The Wrong Question About AI Cost

When Dwarkesh Patel toured Jane Street's new Texas datacenter last week (4,032 liquid-cooled GPUs, 56 racks, 8,000 km of internal fiber) he asked the obvious question: what does all of this cost?

Ron Minsky, who co-heads Jane Street's technology group, gave the answer that should be on every mid-market CFO's wall: the cost of the compute is trivial next to the cost of not having it.

That's opportunity cost. And it's the frame most enterprises still get wrong.

The math behind the answer

Jane Street did $10.1B in net trading revenue in Q2 2025 alone, with roughly 3,000 employees. They've reportedly committed $6B+ to CoreWeave for compute access and $1B+ in equity, plus stakes in Anthropic and the cluster Dwarkesh toured. Compare that to Goldman Sachs: 46,000 employees, a fraction of the net income.

Minsky's answer isn't "compute is cheap." It's that the alternative, losing a signal because your researchers had to wait three days for a training run, or shipping a worse trading model because you couldn't afford the experiment, costs orders of magnitude more than the GPUs.

AI isn't cheap, but doing nothing is the most expensive option on the menu.

Most mid-market enterprises ask the wrong question

The default boardroom question is "what does AI cost?" That's the wrong question. It anchors the conversation on a line item: model fees, infra, headcount. It treats "do nothing" as the zero-cost baseline.

It isn't. MIT's NANDA initiative found that 95% of enterprise GenAI pilots produce no measurable P&L impact, and that the 5% that work share three traits: tightly scoped use cases, back-office focus, and a deployment harness that integrates into actual workflows. The failure mode isn't model quality. It's that companies bolt AI onto a process and call it done.

PwC's 2026 Global CEO Survey found 56% of CEOs report no revenue or cost benefit from their AI spend yet. That number gets thrown around as evidence AI is overhyped. The better reading: most of that spend went to pilots that were never set up to run agents in production against real data, on infrastructure the team controls.

Meanwhile, Deloitte reports 78% of US mid-market leaders have moved at least one AI project into full production. The gap between those two numbers is where opportunity cost compounds. Early movers in your segment are training on workflows you haven't even instrumented yet.

What opportunity cost actually looks like in the mid-market

You're not Jane Street. You don't have $10B in quarterly revenue justifying a $6B compute bill. But the structural question is identical:

• Every workflow you don't automate is a recurring tax. The procurement team manually reconciling vendor invoices, the analyst rebuilding the same dashboard every Monday, the support engineer re-triaging tickets. Those hours don't show up as a line on the P&L, which is exactly why they get ignored. Multiply them by 52 weeks and they dwarf any reasonable agent infra bill.

• Your competitors' training data compounds. The MIT report and follow-on analyses keep landing on the same point: organizations that deploy AI into real workflows accumulate proprietary signal that laggards can't buy later. That gap is not linear. It's exponential, and it has roughly an 18-month window before it locks in.

• The talent question flips. Engineers and analysts who use agents daily are 2 to 3x more effective, and they will not stay at companies that won't let them. Your retention cost is your AI cost in disguise.

The do-nothing path doesn't preserve capital. It bleeds it through channels that don't appear on a budget review.

The harder part: controlling OpEx once you're in

Here's where most companies who do take the leap fall on the other edge of the sword. They sign a frontier-model API contract, give the engineering team a credit card, and discover six weeks later that sustained agent workloads burn through a quarter's budget. Or that token spend on a single agent surface scaled with usage in ways finance didn't model. Or that the "AI line" on the cloud bill is now bigger than the rest of the cloud bill (which will happen).

This is the second half of the opportunity-cost argument that most vendors skip. Saying "yes" to AI doesn't mean saying "yes" to uncapped OpEx. Sustained agentic work, agents running on schedules, against your data, producing real artifacts, has to live somewhere you can govern.

That's the gap Nightshift was built for. Nightshift gives organizations the escape hatch out of the frontier providers. Start with Claude, instrument the workflows, and produce proprietary work trajectories that you can distill your own models on. Once you have models that perform at par with the frontier on your tasks (and in many cases out-perform the frontier), swap your proprietary models in, running on fixed-price compute.

The point isn't to make AI cheaper. It's to make the cost legible and controllable, so the opportunity-cost math actually works in your favor instead of running away from you.

The frame to bring to your next board meeting

Stop asking "what does AI cost." Start asking two questions:

1. What does it cost us to keep doing this work the way we do it now, for the next four quarters? Include the labor hours, the error rate, the lost speed, the talent attrition, and the deals you won't close because a competitor moved first.

2. What does it cost to run this work as sustained agentic operations, on infra we control, with hard ceilings on OpEx?

The first number is almost always larger. That's Minsky's answer in mid-market clothes. The companies that internalize it in 2026 will look like the 5% in next year's MIT report. The ones who keep asking "what does AI cost" will be quoted in the 95%.

Minsky's answer at Jane Street wasn't a flex about the size of their budget. It was a discipline. Ask the right question, the math becomes obvious. Ask the wrong one, and you spend the next four quarters defending a budget for work that was never going to compound anyway.

The cost of compute is not the question. The cost of going without is. Nightshift exists so the answer is actionable for the companies that aren't running $10B quarters.