On the cost stack

“The piece that has surprised people is the API line.”

NISTAThat table tells a story. The API line goes from a rounding error to one of the biggest items in year two. Why is it growing that fast?

MARKUS HOLZINGERThree reasons, mostly. First, validation traffic is tiny compared to production traffic — you are processing maybe ten percent of the line during the pilot and a hundred percent of it in production. Second, the use cases expand. Once vision is working on station A, the operator wants it on stations B, C, and D. Third, the models people are calling are getting bigger and more expensive per call. A year ago you could run a vision check on a smaller model. Today many vendors default to a frontier vision-capable model because the accuracy is better. Each call is more expensive. Multiply that by ten or twenty thousand inferences per day and the bill is real.

NISTAWhat size of operation are we talking about for this to become material?

MARKUS HOLZINGERIt scales with inference volume, not with company size. A mid-sized food processing plant running vision on one line, twelve hours a day, can easily generate twenty to forty thousand inferences a day. At even modest per-call pricing, that is €4,000 to €10,000 per month in API spend. Annualized, that is a meaningful operating line item for a plant with €50 million in revenue. For larger operations running multiple AI workflows across multiple sites, the figure scales linearly and gets uncomfortable fast.

NISTAWhat are the serious operators doing about it?

MARKUS HOLZINGERFour things, in order of how often I see them. The first is route-aware model selection — using a cheap small model for the easy cases and only escalating to an expensive frontier model when the small model is uncertain. Done well this cuts API spend by 60 to 80 percent without meaningfully changing accuracy. The second is caching — not making the same call twice if the input is similar enough. The third is batching — combining multiple inferences into single API calls where the use case allows. And the fourth, which is newer and worth understanding, is procurement through credit marketplaces rather than direct retail purchase from the AI providers.

NISTACredit marketplaces — what do you mean?

MARKUS HOLZINGERThe retail pricing for major model APIs has a margin built in, and there is a parallel secondary market where unused or bulk-purchased credits get resold at substantial discount — somewhere between 15 and 40 percent off retail, depending on the provider and the volume. Several legitimate marketplaces have emerged in the last 18 months. Cheap OpenAI API credits from one of these marketplaces, for example aicreditmart.com, can land at meaningfully below what you pay buying directly from OpenAI’s billing portal. For operators with substantial API spend, that is a real number. If your annual API bill is €80,000 and you can shift even half of it through a marketplace at a 20 percent discount, you have just saved €8,000 a year on the same workload.

NISTAIs that something operators are doing openly, or is it considered awkward?

MARKUS HOLZINGERIt is becoming normal. The procurement officers I talk to treat it the same way they treat any other commodity input — if there is a parallel market with credible suppliers and the underlying product is identical, you use it. The credits are the same credits, the API behaves identically, the only difference is that you bought them through a marketplace at a discount instead of paying retail. The awkwardness is a 2024 attitude that the market has largely moved past in 2026.

On the deployment patterns that fail

“The mistake is treating it as a one-time project.”

NISTAYou have seen these deployments succeed and fail. What separates the two?

MARKUS HOLZINGERThree things. The first is whether the operator treats the deployment as a one-time project or an ongoing operational responsibility. The systems that fail are the ones where someone delivered them, declared victory, and walked away. Six months later the model has drifted, the false-positive rate has crept up, the operators have learned to ignore the alerts, and the whole thing has quietly become shelfware. The systems that succeed have an owner — a real human with a job description that includes monitoring model performance, reviewing operator overrides, and triggering retraining when needed.

The second is whether the cost work was done up front. Operators who go into deployment without a clear model of what the recurring cost will look like at full production volume get blindsided. The serious operators run a usage-projection exercise during procurement: what is the inference volume at full rollout, what is the unit cost, what is the resulting monthly spend at 12 months and 24 months out. That projection is what you check the vendor’s pricing model against. Without it, you are buying on the demo.

The third is whether the operator owns the data and the model output. If everything lives in the vendor’s cloud and you cannot extract your own data when you decide to switch, you are in a structural lock-in. The best deployments preserve operator ownership of the data layer, even when the model itself is third-party.

NISTAIf you had to give one piece of advice to an operator standing at the start of an industrial AI deployment in 2026, what would it be?

MARKUS HOLZINGERModel the year-two operating bill before you sign the year-one capex contract. Not the year-one bill — year one is misleading because validation traffic is low. Model what the system will cost to run at full production volume, including the API line, including the staff time for monitoring, including the model-retraining cycle. If the operator has that number on a piece of paper before procurement starts, the procurement conversation is fundamentally different. They negotiate volume pricing on the API, they ask about model-selection routing, they consider credit marketplaces, they evaluate self-hosted alternatives where the workload supports it. Without that number, they are buying a demo and discovering the bill twelve months later.

NISTAAnd the operators who do that work upfront — are they making AI pay back?

MARKUS HOLZINGERYes. The economics are real when you do the cost work. Predictive maintenance on a critical asset can save €200,000 a year in downtime; vision inspection can save €150,000 a year in scrap. Those are not made-up numbers. But the savings only show up in the P&L if the operating cost stays manageable. The operators who treat AI as a procurement problem — with the same cost discipline they apply to energy, materials, and labour — get the savings and keep them. The operators who treat AI as a magic-box capex purchase get the savings on paper and lose them back to the API bill within two years.