Volume 9
-
The technology for nuclear-powered AI data centers exists; a nuclear industry that learns from one reactor to the next does not. Tony Roulstone of the University of Cambridge set out to bring realism to the demand question, not to dampen the enthusiasm but to locate the obstacles.
Demand is real: 70 gigawatts of data centers by 2030, some say 200 by 2035. Against it stands a delivery record of ten-to-fifteen-billion-dollar gigawatt reactors, planning and construction each taking eight to ten years, and 13,000 workers at Hinkley Point. The US built 101 reactors in the 1970s and 80s, almost all different, even ones on the same site from the same vendor; every industry shows production learning except nuclear. Roulstone’s research found SMRs built conventionally will cost as much as or more than large units. Built as standardized products with production learning, modeled early units undercut a gigawatt-scale PWR on capital cost and later compete with wind and solar.
Interface realities got attention too: an AI training load swings between zero and full power faster than any reactor or gas turbine can follow, so a battery-backed buffer must sit between the two. Roulstone’s closing point widened the frame: these challenges yield only to a whole-systems view, reactor, power profile, fuel, supply chain, workforce, and financing together.
-
Existing project metrics do not predict project performance; production measurements can. Roberto Arbulu of Strategic Project Solutions argued that project outcomes emerge from the behavior of production systems, not from the schedules, budgets, and management reports used to observe them.
The paper defines six fundamental production system measurements, four rates (demand, capacity, throughput, and load) and two ratios (rework and scrap), alongside work in process and capacity utilization, two quantities teams can deliberately design and control. A worked piling illustration carries the logic: 10,000 piles in ten weeks sets demand at 200 per day, the slowest operation caps system capacity at 175, and variability drops actual throughput to 150. Add 20 percent rework at one operation and the reentering piles push its utilization past 100 percent, quietly moving the bottleneck. By Little’s Law, WIP is the leading indicator; cycle time trails it.
Arbulu closed with nine questions for project teams, from whether demand is understood to whether scrap is being generated. Asked how these measurements transfer to engineering work, Arbulu answered that the same laws apply, with engineers, and sometimes software licenses, setting the system’s capacity. The standing challenge: “are you observing your production systems or are you proactively managing them?”
-
A resource-loaded schedule is not a production system model, and the difference is measured in weeks. Chet Carlson of Factory Physics and Strategic Project Solutions built his session around the question project teams ask most, what a model adds that the schedule cannot, taking both apart: inputs, assumptions, and what each can predict.
The test case: 620 pipe spools across five areas of a pipe rack, planned at six weeks on a resource-loaded schedule with adequate resources and material. The same scope, modeled and run 100 times, showed almost nothing exiting the system in week one and the work finishing in nearly twelve weeks. Schedules cannot show this because they assume infinite capacity, deterministic durations, sequential dependencies, constant productivity, and interchangeable resources. Waiting time grows nonlinearly with utilization, and the contractor’s practice of batching spools five at a time built inventory between operations. In Carlson’s summary, “schedules are essentially demand and cost models”: they state what should happen, not how work moves.
Carlson closed by asking whether the standard set of project artifacts should include production system models. The recorded Q&A takes up who owns such a model on a live project and where its data comes from. Most of it already exists, Carlson argued, spread across project documents and the people closest to the work.
-
Ask why capital projects are hard to control and the answer is usually one word: uncertainty. James Choo of the Project Production Institute argued that the word covers six distinct behaviors and using it loosely is costly: one label draws one remedy, more contingency, more planning, more pressure, then longer durations, higher cost, and more tied-up cash.
The paper names six sources of demand-capability mismatch: uncertainty, randomness, statistical variability, structural heterogeneity, rate imbalance, and nonstationarity. Uncertainty is incomplete knowledge of a future state; randomness is a known distribution with an unpredictable next draw, the die whose rolls average 3.5, a number on no face; heterogeneity is physical difference, an 8-inch weld averaged with a 36-inch one; nonstationarity means the system changes as it ramps. Each calls for a different response; none is answered by contingency alone. Stock and time buffers trade against each other; only added capacity reduces their total.
Choo’s closing rule was diagnose before buffering: buffer design is an economic choice, not a slogan. The Q&A asked how to keep closed-loop control stable with noisy data, watch whether queues grow before acting, Choo answered, and how the framework transfers to engineering work. Every mismatch lands somewhere: “It’s not just an operational nuisance. This is a capital productivity issue.”
-
Every project buffers variability somewhere; few choose where deliberately. Iris Tommelein of UC Berkeley, presenting with James Choo of the Project Production Institute on work co-authored with Gregory Saragih, used the Parade of Trades to turn that choice into a measurable trade-off between trade time on site and total project duration.
Five trades pass 100 units at rates set by dice. With no variability, an all-five die, the work takes 24 time units; replace it with a 3-7 die, same average of five, and duration climbs past that, because each trade is capped by whatever its predecessor releases. Delaying a trade’s mobilization inserts a time buffer that decouples it and cuts its time on site, while total duration grows. Computer experiments sweeping mobilization delays and dice variants trace the Pareto frontier between the two; a lower-variability 4-6 die shifts the whole frontier, and placement matters as much as size.
Choo grounded the choice at Heathrow Terminal 5: 200 tons of cut-and-bent rebar per day, storage limited to one day, and a deliberately capacity-buffered system, min-max on rail, trucks as backup. Asked whether the Last Planner pull plan is where buffers get designed, Tommelein answered it starts earlier, at the master schedule. Buffering happens either way; design decides its cost.
-
Stage-gate governance reviews a project quarterly; risk arrives daily. Keith Magowan, a major capital projects executive, argued that the capital value process governing oil and gas megaprojects was designed for simpler supply chains and slower information; in Magowan’s framing, “governance latency is the root cause of drift.”
Magowan cited the record: 80 percent of major capital projects exceed sanctioned cost or schedule, and 1.6 trillion dollars of value is destroyed annually. The Adaptive Project Delivery Framework layers domain AI agents and a synthesizing agent, a virtual project director, over project production management: PPM defines how the production system should behave, APDF signals when it is not, condensed into a continuously refreshed Project Confidence Index. A demo on a synthetic project injected a 14-week compressor fabrication delay: the index fell from 74 to 56, exposure reached 67.2 million dollars, and a Tier 3 escalation convened an emergency supply chain session within 72 hours.
Magowan closed on validation: whether a confidence index transfers across industries is open, and replacing the capital value process is what he called heart surgery for an oil company, run in parallel for about two years while trust builds. AI augments leadership rather than replacing it, humans must decide; the advantage goes to companies that fix governance velocity first.
-
Proper Selection of a Project Delivery Method: Reconciling Deterministic Frameworks with Operations Science and Project Production Management
Delivery method selection is still treated as commercial risk transfer. Victor Viteri of CONEXIG argued it decides whether the production system can be optimized. Risk moved on paper does not change how work flows, yet traditional frameworks evaluate commercial characteristics and largely ignore production physics.
The empirical base is Construction Industry Institute analysis of 351 projects: delivery methods produce statistically different outcomes, with construction management at risk 4.5 percent lower cost than design-bid-build and design-build 6.0 percent lower. “Commercial risk transfer does not eliminate production risk.” Viteri grounded the claim in Little’s Law: a crew completing two apartments a week with ten open fronts carries a five-week cycle time; opening five more fronts stretches it to 7.5 weeks, because throughput never changed. The framework keeps deterministic tools, CII decision models and Analytic Hierarchy Process pairwise comparisons on the Saaty scale, then adds a production test: whether the selected model enables production system optimization.
Viteri closed by rejecting the search for a universally superior method; the contribution is a third stage of delivery selection, after contract-centered and deterministic stages, in which owner strategy and production capability are evaluated together. Whether owners will add production criteria to frameworks built on commercial characteristics alone is the question the session left open.
-
Lean, AWP, and PPM share one origin and push in uncoordinated directions. John Strickland of Collaborative Flow presented Advanced Project Delivery as the reconnection. APD is a forum, not a merger: each institute keeps its mission while the three learn from each other and combine their influence.
Strickland traced the fifteen-year divergence: lean construction began from the tenet that projects are production systems, its planning practices fed Alberta’s workface planning, which became Advanced Work Packaging, and PPI formed to refocus on the original tenet through Operations Science. Strickland’s image is an industry that took a contractual approach generations ago and now sits up to its axles in mud, with well-intentioned institutions pushing without coordinating the push. The foundations, a deliberately non-definitive list, include early engagement and aligned commercial interests, robust production system design and WIP control, respect for human factors, distributed decisions and governance, removal of field constraints, coordinated supply chains, and rapid scope validation.
Individual paradigms change readily, Strickland argued; institutional ones do not, so “we’re going to need institutions to challenge institutional paradigms.” Strickland announced the CII-sponsored Advanced Project Delivery Association, meeting first in November. In the closing exchange, James Choo noted customers already ask whether AWP is compatible with PPM; whether the institutes answer with one voice is now the test.
-
Manual production reporting is too slow and too subjective to trust. Dan Rahill of Pirimid asked whether fabrication data can collect itself. Owners re-check every layer of manually reported progress until monthly reports land six weeks late; advanced manufacturing runs on continuous, objective data from a sensor and an algorithm.
The test case was simple: one week at an oil and gas fabrication facility in Texas, a three-inch filtration housing product, and sensors reading power consumption off the welding machines over long-range wireless. The work did not change. Calibrated against half a day of manually logged welds, the machine-learning model classified welds with 99 percent accuracy, and the two misses had already been flagged as anomalies. Task-level data reordered the diagnosis of variability: active weld time varied about 1 percent day to day, while the number of working hours in the day varied 27 percent, the bulk of the daily swing. The model refreshes every fifteen minutes.
Rahill closed by claiming welding is only the proof: any task drawing power has a characteristic energy signature. Pressed by Gary Fischer, Rahill said the barrier to scale is finding a project willing to try it, not the technology. When a process change shows its effect in five hours rather than a week, experiments become cheap enough to run routinely.
-
A Decision Framework for Early-Stage Construction Robotics Deployment: Evidence from 10 Commercial Envelope-Installation Projects
Every robot deployment forces the same choice: fix the product, or hand the crew a procedure. Kenrick Tjandra of Raise Robotics built a decision framework for that choice. With co-author Rishabh Aggarwal, Tjandra opened on a robot arm post too tall for the construction hoist: folded by hand on a loading dock, then lowered by default in the next release.
Field evidence spans ten deployments, nine general contractors, and eight states: more than 5,000 robot hours, zero safety incidents, and roughly 13,500 worker-hours of fall exposure eliminated. An autonomous robot touches all five production levers, and variability is “the lever the robot moves most, and the one the jobsite moves back.” The framework sorts each problem through a three-class resource taxonomy, three response tiers running from permanent design change to crew procedure, and an overlay of the robot’s lab-characterized capability envelope against the observed field distribution, which is wider, owned by other parties, and rarely measured until it disrupts a deployment.
Tjandra closed on the limits: ten deployments of one application. Asked by Gary Fischer what role the GC or owner plays in workaround decisions they ultimately bear, Aggarwal answered that field-level buy-in is the deployment bottleneck. Which fixes belong on the product roadmap, and which the crew keeps paying for, is the question every deployment re-answers.
-
From Visual Perception to Flow Observability in Anchored Shotcrete Retaining Walls: A Preliminary Computer Vision-Based PPM Framework
Construction sites photograph everything, yet images are documents, not production data. William Yalico Arango of COSAPI turned drone images into production events. A photo shows the site but not when work started or how long a wall waited; the preliminary framework closes that gap for anchored shotcrete retaining walls in deep urban excavation.
The dataset is 13,760 drone images from daily flights over a healthcare project in Lima: two sectors, six levels, 46 wall panels per level. Field reality was the first lesson: a baseline classifier scoring 99 percent under controlled conditions fell to 15 percent on real site imagery, forcing an optimized architecture checked with Grad-CAM to confirm attention on the wall itself. A crew detector performing above 0.95 across five categories lets simple rules infer production states, and the central finding follows: “Same visual phase, different production state.” In the observed sequence, a proof of concept, 40.8 percent of time was productive and 34.3 percent waiting.
Future work extends detection to intermediate activities, links the model to BIM, and scales to other processes. Gary Fischer pressed on the drone flybys: separating waiting, inspection, and support work may need continuous capture, and Yalico Arango pointed to fixed cameras as the next step. Whether flow observability can come from imagery sites already collect is the test ahead.
-
Asked to pick a wall system for a data center, AI answered a better question. Stephanie Ware, an independent consultant, tested AI inside the PPM framework. The structured sequence began with product selection and ended in production system design, contracting strategy, and variability, which Ware presented as the core insight of the work.
The method is the contribution: a five-step prompt sequence, four large language models cross-checking one another as a built-in peer review, and foundational prompts grounded in engineering references and PPM to keep output verifiable. The sequence singled out tilt-up and precast concrete, then described them in production terms: tilt-up as a short-duration on-site factory, precast as off-site manufacturing dependent on logistics. In the AI-modeled variability step, a baseline of 3 hours per panel stretched to between 3.6 and 4.8 hours under 20 percent variability. Ware reported the limits: time windows missed production logic, and outputs stayed linear despite repeated prompts.
Ware’s summary places AI’s value inside the framework: “a traceable chain of reasoning, linking early design choices all the way to the production system performance.” The Q&A turned to teams browbeaten into thinner budgets; Gary Fischer reflected a production system design would have armed him to push back. The value holds, Ware cautioned, only where teams think in production terms.
-
Most 4D BIM models stop being used the day they are first shown. Juan Diego Delgado of the University of Houston asked whether AI can keep them alive. A scope change cannot reach the model until someone revises the schedule, re-exports, relinks, and reruns, so the model falls behind the project it depicts.
The baseline is the Penn State BIM guide workflow, whose model-follows-schedule dependency Delgado called “a death sentence for the 4D models.” Four bottlenecks followed: a rework loop, isolated validations, a schedule-first structure, and static data. AI alone fails: Microsoft research Delgado cited found 25 percent of document content corrupted after more than twenty edits. The prototype therefore confines the AI to extracting intent from natural language; a deterministic Python engine changes the files, and the planner validates a changelog. In the pilot, on one basic model, the time to make a 4D change fell substantially and the flow inverted: the 4D model became the input, the schedule the output.
Delgado’s closing claim was that AI’s role is extraction: pulling information from complex environments planning could not reach before, and feeding it into production processes. Asked whether the tool runs outside Autodesk, Delgado answered that the process logic, not the software link, is the contribution. Whether the inversion holds on real projects is the open question.
-
Turnaround delays are rarely caused by the craft task itself. Keith Grimes of Turnaround EPC blamed the system around the task. More than 65 percent of turnarounds exceed planned cost or schedule, and the causes sit in permits, inspections, scaffold access, and decision delays, not in how fast crews work.
The headline measurement comes from a recent shutdown where waiting accounted for 60 to 80 percent of total cycle time across welding workflows: a pipe weld needs limited touch time but queues behind permits, scaffolding, and inspection hold points. Bottlenecks sit in shared services rather than craft execution; a saturated resource turns small additional demand into disproportionate delay across work fronts. The prescriptions follow the flow: release work no faster than the system can absorb it, deliver prefabricated spools at the last responsible moment rather than staging them months early, and evaluate emergent work in system context, since it competes for the same constrained resources as planned work.
Grimes closed with the paper’s central message: “Production system optimization transforms turnarounds from schedule-driven plans into flow-driven systems.” In the Q&A with Gary Fischer, Grimes called the shift cultural: one turnaround run with PSO cut duration 25 percent; the client’s next outage without it went 30 days over. For owners who budget for overruns, predictability is the advantage still unclaimed.