Shuzhi Maijin Seed

Shuzhi Maijin Deep Dive: A Shenzhen Fully Self-Developed SNN-plus-LBM CFD Startup Claiming Second-Level Results on Hundreds-of-Millions of Grids on Consumer Hardware, With a Tens-of-Millions-RMB Angel Round From a Ministry-of-Finance-Affiliated Fund and an FPGA Migration Plan

A Shenzhen AI-plus-physics CFD startup founded in June 2025 that fuses spiking neural networks with the lattice-Boltzmann method in a fully self-developed stack, claiming second-level results on hundreds-of-millions of grids on consumer hardware with 0.01 to 0.5 percent relative error; its FlowStudio flow-and-noise module is released and a tens-of-millions-RMB angel round led by a Ministry-of-Finance-affiliated fund closed in March 2026, with an FPGA migration planned.

Executive Summary

Shuzhi Maijin, formally Shuzhi Maijin Artificial Intelligence Basic Technology Research Shenzhen Co., Ltd., was founded in June 2025 in Shenzhen and is incubated in the national incubator at Tsinghua Information Port in Nanshan, led by a national Thousand Talents Plan expert with a core team holding nearly a decade of low-level algorithm research. Between March and April 2026 it announced an angel round in the tens of millions of renminbi from Zhongcai Caitou Hunan Private Fund Management, an investment platform under Zhongcai Jinkong affiliated with the Ministry of Finance whose strategy targets import substitution in critical technologies.

The technical claim combines spiking neural networks with the lattice Boltzmann method into an AI plus physics CFD model, with three assertions: extreme speed, since hundreds of millions of grid cells take seconds on consumer hardware; extreme accuracy, with relative error of 0.01 to 0.5 percent against an industry norm of 3 to 5 percent; and full self-reliance, since the SNN operators, LBM model and applications are 100 percent in-house without third-party commercial libraries.

The product is FlowStudio with flow and noise modules, automotive aerodynamics software packaging CFD into an easy tool that delivers second-level results on ordinary laptops and integrates steady and transient analysis in a single model. Applications cover drag optimization and digital wind tunnels, with the company reporting a strategic partnership with a leading automaker and blind tests matching wind tunnel data.

The roadmap has two further steps: migrating core algorithms to FPGAs with a full hardware and software platform to build an open ecosystem around a fluid compute core module, which the company compares to CUDA for industrial simulation, and delivering CPU general and FPGA high-performance versions in 2026 under a 2025-2026 plan. Adjacent domains include aerospace for hypersonic flow, power and grids, oil and gas pipelines, rail with crosswinds and tunnels, nuclear cooling and renewable energy climate modeling.

In physical AI terms Shuzhi Maijin is rewriting the CFD solving kernel with unconventional AI. It builds neither surrogates nor workflow orchestration but attempts to reconstruct the solving process physically through SNN and LBM, a more ambitious and riskier position.

One framing point matters. This is an early company with extremely strong technical claims and minimal verification. If the three assertions hold they are generational breakthroughs, yet they rest on company statements and a qualitative blind-test match with no third-party reproduction, public benchmark or paper. The key questions are what problem the 0.01 percent error was measured on and at what scale the consumer-hardware seconds apply.

Company Overview

Origins and Team

Founded in June 2025 in Shenzhen and incubated at Tsinghua Information Port, the company is led by an unnamed Thousand Talents Plan expert with a core team claiming nearly a decade of low-level algorithm work validated in projects for international automotive brands and leading domestic electric vehicle makers.

The decade-of-accumulation framing suggests the technology predates the company, with 2025 incorporation serving commercialization, which explains how a company less than a year old presents a fully self-developed SNN and LBM stack.

Non-Classical AI Positioning

The company emphasizes artificial intelligence using non-classical algorithms, meaning spiking neural networks rather than the mainstream transformer route, a biologically inspired model computing only when neurons fire.

Technically this means event-driven computation with theoretical energy advantages and fit for neuromorphic hardware, while narratively it differentiates in a homogenized large model landscape and aligns with self-reliance policy.

Funding and Investor

The angel round in the tens of millions of renminbi came solely from Zhongcai Caitou, whose strategy targets import substitution and bottleneck technologies, indicating value measured partly by industrial mission rather than financial return alone. Sole investment means a modest amount absorbable by one institution and strong investor influence.

Technology

Why LBM Rather Than Conventional CFD

Mainstream CFD uses finite volume methods discretizing the domain and solving integral forms of Navier-Stokes equations, mature and well validated but struggling with meshing in complex geometry.

The lattice Boltzmann method simulates mesoscopic distribution functions through collision and streaming, deriving macroscopic density and velocity from moments, with advantages in simple boundary treatment for complex geometry, natural parallelism and suitability for multiphase and porous media, and disadvantages in compressibility, high Reynolds stability and memory use.

Choosing LBM fits automotive external aerodynamics as a complex-geometry, moderate-Reynolds problem, which is why commercial LBM software such as PowerFLOW holds a firm position there.

What Combining SNN and LBM Means

Three technical paths are possible.

First, replacing the LBM collision operator with an SNN, since collision is local and nonlinear and maps well to event-driven computation.

Second, using SNN for adaptive meshing or subgrid modeling, compensating unresolved physics in coarse regions similar to subgrid closure in large eddy simulation.

Third, using SNN for multiscale coupling, computing on fine grids with LBM and learning corrections on coarse grids.

The paths differ greatly in difficulty and effect, and the choice is undisclosed. The combination itself is not new academically, so the innovation lies in how it is combined and the resulting performance.

Reading the 0.01 Percent Error

A relative error of 0.01 to 0.5 percent is an extraordinary claim. Commercial CFD validation against wind tunnel data for automotive aerodynamics typically runs 3 to 5 percent, academic benchmarks such as Ahmed body and DrivAer land in the same range, and wind tunnel measurement uncertainty is about 1 to 2 percent.

If 0.01 percent refers to agreement with another simulation it proves two solvers agree rather than correctness, and if against experiment it falls below wind tunnel uncertainty and requires very strict conditions. The benchmark, operating conditions and error norm are undisclosed, so the figure cannot be read correctly.

Reading Consumer-Hardware Seconds

Hundreds of millions of cells in seconds on consumer hardware is equally strong. For reference, such models take hours to days on supercomputers and tens of minutes with GPU acceleration, so a laptop-scale seconds result implies three to four orders of magnitude efficiency gain, possibly from LBM parallelism on GPUs, SNN event-driven sparsity, or model simplification through steady assumptions, reduced order or surrogates.

Here too the definitions matter: whether the count is cells or degrees of freedom, whether seconds means one step or convergence, and what hardware qualifies as consumer grade.

Boundaries of Full Self-Reliance

The claim of 100 percent in-house development without third-party commercial libraries carries value for domestic substitution, but algorithm self-development must be distinguished from framework self-development, since building on free open-source frameworks such as PyTorch would still count as algorithmic rather than framework self-reliance.

The FPGA Logic

Migrating to FPGAs makes sense since custom datapaths and low power suit event-driven spiking computation where sparse firing avoids dense matrix multiplication, providing a hardware basis for the seconds claim if real.

Costs include high development effort, poor portability across FPGA families and a weak ecosystem compared with CUDA. Building a CUDA-like open ecosystem is a decade-scale goal.

Technical Shortfalls

First, no papers, public benchmarks or third-party reproduction.

Second, SNN training difficulty, since spikes are non-differentiable and require surrogate gradients, with stable industrial-accuracy training an open problem.

Third, inherent LBM limits at high Reynolds numbers, compressibility and strong turbulence, where real vehicles with rotating wheels and complex wakes remain challenging.

Fourth, product maturity, since FlowStudio is released without public user counts, detailed customer cases or comparison tests against commercial software.

Product and Commercial Model

FlowStudio Product Form

FlowStudio has flow and noise modules for external aerodynamics and aeroacoustics, packing CFD into an easy tool running on laptops with second-level results and integrating steady and transient analysis in one model.

The integration has engineering value since conventional tools separate steady RANS from transient LES or DES workflows, while a single model lets users check trends quickly and switch to transients without rebuilding.

Customers and Validation

The company reports a strategic partnership with a leading automaker and blind tests matching wind tunnel data. Blind testing carries weight if predictions were truly made without knowledge and then compared, but the automaker, model, conditions and error magnitudes are undisclosed.

Business Model

Undisclosed, potentially seat licenses, cloud usage or project services. The laptop-level positioning suggests local licensing, attractive to customers for data residency and offline operation.

Roadmap Commitments

Delivery of CPU general and FPGA high-performance versions in 2026 is a checkable milestone, and failure by year end would indicate slower progress than claimed.

Competitive Landscape

Established CFD Rivals

Automotive external aerodynamics tools include Fluent, STAR-CCM+, PowerFLOW on the LBM route, CFX and open-source OpenFOAM. Shuzhi Maijin differentiates through LBM plus SNN plus consumer hardware, with PowerFLOW already proving LBM feasibility and the company adding neural acceleration.

New AI CFD Entrants

AI CFD splits into surrogates such as PhysicsX, Vinci, Neural Concept and domestically Supreium and ShonCloud, and solver rewrites such as Shuzhi Maijin and a few PINN-based efforts. The second group carries higher risk and higher ceiling, since rewriting the kernel produces a different tool rather than a faster one.

Against Domestic CFD

Domestic CFD companies mostly pursue finite volume routes such as Tenfong and Yundao, while LBM avoids direct competition with Fluent and STAR-CCM+ at the cost of a smaller validation and ecosystem base.

Financing and Capital Structure

Angel Round

Tens of millions of renminbi from a sole investor, ample for a company under a year old, supporting 30 to 50 people for 18 to 24 months or a leaner team for three years.

Use of Proceeds

Deepening AI plus physics research, accelerating product delivery and ecosystem expansion, and delivering CPU and FPGA versions in 2026.

Valuation and Revenue

Both undisclosed, with the company between product release and customer validation.

DISCREPANCIES

1. Missing baseline for the 0.01 percent error. The company cites 0.01 to 0.5 percent against an industry norm of 3 to 5 percent without stating the benchmark, conditions or error norm. This report presents both simulation-comparison and experiment-comparison readings. Source tier: company claim, UNVERIFIED.

2. Undefined hardware and scale for the seconds claim. Consumer hardware, hundreds of millions of cells and seconds lack configuration, cell versus degree-of-freedom definitions and step versus convergence clarification. This report flags the ambiguity. Source tier: company claim, UNVERIFIED.

3. Boundaries of 100 percent in-house. The company claims full in-house development without third-party commercial libraries without stating whether open-source frameworks such as PyTorch or CUDA are used. This report separates algorithmic from framework self-reliance. Source tier: OFFICIAL (statement) with detail absent.

4. Framing of the investor. The company describes Zhongcai Caitou as a professional investment institution under Zhongcai Jinkong affiliated with the Ministry of Finance while media call it a central state fund. This report uses the company framing. Source tier: OFFICIAL (company statement).

5. Imprecision of the round size. Tens of millions of renminbi is stated without a precise figure, consistent across media. This report uses the range and notes imprecision. Source tier: OFFICIAL.

6. Missing detail on customers and validation. The company cites a strategic automaker partnership and blind-test agreement without naming the automaker, model, conditions or error values. This report labels it qualitative. Source tier: company claim, UNVERIFIED.

7. Founding date. June 2025 is stated and consistent with registry data, while some media say 2025. This report uses June 2025. Source tier: OFFICIAL.

8. Founder identity. The company cites a Thousand Talents Plan expert without publishing name or background. This report does not speculate. Source tier: OFFICIAL (statement) with name absent.

Risks and Diligence Gaps

First, missing technical validation, since all performance claims lack reproduction, papers or benchmarks, the largest risk for a technology-driven company.

Second, SNN training difficulty, an open academic problem at industrial accuracy.

Third, inherent LBM limits at high Reynolds, strong turbulence and compressibility.

Fourth, FPGA execution risk with high cost and tight 2026 timing.

Fifth, shallow customer validation, since strategic partnership and blind-test agreement are qualitative without payment or production adoption evidence.

Sixth, single-scenario dependence, since expansion to aerospace, energy and nuclear requires different validation and certification.

Seventh, insufficient team disclosure, since founder name and team composition are unpublished, making technical credibility hard to judge.

Composite Assessment

Technology scores 6.5. SNN plus LBM has academic grounding, LBM suits complex geometry in automotive aerodynamics and FPGA migration fits event-driven spiking computation. Deductions for unverified claims, SNN training difficulty and inherent LBM limits.

Business scores 5.5. FlowStudio is released with steady and transient integration carrying engineering value and local licensing suiting automakers on data residency. Deductions for undisclosed customers and revenue, unknown pricing and single-scenario dependence.

Finance scores 6.0. An angel round in the tens of millions from a sole investor provides moderate runway. Deductions for undisclosed valuation and revenue and dependence on technical validation for follow-on funding.

Product scores 6.0. FlowStudio's definition is clear for automotive aerodynamics and noise on laptops, but public usage data and comparison tests are absent. Deductions for unverified maturity and undisclosed interface.

Market scores 7.0. Domestic CFD substitution is policy-driven, automotive aerodynamics is a defined scenario and the Ministry of Finance-backed investor provides policy resources. Deductions for high incumbent barriers, extreme customer conservatism about new solvers and long validation cycles.

Overall 6.2. This is the most ambitious and least verified company in this series: it attempts to rewrite the CFD solving kernel with SNN and LBM, a generational breakthrough if successful and hard to compete with established solvers if not. Its weakness is that every key claim lacks verifiable evidence. The judgment on physical AI is that rewriting physics solvers with unconventional neural networks is imaginative and aligns with self-reliance goals, but simulation credibility rests on extensive validation, and a company under a year old claiming 0.01 percent error and laptop-speed results needs public benchmarks, papers and third-party reproduction that do not yet exist.

References

  1. Angel round announcements from March 30 to April 5, 2026 across China Development Network, China Enterprise News, Xinhua Guangdong and China News Service: tens of millions of renminbi from Zhongcai Caitou Hunan Private Fund Management as sole investor, with proceeds for AI plus physics research, product delivery and ecosystem expansion and delivery of CPU general and FPGA high-performance versions in 2026 under a 2025-2026 plan
  2. Company background from the same reports: founded June 2025 focused on manufacturing intelligence starting with automotive, led by a national Thousand Talents Plan expert with nearly a decade of low-level algorithm work validated in projects for international automotive brands and leading domestic electric vehicle makers, incubated at Tsinghua Information Port in Nanshan
  3. Technical claims from the same reports: fusing spiking neural networks with the lattice Boltzmann method into an AI plus physics CFD model with extreme speed at hundreds of millions of cells in seconds on consumer hardware, relative error of 0.01 to 0.5 percent against an industry norm of 3 to 5 percent with micrometre spatial accuracy, generalization without retraining across vehicle types and scenarios, and 100 percent in-house development from SNN operators through LBM models to applications without third-party commercial libraries
  4. Product information from the same reports: FlowStudio with flow and noise modules as automotive aerodynamics software packaging CFD into an easy tool with laptop second-level results and steady plus transient analysis in a single model, applied to drag optimization and digital wind tunnels with a strategic automaker partnership and blind tests matching wind tunnel data
  5. Ecosystem plans from the same reports: building an open ecosystem around a fluid compute core hardware module, migrating algorithms to FPGAs with a full hardware and software platform, compared to a CUDA-style ecosystem for industrial simulation
  6. Adjacent domains from the same reports: aerospace for hypersonic flow, power including hydro and grid optimization, oil and gas pipelines, rail with crosswinds, tunnels and passing trains, nuclear cooling systems and renewable energy climate modeling
  7. Investor background from the same reports: Zhongcai Caitou under Zhongcai Jinkong affiliated with the Ministry of Finance, targeting national import substitution and companies able to break bottleneck technologies
  8. Technical background from public sources: the lattice Boltzmann method simulates mesoscopic distribution functions through collision and streaming with macroscopic quantities from moments, advantaged in complex boundary treatment, parallelism and multiphase flows, disadvantaged in compressibility, high Reynolds stability and memory, with commercial LBM software such as PowerFLOW established in automotive aerodynamics
  9. Technical background from public sources: finite volume methods dominate conventional CFD, mature and well validated but struggling with meshing in complex geometry, while automotive external aerodynamics as a complex-geometry moderate-Reynolds problem suits LBM
  10. Technical background from public sources: spiking neural networks compute only on firing with theoretical energy advantages and fit for neuromorphic hardware, but training is harder than conventional networks since spikes are non-differentiable and require surrogate gradients, with possible SNN and LBM combinations including collision operator replacement, subgrid modeling and multiscale coupling
  11. Technical background from public sources: commercial CFD validation accuracy against wind tunnel data for automotive aerodynamics typically runs 3 to 5 percent, academic benchmarks such as Ahmed body and DrivAer fall in the same range, and wind tunnel measurement uncertainty is about 1 to 2 percent
  12. Competitive landscape from public sources: automotive external aerodynamics tools include Fluent, STAR-CCM+, PowerFLOW on the LBM route, CFX and open-source OpenFOAM; AI CFD competitors split into surrogates such as PhysicsX, Vinci, Neural Concept, Supreium and ShonCloud and solver rewrites such as Shuzhi Maijin and a few PINN approaches; domestic CFD companies mostly pursue finite volume routes such as Tenfong and Yundao
This report is compiled from public sources for informational purposes only and is not investment advice. The company is private and revenue and customers are undisclosed; accuracy, speed and precision claims are company statements without third-party reproduction or public benchmarks, 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.