Every AI services proposal hides a rate card, and the rate card is where the real strategy lives. For a multi-year subsurface machine-learning programme with a carbonate operator in Oman, ours was a single spreadsheet with two day-rate ladders sitting next to each other. One was the commercial ladder, the day rates a Berlin or Paris consultancy would quote, and the ladder the bench was actually priced at. The other was the academic ladder, sitting near half of it at every tier, a university-linked R&D discount held in reserve. Reading the two columns together explains how a small lab can underbid Big-4 AI consulting and still hold its margin. This piece is the anatomy of that pricing, tier by tier, from a named bench of engineers down to the monthly cost of a GPU server.
Two ladders, one near half the other
Start with the numbers, because the whole argument is in them. The commercial day rates ran 3,200 for a CEO or partner, 2,018 for a head of AI, 1,727 for a senior researcher or engineering lead, 1,200 for an AI researcher, and 800 for an ML developer. The academic ladder, tier for tier, ran 1,600, 1,000, 800, 640, and 480. Line them up and the pattern is deliberate: the academic rung sits near half its commercial rung at every level. The CEO rate halves exactly. The middle tiers land between 45 and 53 percent. The junior developer holds at 60 percent, which is the floor, because below that the rate stops covering a working engineer's cost.
That halving is not a fire sale. It is the entry weapon. A discount that arrives labelled as an academic rate reads as a rate, not a concession, and a rate is much harder to argue down further than an ad hoc discount is. The client sees a defensible number tied to an R&D framing, not a vendor blinking first on price. When the counterparty is comparing against a Big-4 bid built on commercial rates, dropping to the academic ladder puts the lab near 50 percent lower at the door and reframes the entire conversation before scope is even discussed.
The bench, priced two ways
The commercial workbook priced the same named team twice, which is the honest way to show a client what they are buying. The first scenario was a 644-man-day bench: a partner on quality and oversight for 20 days, a domain-expert senior researcher for 64, a project-managing senior researcher for 80, two AI researchers at 100 days each, an engineering lead for 80, a senior ML engineer for 80, and two ML developers at 60 days each. Priced line by line on the commercial ladder, that bench totalled 869,472 EUR. A 10 percent contingency lifted it to 956,419.20. Across 644 man-days that is a blended 1,350 EUR a day, the effective rate the commercial bid carried.
The second scenario compressed the same roles into a leaner 484-man-day build totalling 701,472 EUR, or 771,619.20 with the contingency, which works out to about 1,449 EUR a day blended. The two scenarios are not good-and-bad options; they are the same team held to two different effort budgets, so the client can see exactly what a smaller commitment removes. The leaner build blends slightly higher per day because trimming days falls hardest on the cheaper junior rungs, leaving a more senior average. Both numbers are commercial-rate totals a procurement officer can hold in their head and compare against a Big-4 bid.
The academic ladder is where the discount lever sits, and the workbook makes it concrete. Re-price that same 644-man-day bench tier for tier on the academic rates, and the total falls to roughly 435,000 EUR, near 676 EUR a day blended. That is not a second team or a smaller scope; it is the identical bench, the identical days, re-run down the academic column. The gap between 869,472 and 435,000 is the whole lever: half the cash bill for exactly the same people, offered as a defensible R&D rate rather than a haggled concession.
Owning the compute is the discipline
A discount without cost discipline is just a loss deferred. The reason the lab can drop to the academic ladder without bleeding margin is that the largest variable cost, compute, was not a rented spread we re-marked to the client. We owned it. The workbook itemised GPU servers as fixed monthly lines: 700 EUR a month for a 48GB-GPU supercompute node and 1,500 EUR a month for an 80GB-GPU server. Those are amortised hardware costs, not on-demand cloud invoices with a markup baked in.
The distinction matters more than it looks. A firm that rents its GPUs has to charge the client the rental plus a margin, and that margin has to survive whatever the cloud provider does to spot prices. A firm that owns its GPUs carries a known monthly number that does not move, which means the discount on labour is not quietly funded by an inflated infrastructure line. When the 2022 energy shock pushed DGX-class hosting from a 12-to-15K band to over 20K a month, owned hardware is what let the programme absorb it as a bounded, pre-agreed draw rather than a contract reopening. We cover that episode in The Contingency Clause You Hope Never to Use; the pricing lesson here is simpler. Owned infrastructure is the balance-sheet reason a discounted day rate can stay disciplined.
The instrument below puts the whole argument on one plate: the paired ladder that shows the near-half gap at every tier, the two bench scenarios and their totals, the owned-GPU line items, and the one lever that turns the aggressive edge on.
The equity option
The most unusual line in the workbook was a single sentence: we would offer a discount in return for equity or joint-project shared revenues. That clause is the aggressive edge, and it is the orange lever in the instrument.
Read plainly, it says even the academic rate is not the floor. If a client is willing to trade equity or a share of downstream revenue, the day rate can go lower still, below the discounted academic blend and far below the commercial number the bench was priced at. This is a deliberate inversion of how consulting usually prices risk. A traditional firm charges more when a project is uncertain, pricing the risk into the rate. The equity clause does the opposite: it lowers the cash rate and takes the upside instead, converting a services engagement into something closer to a co-investment. For a small lab with more conviction than cash flow, that is a rational trade. It swaps margin today for a claim on the thing being built.
It is also a filter. A client that flinches at an equity conversation is telling you something about how much they believe in the work, and a client that engages with it is a different kind of counterparty than one shopping for the lowest day rate. The clause does not have to be exercised to be useful. Sitting in the proposal, it signals that the lab prices like a principal, not a vendor.
Why this underbids Big-4 without a race to the bottom
Put the three moves together and the strategy is coherent. The academic ladder gets the lab in the door at half the commercial rate, which no Big-4 bench built on commercial rates can match. Owning the GPU infrastructure keeps that discount from eating margin, because the biggest variable cost is a fixed monthly line instead of a marked-up rental. And the equity-for-discount clause offers a floor below the floor for the clients who will trade upside for a lower cash bill.
None of it is underpricing in the reckless sense. The academic rates still cover working engineers. The 10 percent contingency is real budget, not a hidden buffer. Both the commercial blend near 1,350 a day and the academic re-price near 676 are legible, arithmetic numbers a client can check against the man-day counts. What the pricing does is separate the two things a Big-4 proposal fuses together: the cost of the people and the markup on everything around them. By holding a defensible academic ladder in reserve and refusing to mark up owned compute, a small lab can drop dramatically cheaper at the door and still run a disciplined book. The equity clause is what makes it aggressive rather than merely cheap.
Limitations
These figures are the internal commercial anatomy of one R&D engagement, not a market rate card. The academic ladder was tied to a specific university-linked framing and would not transfer to a purely commercial bid. The blended day rate and the near-half ratios are arithmetic on the sourced numbers; the depth of any equity-for-discount trade was never fixed in the workbook and would be negotiated case by case, so the clause lever in the instrument sweeps an illustrative range rather than a contracted one. The GPU-server line items reflect owned hardware amortised at a point in time and predate the 2022 energy shock that reset hosting costs. Treat the instrument as engagement economics, not a quote.
References
[1] Machine Learning for Well Data Analytics, engagement commercial workbook (internal day-rate ladder, bench scenarios, and GPU-server line items).
[2] Machine Learning for Well Data Analytics, signed service-contract draft (equity-for-discount and shared-revenue clause).
