Industries / Construction and property
EPC contractors
EPC contractors are the firms that engineer an industrial plant, buy its equipment and build it for an owner (engineering, procurement and construction), often for one price and by one date: refineries, LNG terminals, chemical plants and power stations.
1889
Charles Stone and Edwin Webster, who had graduated in electrical engineering from MIT the year before, each borrowed $2,000 and opened an office on Post Office Square in Boston. TIME reported in 1929 that their next large contract after the first was to build a power plant for the paper maker S. D. Warren that carried 1,000 horsepower seven miles at 1,000 volts; by 1903, a Boston history site says, Stone & Webster offered a full range of design and construction services.
40
Milestones achieved across 7 domains.
10
To watch for: announced, not yet confirmed outside.
14
Open deliverables. Nobody has shipped them.
In one minute
Where it started
In 1889 two young engineers in Boston started a firm that went on to design and build power plants for its clients, and in 1922 Fluor built its first refinery project, a gasoline plant in California. In the 1950s large plants were laid out on scale models, with most of the piping placed on the model before the final drawings; in 1976 CADCentre in Cambridge, England, moved that model onto a computer screen. Shared rules followed: indexes and estimate classes for pricing a plant, standard specifications for buying its equipment, and standard forms of contract such as FIDIC's 1999 Silver Book for turnkey plants.
Where it stands
Independent Project Analysis, which benchmarks owners' projects, reported in 2012 that only 22% of oil and gas megaprojects in its database succeeded; the rest averaged a 33% real cost overrun. After losses on fixed-price plants, several large contractors said in 2019 they would stop or limit lump-sum bids, CB&I's owner McDermott went into bankruptcy in 2020, Chiyoda was rescued by its largest shareholder and a bank, and Petrofac's holding company went into administration in 2025. Owners still buy fixed prices from the few contractors that will give them, as Woodside did from Bechtel in 2025, while plant sections are built as modules in Asian yards and buyers pay deposits for a place in the queue for gas turbines.
What is open
Many records that EPC work runs on are now public: the US government's detailed cost estimates for power plants, the oil industry's shared purchase and handover specifications, and signed LNG contracts and their change orders filed with US regulators. A small team can use them to build answer keys and test AI estimators and document checkers, track how fixed prices move and how close contractors run to the limit of their guarantees, publish open models of module yards and plant cost indexes, and score forecasts of site workforces and startup emissions against what later happened, all from drawings, schedules and data rather than live equipment.
Domains
AchievedWatch forOpen
Domain
Progress
Last achieved
Watch for
Open
Front-end design and estimating
Before an owner decides to build, engineers define the plant well enough to price it and judge whether it is worth building; this phase is called front-end loading, and its last step is usually FEED (front-end engineering design).
7 · 1 · 2
Last achievedMost megaprojects miss2011
Watch forFeasibility models in hours
2
openDetailed engineering
After the decision to build, engineers turn the front-end design into the full set of drawings, 3D models, specifications and lists from which equipment is bought and the plant is built.
5 · 2 · 2
Last achievedBuilding before design is done2016
Watch forAI-drawn P&IDs
2
openBuying and expediting equipment
This is the work of buying a plant's equipment and bulk materials, chasing suppliers so that each item arrives on time (expediting), inspecting work in suppliers' factories, and collecting the documents each supplier owes.
5 · 2 · 2
Last achievedPaying for a factory slot2025
Watch forTwenty-four gigawatts a year
2
openModules and fabrication
This is the work of building sections of a plant, called modules, in fabrication yards far from the site, shipping them, and setting and joining them at the site, along with shop-built sections of pipe (spools) and steel.
6 · 1 · 2
Last achievedSanctions turn modules around2024
Watch forModules built near the site
2
openConstruction management
Construction management is running the site of a large plant: splitting the work into packages that crews can finish, housing and supplying thousands of workers, and measuring how much of their day goes into building.
6 · 1 · 2
Last achievedOutside data on work packaging2025
Watch forPackages drawn from the model
2
openCommissioning and handover
Commissioning is the run of checks and trials that takes a plant from mechanical completion, when everything is installed, through pre-commissioning checks, commissioning and first production, and ends with handing the plant and its records to the owner.
5 · 2 · 2
Last achievedCommissioning breaks a permit2023
Watch forHandover compiled 98% faster
2
openContracts and risk
The contract decides who pays when a plant costs more than planned, finishes late or fails its tests: the owner, the contractor, or the partners in a contractor joint venture.
6 · 1 · 2
Last achievedA partner's debts come due2025
Watch forLump sum returns to LNG
2
open
Advances to build with
3 research notes

Claude Fable 5.2
Three working devices
Anthropic says its AI model Claude Fable 5.2 designed three prototype devices that worked without electrical modifications, with no outside test reported.

Muse Ember
The jobs agents repeat
Meta's Muse Ember matched Claude Haiku 5.5 on an outside firm's tests at a third of the cost per task, though both list the same prices for text in and out.

Microsoft ThinkingBox
All twenty times
By Microsoft's own count, 50 steps of training took an AI model from passing 38 of 507 test tasks on all 20 tries to passing 107.
Open in epc contractors
3 of 14 open deliverables
Open · Front-end design and estimating
Agents price NETL's plants
A public test of AI agents as cost estimators. Each agent gets the design basis, block flow diagram, stream table and major equipment list of one of the 20 designs in NETL's May 2025 baseline, with the case name removed, and must return the total plant cost by NETL's account structure. Scores compare each answer with NETL's own estimate. A second, harder track gives the block flow diagram only, as for a Class 5 estimate.
Built with what exists
NETL's 2025 Volume 1 report (public, on OSTI) as the answer key; any agent framework with a spreadsheet or Python tool; open-weights and commercial models. Run every agent several times per case, since a model that passes once may fail on a repeat (see the Note "All twenty times"). A human baseline using Lang factors on the same inputs.
Open · Detailed engineering
An open plant design set
A complete, openly licensed detailed-design package for one plant, for anyone to test tools on: P&IDs in DEXPI (an open file format for P&IDs), a line list giving each pipe's size, material class and design conditions, a 3D model of equipment, steel and main piping, and isometrics as PCF files (the plain-text piping component format that ISOGEN reads). Base it on NETL's natural gas combined-cycle case B31A, whose process data are public.
Built with what exists
pyDEXPI and DEXPI Studio for the P&IDs; FreeCAD with IfcOpenShell for the 3D model; plain-text PCF files for isometrics; NETL's 2025 Volume 1 data for case B31A; GitHub or Zenodo for release under CC BY 4.0.
Open · Buying and expediting equipment
Check supplier packages against JIP33
A public test of AI document reviewers on a job an EPC contractor's engineers and document controllers do for every order: checking that a supplier's documents meet the buyer's specification. Take five JIP33 specifications, such as S-615 for centrifugal pumps and S-562 for ball valves. For each, write 20 made-up supplier packages (data sheets, drawing lists, inspection and test plans, certificate lists) that follow the specification's information and quality requirements, and plant known gaps in them: a missing deliverable, a document sent "for information" that must be sent "for acceptance", a missing hold or witness point for the chosen risk level, a data sheet value outside the buyer's selection. Publish the packages, the answer key and a scoring script.
Built with what exists
The JIP33 specifications (free PDFs from IOGP); Python with PyMuPDF to read PDFs; AI models such as Claude Haiku 5.5 or Meta's Muse Ember (see the Note "The jobs agents repeat") and open models run through Ollama; a spreadsheet for the answer key.