Why AI Is So Expensive to Run, and Who Pays for It
The price a company pays for AI today is not what it costs to serve. Ryan Cawood, Co-Founder and CEO of Lab Thread, has to decide how much AI to build into his own product, which means he has looked closely at the economics underneath the invoice. His estimate of the gap is large.
these companies are actually underwriting the cost of service serving you the AI result to the degree of 90%. So you know we're paying a couple hundred dollars to use whatever the program might be that you're using on a monthly basis. the genuine cost to the company that is providing that service is closer to $1,000. You if you're pay $100, they're actually paying $1,000
That is his arithmetic, not an inference from it: the difference is absorbed by the provider to win adoption. That is a normal and well-understood growth strategy. It is also a liability if you are building a product on top of it, because the subsidy is a decision somebody else can reverse.
Why is the cost being absorbed at all?
Because the market is still being taken rather than served. Cawood is blunt about the phase everyone is operating in.
I don't disagree that it's the wild west right now
In a land-grab, price is a customer-acquisition tool rather than a reflection of cost. The strategic question for a buyer is not whether that is happening, but what your own economics look like on the day it stops.
What does the price change look like when it comes?
Cawood's argument is that it has already started, and that it is not arriving as a headline price rise. It arrives as a quieter adjustment to what you get for the same money.
they're releasing new models which are better than the previous ones but they use that as a ratchet mechanism to provide you with less tokens for the same money
Over the last eighteen months to two years he has watched the tokens per dollar shrink while the headline subscription holds steady. A better model for the same price sounds like progress, and it is, but if the usage allowance falls at the same time then the effective cost per unit of work has risen without anyone announcing an increase. For a company whose product calls the model on every customer action, that is a margin problem that compounds silently.
What happens to a product built on a subsidized input?
This is the risk Cawood is actually managing, and it is a design decision rather than a procurement one. If AI is deployed everywhere in a product, every customer interaction burns tokens the company pays for. If the unit economics were set when the input was subsidized, the product can become structurally unprofitable without a single thing changing on the company's side.
His concern is not this quarter's bill. It is dependency: building a business whose cost base is set by a provider currently choosing to lose money, and who will eventually choose not to. The same reasoning is why he is selective about where AI sits in the product at all, rather than adding it everywhere it could technically go.
Is anyone pricing the energy?
Raphaël Ognar, President, CEO, Co-Founder and Chairman of NKILT Therapeutics, extends the accounting past the invoice entirely. His objection is that the dollar figure is only the visible layer.
We don't talk about the overall environmental and energy cost that AI is leading us to as an entire society
His point is about the physical buildout behind the invoice: the data centers going up everywhere are, in his description, consuming energy at a pace that is completely out of control, and doing it inside systems that were never taught to conserve.
He treats this as a business risk rather than only an environmental one, and the distinction matters for anyone modeling long-term AI spend. If the systems are not currently optimized to conserve energy, and the constraint eventually binds, then the cost curve people are extrapolating from is the wrong curve. US data centers consumed about 4.4 percent of national electricity in 2023 and are projected by the Department of Energy, in its December 2024 report, to reach between 6.7 and 12 percent by 2028, which is the kind of number that turns an externality into a line item.
Cawood's response is the driest line in the episode.
It'll be quite ironic if we stop burning oil and then start burning water which I think is where we're heading
What is the AI business optimized for?
Ognar's broader objection is not about price at all. It is about intent, and he states it as a design principle rather than a complaint.
we need to try to think about designing systems and leveraging processes and AI technology not just based on greed but based on outcome
His position is that a great deal of what is being built is optimized for extraction rather than for results, and that the people building it are not naive about which one they are doing. For a biotech founder deciding where to spend limited capital, that is a practical filter and not a philosophical one: it asks what a tool actually delivers before asking what it costs. Getting a tool to actually deliver is its own discipline, and both founders describe it in how to prompt AI for scientific research.
That filter has consequences elsewhere in the same conversation. It is the same discipline behind Ognar's refusal to claim AI he does not need to claim, which he sets out in should startups put AI in their pitch deck, and it sits alongside the harder operational question of who is liable when an AI mistake reaches a submission, and the slower one of why those errors take months to surface.
The convener's read
The two objections here come from opposite directions and land in the same place. Cawood is worried as a builder about a cost base he does not control. Ognar is worried as a founder about a system optimized for the wrong thing. Neither is arguing that AI is too expensive to use. Both are arguing that the number in front of you is not the real number, which is a different and more useful claim.
The full conversation is on the episode page, along with more from Ryan Cawood and Raphaël Ognar. If your organization wants to reach the people making these calls, partner with Open Door Salon.
Methodology: every quotation above is drawn verbatim from the recorded, on-the-record conversation between Ryan Cawood, Raphaël Ognar and host Lori Ellis on Open Door Salon, and was checked against the episode transcript before publication.