Eulia, Inc. | for Janika and Munim

Business plan for the data-centre scheduler

What it costs to run, what to charge and how, how big the market is, how the first product grows into a platform, how to sell it, and how to pitch each funder. Answer first; detail and calculators below.

From research run on 10 and 11 October 2026: six research agents, one blind agent who saw none of their work, and two independent reviewers who attacked the result and re-checked it. Product name still open; "the product" below.

The answer

AI cost is small; price, people time and usage decide the margin; the price decides the market, and data centres alone are a beachhead.

AI and hosting, one project, a month

about $20 to $50

At about one big delay a month. Lean design about $2.20 a delay; the deeper design the rulebook allows $8 to $11. Under 2% of any band from B up. Modelled, not measured.

estimate

Price per project

$12,000 pilot

Paid, 90 days, credited to year one. Then about $1,500 to $15,000+ a month by contract size, unlimited users. US forms quote these dollar numbers.

founders' draft, untested with buyers

Market, data centres alone

$53M vs $3M

$53M a year at our prices, $3M at a competitor's (SmartPM's $417 a project-month), on the same 586 projects. The price is the market.

estimate

1AI cost does not decide this business; price, people time and usage do.

On the lean design (code computes, the AI reads small packets) one delay event costs about $2.20 in AI, and one project about $20 to $50 a month in AI plus hosting at about one big delay a month. The lower figure holds routine AI and hosting at their central values; a wider spread of those puts the middle near $47. That is under 2% of any band from B up. The blind agent got the same order ($50 to $100). Every figure is modelled, not measured: token volumes come from a 138-activity demo schedule and usage has never been counted on a real project.

estimate

2Multi-agent work is dearer, and the version the rulebook allows stays cheap.

A tool-using "deep search" (about $62 a delay event on Opus 5.5) is not allowed: the rulebook says "the AI has no tools and never acts". A deeper design that obeys it costs about $11 a delay event on Opus 5.5 high, $8 with Sonnet workers, against $2.20 lean. It does not fit today's $5 per analysis and $3 per round caps (only 3% of simulated events fit both on Opus); it would need about $9 and $10, a founders' decision. At the stress case (15 events a project-month) it costs about $190 a project-month against $50 lean.

rulebook: checkedcosts: estimate

3Fable 5.1 cannot be used on customer data; prompt caching can.

Fable 5.1 requires 30-day retention and is not available under zero retention. Prompt caching is allowed under zero retention.

checked: Anthropic data-retention page, Oct 2026

4Price per project, banded by contract size, unlimited users.

One build or phase, sold for the project's term. Never per user; country sets only currency and price list; "per company" comes later as the contract wrapper. The blind agent reached this independently; the pricing research tested the founders' draft bands against competitors and kept them, so its agreement is a review, not an independent finding.

5The price points are the founders' draft, untested with any buyer.

A paid 90-day pilot at $12,000, then per project a month A $1,500 / B $4,000 / C $8,000 / D $15,000+ if the euro numbers are used as dollars, or about $1,700 / $4,500 / $9,000 / $16,800 converted. One rule must be picked and quoted the same way on every form. The blind agent priced the floor at $3,000 and the pilot at $15,000 to $30,000.

founders' decision

6Cheap comparables already sell part of this.

SmartPM's Controls plan, about $417 a project-month, lists "What-if Scenarios & Time Impact Analysis"; Sitedrive is €538 to €690. The bands are 2 to 40 times those, and the premium rests on three things no buyer has yet confirmed: knowing which schedule links can really move, costing each option against the budget, and same-day speed.

checked: SmartPM and Sitedrive price pages

7Who pays is not who loses.

The famous delay figures are the owner's loss (up to $14.2M a month on a 60 MW site, STL Partners: "can cost developers"). The contractor pays the fee; its exposure is its own delay damages (rate and cap not found; under the contract, nil when the owner caused the delay, which the product marks as client risk) and its site overheads. In Janika's experience: "Most data center projects care more about days than cost. They are often willing to throw a lot of money at recovering a delay" (money spent on the recovery itself, crews and overtime, not on software), and "sometimes project director, sometimes (if we're lucky) client asks for it". So the contractor's reason to buy is recovering the project days its client presses for; that is her judgement, not yet buyer evidence. The money case on the contractor's own figures is the first thing buyer interviews must supply.

STL wording: checked

8Margin is set by the band and by people time per project, not by the AI design.

At one delay a month: band B about 37% to 85% as people cost to serve falls from $2,500 to $600 a project-month; band C 69% to 92%; band D 83% to 96%; band A negative at $2,500 and under 60% even at $600. Band A reaches 75% only below about $350 of people time a month. Moving people cost from $2,500 to $600 saves about 14 times more than every AI lever combined. The thing to avoid is margin that falls as customers use it more.

estimate on judgement people cost

9The market is a beachhead, and its size is the price.

About 586 live data-centre projects with a P6-scale schedule outside China (320 to 921). At the founders' bands that is $53M a year for delay recovery and $127M end to end; at SmartPM's price the same projects are $3M. The blind agent's first-product figure is $125M, 2.4 times ours, because it counts more projects (likely including China) and 1.3 paying accounts per project. Both say data centres alone are low hundreds of millions at most. Revenue in 2031: about $4M to $5M from about 48 live projects; because fees end with each project, holding that needs about 25 to 30 new projects a year by then.

market: estimate2031: speculative

10How big can this get. internal: not for applications

Not a billion on data centres alone. The $2.3M a project-year of planning, commissioning-paperwork, consultant and reporting labour is spread across four payers (contractor, owner, trades, the owner's commissioning agent); taking 10% of all of it is about $133M, the same as the software TAM, so it is not a separate market. How much of the contractor's own spend the first product replaces is unknown. Venture scale needs three things, each with a first proof: the bands holding with buyers; a second vertical sharing the same contractors (chip fabs first: Turner, DPR and Mortenson build both); and moving up to the owner's budget. The claim to make outward: the product starts where every P6-scheduled megaproject hurts (a reported delay) for contractors who build data centres, fabs, battery plants, power and LNG; if the price holds across those, the market is about $340M a year at software prices (range $45M to $1.0B), growing with the build-out, and the high case reaches about $47M of revenue in 2031 on data centres alone.

speculative

11A narrow first product with a visible ladder, never "replace P6" first.

For YC, Arc, speedrun and construction VCs. It is now the most common shape of application, so it is not a differentiator: the team, the demo and buyer evidence decide (YC partner: team 70, idea 20, traction 10). Say what the product makes as the specification names it: a time impact analysis of each delay (report to decision), a delay impact statement for the client measured against the accepted schedule with recovery shown separately, and costed recovery options. Whether clients accept the statement as the contract's analysis, run on the product's own engine rather than in P6, is a first-pilot question. Speed: Janika's own experience is attributable (a planner "has to manually think about it for multiple days"); "we do it in a working day" waits for the demo and a timed test.

high confidence

12One premise to correct before pitching.

Commissioning is not run on spreadsheets "because there's no platform". At least five commissioning tools sell to data centres (CxAlloy, Facility Grid, Bluerithm, CxPlanner, Quicx), and Facility Grid already syncs P6 with commissioning. The real gap: none reasons about schedule logic, and the schedule, test status and equipment dates sit in three systems owned by three parties.

tools: high confidencethe gap: medium confidence

13Why we win, and what is not yet defensible.

Outward: we start from the reported delay where others start from the whole schedule; we know which links can really move, from a co-founder who plans these jobs, written into the founders' shared rules library; every option is costed against the budget and checked by code, so a planner can trust it the same day; and each project keeps its decision record.

internal: not for applications None of this is a moat yet: a funded competitor with a senior planner could copy the rules, and by design each customer's confirmed rules stay theirs and never improve the shared library without written permission, so never claim a "compounding" library.

rulebook: checked

AI running cost

Lean costs about $2 a delay; the deeper design the rulebook allows, $8 to $11; the tool-using designs are not allowed at all.

Three words first. A token is the unit AI is billed by, roughly three-quarters of a word. The lean design has ordinary code do the heavy work and asks the AI to read small packets. A deeper (multi-agent) design sends many small AI calls at once and has code check every answer. The cache lets the AI re-read text it has just seen at a fraction of the price.

A 9,000-activity schedule is about 800,000 tokens; 25,000 activities (about 2.2 million) does not fit any Claude model's working memory. So "code computes, the AI reads small packets" is not only cheaper: it is the only way large schedules work. estimate

Cost per delay event (build, round re-runs and documents)

Likely case (middle of a simulation), with the 10th to 90th percentile in brackets. Grey rows break the rulebook. estimate
Design	9,000 activities	25,000 activities	Fits today's $5 / $3 caps?
Lean, tiered models (Haiku, Sonnet, Opus)	$1.49 ($1.17 to $1.90)	$1.93	Yes, always
Lean, Opus 5.5 medium (the specification's default)	$2.23 ($1.79 to $2.79)	$3.03	Yes, always
Lean, Opus 5.5 high	$3.50 ($2.72 to $4.53)	$4.70	Yes at 9,000; 89% at 25,000
Deeper, Sonnet workers under an Opus lead (no model-held tools)	$7.96 ($6.07 to $10.41)	$14.06	Build 98%, round 27% at 9,000
Deeper, Opus 5.5 high (no model-held tools)	$11.41 ($8.74 to $14.94)	$20.28	3% at 9,000; never at 25,000
Not allowed: tool-using lead agent, Opus 5.5 high	$62 ($36 to $110)	$86	Model-held tools
Not allowed: the same on Fable 5.1	$148	$205	Retention
Not allowed: naive agent pulling the schedule in	$223	$600	Model-held tools

Allowed by the rulebookNot allowedWhisker: 10th to 90th percentile

The deeper design that obeys the rulebook

Code builds every candidate option the rules allow and slices them into packets.
Each packet goes to a worker in one structured call with no tools.
Code builds and engine-tests the workers' suggestions for a second wave.
One lead call picks, ranks and explains up to three options; critic calls return fixed objections that code checks.
The link judge votes five times instead of three; three document writers split by audience.

No model calls another model. It costs a fifth of the tool-using design because no call re-reads a growing conversation. Whether several waves of AI-suggested changes stay within the rulebook's "the AI may suggest further changes; code builds and checks every suggestion" is a reading for the founders.

Caps it would need (founders' decision, set only in the specification): about $9 build and $10 round at 9,000 activities on Opus high (99 events in 100); about $16 and $21 at 25,000. estimate

Blind cross-check: about $35 a pass and three passes, about $105 a delay event for its tool-using design, above our $62 for the same kind; both are the kind the rulebook does not allow.

Bounding a month

Per-run caps bound each run, not the number of runs, and today's monthly counter (50 analyses a project, a placeholder) alerts and does not block. checked

An allowance of included deeper runs per band with paid overage. Each included deeper run adds at most about $12 over a lean one, so the allowance size barely moves margin; to earn 80% on the overage alone it would sell at about $100 a run.
A hard block at a monthly AI budget, which needs a rulebook change.

Testing cost is a budget fact

The rulebook lets no model, prompt, engine or shared-rule change go live until the graded test set passes 10 runs. checked For the tool-using design that is about $31,000 per change, against $1,000 to $1,750 lean; for the compliant deeper design it is lower but not yet costed. At pre-seed this argues for lean through the pilots. estimate

Munim's cost dashboard, checked internal: not for applications

It reproduces both cost models exactly at the central case and every price is current. In its costlier case it understates infrastructure by about a third (one activities input feeds both models); it has no multi-agent design and no routine AI; one of its options relies on a confidence source the rulebook forbids; and its headline costs carry $25 of hosting where the hosting model gives $1 to $10 at the central case. Engineering detail for the scheduler, never application text. high confidence

Measure first

Tokens per run on a realistic schedule (a stub test rig, a few hundred dollars).
Cache-hit rate.
Whether a deeper design beats lean on quality at all.
Delay events per project per month.

Margin, and the route to software margins

People time per project sets the margin. Bands C and D clear software norms; band B needs productised onboarding; band A does not work in year one.

Gross margin per project-month, by people cost to serve

One delay event a month, lean design (the specification's default), AI at the likely case, 9,000 activities; payment fees not included. Each cell: same-numbers rule / converted rule.

Green clears 75%, ochre sits between 60% and 75%, red is below 60%. $2,500 is one person per about 5 projects (first pilots); $600 is one person per about 20 projects. estimate on judgement people cost
People cost a project-month	A $1,500 / $1,681	B $4,000 / $4,482	C $8,000 / $8,965	D $15,000 / $16,809	Pilot $4,000

$2,500 first pilots$1,250$600 productisedDashed line: 75%

Reading the table

The deeper design changes these by at most a point at one event a month, and by 3 to 8 points at 15 events a month (band B at $600: 80% / 82% instead of 84% / 85%). The AI choice moves margin per event by about $9; people cost moves it by up to $1,900.
The pilot stage is where the company will be through YC and StartX: read the $2,500 row. 85% at band B is a claim for when onboarding is productised.
What 75% to 80% needs: C and D get there at every people cost modelled except C at $2,500; B and the pilot need people cost at or under about $600 to $1,100 a project-month; A needs it under about $250 to $400, below anything modelled. No AI lever closes band A's gap.

What others report

Procore, the closest public comparable: 80% gross margin (84% adjusted) in 2025. high confidence
AI products averaged 45% in 2025, 53% projected for 2026 (ICONIQ). high confidence
Business software median about 75% to 80%. medium confidence
Seed investors accept about 60%+ now with a credible path to 75% to 80%; what loses them is margin that falls as usage rises. judgement

The route to software margins, largest saving first

Starting from the worst allowed case: the deeper design on Opus high for all 15 events a project-month, people $2,500, about $2,690 a project-month.

Likely case. estimate Productised onboarding (a code-mapping wizard, rule templates, self-serve P6 round trips) is the margin plan. The blind agent: one solutions person per 5 projects gives about 65% at its own $7,500 band B (about 42% at our band B); one per 20 projects, above 90%.
Step	Saves a project-month
People cost $2,500 to $1,250	$1,250
People cost $1,250 to $600	$650
Send only 3 of 15 events to the deeper design	$110
Medium effort and cheaper models where the test set allows	$14
Sonnet workers	$10
Warm caching	$2

"AI gets cheaper", honestly

Plan on the price per token of the tier we use staying flat: top-tier list prices have been flat to rising since 2025, a dearer tier appeared above it, and the newer way of counting tokens adds about 30%. What falls, about 3 times a year planned (Epoch AI measures 5 to 10 times), is the cost of a fixed quality bar, and only if each step is re-tested and moved to a cheaper model. Never price on the decline; treat it as margin.

Unmeasured, and not levers

Also unmeasured and able to move small-band margin: hosting while there are only one or two customers (the whole platform floor, about $100 a month, falls on the first one); routine AI spread (chat and documents, central $6.50, about $30 at the 90th percentile); payment fees and test-set runs (not modelled).
Not a margin lever: distilling a cheaper model from customers' approved answers, or pooling customer rules. Both need written customer permission under the rulebook. checked

Pricing: the unit and the numbers

Per project, banded by contract size, unlimited users. The numbers are the founders' draft; buyers have not seen them.

Munim's options, judged

Option	Verdict
Per company	Later, as the contract wrapper once a customer runs 2 to 3+ projects (a company agreement on yearly data-centre volume, Procore's model)
Per country	No: no precedent in construction software; country sets currency and price list only
Per project per country	Keep "per project"; country only in the price list
Per user per project per country	No: the recovery round needs area leads and subcontractors; Procore says 60%+ of its users pay nothing; PlanGrid's founders say per-seat pricing was one reason they sold
Per project, banded by size, unlimited users	Yes, now (score 28 of 35)
Platform fee plus per project	Later, once there is a platform (27 of 35)

One price table: founders' draft, untested

Monthly, per project. The pilot is 90 days, credited to year one, never free; the draft already uses $12,000 in both US options. founders' decision
	Contract value	Euro draft	US option 1: same numbers	US option 2: converted (1.12)
Pilot	any	€12,000	$12,000	$12,000
A	under 50M	€1,500	$1,500	about $1,700
B	50M to 250M	€4,000	$4,000	about $4,500
C	250M to 1B	€8,000	$8,000	about $9,000
D	over 1B	from €15,000	from $15,000	from about $16,800

Founders' decisions before any form quotes a price

US rule chosen (Munim, 11 Oct): same numbers in dollars.
Whether band A exists in year one: it loses money at pilot-stage people cost; the blind agent's floor is $3,000.
Whether the pilot is a discount on the core market: a $12,000 pilot is band B's price, while most live projects are band C or D.
Under option 1 a $1B job sits exactly on the C/D boundary; under option 2 it is band C, about $9,000.

Usage, and what the product replaces

Plan on about 1 big delay event per project per month (the market model used 10 a year; Munim's relayed figure is 1 to 2). Any per-band allowance (4 / 8 / 15 / 30 events was suggested) is a fair-use ceiling, not expected use. At 1 event a month, band B is about $4,000 to $4,500 per event.

The product does not replace the formal time impact analysis (which may still have to be done by the contract's method); what it replaces is the planner time spent finding and costing recovery options, which no source measures. The only cost on file is an in-house formal analysis at $8,000 to $25,000, context rather than the saving. speculative

The contractor-side anchor

The fee directs recovery money the project spends anyway: in Janika's experience data-centre projects "are often willing to throw a lot of money at recovering a delay", and a contractor's site overheads alone are about $41K to $82K a day, so a month of band B is about a tenth of one day's overheads, or less. The amounts need buyers' own figures. speculative

Value, on the right party: the owner's loss (STL $14.2M a month on a 60 MW site, a modelled scenario; CBRE asking rents of $6,000 to $9,000 per MW per day, retail rents for 250 to 500 kW, wholesale lower) explains why owners push urgency down to contractors. The contractor's own reasons to pay are its damages exposure (rate and cap: an interview question), its site overheads, and keeping a client that presses for days. No consultancy publishes a fee for a time impact analysis: the biggest gap in the value case.

The comparison buyers will make

SmartPM reports delays and runs what-ifs and time impact analysis for about $417 a project-month. The one-line answer:

"SmartPM tells you what the delay did; we find which links can really move and cost each way out."

Test willingness to pay on that layer alone.

Who signs

In Janika's experience: "sometimes project director, sometimes (if we're lucky) client asks for it", and sometimes a company uses the same software on all its projects. On how a company standard is set she says "Im not sure. Likely good practice on one project", then senior directors and head office.

A band C or D fee (about $96,000 to $180,000+ a year) is likely above a project director's own limit; the pilot may fit inside it. speculative

Buyer questions, in this order

Who approves how much, from which budget line?
How many delay events serious enough for a time impact analysis does a project have a month?
What did a consultant charge for one time impact analysis?
What is the delay-damages rate and cap, and how much of the owner's loss reaches the contractor?
Would they reduce planner or consultant spend?
Contract value or megawatts to set the band?
Does the client contract allow an AI tool on the schedule?

The close: "Would you run a paid 90-day pilot at $12,000, credited to year one?"

Market size, bottom up

TAM is live projects x price per project-month x 12, so the price is the market. Data centres alone are low hundreds of millions at most.

The price is the driver

First product (delay recovery on P6), 586 projects	Price a project-month	TAM a year
At SmartPM's price	$417	$2.9M
At Sitedrive's price	about $700	$4.9M
At the founders' bands, weighted by project mix	$7,500	$53M

TAM is everything that could be spent on this a year; SAM is the part we can reach in 3 to 5 years; the last column is revenue we could realistically earn in 2031.

Worldwide outside China, a year. About 586 live data-centre projects with a P6-scale schedule (320 to 921), about 375 in the US; about 306 reachable after geography, P6 share and an assumed 80% that allow a cloud AI tool (no source; the first thing to ask owners). The 40% what-if has no source either; the market model's low case is 70%. TAM and SAM: estimate 2031: speculative
Scope	TAM today	SAM	Revenue 2031
1. Delay recovery on P6 (first product)	$53M (8M to 166M)	$28M ($14M if only 40% of projects allow a cloud AI tool)	$4.3M about 48 projects
2. Plus replacing P6 on data-centre projects	$84M	$44M	$4.6M
3. End to end (plus commissioning and stakeholder reports)	$127M	$66M	$4.3M
3 plus adjacent megaprojects (fabs, batteries, power, nuclear, LNG, pharma)	$338M (45M to 1.04B)	not sized	little before 2031

TAMSAMRevenue 2031Bottom row follows the calculator

Blind cross-check: first-product TAM about $125M ($54M to $252M; about $90,000 a year x 1.3 accounts x 1,070 projects), SAM about $49M, revenue $2.5M to $5M a year in years 3 to 5. The 2.4 times gap with our $53M is in the units (1,070 projects x 1.3 accounts against 586 x 1.0), not the price; both land in the low hundreds of millions at most.

Inputs, and how sure they are

US private data-centre construction $84.95B a year, August 2026, up 73.2% on a year (US Census, release 1 Oct 2026). checked
ConstructConnect: $84.1B of US data-centre starts January to July 2026 over 136 starts, and an average of $1.06 billion among projects with a known price and size (the source's words: "the average project cost $1.06 billion"). checked
The $618M used in the market model is our own division of $84.1B by 136 starts, which includes starts with no stated value. estimate At $1.06B the dollars route gives about 160 US projects instead of 275 and does not agree with the starts route (about 440); more projects fall in band D (a 10 / 25 / 35 / 30 mix gives about $9,500 a project-month against the $7,500 used) and spend per project rises. The count and the TAM are less certain than the tables show, and the direction of the net change is unknown.
Cushman & Wakefield: 37.7 GW under construction in the Americas, 3.8 GW in Europe, Middle East and Africa, 4.8 GW in Asia Pacific outside China. checked

Revenue ends with each project

Projects run 18 to 30 months. An illustrative path to about the 2031 figure: two-year projects at $7,500 a month, no churn assumed, wins per year unsourced. speculative, our arithmetic

Year end	New projects won that year	Live projects	Year-end run rate (not revenue in the year)
2027	3 (paid pilots at $4,000 a month)	3	$140K
2028	8	11	$1M
2029	14	22	$2.0M
2030	22	36	$3.2M
2031	30	52	$4.7M

internal: not for applications This ends at 52 live projects and $4.7M, against the market model's 48 and $4.3M; revenue across 2031 would be about $4.0M (about 44 live on average). The high case, if prices, reach and win rate all land high, is about 260 projects and $47M in 2031. The test of the path is wins per year and the share of customers who start a second project (the company agreement).

Spend the product touches

About $2.27M per project-year: planners $880K, commissioning tracking and documentation $927K, consultant delay analyses $200K, reporting $220K; software under 3%. speculative on the dollar figure

It belongs to four payers; the commissioning documentation is mostly the owner's commissioning agent's. Never quote the total across all projects as a market.

Buyers are concentrated

About 15 US general contractors build most of the volume (HITT, Holder, DPR, Clayco, Turner, Whiting-Turner, Fortis, JE Dunn, Mortenson and others). Good for sales, small for the size story. medium confidence

Downturn: 75 projects worth $130B were blocked or delayed in Q1 2026 and 14 states have pause bills. checked (via Bain) The market model's low case puts the 2031 market 30% below today and about 46% below its central case.

First proofs on the ladder (they show pull, not size)

Price: paid pilots convert at their own project's band.
Second vertical: a contractor that builds both data centres and fabs asks for the product on a fab.
Owner budget: an owner pays for the product, or requires its contractors to use it ("client asks for it").

The size test is separate: a project-database count of the adjacent verticals. We count 975 adjacent P6-scale projects; the blind agent's 20,000 to 40,000 "large capital projects" is unsourced. The size of the expansion market is the least-supported number here; firm it with a project-database count before quoting it.

What would change the answer most: the price per project-month (moves every scope 2 to 7 times); the US live-project count (200 to 600 in the model; the dollars route at $1.06B gives about 160); the share whose owners allow a cloud AI tool, especially on hyperscale jobs, where bands C and D sit and data access is hardest.

The path from first product to platform

Four rungs, schedule of record last. Each needs proof before investors believe the next.

The order is speculative, built on sourced facts; the blind agent agrees. Commissioning means testing every system, from single pieces of equipment to the whole building under load, before handover. speculative as an order
Stage	Adds	Buyer	Revenue per project-year	Proof a VC wants before believing the next
1. Delay recovery on top of P6 (now)	Link meanings, costed recovery options, a delay impact statement, the file back to P6	Contractor project director; planners champion it	the band: about $18K to $200K	Paid pilots; options approved and imported back; time from delay to issued statement
2. Commissioning readiness tied to the schedule	Reads commissioning status and equipment trackers; forecasts which tests will be ready; models witnesses, makers' engineers, load banks	Still the contractor (commissioning and MEP managers); owner's commissioning lead second	unresolved: extra price, or only a reason to stay	Weekly unprompted use; fewer idle test days; one commissioning tool connected
3. Reports per stakeholder	One model of the job, a report per reader (owner, contractor, commissioning agent, lender's adviser, trades)	Shifts to the owner or developer	$150K to $500K or an owner programme fee	An owner accepts a report as a contract submittal
4. The schedule of record (replacing P6)	Native planning with link meanings built in; exports P6 files for owners who still require them	Owner sets the specification	Platform fee on construction volume	An owner writes it into its schedule specification

Data across companies

Stage 2 needs only the contractor's own uploads of commissioning status. Stage 3 works today as exports the contractor sends its client (the delay impact statement already is one), and Janika has named a later client role that reads statements. A shared view that several companies sign in to is not in the rulebook today (every record belongs to one company) and would be a founders' design change. checked A report built from the contractor's delay statement is a claim document, so the owner may not treat it as neutral.

Commissioning: where the money and pain sit

Practitioners put fit-out and commissioning at about 70% of a data-centre job; commissioning fees about 0.75% to 3% of construction cost ($8M to $30M on a $1.06B project). not yet checked
Fenwick Elliott and FTI name a missing logic-linked commissioning plan as a frequent cause of data-centre delay. high confidence
But commissioning software is cheap (CxPlanner from $430 a month per company checked), its buyer is usually the owner's commissioning agent, and a contractor scheduler says the contractor has "no say" in commissioning dates. medium confidence

Keep delay recovery first and build commissioning into it (the commissioning manager as an area lead; witnesses, makers' engineers, load banks as constraints). Flip only if 20 to 30 pilot conversations say "will this test be ready next week" is what contractors pay for first.

Replacing P6: last and hardest, 5 to 10 years

Owners name P6 (or a format) in their schedule specifications; claims run on the accepted P6 file; Outbuild built P6 export because owners require it. Route: author in our tool, export to P6, then win an owner that writes its own specification (a neocloud or hyperscaler). speculative

The blind agent's extra rung, worth keeping: make the product a weekly habit (schedule health on every update) before commissioning, because about one use a month invites churn. It adds usage nobody has costed yet.

Everyone is converging

Procore Scheduling, Feb 2026, imports P6 files
Planera, $8M Oct 2025, data-centre scheduling "through commissioning"
Foresight, $25M Series A Mar 2026, owner side
Facility Grid, commissioning plus P6 sync
Nodes & Links ($12M), nPlan ($16M), ALICE, SmartPM

Nobody sells the whole chain for data centres; every link has a funded owner.

Go-to-market, money and milestones

One project at a time, sold to the project team, made standard by head office; the first round buys a demo, interviews, a timed test and paid pilots.

First sale and first accounts

First sale: one project, bought by its project director, or because the client asked for it; then, in Janika's experience (she says "Im not sure"), the project team shows its directors and head office makes it standard. The pitch must give a project team something to show its directors: days recovered on key dates first, then cost.

First accounts: the 15 or so US general contractors above, plus the UK and European data-centre contractors Janika knows; a corporate fund tied to a contractor (Suffolk, Zachry) can bring a first customer.

Gates on every project

The client contract must allow an AI tool on its schedule.
A security review.
SOC 2 tooling and audit at about $15,000 to $40,000 a year once enterprise customers ask. judgement
Sales cycle and approval path: unknown, interview question one.

Liability: when an approved option proves physically wrong, a person approved it, but the contract terms and insurance need counsel before the first paid pilot.

Why now

US data-centre construction up 73% in a year (US Census). checked
Long equipment lead times make late deliveries routine: transformers 68 to 113 weeks (Cushman & Wakefield 2026). medium confidence
Models are now cheap enough to read every delay while code checks every date they touch.

What the first round must buy, in order

The demo on a realistic schedule (due 28 Oct 2026).
15 to 25 planner and project-director interviews with numbers.
A timed head-to-head against planners.
Three to five paid pilots.
Measured token cost and usage; SOC 2 started; the product name.

The amount and runway need the founders' monthly costs, which are not in this research.

Programme terms on file

a16z speedrun: $500K for 10% plus $500K into the next round within 18 months, in person January to April 2027, overlapping YC. high confidence: speedrun FAQ and apply page The blind agent calls 10% for $500K expensive next to an ordinary pre-seed.
YC's standard deal: $500,000, as $125,000 on a post-money SAFE for 7% plus $375,000 on an uncapped SAFE with most-favoured-nation terms. checked: ycombinator.com/deal, 11 Oct 2026

Pitch per funder

One story everywhere: a narrow first product with a visible ladder. The team, the demo and buyer evidence decide.

Target	Lead with	Biggest objection, and the answer
YC W27 2 Nov, 8pm Pacific	Janika's construction-planning career and the problem in her words; Munim builds it, with proof (demo, coding-agent session). YC's Fall 2026 requests include "New Operating Systems for the Physical World", which names scheduling teams in construction. checked	"How big can this get?": the honest size (data centres alone are low hundreds of millions at most; the next verticals named but not yet sized), then the first proofs on the ladder; never the total labour spend. "Narrow, integration-dependent start" (a YC partner's stated dislike): starting on top of P6 is deliberate, because owners require P6 files and claims run on them; native planning follows once an owner writes it into a specification. "Fourth idea; will they stick, can they build?": say the pivot first in one line; demo plus coding session.
Sequoia Arc no 2026 call seen	The problem as a hard fact with commercial pit stops	"A tool for one task in a niche": the ladder with the proof each rung needs, and the second vertical.
a16z speedrun SR008 priority window 12 Oct to 1 Nov	Janika's earned secret in sentence one; short, founder first	"Why you, and what have you shipped?" Demo, interview count.
StartX closes Sun 11 Oct, 11:59 PM Pacific	Founder quality, the Stanford tie, progress since the last application (it is equity-free and does not buy pilots)	"Why will this attempt stick?": both full time "from the first day of the session" (the draft's own wording), a co-founder from the field. Before submitting, two lines of the draft conflict with this plan (the owner's loss as the contractor's damages; "under 1%" holds only for the lean design).
Pilot-buying accelerators Suffolk BOOST, Maria 01, Techstars	The first product, one line of ladder, the pilot you want with them	"What will you pilot with us in 8 to 12 weeks?" Name the delays, the team, two measures.
Construction-tech VCs	Works inside P6; nobody changes how they use it	"SmartPM, ALICE, Foresight, Nodes & Links or nPlan can add this": name each and the one-line difference; the reasons we win, never "compounding". Expect conflict passes from funds already in Planera, nPlan or SmartPM.
Grants NSF SBIR, Business Finland, Innovate UK; EIC later	The first product told as a research problem: can software infer whether each dependency is physical, cost-dependent or preference?	Eligibility conditions are quoted word for word in the research and are for counsel; pitch NSF only after counsel's read; EIC waits for pilots.

Before reusing the drafted pitch paragraphs

Use the ladder order above (schedule of record last, not second).
Replace the "monday.com, Autodesk and spreadsheets" line with the commissioning premise correction.
Strike "We built the tool": no product code exists yet; the scheduler holds the specification, its checks and research; the demo is due 28 Oct.
Keep "each customer's own confirmed rules, which stay theirs".

The way out of "too small or too big"

Small in scope, judged on project days. The first product does one job (recovering from a reported delay) for a buyer whose client presses for days (in Janika's experience, data-centre projects "are often willing to throw a lot of money at recovering a delay"); the ladder shows where it goes. Name P6 once, as the end state the ladder earns, with no date on the later rungs.

A sharper YC variant, for the founders to decide: sell the first product as a priced service delivered with software (the recovery plan and delay statement done for the contractor), as Marengo (YC S26) sells data-centre engineering. It matches YC's recent "service company" requests and earns revenue sooner, at the risk of services pulling margin down.

Doubts to answer with evidence

Can a physician-MBA build it? Never imply Munim wrote Clea's code; show this product and the coding session.
Serial pivots. Say it first; Janika's construction career is the constant.
Depth of data-centre experience. Exact dates and tense on every form, since both founders go full time from the day of applying; add one or two named data-centre planning advisers.
Number hygiene. Leave out every unsourced figure: interview counts of 50, 70 or 100; the unsourced per-day delay cost in the founder-video notes; "over GBP 4 billion"; the product-speed half of "a week to a working day" until timed. Janika's own "multiple days" is attributable.

Founders' to-do: our files hold two versions of how the Clea YC place ended (the founders declined it after a co-founder left, and a later note calling it a rescission); agree one plain, accurate version before any form mentions the acceptance.

Traction before 2 November, best value first

A coding-agent session upload or builder report, if the W27 form offers it.
15 to 25 structured planner interviews, led by Janika, with numbers.
A working demo on a realistic anonymised schedule.
A timed head-to-head against planners.
Two or three design-partner letters.

A paid pilot is the biggest signal but unlikely to close by 2 November; start it for the interview stage (decisions by 11 December).

Risks investors will cite

Seven, in the order they will come up.

1Team

Full-time status, IP ownership, the technical co-founder's depth. The blind agent raises the co-founder's current employment as an IP question; that is Munim's to handle, recorded here only as an investor question.

2Buyer access and willingness to pay

Hyperscale owners restrict schedule data; about one use a month is thin; contractor sales are slow and pass security review; the client contract must allow AI on every project.

3Crowded and small

Planera, Foresight, Nodes & Links, nPlan, ALICE, SmartPM, Procore Scheduling; data centres alone are low hundreds of millions at most.

internal: not for applications Platforms buy point tools (Trimble bought Document Crunch for $246.4M), so a focused product's likeliest exit is acquisition.

4Accuracy on real 9,000-activity schedules

One impossible option, or dates that do not match P6, ends a pilot.

5Revenue ends with each project

And data-centre starts can stall (see the downturn under market size).

6One model provider

Fable is excluded by its retention terms and top-tier prices are flat; a price or terms change hits every project.

7Liability

For an approved option that proves wrong (see go-to-market).

Founders' decisions this plan surfaces

Six calls only the founders can make; the first is due tonight.

StartX, tonight (11:59 PM Pacific). The draft still says the owner's loss "returns to the contractor as damages" and "drives the contractor's delay damages"; it says "under 1%" for AI and hosting (that holds at about one big delay a month on both designs the rulebook allows, about $19 to $28 on a $9,000 fee, and reaches about 2% only if all 15 delays in a stress month ran the deeper design); it says "about $9,000 a month on a $1B job" (under the chosen rule, $8,000 up to $1B); and it uses the name being dropped 15 times.
Band A in year one, and whether the pilot is a stated discount. The US price rule is decided: same numbers in dollars (Munim, 11 Oct).
Whether to offer a deeper AI tier at all, which needs higher caps (about $9 and $10) and possibly a reading of the rulebook's line on AI-suggested changes, or stay lean-only through the pilots; and whether a month is bounded by an allowance with paid overage or a hard block.
The product name, before the YC form.
One accurate tense for Janika's role on every form.
The service variant for YC.

Where the agents disagree

Where our research and the blind agent (or another source) land apart, and how we read it.

Point	Ours	Blind agent or other source	Reading
Deeper run cost per delay event	Allowed design: $11 Opus, $8 Sonnet workers; tool-using design $62	Blind tool-using design: about $105 (three passes of $35)	The tool-using kind is not allowed; the allowed kind is a fifth of the cost
Delay events per project-month	Market model: about 0.8; Munim: 1 to 2	Pricing research allowance 4 to 30; cost study stress 15; rulebook placeholder 50	Plan on about 1; the rest are ceilings and stress cases; interview question
Year-one people cost	Cost study: $600	Blind: about $2,500 early (one person per 5 projects)	Neither researched; show the range; measure solutions hours per project
Pilot price	$12,000 (the founders' draft, reviewed)	$15,000 to $30,000	Interview question
Lowest band	€1,500 under €50M	about $3,000 under $250M	Founders' decision with buyers' answers
First-product TAM	$53M (586 projects x 1.0 account)	about $125M (1,070 x 1.3)	2.4 times apart on units; both low hundreds of millions at most
Wider expansion market	$338M (975 named projects)	$1.2B to $2.4B (20,000 to 40,000, unsourced)	Unresolved; least-supported number
Ladder order	Platform research and blind agent: schedule of record last	Drafted pitch paragraphs: second	The platform research is better grounded; fix the paragraphs before reuse
Commissioning revenue	Market model and platform research: adds price per project	Pricing research: a reason to buy, not a revenue line	Unresolved; interview question
Average US project value	ConstructConnect's stated $1.06B	Market model's $618M (our division)	Use $1.06B; the model's count, band mix and spend per project rest on $618M, so they carry more doubt than shown
Fable for hard cases	Excluded (retention)	Blind agent suggested it	Excluded
Batch processing for tagging	Ruled out for customer data	Blind agent suggested it	The rulebook wins
speedrun timing	SR008 priority window 12 Oct to 1 Nov	Blind agent read the SR007 page	Ours is current

What we do not know yet

No buyer has been asked anything and no product code exists yet; most open numbers on this page wait on both.

No buyer interview yet: willingness to pay, who signs at what limit, consultant fee per time impact analysis, delay events a month, damages rate and cap, AI acceptance by owners.
No product code yet: specification, checks and research only; demo due 28 Oct.
No measured token counts on a real schedule.
Whether the deeper design gives better answers than lean at all.
People cost to serve per project.
No hyperscaler schedule specification read.
Paid project databases not consulted.
Competitors' revenue unpublished (estimates $2M to $20M).
The founders' monthly costs and runway.