The invoice that could have killed the project never went out. In the middle of 2022, with a Middle East carbonate operator's multi-year AI programme roughly at its halfway mark, the war in Ukraine turned Europe's energy market into something we could not budget against. Our compute ran on servers whose electricity we paid for, and the electricity was moving in a direction that had no ceiling anyone could name. The honest engineering fear was not that compute would get expensive. It was that compute would get expensive without a bound, and that we would have to walk into a board meeting every quarter with a bigger number and no story for why it stopped there.
It did stop there. This is the account of why, and it has almost nothing to do with the models. It has to do with three decisions we made before the war started, back when locking hardware and a data-center deal looked like ordinary procurement rather than a hedge.
What the shock actually did to the bill
The numbers were not subtle. Energy prices sat about 400 percent above their pre-war level, and the operator's own electricity bills came in 394 percent higher. Gas, which sets the marginal price of European power, was up 600 percent since the project kicked off in December 2021 and 1300 percent since September 2021, per ICE TTF settlement data. Those last two figures are the ones that matter for a compute budget, because they describe a curve that was still bending when we had to commit to a plan.
The pass-through into our world was direct. Supercompute-class hosting, the kind a rack of high-memory training servers needs, moved from roughly USD 12,000 to 15,000 a month before the war to more than USD 20,000 a month after it. A vendor list price near USD 9,000 looked cheaper on paper and then reconciled to 11,000 to 15,000 once configuration and electricity were added, exactly the kind of opaque, config-plus-energy pricing that makes a runaway feel inevitable. Left alone, the compute line of a multi-year programme reprices every month with the power market, with no natural place to settle.
That is the exposure. An uncapped, market-tracking cost sitting on the critical path of a research programme the client had already approved and budgeted.
The three hedges we had already locked
None of the three moves below were made in response to the war. That is the whole point of a hedge. They were in place before the shock, and they are why the shock hit an absorber instead of the invoice.
The first was owned hardware. We had locked a multi-year partnership with a GPU maker back in 2020, and the practical effect was that our core accelerators were owned capacity rather than rented capacity. When rental markets repriced with electricity, the owned tier did not. A rented instance is a claim on a market that had just tripled. A card you already own is a fixed asset whose only variable cost is the power to run it, and even that we had partly addressed.
The second was a local data-center deal. Hosting through a local partner rather than a hyperscaler's on-demand tier bought us full cost visibility. We could see the electricity component as a line item and reason about it, instead of receiving a blended bill that moved for reasons we could not audit. Visibility is not the same as a discount, but during a shock, being able to say precisely which part of the bill moved and by how much is what lets you draw a bound around it.
The third was academic-rate pricing. The programme ran as an R&D engagement, and the infrastructure carried discounted academic rates rather than commercial ones. That kept the base off which every percentage increase compounded far lower than a peer paying commercial war-time rates would have seen.
Stacked together, these let us make two claims to the client that would sound like marketing if they were not arithmetic: we remained about 140 percent more cost efficient than the market alternative, and a peer procuring the same capability at war-time commercial rates was paying on the order of 300 percent more. The hedges did not make the energy crisis disappear. They converted it from a cost we absorbed into a cost we could describe.
One line the board could approve in a single decision
Here is where the hedges stopped being an infrastructure detail and became a governance instrument. Because the shock hit absorbers, the ask we finally put in front of the client was bounded and small relative to the programme.
The additional funds request, sent in October 2022 with Phase 1 complete and Phase 2 due, added USD 104,300 to an already-approved budget of USD 533,000. That total split into two buckets. USD 53,300 was additional AI infrastructure, driven mostly by a deliberate technical choice to run supervised and unsupervised tracks in parallel, which doubled the compute paths, plus the Phase-3 well-to-well models. The second bucket, and the one the hedges made possible, was USD 51,000, billed as an energy and utility contingency for twelve months.
That USD 51,000 line is the artifact I would show any board wrestling with AI infrastructure risk. It is a single, pre-agreed, time-boxed number. It is not a quarterly re-forecast that grows with the power market. It is not an open-ended pass-through clause. It is one decision, because the hedges had already absorbed the part of the shock that would otherwise have been unbounded, leaving only a modest, nameable residual for the client to cover. A twelve-month energy contingency of USD 51,000 against a USD 533,000 programme is the difference between a board approving a contingency and a board reopening the whole engagement.
The compute reality underneath was not calm. Phase 2 had been budgeted at 1,200 GPU-hours and about USD 20,000, and it actually ran to 2,600 hours and USD 44,700, an overrun of USD 24,700 that came from the dual-track strategy rather than from energy. Phase 3 was forecast to grow from 5,800 hours to 7,500, another USD 28,600, with parallel runs expected to reach sixty to ninety concurrent jobs at 80 to 90 percent sustained utilisation, contained by containerising onto up to 28 partitioned accelerator instances. Even with all that real compute growth, the energy-specific exposure the war created stayed a bounded line, because the hedges, not a spreadsheet, held it there.
Why this belongs in a board pack, not just an infra review
The cloud-versus-owned trade-off is well-trodden, and we have written the fixed-cost private-infrastructure case up in Why Energy Companies Keep Their Models On-Premises. What that piece does not say, and what the 2022 shock taught us directly, is that the value of owning your compute is not only steady-state unit cost. It is optionality during a shock.
The board lesson is one sentence: AI capability economics are energy economics. A training programme is, at the margin, a machine for converting electricity into model weights, so anyone who signs a multi-year AI budget is signing an implicit position on power prices whether they know it or not. You can take that position blind, by renting on-demand and repricing with the market, or you can hedge it in advance with owned hardware, cost visibility, and a favourable rate base. The hedge is boring to buy and invisible until the day it matters.
So if I were advising a subsurface or industrial team standing up an AI programme today, I would ask one question before the model architecture: what is your position on the price of the electricity this will burn over the next three years, and where is it written down. If the answer is a shrug, the programme is carrying an uncapped energy option it has not priced. Ours was priced in 2020, and that is the only reason the 2022 invoice stayed a single line the board could sign.
Limitations
These figures come from one confidential engagement and one moment in the 2022 energy shock; the price moves and dollar amounts are from our own contingency memo and additional-funds request to the client, not a market benchmark. The 140 percent and 300 percent efficiency claims are our own positioning against market and peer rates at the time, not audited third-party comparisons. The hedges worked here because they were already in place before the shock; a team standing up infrastructure mid-crisis does not have the owned-hardware option on the same terms, so the transferable lesson is about timing as much as structure. The runaway counterfactual in the exhibit is illustrative, drawn to show the shape of an unbounded pass-through, and is not a billed number.
References
- Intercontinental Exchange (ICE), Dutch TTF Natural Gas Futures settlement price series, cited in our August 2022 contingency memo as the source for the gas price moves (+600 percent since December 2021, +1300 percent since September 2021). https://www.theice.com/products/27996665/Dutch-TTF-Gas-Futures
