“The model being good is a precondition, not the business case. A digitiser can hit the published accuracy and still lose money on every seat if the serving cost was never worked out. The business case is a small number of divisions, and each one has to survive contact with the real figures.
”
The two questions
Good research and a good business are not the same question
Why this document is about arithmetic, not architecture
We built a curve-digitisation model for subsurface teams and published its architecture and its accuracy in a peer-reviewed venue [1]. That document answers one question: does the model work. This document answers a different one that the first cannot touch: does the thing built around the model make money per seat, and is the plan to grow it honest. These are separate questions, and a team can pass the first and fail the second without noticing, because the accuracy figure feels like a business result and is not.
The reason to keep them apart is that they fail for unrelated reasons. A model fails on a benchmark. A business fails on a division: the price a customer pays turns out to be smaller than the cost of serving them, or the revenue target on the slide turns out to require a share of the market nobody has ever reached, or the money raised turns out to buy fewer months than the build takes. None of those failures shows up in an F1 score. All of them show up the moment you write the arithmetic down and refuse to round in your own favour.
So this is a product-and-go-to-market whitepaper, written from the position of having built the digitiser and then having to price it, staff it, and raise against it. It is grounded in one reference programme, anonymised throughout: a digitisation effort for a US energy-technology client, whose project archive supplies the serving-compute rents, the market funnel, the runway and the two build tracks we will read. The archive also records a seat price, revenue targets, a seed amount and build costs; we keep those confidential and work in shares, indices, months and weeks instead, which is enough to carry every division. Every number in the argument traces to that archive or to the published sources in the references, and where an input is a modelling assumption rather than a sourced figure we say so on the instrument and in the prose.
The business case, as four sourced numbers
GPU rent, EUR per month, the serving-cost source
Serviceable slice of the total market, and the share of it the plan targets by year five
Runway the seed was sized to fund
Head start the accelerated build track buys
The rest of the document takes those numbers and works the divisions through in relative terms. First, the margin per seat, which is the question of whether the business works at all. Then the seed as a burn ceiling, which is the question of whether the build was sized to the money. Then the delivery-track fork, which is the question of whether it is worth paying to arrive sooner. The market funnel sits underneath all three, because it is what decides whether the seat count the divisions assume is a plan or a wish.
The unit that repeats
A subscription business is judged on the unit that repeats, not on the total it hopes to reach [3]. For a per-seat SaaS the unit is one seat: the annual revenue it returns and the direct cost of serving it for that year. If that unit is healthy, growth compounds it; if the unit loses money, growth multiplies the loss, which is the specific way a business with good-looking topline and bad unit economics dies. So the first thing to establish is not the market size or the revenue projection. It is whether one seat, served for one year, contributes more than it costs.
That is a favourable comparison for a curve-digitiser, for a structural reason worth stating before the numbers. The expensive part of a machine-learning product is building and training the model, and that cost is paid once. Serving is cheap and gets cheaper per unit as volume rises, because one trained model serves every log rather than being rebuilt for each one. The consequence is that the marginal cost of a curve-digitiser is dominated by serving compute, which is small, and the fixed cost is a one-off the seat price does not have to carry per unit. The margin question, then, is whether the seat price clears the serving cost of the logs that seat runs. It does, and by a wide margin, which is the first instrument.
The margin
A seat pays far more than it costs to serve
What one seat actually costs to serve
The serving cost of a seat is built from the ground up. The model serves on a rented GPU. The project archive records two rentable GPU tiers, 750 EUR per month for a high-end card and 1,800 EUR per month for an advanced one. That monthly rent, divided into an hourly rate, is the numerator of the per-log serving cost. The denominator is throughput: the number of logs the served model clears in an hour. Divide the hourly rent by the logs cleared that hour and you have the compute cost of digitising one more log, and it is small, because a rented GPU hour is cheap next to what a seat pays and a served model clears many logs in that hour.
The seat's yearly serving cost is that per-log cost times the number of logs one user runs in a year. This is where the structure of the product does the work. Because a single trained model serves every scan width, the per-log cost does not rise with the seat's usage, so the seat's yearly serving cost grows linearly in usage while the revenue it returns stays fixed. Even a heavy user, running thousands of logs a year, imposes a serving cost well below what the seat pays. The gap between the two is the gross contribution margin, and it is the whole business.
The three inputs we can vary honestly are the GPU tier, the seat's annual usage, and, buried in the derivation, the serving throughput. The tier and usage are levers on the instrument; the throughput and the usage band are modelling inputs flagged as illustrative, because the archive gives us the tier rents but not a measured serving throughput at scale. To keep the seat price off the page, the ledger indexes one seat's annual revenue to 100 and reads every cost as a share of it, which is all the margin question needs. What the instrument argues does not depend on the illustrative inputs being exact. It depends on the shape: a fixed revenue column that a growing cost column cannot come close to reaching at any realistic usage.
The ledger reads the margin directly. The teal column is the revenue a seat returns each year, fixed at an index of 100. The dim column is the serving cost that seat imposes on the same scale, growing as you drag its yearly usage up. The orange bracket between them is the only element that argues, and it is the contribution margin: at any realistic serving throughput it stays wide, because the serving cost of even a heavy seat stays well below what the seat pays. Toggle to the advanced GPU tier and the cost column rises, because the rent is 2.4 times higher, and at typical usage the margin moves only a little, because the rent was small relative to the seat's revenue to begin with. That is the thesis of the whole whitepaper stated as one picture: a vertical curve-digitisation SaaS is attractive precisely because the price sits far above the marginal serving cost per log, and it stays attractive as usage grows because one model serves everything.
The posture on margin
Price the seat against the marginal serving cost, not against the cost of building the model. The build is a one-off that the seat price does not carry per unit; the serving cost is what recurs, and for a curve-digitiser it is small and roughly flat in usage. If a vertical SaaS cannot show a wide gap between seat price and marginal serving cost, no amount of topline growth will fix it, because growth multiplies the unit.
The margin is necessary, not sufficient
A wide per-seat margin is the entry ticket, not the whole game. It says the unit is healthy, which means growth will compound rather than dilute it. It says nothing about whether you can acquire the seats, whether the money you raised sizes the build, or whether the revenue projection you are selling against is reachable. Those are the next three questions, and the first of them is the one a wide margin can seduce a team into skipping: how big is the reachable market, really, and is the revenue target a share of it that anyone has ever obtained.
The funnel that keeps the target honest
The reference deck sizes the market in the conventional three steps. The total addressable market covers the transactions the product touches. The serviceable slice, the oil-and-gas technology market the product can actually address, is about 5% of that total. The obtainable share the plan commits to is 3% of that serviceable slice by end of year five, which becomes the year-five revenue projection, with a first-year target set for the twelve months after the seed. The walk-down deserves its own treatment; the point that matters for the margin argument is narrower.
It is this. The year-five revenue projection is not a number to lead with; it is a number to check against the seat price. Divide each revenue target by what one seat pays and you have the real targets, expressed in the unit the business actually sells. When we did that division on the reference deck, year five needed more than an order of magnitude more seats than year one. A revenue projection stated in currency can float free of reality; the same projection stated in seats has to answer for whether a sales organisation can close the year-one seat count in a year and multiply it many times over by year five. Keeping the target in seats, and demanding that the funnel from market size down to obtainable share be a share someone has plausibly reached, is what stops the projection from becoming a spreadsheet fantasy. The wide margin makes each seat worth selling; the honest funnel makes the seat count a plan rather than a wish.
The money
A seed is a burn ceiling, not a lump sum
Eighteen months of runway is a spending rate
The seed round was sized for an 18-month runway. The mistake a team makes with a round like that is to read it as a pile of money to spend as opportunities arise. It is not. A seed raised for a fixed runway is a fixed maximum monthly burn, because the capital divided by the runway is the most you can spend each month and still reach the end of the runway with the company alive. Over 18 months that is one eighteenth of the round each month, about 5.6%, and that rate, not the size of the round, is the constraint every hiring and build decision has to fit inside.
This reframing matters because it converts a vague sense of having money into a hard ceiling that the build plan must respect. If the plan implies a burn above the ceiling, the runway ends before the milestones the round was raised to reach, and the company is raising its next round from a position of weakness or not raising it at all. If the plan sits below the ceiling, there is slack to absorb the things that always go wrong. The seed does not fund a wish list; it funds a rate, and the rate is what the second instrument makes visible.
The glidepath draws the capital, indexed to 100% of the round, falling to zero as burn times months eats it. The teal line is the share of the round still on hand; the teal dashed line is the 18-month target the round was sized to reach; the orange marker is the only element that argues, and it is the month the cash runs out. Drag the burn below the ceiling of about 5.6% of the round a month and the line reaches month 18 with cash to spare. Drag it to the ceiling exactly and the line lands on month 18, which is what the round was sized for. Drag it above and the orange marker slides left of the target, and you can watch the round fail to buy the runway it promised. The instrument argues one thing: the seed sets a burn ceiling, and the ceiling is what keeps the revenue plan honest, because a plan that requires spending above the ceiling is a plan that runs out of money before it can be tested.
Why the ceiling and the margin talk to each other
The burn ceiling and the per-seat margin are not separate facts; they constrain each other. The wide margin means every seat sold reduces the net burn, because the revenue a seat returns is almost all contribution. So the burn ceiling is not a fixed sentence; it softens as seats come on. But that only helps if the seats arrive inside the runway, which is why the timing of first revenue is the next question, and why paying to ship sooner is a real lever on survival rather than a vanity. A team that ships late burns through the runway with no offsetting revenue; a team that ships early starts converting the wide margin into cash while there is still runway to protect. That is the bridge from the money question to the delivery question.
“A seed does not buy a pile of money to spend as you like. It buys eighteen months at a fixed monthly rate, and the only way to loosen that rate is to start selling seats before the runway runs down.”
— Paraphrased from our own planning notes
The timing
Paying to ship sooner buys weeks of revenue
Two ways to build the same product
The project archive records two build tracks for the product. The accelerated track staffs 6 FTE and finishes in 16 weeks. The standard track staffs 4 FTE and finishes in 32 weeks. The accelerated track costs more, a premium for compressing the schedule, and ships 16 weeks sooner. The naive reading is that the standard track is cheaper and therefore the safer choice for a company watching a burn ceiling. The correct reading is that the two tracks are not being compared on cost; they are being compared on when the revenue starts, and 16 weeks of earlier seat revenue can be worth far more than the premium.
The reasoning is the payback question. Both tracks build the same product, so both earn the same revenue per seat once they ship. The only difference is the week the earning begins. The accelerated track starts earning 16 weeks before the standard one, and during those 16 weeks it accumulates revenue the standard track earns nothing against. The premium is repaid at the moment that head-start revenue clears it, and whether that moment arrives depends only on how much revenue the live seats bring in each week. Enough live seats and the premium repays itself before the standard track has even shipped; too few, and in this simple model it is never repaid, because once both tracks are selling they earn at the same rate and the gap stops closing.
The fork plots each track's cumulative net cash from its own finish week, in multiples of the premium and measured after the standard track's build cost, so the standard track starts at zero and the accelerated track starts one premium below it. The teal line is the accelerated track, climbing from week 16. The dim dashed line is the standard track, climbing only from week 32. The orange marker is the only element that argues, and it sits at the week the accelerated track's head-start revenue has repaid its premium and the two lines cross. The lever sets what one live week of seat revenue earns as a share of the premium, which rises with the number of live seats. Drag it up and the marker slides left, because each week of the head start is worth more and the premium repays faster. The threshold is simple: if a live week earns at least a sixteenth of the premium, the 16-week head start repays it before the standard track ships, and the accelerated track stays ahead from then on. Below the threshold the marker becomes a permanent shortfall. That is the argument in its honest form: paying to arrive sooner is rational exactly when the revenue the head start unlocks exceeds the premium, and a wide-margin product with live demand makes that easy to clear.
When the standard track is still the right call
The fork does not say always pay the premium. It says the premium is worth paying when the head-start revenue clears it inside a horizon you care about, and that condition can fail. If demand is thin at launch, few seats are live, the head start is worth little per week, and the premium takes a long time to repay, which argues for the cheaper track and its lighter draw on the burn ceiling. If the product is not yet validated and the risk is building the wrong thing fast, the standard track buys time to learn at a lower cost. The instrument is a decision aid, not a verdict: it tells you, at a given level of weekly revenue, whether the head start repays the premium, and it lets you see how much live demand it takes to make speed the right call. The honest answer depends on the demand you actually have, which loops back to the funnel: speed only pays when there are seats to earn against.
The synthesis
The four questions are one system
How the pieces constrain each other
Read separately, the four questions look like a checklist. Read together, they are a single system where each answer sets the terms for the next. The per-seat margin decides whether the unit is worth selling at all; because it is wide, every seat sold is almost pure contribution. That wide margin is what makes the seed's burn ceiling survivable, because seats coming on soften the burn faster than they would for a thin-margin product. The burn ceiling is what makes the timing of first revenue matter, because a runway with no offsetting revenue is a runway spent; and the timing is what makes the delivery-track premium rational, because shipping 16 weeks sooner converts the wide margin into cash while there is still runway to protect. And the market funnel sits underneath all of it, because every one of these arguments assumes seats can be acquired, and the funnel is what turns the revenue target from a currency figure into a defensible seat count.
The failure of any one piece propagates. A thin margin makes the burn ceiling unsurvivable no matter how the build is timed. A dishonest funnel makes the seat count a fantasy, and then the payback fork is repaying a premium against revenue that never arrives. A build sized above the burn ceiling runs out of money before the margin can prove itself. The reason to write all four down as arithmetic, and to refuse to round any of them in your own favour, is that they are load-bearing on each other, and a soft number in one place quietly weakens the argument everywhere.
What this looks like for a subsurface team specifically
Nothing in the argument so far is specific to subsurface work, which is deliberate: the unit-economics reasoning is the same for any vertical SaaS. What is specific is why the margin is so wide in this case, and it comes from the domain. A subsurface team sits on a corpus of raster logs that is expensive to interpret by hand, and having that corpus digitised is valuable enough to be worth paying for on its own terms [2]. That means the value a seat receives is high, which supports a seat price set well above the cost of serving it, while the cost of delivering it is dominated by cheap serving compute, which keeps the marginal cost tiny. The wide margin is not an accident of pricing; it is the domain's expensive manual baseline meeting the product's cheap automated serving, and the gap between them is the business.
The corollary is a warning. The same reasoning that makes the margin wide makes it tempting to skip the harder questions, because a wide margin feels like the business is already won. It is not. A wide margin with an unreachable seat target, or a build that overruns the burn ceiling, or a launch timed so late the runway is gone before revenue starts, still fails. The margin is the reason to do the work, not a substitute for it.
What to carry out of this
- The model being good is a precondition, not the business case. A curve-digitiser can hit its published accuracy and still lose money per seat if the serving cost was never priced. The business case is a small number of divisions, and each has to survive contact with the real figures.
- A vertical SaaS works only when the seat price sits far above the marginal serving cost. With one seat's annual revenue indexed to 100, serving a typical user's logs on rented GPUs costs a few points of it at the ledger's illustrative serving rate, so the contribution margin is wide, and it stays wide as usage grows because one trained model serves every scan.
- The revenue projection has to be read in seats, not currency. Divide each revenue target by what one seat pays; a funnel from the total market through a serviceable slice of about 5 percent to a 3 percent obtainable share only holds if that share is one someone has reached, and if a sales organisation can close the seat count it implies.
- A seed sized for 18 months is a burn ceiling of one eighteenth of the round, about 5.6 percent, per month, not a lump sum. Every build and hiring decision has to fit inside the rate, or the runway ends before the milestones the round funded.
- Paying a premium to ship 16 weeks sooner buys 16 weeks of seat revenue. The premium repays itself when the head-start revenue clears it, which happens before the standard track ships once a live week earns a sixteenth of the premium, so speed to revenue is a survival lever, not a vanity.
Limitations
The per-seat serving cost in the margin ledger is built from the two GPU rental tiers, 750 and 1,800 EUR per month, and read against one seat's annual revenue indexed to 100; the seat price itself is not printed. The serving cost of 1,000 logs as a share of seat revenue folds in an illustrative serving throughput, and the logs-per-seat-per-year usage band is illustrative too, both flagged on the instrument, because the project archive supplies the rents but not a measured serving throughput at scale. The argument the ledger makes, that the margin stays wide at any realistic usage, does not depend on those inputs being exact; it depends on the serving cost being small relative to what a seat pays. Reading a EUR-denominated serving cost as a simple share of seat revenue is a simplification, not an accounting treatment; a precise statement would carry the exchange rate and the fixed company costs the gross margin sits above. The seed glidepath treats burn as constant across the runway, which real burn is not; it reads the round as 100%, the sustainable monthly burn is the exact division of the round by the sourced 18-month runway, and the burn-lever range around it is presentational. The delivery-track fork uses the sourced FTE counts and timelines, draws cash in multiples of the premium, and treats one live week's revenue as a share of the premium, a presentational input; it ignores ramp, churn and any advantage from arriving first, so it argues the direction of the timing trade and the threshold that decides it, not a precise payback date. The market funnel shares are the reference deck's own numbers; whether a 3% obtainable share is reachable is a judgement the arithmetic cannot settle, and the whitepaper's claim is only that the target should be checked in seats, not that the share is assured. Finally, the wide margin is specific to a domain where the manual baseline is expensive and the automated serving is cheap; a vertical where that gap is narrower would not support the same reasoning, and the method here is the reasoning, not the particular numbers.
References
- Nasim, M. Q., Patwardhan, N., Maiti, T., Marrone, S., Singh, T. (2023). VeerNet: Using Deep Neural Networks for Curve Classification and Digitization of Raster Well-Log Images. MDPI Journal of Imaging, 9(7), 136. DOI 10.3390/jimaging9070136. The published architecture whose serving compute sets the marginal cost this whitepaper prices against. https://www.mdpi.com/2313-433X/9/7/136
- Koroteev, D., Tekic, Z. (2021). Artificial intelligence in oil and gas upstream: Trends, challenges, and scenarios for the future. Energy and AI. The context for why a subsurface team has a corpus worth digitising, and why the manual baseline is expensive enough for a digitised corpus to be worth paying for. https://www.sciencedirect.com/science/article/pii/S2666546820300033
- Skok, D. (2013). SaaS Metrics 2.0: A Guide to Measuring and Improving What Matters. For Entrepreneurs. The reference framing for reading a subscription business through the unit that repeats: the cost to acquire and serve a customer against the revenue that customer returns. https://www.forentrepreneurs.com/saas-metrics-2/
Get the full whitepaper
This page is the long-form summary. The PDF edition carries the full argument with the three instruments printed as exhibits, the per-seat margin ledger, the runway glidepath and the delivery-track payback fork, for reading offline or passing to a finance or product lead.


