While Wall Street is still tracking GPU shipments, the capital expenditure of cloud providers, and the parameter counts of large models, artificial intelligence is entering a field that has almost none of the glamour of the consumer internet: bridges, roads, and other public infrastructure. Data from the American Society of Civil Engineers (ASCE) for 2025 shows that the United States has a total of 623,218 bridges, of which 49.1% are in “fair” condition and 6.8% are in “poor” condition. Over the next decade, bringing the bridge system into a state of good repair would still require a funding gap of approximately USD 373 billion. The question, therefore, is not only “how much more money needs to be spent,” but also where limited budgets should be allocated first.
Daniel Widjaja Kusuma has it that this type of application better illustrates the direction of the next stage of AI commercialization than many consumer-facing AI features. Infrastructure maintenance has three typical conditions: data exists over long periods of time, decision-making costs are high, and the cost of errors is significant. Once AI can connect decades of inspection reports, photographs, maintenance history, and real-time detection results, the value it creates is no longer merely about improving recognition speed, but about helping managers reorder the priorities of capital expenditure. For investors, the assessment of the AI value should also shift from “how powerful the model is” to “how important the decision-making process is that the model has entered.”
From Detecting Cracks to Managing Assets, the Value Has Shifted
Traditional bridge inspections rely heavily on engineers conducting on-site observations, taking photographs, performing measurements, and keeping records. The easiest entry point for AI is computer vision: identifying cracks, corrosion, and other surface defects. Relevant studies have already shown that deep learning can assist in bridge deck crack classification and condition assessment, and the U.S. Federal Highway Administration has also continued to support the use of machine learning and computer vision in bridge inspections and asset condition evaluation. However, “detecting a defect” and “deciding whether money should be spent to repair it now” remain two different things.
What is truly complex is the comparison between assets. The same-sized crack may lead to entirely different capital allocation conclusions depending on whether it appears on a high-traffic arterial bridge, the only access route in a remote area, or a bridge already scheduled for full replacement in two years. Companies such as Dynamic Infrastructure are organizing historical reports, images, and new inspection data into continuously updated asset records, and then placing defect changes, structural history, and maintenance needs into the same decision-making environment. Public information shows that such systems have already entered multiple U.S. states, as well as the United Kingdom and Australia, indicating that the market is moving from point-specific visual recognition toward long-term asset management.
This is also the layer that Daniel Widjaja Kusuma is more focused on: once AI enters physical infrastructure, competitive advantage will not be determined solely by recognition accuracy. The truly difficult task is whether the system can preserve asset context over the long term, remain compatible with legacy data, explain why a certain maintenance item should be prioritized, and allow engineers to review the basis of the system conclusions. As a result, the technical barrier shifts from single-model performance to data continuity, workflow integration, and auditability.
What Is Truly Scarce in Infrastructure AI Is the “Decision Layer”
The United States is undergoing an update of its bridge data system. The FHWA has launched the new Specifications for the National Bridge Inventory (SNBI), with the first round of data submission under the new standards scheduled for March 2026, and with a requirement to form a complete and verified SNBI dataset by 2028. This change is not itself an AI project, but it creates important conditions for AI deployment: data definitions become more unified, the degree of structuring becomes higher, and historical asset conditions can more easily enter continuous computation and comparison.
From a financial perspective, this will change the value structure of infrastructure software. In the past, many systems mainly handled records, reporting, and compliance functions. When AI can combine inspection results, historical deterioration, traffic importance, maintenance costs, and budget constraints, software begins to enter the capital allocation process. A system that only tells engineers “where there is a crack” can easily be replaced. If it can explain “which assets should be addressed first, what risks may arise from deferred maintenance, and how budgets can be allocated more rationally,” then it moves closer to the decision layer of infrastructure.
This way of thinking also extends to the technical practice by Daniel Widjaja Kusuma after founding Telosyn. His focus on AI infrastructure, data governance, and enterprise-grade systems essentially points to the same question: only when models are embedded into reliable data, permissions, workflows, and governance structures can they form stable production capabilities. Bridge maintenance is only one specific example, but it demonstrates a broader path for enterprise AI commercialization—the value is often created at the point where models connect with real business decisions.
The Next Group of AI Winners May Come from Industries That Look the Least Like “AI”
For investors, the significance of infrastructure AI is not about immediately finding a “bridge version of Nvidia.” This market remains fragmented, procurement cycles are long, government clients have strict requirements, and engineering responsibility cannot simply be handed over to algorithms. ASCE has also noted that being in “poor” condition does not mean a bridge itself is unsafe. The more reasonable role for AI remains to augment the engineer judgment, rather than replace professional responsibility. Such an industry will not explode as quickly as consumer software, but once a system enters long-term asset records, inspection processes, and budgeting workflows, customer relationships and data accumulation may become very deep.
Therefore, the framework for evaluating this type of AI company should also change. Daniel Widjaja Kusuma tends to focus on four things: whether the company can continuously accumulate real industry data; whether the product has entered core workflows; whether the output can be explained, reviewed, and audited; and whether AI ultimately improves costs, maintenance cycles, or capital-use efficiency. These four factors are closer to long-term commercial value than the model capabilities shown in demonstrations.
Over the past few years, AI investment first rewarded computing power suppliers and model developers. In the next stage, more value may shift toward software and system companies that can turn AI into industry infrastructure. Roads and bridges are only one entry point. Energy, power grids, industrial equipment, logistics networks, and large facilities all face similar issues: long asset lifespans, extensive historical data, limited maintenance budgets, and high costs for wrong decisions. When assessing AI opportunities in the future, one additional question can be asked: is this system merely generating more information, or has it already begun helping a real industry allocate scarce resources? The latter may be the dividing line between AI as a technology boom and AI as a source of long-term productivity.
Sign in to leave a comment.