UNVERIFIED.

5. Nature of the demonstration. The rack cooling case with 2,400 simulations, 12 parameters and 0.8 seconds is a scripted demo rather than a published benchmark, proving the system runs without proving accuracy. This report notes the distinction. Source tier: OFFICIAL (company demo) distinguished from technical validation.

6. Undisclosed physics constraint implementation. The company calls the model physics-constrained without specifying loss-based, architectural or post-processing constraints. This report lists all three possibilities. Source tier: OFFICIAL (statement) with implementation absent.

7. Meaning of the pilot model. Access by pilot request without self-serve signup, pricing, documentation or security notes leads some to read the product as not ready. This report labels it an enterprise sales model without inferring maturity. Source tier: OFFICIAL (access method).

8. Missing funding amount. The seed round discloses investors and date but not amount, and YC standard terms suggest a total between one and three million dollars. This report labels the figure as inference. Source tier: OFFICIAL (investors and date) plus inference (amount).

Risks and Diligence Gaps

First, missing accuracy validation, with no public error comparison against commercial solvers while simulation-quality is a strong claim requiring data.

Second, data dependence limiting scenarios, since requiring customer simulation data excludes new designs, materials or conditions without history, contrasting with Fainite's limited-data route.

Third, unconverged positioning, with replacement and home narratives coexisting and affecting sales messaging and product roadmap.

Fourth, a very small team of ten, and possibly in flux, handling product, sales, delivery and research simultaneously. The three open founding-level roles suggest both that headcount is growing and that the founding team still covers most functions directly.

Fifth, competitive pressure from funded upper-tier companies with customers, from corporate AI features and from peers such as Fainite.

Sixth, the customer trust threshold, since engineering buyers are extremely conservative especially where a wrong thermal answer damages hardware.

Seventh, the gap between demo and production, since 0.8 seconds assumes a trained model and prepared inputs while production involves data cleaning, format conversion, boundary setting and outlier handling whose workload may exceed inference.

Composite Assessment

Technology scores 6.5. The three-step loop of training, exploration and optimization is a clear product logic, the physics constraint direction is correct and the founding team combines physics, machine learning and industry. Deductions for undisclosed constraint implementation, unvalidated accuracy and unclear cross-domain generalization.

Business scores 5.0. Five domains cover major engineering scenarios with data centers and electronics cooling as clear entry points, and the pilot model fits industrial sales logic. Deductions for zero public customers, zero pricing, zero documentation and unconverged positioning.

Finance scores 5.5. A seed round with YC and six investors provides 18 to 24 months of runway. Deductions for undisclosed amount, unreliable revenue estimates and unknown valuation.

Product scores 6.0. The three-step interaction with natural language what-if suits engineers and mixed CAD and experimental inputs add flexibility. Deductions for no documentation, no trial, no security notes and a large demo-to-production gap.

Market scores 6.5. Physical AI simulation is hot, data center and electronics cooling demand grows quickly with AI compute and the YC network provides early customers. Deductions for intense competition, conservative buyers and high education cost.

Overall 5.9. This is an early company with a YC halo, clear logic and the least evidence: it turned asking physics questions in natural language into a product, built training, exploration and optimization into a loop and has a matching team. Yet it publishes no accuracy validation, no customer case and no pricing, and its positioning between replacing simulation and becoming its home is unresolved. The judgment on physical AI is that training physics-constrained models on customer data is currently the most pragmatic surrogate route and Godela's productization with natural language interfaces and optimization search fits how engineers work, while moving from demo to production and from pilot to scale requires accuracy data and customer evidence.

References

  1. Y Combinator company page, accessed November 2, 2026: Godela in the Spring 2025 batch, X25, founded 2025, headquartered in San Francisco, status active, primary partner Aaron Epstein, listed team size of ten, founders Cinnamon Sipper and Abhijit Pranav Pamarty; positioned as an AI physics engine replacing simulations and prototypes; capabilities of uploading CAD, experimental and simulation data, asking questions in natural language and obtaining simulation-quality results; and three open founding-level roles in San Francisco covering machine learning systems, simulation engineering and software engineering, on equity bands of 0.08 to 1.50 percent at salaries of 120,000 to 180,000 dollars
  2. YC launch post, June 2025: founders as former Apple, Google and Intel engineers turning Stanford research into a product; the three-step loop of Train, Explore and Optimize; the statement that Godela does not replace simulations but becomes their intelligent adaptive home
  3. Company website, accessed 2026: the three-step loop, five domains in data center thermals, electronics cooling, aerodynamics, electromagnetics and structures plus a build-your-own path; a rack cooling demonstration with 2,400 simulations, 12 parameters and 0.8 second answers; access by pilot request
  4. Third-party company profile from fundraisingfox: founded 2025 in San Francisco with ten employees; Sipper as CEO with Apple hardware and Google AI product background; Pamarty as CTO with Harvard physics-driven optimization, Stanford fluid generative models and Intel semiconductor surrogates; seven investors including Aera VC, Imagination Capital, MGV.VC, Swell Partners, Y Combinator, CLAI Ventures and Network VC
  5. Third-party tool assessment from hokai.io: training on the customer's own simulation data with switch costs from per-customer models; no public pricing, documentation or trial with access by pilot request; no public accuracy comparison against commercial solvers; the rack cooling case as a scripted demo rather than a published benchmark; simulation-quality as a vendor claim
  6. Third-party database from seedtable: YC X25, San Francisco, ten employees, founded 2025; seed round on June 3, 2025 with Break VC, CLAI Ventures, Network VC and others with amount undisclosed; estimated annual revenue of two to four million dollars as an estimate
  7. Investor note from CLAI Ventures, July 2025: describing the physics-aware AI engine as running 4,500 times faster than legacy simulation tools
  8. Third-party introduction from promptloop: team of two to ten; positioned as an AI-native physics engine accepting CAD, experimental and simulation data; target industries including manufacturing, robotics, chips, chemicals and industrial automation
  9. Technical background from public sources: the two surrogate paradigms of data-driven and physics-constrained; three implementations of physics constraints through physics-informed neural network loss terms, physics-enforced architectural hard constraints and post-processing correction, differing in generalization and difficulty
  10. Technical background from public sources: the trade-off between front-loaded surrogate investment in data generation, training and validation and later gains from second-level inference, and the misleading effect of advertising speedup multiples without front-loaded cost
  11. Competitive landscape from public sources: tiering of physical AI simulation into funded shippers such as PhysicsX at 300 million dollars, Vinci at 250 million, Neural Concept at 100 million and Luminary Cloud at 72 million, early startups including Fainite and Godela, and corporate AI features from Ansys SimAI and Siemens PhysicsAI
  12. Business model background from public sources: the typical industrial AI path of pilot, proof of concept and scaled procurement, because engineering customers are conservative about unvalidated tools especially where failure is costly such as data center thermal management
  13. Third-party company directory gpagent.ai, self-dated October 5, 2026: YC Spring 2025 batch, San Francisco, team size of four, two founders, early stage, engineer-led pilot request access motion
This report is compiled from public sources for informational purposes only and is not investment advice. The company is early with revenue, customers and pricing undisclosed; third-party revenue estimates and speedup figures come from non-company sources without independent verification, and related judgments are labeled as uncertain.

Related reports66

This report is compiled from public sources. It is for reference only and does not constitute investment advice. Data is current as of the publication date; please refer to the company’s official filings.