Most budgets are written as if the world will hold still. The scope is fixed, the line items add up, the total is defended to the second decimal, and everyone signs. Then something moves that no spreadsheet cell anticipated, and the project stops delivering while the two parties reopen the contract to argue about who pays. We wrote our R&D proposal for a multi-year subsurface AI programme with a mid-sized Middle East carbonate operator the other way. It carried an explicit contingency line, and next to it a plain written sentence: the requirements would change. That sentence is the whole point of this post. The reserve on its own is a number; the acknowledgement is what makes the number legitimate to draw down later without a fight.
Eight months into delivery, we got to test whether the phrasing held.
What the proposal said before anything went wrong
An R&D engagement is not a fixed-scope build. You are agreeing to learn something, and learning reroutes the plan. Our commercial model reflected that from the first version. The manday and infrastructure sheets carried a named contingency percentage on top of the priced work, and the founding scope document set the budget at USD 533,000 with the phases sized in months, not fixed deliverables. The proposal text did not hedge about it. It stated, in the client's language and ours, that the technical direction of an R&D project would evolve and that a reserve existed for exactly that reason.
The distinction between padding and contingency lives entirely in how you say it. Padding is an unnamed cushion buried inside every line item, discovered only when someone audits the rates. Contingency is a single visible line with a stated trigger. It reads as prudence because it is legible: the client can see precisely how large the reserve is, what class of event it covers, and that nothing else in the budget is quietly inflated to hide it. We priced the reserve as a percentage of the committed work and put the word "contingency" on it in plain sight.
Nine and a half percent is not a rounding error and it is not a blank cheque. It is a bounded number, agreed in advance, that a client can reason about at signing time instead of during a crisis.
The shock that arrived on schedule for nobody
In 2022 the energy market moved in a way no subsurface AI proposal had modelled. Against pre-war levels, energy prices rose about 400 percent and electricity bills rose 394 percent. For a programme whose core cost is GPU compute in a data centre, that is not a background inconvenience; it is a direct hit on the unit economics. The market for DGX-class hosting moved with it, from roughly USD 12,000 to 15,000 per month before the war to over USD 20,000 per month after. The compute we needed had not changed. The price of running it had.
This is the moment a fixed-scope budget breaks. The work is proceeding, the models are training, and an external cost has stepped outside every assumption in the sheet. With no reserve and no prior acknowledgement, the only move is an emergency conversation: stop, reopen the contract, and negotiate an unscheduled increase while the client wonders whether the original estimate was ever sound.
Absorbing it as a draw, not a crisis
Because the reserve was already named and the acknowledgement already signed, the energy shock had somewhere to go. We raised the contingency against the pre-agreed line rather than against the client's patience. The energy component was USD 51,000, scoped explicitly over 12 months, and it sat inside a larger USD 104,300 additional request against the USD 533,000 approved budget. The rest of that request covered infrastructure the dual supervised-and-unsupervised model track had made necessary, a separate story from the compute-hour side that we treat in the GPU-hour ledger piece and will not re-derive here.
The instrument below is the argument in one frame. Fire the shock and watch where it lands.
The bounded draw is the orange bar riding on the approved-budget track: a 9.6 percent top-up on a line the client had already agreed could move. The counterfactual, the muted path on the right, is the same energy spike hitting a budget with no clause, forcing an unscheduled reopening of the contract with an unbounded ask and delivery stalled while the two sides argue it out. Same shock, two paths. The clause decided which one we walked.
Why the draw held up as prudence
A contingency request lands better when it is not the only thing in the letter. We framed the ask against structural moves that had already insulated the programme, so the reserve read as the last line of a defence rather than the first. We had owned hardware through a multi-year Nvidia partnership locked in 2020, which meant we were not fully exposed to the rental market that had just moved to USD 20K a month. We had a local data-centre arrangement that gave full cost visibility. We had academic rates on the R&D phase. On top of all that, the energy line still moved enough to need the reserve. Presented that way, the USD 51K draw was clearly the residual after every cheaper hedge had already been used, not a convenience.
That is the difference a written acknowledgement makes at draw time. The client was not being asked to accept that the estimate was wrong. They were being asked to honour a mechanism they had already approved for precisely this class of event, an external macro shock rather than internal scope creep. The word "contingency" on the original line, and the sentence next to it about requirements changing, did the persuading months before the invoice.
How to price and phrase it
Three things travel from this to any R&D budget. First, size the reserve as a visible percentage of the committed work and label it. A single-digit-to-low-double-digit percentage, named and bounded, is defensible; an unnamed cushion is not, and gets stripped in the first negotiation. Second, write the acknowledgement into the proposal prose, not just the numbers. State that the requirements will change and that the reserve is what covers it. The sentence is cheap to write and expensive to lack. Third, define the trigger class. Our reserve was scoped to external macro events, which is why an energy shock qualified cleanly and ordinary scope drift would not have. A contingency line with no stated trigger invites the client to treat every draw as an admission of poor planning.
The clause you hope never to use is the one that lets you keep delivering when the thing you could not predict shows up anyway. We used ours once, for a bounded 9.6 percent, and the programme never stopped.
Limitations
The counterfactual "emergency renegotiation" path in the instrument is a labelled contrast, not a measured outcome; we did not run the no-clause world, so its cost is illustrative of the mechanism rather than a figure from the archive. The percentages here are drawn from a single engagement with one client in one market during one specific macro shock, and the right reserve size for another programme depends on its risk profile, contract type, and client relationship. The energy and hosting figures are period-specific to 2022 and should not be read as a general cost model. Finally, a contingency line only works when the acknowledgement is genuinely pre-agreed; retrofitting the language after a shock arrives does not carry the same weight.
References
[1] The GPU-Hour Ledger: What Three Phases of Subsurface AI Actually Cost in Compute. Earthscan insights, 2025-11-11. https://earthscan.io/insights/gpu-hour-ledger-three-phases-subsurface-ai-compute
