Your Data Left Ten Years Ago. Your Judgment Is Leaving Now.

The office is almost finished. Most of the furniture is in, a few details are still missing, but it has finally become what I wanted it to be: a corner of quiet where I can work at full speed, surrounded by a paradox I built on purpose. Things from my childhood and things from this year, side by side. When I opened the door to Milan, only the childhood was showing.

Milan Toth is a friend, a former colleague from IBM, and a rider. We had not managed a single ride together all summer, so the hug lasted a little longer than usual. He looked around the room. I had just put down the books I was carrying, next to the records and the CDs. The Rolling Stones, the Beatles, AC/DC, Metallica. Ken Follett, Tolkien, Stephen King. Then we went to make coffee, the Italian way, and we turned the music down. The music was coming out of a Bluetooth speaker streaming from Spotify. Five centimetres from the speaker, two Kindles. Two men who grew up with mud on their knees and vinyl on the shelf, standing in a room where everything they own is next to everything they rent, talking about IT, data centres and AI. I will come back to that speaker at the end. First the coffee, because that is where the conversation went.

Milan’s company lives on infrastructure. DevOps, cloud, Kubernetes, the work of moving what an organisation runs from one place to another. So the question over the coffee was the one we ask each other every time: after a year of the whole industry moving to AI, what do we see in the organisations we work with? We had the same first observation, and we had both got it on our own skin. The cloud is being used more consciously than it was, because the abuse of it produces bills that are neither sustainable nor predictable. And the AI bill is starting to behave exactly the same way. It is becoming the joy and the sorrow of the CEOs and CFOs we talk to. Milan sees it from the racks. I see it from the software and the tokens. Same photograph.

Before I go on, one thing about where I stand, because the rest of this piece can be misread without it. In 2014 I was part of a team that took a core asset of a European bank into AWS, in the Frankfurt region the week it opened, after developing in Ireland because Ireland was all Europe had. That was the cloud as revolution, and the SW Craftsmanship Dojo® was born on that project. I have used the cloud aggressively ever since, and I still do. For creating and destroying machines, for testing, for prototyping, there is nothing better, and there will not be. This is not a piece about fleeing the cloud. It is a piece about what the cloud became when it stopped being a tool and became the single place where everything lived.

The first meter, and the tap nobody fitted

The people like us who measured it saw the fork early. In 2019 DORA, the research programme behind the State of DevOps reports, found that what predicted performance was not being on a cloud provider but having the five characteristics the NIST definition of cloud describes: on-demand self-service, broad network access, resource pooling, rapid elasticity, measured service. Only 29% of the respondents who said they had adopted cloud agreed they had all five.[*1] The market bought the provider and skipped the discipline that earns the characteristics. And the fifth one is the whole story. DORA’s own text on measured service warns that “this change from fixed-cost infrastructure to variable cost can have significant implications for both procurement and finance.”[*1] A variable cost is a meter. Nobody said the meter came without a tap.

Then the meter concentrated. Three companies now take almost two thirds of everything the world spends on cloud infrastructure.[*2] Britain’s competition authority finished its investigation in July 2025 with a sentence I could not have written without being accused of exaggeration: the two largest each hold 30 to 40% of the supply, and possess “significant unilateral market power.”[*3] The regulator’s words, not mine. An economist wrote a whole book about what that does to the people who depend on it. Yanis Varoufakis calls what the platforms take from them cloud rent,[*4] and you do not have to sign the rest of his argument for the word to stay with you. Rent is what you pay for a room that is never going to be yours. What convinced me the word fits is not the book. It is that Europe had to legislate the way out. The Data Act removes switching charges, egress included, from 12 January 2027:[*5] a law, so that walking away would not cost you. Until that law lands, the price of leaving is set by the company you are leaving, by the gigabyte, on your own data. Except that AWS waived it in March 2024, nearly three years ahead of the deadline. You ask their support, they approve the account, and you have ninety days to carry everything out for nothing. A real concession, and I would use it. It is also, in their own words, a waiver that “follows the direction set by the European Data Act.”[*6] The door opened while the law was still being written. You do not pass a law to open a door that was already open…

The second meter, the one your board approved

As you might have noticed, this year the price of a token has collapsed. For a fixed level of capability, Epoch AI measured the price falling somewhere between nine times and nine hundred times per year, depending on the task.[*7] On that curve the AI budget should have been the easiest line of the year. Instead enterprise spending on generative AI reached 37 billion dollars in 2025, 3.2 times the year before.[*8] Satya Nadella named the mechanism in January 2025: “Jevons paradox strikes again! As AI gets more efficient and accessible, we will see its use skyrocket.”[*9] Cheaper units, many more units, bigger bill. The frontier models have made it worse in a way that is easy to miss on an invoice: the reasoning they do before answering is billed as output tokens, whether or not you ever read it.[*10] You pay for the thinking, and the thinking got longer.

And that second meter runs on top of the first. Inference is compute, compute is cloud, cloud is measured service. A variable cost feeding a variable cost. That is the mechanism behind the sentence I promised you last week: this is why the AI bill went up while the roadmap stayed the same. The FinOps Foundation asked 1,192 practitioners this year: 98% now manage AI spend, up from 31% two years ago; 73% exceeded their AI budget; only one in five predicted it within ten percent.[*11] Flexera’s 753 cloud decision-makers put their own estimate of wasted cloud spend at 29%, the first rise in five years, and attribute the rise to AI workloads.[*12] Two meters, no tap.

The context you cannot afford to send outside

Here is what changed in nature, and not just in degree. When the cloud held your servers, a provider could see your workloads. To make AI reason about your business without inventing it, you have to feed it your processes, your people, your organisation. The context that stops the hallucination is the context you least want to leave the building. And the building is no longer yours. In June 2025 a senior Microsoft executive sat before a committee of the French Senate and was asked, under oath, whether he could guarantee that French data would never be handed to the US authorities without French consent. His answer, verbatim: “Non, je ne peux pas le garantir.”[*13] He was not confessing a breach. He was describing the CLOUD Act, which obliges providers under US jurisdiction to hand over data “regardless of whether the data is located in the United States.”[*14] Nobody had to break anything. It is simply the terms.

What you rent is not what you own

You already live under such terms at home. The Kindle five centimetres from my speaker holds books I paid for and do not own; the store terms say I “may not sell, rent, lease, distribute, broadcast, sublicense, or otherwise assign any rights” to them.[*15] The speaker plays a catalogue that “may remove access to particular songs, videos, podcasts, audiobooks and other Content without notice.”[*16] The vinyl next to them does not have terms. Europe has noticed the difference at scale: a proposed Cloud and AI Development Act now states that “over reliance on non-EU cloud service providers poses a significant risk to Europe’s digital autonomy and resilience,”[*17] and a German retailer’s cloud arm now sells itself as a German hyperscaler, built for organisations that do not want their processes under another jurisdiction’s legal framework.[*18] I am not telling you to move. I am telling you that the direction of the market has already been decided, and it was decided by the bills, by the lack of transparency, and by a question nobody was asked: who the data belongs to.

What a modern AI software factory does about owning instead of renting

So, the last two axes on the board I promised you: infrastructure, and the token economy.

On infrastructure, what runs after the release is where the AI year is landing. DORA’s 2025 research is direct: “AI adoption does continue to have a negative relationship with software delivery stability.”[*19] More change, faster, into systems whose operations were sized for the old speed. With nWave open source you get the workflow, the gates and the delivery record, and what happens after the release is still yours to watch. In the Software Factory it is watched for you, by the discipline Milan has built his company on: modern DevOps, DevSecOps at every gate, and SRE agents that know what your software is doing in production before your customers tell you. Not a dashboard. Closer to a night shift that never gets tired and never gets bored, reading every line of the logs for the thing a human skims past: the drift in latency, the shape of a response time an hour before it becomes an outage. And because it is a team and not a subscription, you can point it at whatever you own. Your servers. Your models. Reliability stops being something you rent from whoever holds the racks, and becomes something you keep.

On the token economy, the lever is architectural, and the research has been public for two years. A router that sends easy work to a cheaper model cut cost by over 85% on one benchmark while keeping 95% of the strongest model’s quality;[*20] cascades reached 50 to 98% depending on the task;[*21] Microsoft’s own routing work made up to 40% fewer calls to the large model “with no drop in response quality.”[*22] Add the levers the vendors publish and few teams use, cached context at a tenth of the input price, batches at half.[*23] Inside the Factory we call it AI-cognitive routing, and it is the daily work of a harness engineer. The frontier model reasons where reasoning is the job. The adversarial reviewer that checks a citation exists, or that page four does not contradict page two, runs on a model built for exactly that, and costs a fraction. Two kinds of waste leave with it: the mind that was too big for the task, and the answer that was longer than anyone was ever going to read. Then you put a gauge on what is left, and that is the part nobody sells you. When you can see the cognitive power leaving the building hour by hour, you can turn it down. The bill stops being weather and becomes a setting. The meter finally has its tap, and your hand is on it.

With these two, the six axes are on the table. Six was never the useful number. They run in pairs, and every pair is a track: two rails, and nothing moves on one of them alone. Quality and security are the first track. They decide whether what you shipped is a product or a liability. Speed and knowing what to build are the second, the user centricity track. It carries a factory from writing code to making something a customer actually wants. Infrastructure and the token economy are the third. It shapes what the factory costs to run, so the number goes into next year’s budget instead of onto a blackjack table where you play every invoice as it lands.

Which is why this took three conversations and not one. Three tracks, three episodes, one product release train: nWave SW Factory.

Two meters, one question

The coffee was finished long before the conversation. Milan and I ended where we always end, on where this goes next, and we agree: what happened to the cloud is about to happen to AI. Hybrid by necessity, not by taste. Specific models instead of one that knows everything. Infrastructure that lets an organisation keep its sensitive data inside its own walls and still move at this year’s speed. That is the shape of a modern hybrid infrastructure with modern DevOps, DevSecOps and SRE, where product teams own the product, the platform and the bill, and the organisation controls its spend without ever being surprised by it. We have been watching the pattern form, and now the first signs of demand are showing. That is a whole episode of its own, still to come: the rise of the hybrid cloud organisation.

Before he left, the music came back up. AC/DC filled the room again with Who Made Who. The two Kindles were still there, five centimetres from a speaker that decides what I hear. The records behind them do not decide anything. I own them. That is not nostalgia. It is the only question this whole year has been asking, and now it is being asked about your infrastructure and about the context you feed your AI: what do you own, what do you rent, and who set the price.

You invested in the cloud when it was the right call, and you invested in AI when it was the right call. Both halves of that were right. What arrived without a note was the meter. I have sat with leaders who discovered the second meter the way you discover it, on an invoice with a number nobody forecast. nWave is open source and you can put it on your laptop this afternoon. When you want the machine, the infrastructure and the people who run it arriving together, so that the bill is a decision rather than a surprise, that is what The Helm is for.

Your timeline. Your choice.

→ The Helm

Next week: the conversation Milan and I did not finish. What happened to the cloud is happening to AI, and what a hybrid cloud organisation looks like when the model lives inside your walls.

Sources

Every number in this piece traces to one of these. They open in a new tab, so the article stays where you left it.