About one US bridge in fifteen is rated Poor: 41,677 bridges in FHWA’s 2025 inventory. Most Poor bridges are small, old structures on local roads, but public exposure is concentrated elsewhere: bridges with a Critical or Serious structural rating carry about 14.3 million vehicle crossings a day, and the largest share of traffic on Poor bridges is on urban Interstates.
How long a bridge lasts depends heavily on what it is made of, where it is, who owns it and how well it was built. Estimated service lives range from about 40 years for timber in the South Central states to over 100 years for concrete on the Pacific coast.
About 56,000 bridges are already past their expected service life. The number coming due each year rises from about 3,700 in the next decade to about 7,900 in 2056 to 2065, against a recent build and reconstruction pace of about 5,700 a year. Aging is only one cause: 52% of ranked Poor bridges wear out faster in their state’s climate, 58% have a failing substructure or culvert, where river flooding and scour do their damage, and the risks differ by region: hurricanes on the Gulf and Atlantic coasts, earthquakes in the Pacific states, Alaska and Hawaii, and collisions on busy Interstates.
Aging is only one cause. Of the Poor bridges ranked in the report, 41% are past their expected service life, 52% are of a material that wears out faster in their state’s climate, and 58% have a failing substructure or culvert, where flood and scour damage appears. Among US bridge collapses, floods and scour cause about half, collisions about 15% and overloads about 13%. The recommendations and state plans address each of these causes, not aging alone.
At FHWA’s official 2025 unit costs, replacing every Poor bridge would cost about $85 billion, or $58 billion if they were rehabilitated instead. The 34,883 Poor bridges ranked individually in the report cost $72.5 billion. Ordering the work by risk to people is highly efficient: the 500 highest-risk bridges cost about $5.0 billion and carry 16.4 million crossings a day. The report closes with an action plan for every state, the District of Columbia, Puerto Rico and the US Virgin Islands, ranked by the size of the problem, naming each state’s at-risk bridges in priority order and whether aging, climate, river flooding, hurricanes, earthquakes or collisions lie behind them.
Roles. I directed the analysis: I chose the questions, set the focus on risk to the public, challenged the assumptions, and decided what went into the report. AI did the technical execution: exploring the data, writing and testing the SQL, fitting the models, building the charts, and drafting the report.
The data. Six National Bridge Inventory tables (bridges, inspections, states, counties, owner agencies, and functional classes) were loaded into a SQLite database: 569,864 bridges and 624,193 inspection records, whose Poor count matches FHWA’s 2025 inventory. The analysis also uses FHWA’s 2025 bridge replacement unit costs, FHWA’s scour program records and published research on US bridge failures.
Questions first, then queries. The work started from the questions that matter most for public safety, each answered with SQL queries against the inventory. Each result was interpreted before the next question was chosen, so the analysis drilled down step by step: condition by state and material, then which component fails, how much traffic crosses deteriorated bridges, who owns them, and when replacement comes due.
From an approximation to detailed service lives. The analysis began with an 80-year lifespan for every bridge as an initial approximation. That was discarded during the analysis in favor of more detailed life data, because lifespans specific to location and construction made more sense. The final model is fitted to the data. For bridges never reconstructed, the chance of a Poor rating was fitted against age for each material, which gives the age at which bridges of that material wear out. Adjustments for region, owner, road type, design load standard, and number of spans were fitted together, so each one is net of the others. Design load standard served as the best available stand-in for construction quality. The resulting lives range from about 40 years for timber in the South Central states to over 100 years for concrete on the Pacific coast.
Ranking by risk to people. All 34,883 Poor bridges were ranked by a score that combines severity (the worst component rating, weighted most heavily), traffic exposure, age, route importance, and whether the bridge was ever reconstructed. Closed bridges were listed separately. Traffic counts that could not be real, such as over 500,000 vehicles a day on a small 1930 bridge, were scored at the typical value for that class of road instead of being trusted.
Cost and schedule. Each Poor bridge was priced at its deck area times its state’s FHWA replacement unit cost, using the National Highway System rate for principal arterials, and the replacement program was scheduled with many projects running at once, with costs escalated at 3% a year. The start rate, project durations, and escalation remained adjustable, so the schedule could be tested under different assumptions.
Checking the numbers. Every figure in the report traces to a query or a model output. Data problems were documented rather than hidden: repeated structure numbers that attach conflicting inspections to about 33,700 bridges, implausible traffic counts, length and area columns that appear to be metric despite feet labels, and missing inspection dates. As a final step, every output was regenerated together so that the report, the queries, and the supporting tables agree.
Every cause, not only aging. Each Poor bridge was checked for aging (past its expected service life), climate (a material that wears out faster in its state) and flood and scour (a failing substructure or culvert). Causes beyond aging were indicated from the data: climate from shorter state service lives, river flooding and scour from failing substructures and culverts, and collision exposure from route type. Hurricane and earthquake exposure was assigned by state from NOAA landfall records and the USGS 2023 seismic hazard model. Collisions, overloads and earthquakes are not recorded in the tables, so each state plan includes a check for them.
Limits. The tables analyzed have no structure-type, scour-vulnerability, overtopping, seismic or load-posting fields; construction and maintenance quality are inferred from proxies; and a single inspection snapshot cannot show bridges already replaced. The report’s Areas for further research section sets out how to close each gap.