Bentley NASDAQ BSY Public

Bentley Systems Deep Dive: iTwin Digital Twin and Generative AI Empower Full-Lifecycle Infrastructure Engineering

Bentley Systems (NASDAQ: BSY) covers the infrastructure lifecycle with the iTwin digital twin and OpenRoads/OpenBridge; 2026 brings OpenSite+ generative design and an OpenSTAAD MCP server.

Executive Summary

Bentley Systems (Nasdaq: BSY) is the largest independent vendor of infrastructure engineering software, founded by engineers in 1984, headquartered in Exton, Pennsylvania, and listed in September 2020. Full year 2025 revenue was 1,501.8 million dollars, up 11.0 percent or 10.1 percent in constant currency, with subscriptions of 1,376.7 million dollars up 12.5 percent, annualized recurring revenue of 1,462.1 million dollars, constant currency annualized recurring revenue growth of 11.5 percent, and free cash flow of 520.2 million dollars. For the second quarter of fiscal 2026 ended June 30, revenue was 410.7 million dollars up 12.8 percent, subscriptions were 378.6 million dollars up 13.6 percent, annualized recurring revenue was 1,536.0 million dollars, constant currency growth was 12 percent, and the last twelve month recurring revenue dollar-based net retention rate was 109 percent.

Its position in the physical AI discussion differs from that of engineering simulation vendors. Bentley does not try to solve partial differential equations more accurately than Ansys. Its value sits in three other things. First, deterministic calculation engines covering the whole infrastructure lifecycle, including STAAD for structural analysis, PLAXIS for geotechnics, SACS for offshore structures, OpenRoads for roads and OpenUtilities for substations, each embedding national design codes. Second, the iTwin platform, which unifies fragmented design, construction and operations data into a semantically consistent asset model. Third, Model Context Protocol servers that expose those deterministic engines to any AI agent, so natural language instructions return reproducible results after being computed by the engine.

This route has been validated in 2026 to the point of quantified outcomes. Bentley published a case in which an AI agent used the STAAD.Pro MCP server to perform a full structural optimization on a production model, varying geometry, section sizes and materials to reach a design with 40 percent less steel weight. A second case automated slab and wall meshing, one of the most tedious and error-prone tasks in structural engineering, from a single natural language prompt. A third internal test had a team with no programming experience complete an entire design-and-analysis workflow in one week using a large language model plus the STAAD MCP server. These figures are company-reported, but they describe reproducible workflows rather than marketing slogans.

One framing point matters. Bentley belongs to infrastructure-side physical AI, where the central problem is engineering capacity. CEO Nicholas Cumins put it directly at the 2025 Year in Infrastructure conference: there simply are not enough engineers in the world to build, maintain, upgrade, harden and decarbonize the infrastructure that is needed. The AI goal is therefore to compress repetitive work and amplify engineering judgment, not to replace engineers in making judgments. That positioning produces a product form of deterministic engine plus agent orchestration plus engineer sign-off, which differs materially from end-to-end generative design narratives.

Company Overview

Four Decades of Engineering Software

Bentley was founded in 1984 by Keith and Barry Bentley, starting from an early form of MicroStation and expanding across roads, rail, bridges, water, grid, offshore, geotechnics and mining. The company listed on Nasdaq in September 2020. Its long-standing self-description is software built by engineers for engineers, with a product naming system centered on the Open prefix covering OpenRoads, OpenRail, OpenBridge, OpenUtilities, OpenPlant and OpenSite, plus MicroStation as the general modeling base, ProjectWise as the collaboration environment, SYNCHRO for construction scheduling, and STAAD plus PLAXIS for analysis and geotechnics, and SACS for offshore structures.

In fiscal 2025 subscriptions accounted for 1,376.7 million dollars, or 91.7 percent of total revenue, with the remainder from perpetual licenses and services. That ratio shows the transition from licenses to subscriptions is complete, and revenue durability rests on annualized recurring revenue rather than on large annual new contracts.

Three Acquisitions That Changed the Structure

The first was Seequent, announced March 11, 2021 and closed June 17, 2021, for 900 million dollars in cash plus 3,141,361 Class B shares, roughly 1.05 billion dollars in total. Seequent is headquartered in Christchurch, New Zealand, with Leapfrog for three-dimensional geological modeling, Geosoft for geoscience data management, and GeoStudio for slope stability and deformation analysis. The strategic significance is often underestimated: subsurface conditions are the single largest source of cost overruns in heavy construction, because tunnels, mines and foundations cannot know in advance what lies ahead of the face. Bringing the subsurface into the digital twin turns invisible risk into a computable model.

The second was Power Line Systems, announced November 16, 2021 and closed January 31, 2022, for approximately 700 million dollars. It brought software for overhead transmission line and structure design, completing the grid segment. On the August 2026 earnings call Cumins offered a checkable proof point: Power Line Systems revenue outside the United States is now as large as the entire business was at acquisition. The value of that statement is that it distinguishes real distribution synergy from narrative synergy, since Bentley took a domestically concentrated product and globalized it through an existing sales machine.

The third was Cesium, acquired September 6, 2024 for undisclosed terms. Cesium created the 3D Tiles standard adopted by the Open Geospatial Consortium, which underpins streaming of very large three-dimensional geospatial datasets. The acquisition lets iTwin carry city-scale three-dimensional scenes in a browser and lets SYNCHRO+ integrate rich geospatial context.

Several smaller deals point in the same direction. In 2023 the iTwin Ventures arm acquired Blyncsy, which derives roadway condition data from crowdsourced dashcam imagery, turning ordinary vehicles into a distributed inspection fleet. In December 2025 Bentley closed the acquisitions of Talon Aerolytics, which provides asset analytics for telecom and utilities, and the technology and technical expertise of Pointivo, covering drone data processing and AI-powered damage detection and geolocation. Both were announced January 5, 2026 and are grouped under Asset Analytics, pointing at operations-side asset performance.

Segment Structure and Growth Sources

Bentley reports by end market. Growth in the first half of 2026 was led by Resources, covering mining and energy, followed by Public Works and Utilities including the electric grid. The company indicates Resources represents roughly 27 percent of sector-attributable annualized recurring revenue and was the fastest-growing sector in the first half of 2026, driven by mining demand tied to critical minerals. That structure connects directly to the Seequent acquisition: without subsurface geological modeling there is no mining customer base.

The grid is another clear growth line. On the August 2026 call Cumins cited a demand backdrop of roughly 35 gigawatts of additional United States grid capacity required by 2030. That figure is not a Bentley forecast, but it explains why orders for Power Line Systems and OpenUtilities are accelerating.

Technology

Three Layers: Deterministic Engines, Data Foundation, Agent Layer

The clearest way to understand Bentley technology is again to separate layers, because their nature differs completely and discussing them together produces wrong conclusions.

The base layer is deterministic calculation engines. STAAD.Pro handles structural analysis and design with built-in codes for the United States, Europe, China, India and others. STAAD Foundation Advanced handles foundation design across structure, soil and loads. SACS handles offshore structures including jackets and topsides. PLAXIS handles geotechnical finite elements. OpenRoads handles road geometry and earthwork. OpenRail handles track. OpenUtilities Substation handles substations. OpenPlant handles plant piping and equipment. The shared property of this layer is that results are reproducible, auditable and compliant with code clauses, with correctness coming from numerical methods and code implementation rather than machine learning.

The middle layer is the iTwin data foundation. iTwin unifies design models, construction schedules, sensor data and operations records into a semantic asset model, so the same data can be read and written by design tools, consumed by analysis tools and queried by AI agents. Bentley describes it as a universal translator that structures fragmented lifecycle data into an AI-ready schema. The value here is not algorithmic but architectural: without a consistent, queryable and semantically rich representation of engineering information, no agent can reason across assets, designs and simulations.

The upper layer is the agent interface, added during 2025 and 2026, with two components. One is Bentley Copilot, a context-aware assistant that guides users through workflows, surfaces documents and can modify models. The other is a family of MCP servers exposing product APIs through a standard protocol. Published servers include STAAD.Pro, STAAD Foundation Advanced, SACS, MicroStation and iTwin IoT. The MicroStation server, for example, gives agents access to the Python API so they can operate any element, level, reference or view in a DGN file.

The division of labor can be stated in one sentence: the agent interprets intent and orchestrates steps, the engine computes, iTwin supplies context, and the engineer judges and signs off. Bentley states the principle in its technical writing as not letting a language model invent answers, and emphasizes that outputs must explain their reasoning, be easy to verify and fail visibly rather than silently.

Why MCP Suits Engineering Software

MCP became the de facto standard for connecting engineering software to AI in 2026, and the reason lies in the accountability structure of the profession. In many business settings a generative system that is mostly right is still useful: summarizing a document, drafting a message, classifying a record or producing a first-pass workflow all tolerate error. Civil and structural engineering operates under a different standard. Bridges, roads, rail systems, water infrastructure and industrial facilities cannot be designed on creative guesses, and engineers need validated outputs, code-compliant calculations, auditable workflows and control over final decisions.

MCP draws the boundary cleanly. The protocol does not validate engineering results; it is only a connection layer. Domain work is done by the engineering application. The agent can interpret intent, invoke tools and orchestrate steps, but STAAD remains the structural analysis environment and the human engineer remains responsible for review, approval and final judgment. The architecture is not novel technically, but the allocation of responsibility is the key design decision.

Adoption also carries an organizational benefit: it supports bring-your-own-agent. Bentley published the STAAD MCP server to the MCP Registry and submitted it as a Claude Connector, which means design firms can connect their preferred assistant, an enterprise agent framework or internal automation without being locked to one model vendor. For a conservative industry, that non-locking posture lowers procurement resistance.

Three Verifiable Engineering Results

Bentley published three cases worth taking apart individually, because together they show how far AI can go in engineering.

Case one, 40 percent steel weight reduction. An agent used the STAAD.Pro MCP server to perform a full structural optimization on a production model, varying geometry, section sizes and materials to reach a design with 40 percent less steel weight. The number needs careful reading: the headroom for structural optimization depends on how conservative the original design was, and a conservative design has more room than one already through several optimization rounds. The case proves that an agent can drive a complete optimization loop, not that any project can cut 40 percent.

Case two, automated slab and wall meshing. Meshing slabs and walls is widely regarded as tedious and error-prone, consuming hours of manual work and risking structural continuity errors. Bentley completed the entire task from one natural language prompt. The technical substance is not prompt engineering, because the meshing algorithm was already deterministic. The AI replaced the interaction layer of clicking menus, setting parameters and checking results.

Case three, one week for a full design and analysis workflow. In an internal test, a team with no programming experience used a large language model plus the STAAD MCP server to complete an entire design-and-analysis workflow in a week. The most important conclusion Bentley drew was not about efficiency but about quality: when the engineer's expertise, including code requirements, rules of thumb and accumulated lessons, was written directly into the prompts, output quality improved markedly. That finding supports a human-in-the-loop design and indicates that AI cannot yet work independently of engineering judgment.

Data Governance Commitments as a Real Differentiator

Bentley's public commitment on AI data governance deserves separate mention: customer design and asset data is never scraped or used for AI training unless the customer chooses, and if a customer opts in, the resulting model belongs to that customer alone. This has real commercial value in engineering, because the core asset of a design firm is decades of project data and methodology, and using it to train a general model would hand competitive advantage to an outside party.

The commitment is also a competitive weapon. When owners or design firms assess the risk of handing model data to an AI vendor, an explicit no-training clause materially reduces friction. Bentley positions itself as a steward rather than an owner of that data, a phrasing that requires concrete contract terms. Public information does not include those terms, which makes this a diligence item.

Technical Shortfalls and Boundaries

First, Bentley has no high-end multiphysics solver. Analysis capability concentrates in structures, geotechnics, offshore and road geometry, while fluid dynamics, heat transfer, electromagnetics and acoustics require tools from Ansys, Siemens, COMSOL or Cadence. In scenarios needing fluid-structure interaction, aeroelasticity or thermal management, Bentley can only supply pre-processing and data management, with the solve coming from a partner.

Second, AI features are not separately billed. Reported annualized recurring revenue growth does not break out an AI contribution, and management said on the second quarter 2026 call that customers are beginning to apply the capabilities on live projects and that the company intends to monetize them in due course. In other words, AI is currently a retention and upsell instrument, not an independent revenue line.

Third, MCP coverage remains limited. Five servers cover structures, foundations, offshore, general modeling and IoT, but the core products for roads, rail, water and grid design have no corresponding MCP coverage yet. The boundary of what an agent can operate is the boundary of where AI can create value.

Product and Commercial Model

Four Product Lines and One Platform

By function the portfolio divides into four groups. Design and modeling centers on MicroStation plus OpenRoads, OpenRail, OpenBridge, OpenUtilities, OpenPlant, OpenSite+ and OpenGround. Analysis and simulation includes STAAD.Pro, STAAD Foundation Advanced, SACS, PLAXIS and GeoStudio. Project delivery includes ProjectWise, SYNCHRO, SYNCHRO+ and Bentley Infrastructure Cloud. Asset operations includes iTwin, iTwin IoT, Asset Analytics, and Seequent's Leapfrog and Geosoft.

The platform layer is iTwin, which belongs to no single line but carries data horizontally across all products. That structure explains why Bentley integrates acquisitions relatively smoothly: once a new product connects to the iTwin data model, it exchanges information with other products without rewriting its foundation.

The New AI-Native Applications

At the October 2025 Year in Infrastructure conference Bentley released three next-generation applications, all architected around digital twins, AI and cloud connectivity.

OpenSite+ is the first engineering application using generative AI for civil site design, entering limited availability in October 2025, with the company claiming projects delivered up to ten times faster without sacrificing accuracy. The tenfold claim deserves caution: site design iteration cycles are measured in days, and a generative approach shifts scheme comparison from drawing a few options by hand to generating dozens and filtering them, so the gain depends on the manual cost of filtering. Brianne Belschner, model-based design lead at VHB, described it in the early access program as opening many doors to move faster and get better answers, which is a qualitative statement.

OpenUtilities Substation+ targets collaborative substation design, with the core improvement being that multiple users can work in the same model simultaneously, reducing errors and rework during construction. It entered early access in November 2025. SYNCHRO+ redefines four-dimensional construction modeling by shifting to a data-centric workflow, integrating Cesium for geospatial context and using AI to explore construction sequences. It entered early access in December 2025.

AI enhancements also reached existing products. An AI agent that automates drawing annotation, one of the most time-consuming steps in design, became generally available for OpenRoads Designer and OpenRail Designer in November 2025. Bentley Copilot was integrated into both products in early 2026. ProjectWise gained AI search that returns instant summaries without opening files.

The Consumption-Based Model

Management describes the commercial model as consumption-based, meaning usage-based billing combined with subscriptions. Greg Bentley made the AI-era significance explicit: AI-agentic automation greatly increases the value generated by modeling and simulation functionality, and that value is ultimately captured through API consumption. In other words, if agents call solvers frequently, the billing base grows with call volume rather than only with seat count.

Financially this has not yet shown up. Second quarter fiscal 2026 subscription revenue grew 13.6 percent, roughly in line with 12 percent annualized recurring revenue growth, with no visible acceleration from call volume. Judging whether the model works requires watching whether subscription revenue growth outpaces annualized recurring revenue growth over coming quarters.

Customer Metrics and Market Size

A global survey released at the 2025 Year in Infrastructure conference, conducted with Pinsent Masons, Mott MacDonald and Turner and Townsend, found that about half of respondents are piloting AI or have implemented it and plan to scale across their organizations, with priority use cases in design and engineering productivity and documentation automation, covering more than 130 industry leaders. In the same year's Going Digital Awards, nearly one third of submissions and almost half of finalists incorporated AI. These figures indicate AI has moved from experiment into practice in infrastructure, with penetration still around half.

Verifiable customer outcomes include Guangdong Airport Authority, which used iTwin and SYNCHRO on China's largest civil aviation reconstruction and expansion project and saved 78 million renminbi while shortening construction by 33 days, and PT Hutama Karya in Jakarta, which used Bentley software for subsurface analysis and construction of two new substations, improving traffic congestion and air quality. The former is a rare public data point with a specific amount and schedule impact.

Competitive Landscape

The Direct Comparison With Autodesk

In civil and infrastructure design, Bentley and Autodesk are long-standing rivals. Autodesk is strong in buildings through Revit, general design through AutoCAD and manufacturing through Fusion, with infrastructure anchored on Civil 3D and Forma and AI assistants promoted as platform capability in recent years. Bentley is strong in the professional depth of heavy infrastructure: roads, rail, bridges, water, grid, offshore, geotechnics and mining, each with a dedicated product line and code implementation.

The commercial structures differ more than the products. Autodesk is primarily seat-based and aimed at a broad population of design professionals, while Bentley is subscription plus consumption and aimed at large engineering firms and owner-operators. The former relies on user base, the latter on depth of value per customer. In the AI era that difference widens: under seat-based models, AI raising per-person productivity compresses seat demand, while under consumption models AI raising engine calls expands the billing base. That is a structural advantage for Bentley, though it also means customers will scrutinize whether consumption billing is fair.

Against Trimble and Hexagon

Trimble holds strong positions in surveying, geospatial, agriculture and construction machine control, and has advanced project collaboration through Trimble Construction One. Hexagon covers reality capture and asset management through Leica instruments, SmartSite and the HxGN platform. Both combine hardware and software and can collect data in the field. Bentley has no hardware business, and its response is to strengthen iTwin as the aggregation layer and to acquire Cesium for carrying three-dimensional geospatial data.

The relationship with Esri is subtler. Esri is the de facto standard in GIS, and Bentley's geospatial capability must interoperate with that ecosystem. The two are simultaneously partners and competitors, overlapping at the boundary of the digital twin.

The Absence of Simulation Giants and How It Is Covered

Bentley does not do high-end multiphysics simulation, a space occupied by Ansys, now part of Synopsys, along with Siemens, Dassault, COMSOL and Cadence. Where projects need structural nonlinearity, fatigue, fluid-structure interaction or thermal management, Bentley's STAAD and SACS handle routine analysis and code checking while complex problems go to specialist solvers. That division is normal in engineering practice, but it weakens the physical AI narrative: Bentley's physical depth concentrates in structural statics and geotechnics, with limited coverage of dynamics, nonlinearity and multiphysics.

Its compensation is on the data side. Once simulation inputs and outputs are unified in the iTwin asset model, Bentley becomes the manager of simulation data even without owning solvers. That position has long-term value, because engineering decisions need cross-discipline, cross-phase information integration rather than higher fidelity at a single point.

Financing and Capital Structure

Dual-Class Shares and Family Control

Bentley uses a dual-class structure in which the Bentley family retains control through Class B shares. At the IPO, outside Class A holders held about 57.4 percent of the economic interest while the Bentley control group held or could control about 67.4 percent of voting power, making the company a controlled company under Nasdaq rules, exempt from independent-majority board and dedicated compensation committee requirements, with those duties handled by fully independent nominating and sustainability committees. The 2026 proxy indicates the family holds about 52 percent of outstanding Class B shares on an as-converted basis.

The structure has one concrete consequence. On April 19, 2024 Bentley confirmed preliminary discussions with Schneider Electric about a potential strategic transaction and formed a special committee to explore options including a joint venture or a merger with Schneider's software business. Five weeks later the talks ended. Reporting at the time indicated the controlling family would not support a structure in which it lost control or transferred too much combined value. That is the clearest practical demonstration of dual-class control: a very large industrial acquirer arrived and the answer was no.

For minority shareholders this means limited governance checks. Historically, however, family control has coincided with long-term reinvestment: the fully diluted share count changed at roughly negative 1 percent annually over the past five years, reflecting continuous buybacks.

Debt and Convertible Notes

For the Seequent acquisition the company issued 690 million dollars of convertible notes due 2026. On January 15, 2026 it repaid the 678 million dollar principal plus accrued interest using cash on hand and 610 million dollars drawn on revolving credit facilities. A further 575 million dollars of 0.375 percent convertible notes due mid-2027 remains outstanding. Net debt leverage stood at 2.1 times at the end of 2025, a four-year low, falling to 1.9 times at June 30, 2026. Repaying the convertibles reduced the fully diluted share count by about 3 percent.

The stated capital allocation framework is up to 400 million dollars annually for programmatic acquisitions alongside dividends and buybacks. Fiscal 2026 guidance puts free cash flow between 500 and 570 million dollars. That level supports organic growth, small acquisitions and shareholder returns in parallel without leverage for large deals.

Fiscal 2026 Guidance

Guidance for fiscal 2026 calls for total revenue of 1,685 to 1,715 million dollars, growth of 11 to 13 percent in constant currency, subscription revenue growth of 11 to 13 percent, perpetual license revenue roughly flat, services revenue growth of 15 to 20 percent, constant currency annualized recurring revenue growth of 10.5 to 12.5 percent including programmatic acquisitions, adjusted operating income less operating stock-based compensation of 495 to 510 million dollars implying about 100 basis points of margin improvement, an effective tax rate near 21 percent, and free cash flow of 500 to 570 million dollars.

The quality of that guidance needs to be read against first-half execution. First half 2026 revenue was 834.9 million dollars up 13.6 percent, subscriptions were 771.1 million dollars up 14.1 percent, operating margin was 25.7 percent against 27.2 percent a year earlier, and free cash flow was 251.7 million dollars against 273.4 million. Revenue growth exceeded the top of guidance while margin and free cash flow declined year over year, because new enterprise-wide finance and quote-to-cash platforms went live in the second quarter with costs absorbed inside the margin commitment. That is a one-time systems investment, but it did depress current-period profit.

DISCREPANCIES

1. Missing denominators for employee and customer scale. Public materials do not disclose a consistent employee count or customer count. Third-party sources contain phrases such as thousands of employees or serving most large design institutes worldwide, none with a citation. This report avoids unlabeled scale figures and uses only financial metrics from filings and case data from official releases. Source tier: OFFICIAL preferred; unverifiable scale claims excluded.

2. The tenfold efficiency claim has no defined baseline. The company states that OpenSite+ delivers projects up to ten times faster without sacrificing accuracy, but does not specify the comparison baseline, whether that is manual scheme comparison, a specific task or the whole delivery cycle, nor sample size or measurement method. This report retains the figure as company framing and does not treat it as an efficiency conclusion. Source tier: company marketing claim, UNVERIFIED.

3. Scope of the 40 percent steel reduction. Bentley's blog states that a structural optimization driven through the STAAD.Pro MCP server achieved a 40 percent reduction in steel weight on one production model. The result depends on the original design's redundancy and constraints and cannot be extrapolated as an industry average. This report labels it as a single case. Source tier: company technical blog, OFFICIAL but a single sample.

4. Contradictory descriptions of the dual-class vote multiple. A third-party governance analysis states that each Class A share carries 29 votes and converts into one Class B share, which inverts the usual high-vote class design and conflicts with the description of family control through Class B shares. This report does not cite a specific vote count and relies only on verifiable conclusions: the control group held about 67.4 percent of voting power at the IPO and the company is a controlled company under Nasdaq rules. Source tier: OFFICIAL (proxy conclusions) over third-party detail.

5. Three numbers for the Seequent consideration. The official announcement states 900 million dollars in cash plus 3,141,361 Class B shares, roughly 1.05 billion dollars in total. Some sources cite 1.05 billion dollars without breaking out components, and others cite 900 million dollars as if it were the whole price. All three are correct with different meanings. This report uses the itemized form. Source tier: OFFICIAL (acquisition announcement).

6. Definition of the Resources share. A third-party analysis states that Resources accounts for roughly 27 percent of sector-attributable annualized recurring revenue, a phrasing the company does not use in its filings. This report retains the qualifier that it is a third-party estimate and does not present it as company disclosure. Source tier: third-party estimate (MEDIA), below OFFICIAL.

7. Whether AI already generates independent revenue. Management stated that customers are beginning to apply the capabilities on live projects and that the company intends to monetize them in due course, which clearly indicates no separate monetization yet. Some coverage interprets this as AI already being a growth engine. This report follows the management wording and lists the AI revenue contribution as undisclosed. Source tier: OFFICIAL (earnings call) over MEDIA.

8. Attribution of the first-half 2026 margin decline. The company attributes it to costs of new finance and quote-to-cash platforms while emphasizing that costs were absorbed within the margin commitment. The filings simultaneously show operating margin falling from 27.2 to 25.7 percent and free cash flow declining year over year. Both statements are accurate, but the first emphasizes controllability while the second is a quantified fact. This report presents them side by side. Source tier: OFFICIAL.

Risks and Diligence Gaps

First, the AI monetization path is not closed. MCP servers, Copilot and generative design are capabilities rather than revenue. Management acknowledges monetization will come in due course without a timetable or pricing mechanism. If AI features remain retention instruments, their contribution to annualized recurring revenue growth will show up as lower churn rather than higher price per customer, and the valuation logic will differ materially from that of an AI platform company.

Second, the consumption model cuts both ways. Usage-based billing is a tailwind in the AI era, but it requires customers to accept paying more for using more. Large owners and design firms are typically sensitive to opaque usage charges, particularly in public works where budget cycles are fixed. Adoption of the model may lag the technical capability.

Third, two-way uncertainty in China. The fourth quarter 2024 call disclosed constant currency annualized recurring revenue growth of 12 percent, or 12.5 percent excluding China, implying a drag from that market. Bentley has a Beijing subsidiary exposed to data localization, technology transfer and export control issues. At the same time China is pushing domestic substitution of industrial software, which creates structural pressure on its China business over time.

Fourth, strategic continuity risk and opportunity under family control. Dual-class shares guarantee long-termism, and outside shareholders cannot redirect strategy through governance. The termination of the Schneider talks in 2024 shows the family prioritizes control over a short-term premium. That may suit long-term holders and is a risk for those expecting an acquisition exit.

Fifth, the imbalance between professional depth and solver breadth. Bentley's analysis capability concentrates in structures and geotechnics, with no high-end solvers for fluid, heat, electromagnetics or acoustics. When projects require multiphysics coupling, it can only play the data management and pre-processing role. Under a physical AI assessment framework this is a substantive shortfall.

Sixth, standard risk in the MCP ecosystem. MCP is currently the de facto standard but is promoted mainly by a single vendor, and its evolution is outside Bentley's control. If the standard fragments or major model vendors move to proprietary protocols, the interface investment would need to be redone. This risk applies to every vendor building integrations on a third-party protocol.

Seventh, the ongoing burden of acquisition integration. The company has a 400 million dollar annual programmatic acquisition budget and a long integration record. Seequent, Power Line Systems and Cesium integrated relatively well, but Asset Analytics absorbed both Talon and Pointivo in December 2025, and the integration outcome has not been validated through a full year.

Composite Assessment

Technology scores 7.5. Code coverage and professional depth in deterministic engines are top tier for infrastructure, iTwin is a sound information architecture, the MCP route defines responsibility boundaries clearly, and the no-training data commitment is genuine differentiation. Deductions for the absence of high-end multiphysics solvers, narrow MCP product coverage, and limited technical disclosure on AI features.

Business scores 8.0. Annualized recurring revenue of 1.536 billion dollars, 12 percent constant currency growth, 109 percent net retention and 91.7 percent subscription mix are high-quality metrics for industrial software. The consumption model leaves room for billing expansion in the AI era. Deductions because AI is not yet separately monetized and growth is increasingly dependent on the cyclical Resources and grid segments.

Finance scores 8.0. Free cash flow of 520.2 million dollars in 2025, a free cash flow margin near 34.6 percent, net leverage of 1.9 times, and a 3 percent share count reduction from repaying convertibles all indicate capital discipline. Deductions because first-half 2026 free cash flow declined and new systems costs depressed current profit.

Product scores 7.5. Four lines cover the full design-to-operations cycle, the new AI-native applications are architected correctly, and the pace of Copilot and MCP delivery is pragmatic. Deductions because the new applications remain in limited availability or early access with scale validation incomplete.

Market scores 8.0. Large global engineering firms and owner-operators are essentially all customers, demand certainty in Resources and grid is high, and quantified cases such as the Guangdong airport project add credibility. Deductions for the China drag on growth, Autodesk pressure in general design, and sensitivity of infrastructure spending to fiscal and rate cycles.

Overall 7.8. Among the companies covered in this series, this is the most pragmatic AI implementation: it does not claim to replace engineers, it turns deterministic engines into agent-callable tools, it writes responsibility boundaries into the architecture, and it backs the claim with reproducible engineering cases. The shortfalls are equally clear, with physical depth concentrated in structures and geotechnics and AI not yet converted into revenue. The judgment on physical AI is that this represents the correct posture for engineering software adopting AI, with the ceiling determined by whether customers accept consumption-based billing.

References

  1. Bentley Systems second quarter 2026 results release, August 6, 2026: total revenues 410.7 million dollars up 12.8 percent or 12.2 percent constant currency, subscription revenues 378.6 million dollars up 13.6 percent, annualized recurring revenues 1,536.0 million dollars versus 1,379.2 million a year earlier, constant currency annualized recurring revenue growth 12 percent, last twelve month recurring revenue dollar-based net retention 109 percent, operating income margin 21.6 percent versus 23.2 percent, adjusted operating income less operating stock-based compensation margin 28.3 percent versus 29.3 percent, diluted earnings per share 0.25 dollars, adjusted earnings per share 0.35 dollars, operating cash flow 71.5 million dollars, free cash flow 63.8 million dollars
  2. Bentley Systems first half 2026 results in the same release: total revenues 834.9 million dollars up 13.6 percent, subscription revenues 771.1 million dollars up 14.1 percent, operating income margin 25.7 percent versus 27.2 percent, free cash flow 251.7 million dollars versus 273.4 million
  3. Bentley Systems fourth quarter and full year 2025 results release, February 26, 2026: full year revenues 1,501.8 million dollars up 11.0 percent or 10.1 percent constant currency, subscription revenues 1,376.7 million dollars up 12.5 percent, annualized recurring revenues 1,462.1 million dollars versus 1,283.3 million at the end of 2024, constant currency annualized recurring revenue growth 11.5 percent, net retention 109 percent versus 110 percent, operating income margin 24.1 percent, adjusted operating income less stock-based compensation margin 28.6 percent, diluted earnings per share 0.85 dollars, adjusted earnings per share 1.21 dollars, operating cash flow 538.5 million dollars, free cash flow 520.2 million dollars
  4. Bentley fiscal 2026 outlook in the same release: total revenues 1,685 to 1,715 million dollars, constant currency annualized recurring revenue growth 10.5 to 12.5 percent, adjusted operating income less operating stock-based compensation 495 to 510 million dollars, effective tax rate near 21 percent, free cash flow 500 to 570 million dollars; net debt leverage 2.1 times at the end of 2025 and 1.9 times at June 30, 2026; repayment of 678 million dollars of 0.125 percent convertible notes on January 15, 2026; 575 million dollars of 0.375 percent convertible notes due mid-2027 outstanding; up to 400 million dollars annually for programmatic acquisitions
  5. Bentley Year in Infrastructure 2025 press release, October 15, 2025: OpenSite+ in limited availability with the company claiming projects up to ten times faster, OpenUtilities Substation+ early access in November 2025, SYNCHRO+ early access in December 2025 with Cesium integration, drawing annotation agent generally available in November 2025 for OpenRoads Designer and OpenRail Designer with Bentley Copilot integrated in early 2026, ProjectWise AI search, Bentley Copilot described as a context-aware assistant that can modify models
  6. Bentley blog, Would You Cross That Bridge, AI and the Trust Problem in Infrastructure: the STAAD.Pro MCP server performed a structural optimization on a production model achieving a 40 percent reduction in steel weight, slab and wall meshing automated from natural language, a team without programming experience completed a full design-and-analysis workflow in one week, output quality improved markedly when engineering expertise was embedded in prompts, customer data is never scraped or used for training unless the customer chooses and resulting models belong to the customer, iTwin described as a universal translator for lifecycle data
  7. Bentley blog, AI Tools Are Being Employed in Bentley Products: capabilities of the STAAD.Pro, STAAD Foundation Advanced, SACS, MicroStation and iTwin IoT MCP servers, MicroStation MCP access to the Python API and any element in a DGN file, iTwin IoT MCP support for live and historical sensor data and structural health monitoring
  8. Bentley article, Bentley's MCP Server Shows How AI Can Work in Engineering Without Guessing: the STAAD MCP server published to the MCP Registry and submitted as a Claude Connector, MCP as a connection layer that does not validate results while the engineering application performs domain work and the engineer reviews and approves, support for bring-your-own-agent
  9. Seequent acquisition announcement and ARC analysis, announced March 11, 2021 and closed June 17, 2021: 900 million dollars in cash plus 3,141,361 Class B shares for roughly 1.05 billion dollars in total, Seequent headquartered in Christchurch with more than 430 colleagues in 16 offices serving over 100 countries, products including Leapfrog, Geosoft and GeoStudio, expected to add over 80 million dollars to annualized recurring revenue and about 10 percent to annualized recurring revenue, revenue and EBITDA
  10. Power Line Systems acquisition and capital history, third-party company research: announced November 16, 2021 and closed January 31, 2022 for approximately 700 million dollars, Cesium acquired September 6, 2024 for undisclosed terms, Blyncsy acquired in 2023 through iTwin Ventures, Talon Aerolytics and Pointivo technology assets closed in December 2025 and announced January 5, 2026, preliminary Schneider Electric discussions confirmed April 19, 2024 and terminated five weeks later, Resources at roughly 27 percent of sector-attributable annualized recurring revenue, Power Line Systems revenue outside the United States now equal to the entire business at acquisition, about 35 gigawatts of additional United States grid capacity required by 2030
  11. Governance research, third-party diligence analysis: dual-class structure, about 57.4 percent of economic interest held by outside Class A holders and about 67.4 percent of voting power held or controllable by the Bentley control group at the IPO, controlled company status under Nasdaq rules with no independent-majority board or dedicated compensation committee, 2026 proxy showing the family holding about 52 percent of outstanding Class B shares, Nicholas Cumins becoming CEO on July 1, 2024 as the first non-family leader
  12. Bentley Infrastructure AI page and customer cases: Guangdong Airport Authority using iTwin and SYNCHRO to save 78 million renminbi and shorten construction by 33 days, PT Hutama Karya substation work in Jakarta, the global survey with Pinsent Masons, Mott MacDonald and Turner and Townsend finding about half of respondents piloting or having implemented AI across more than 130 industry leaders, and nearly one third of Going Digital Awards submissions with almost half of finalists incorporating AI
This report is compiled from public sources for informational purposes only and is not investment advice. Financial figures follow company announcements and SEC filings, case data reflect company disclosure, and third-party estimates are labeled with their sources.

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.