Inside Forbes Bulgaria’s profile of DiscreteStack CEO Hristo Todorov – and why the rent-vs-own question just became a boardroom decision.
Walk into the DiscreteStack office and there is not much to give the game away: a meeting room, a few desks, and one large screen tracking client systems in real time. Nothing that announces a product its founder believes can reset how European companies buy artificial intelligence.
That was the scene Forbes Bulgaria found when it profiled DiscreteStack CEO Hristo Todorov for its feature “Дигитален суверенитет” (“Digital Sovereignty”), published in Forbes Insights on May 27, 2026.
“We believe we have something with enormous potential,” Todorov told Forbes.
The product behind that line is DiscreteStack, a system designed to run artificial intelligence on infrastructure, controlled by the customer. And it lands on a question most enterprises have not yet stopped to answer: once AI is embedded in how your business operates, whose business is it actually running on? It is easy to defer while AI is still a pilot. It gets harder to ignore the moment AI starts touching the work that keeps the company running.
From Experiment to Operating System
Enterprise AI has moved from something companies pilot at the edges to something they run at the core. As it moves inward, the relevant question changes. It is no longer whether to use AI. It is how much of the company should depend on something it does not own.
Most organizations got here by renting. They draw their AI from one of a few large providers like Anthropic or OpenAI, pay for what they use, and accept the provider’s terms in exchange for not having to build anything. That trade is easy to make when AI is a convenience and expensive to unwind once it is load-bearing.
The Risks of Renting AI From a Hyperscaler
Renting AI from a third-party provider exposes a company to three connected risks: data leaving its perimeter, dependence on a model it does not control, and metered costs it cannot predict.
Start with the data. Anything sent to an external AI platform, whether a prompt, a contract, or a customer record, travels outside the walls the company has built to protect it. A firm that has assured its clients their information stays in-house is, in practice, no longer the only party handling it.
Then there is dependence on someone else’s roadmap. When a core process is built on a model the company neither owns nor operates, the provider’s decisions become the company’s problems. A model can be retired, its behavior can shift, and the customer absorbs all of it, usually without warning and without a vote.
The third problem is the one finance feels first: the bill. Rented AI is metered, and metering ties spend to things the business cannot steer, like the volume of tokens a model produces or the number of moves an agent makes to finish a task. Forecasts get written against numbers nobody can predict, and the more deeply AI is woven into daily operations, the less containable the total becomes.
None of this is hypothetical, and it does not land evenly. It lands hardest on the companies that can least afford it, which is exactly where Todorov places his customer, as he described to Forbes:
“Anyone for whom full control over their operations matters, and anyone who carries more serious responsibility for regulatory compliance and for their clients’ data.”
For those organizations, these stop being engineering concerns. They become questions for whoever signs off on risk and budget.
Built From the Inside of the Problem
The Forbes profile does not treat DiscreteStack as a theory. It traces the company back to a specific bottleneck Todorov kept running into as a founder.
The idea dates to 2023. At the time, Todorov – Co-Founder of the Bulgarian software company CleverPine – was working with large international clients, including Lufthansa Technik and Software AG. He kept meeting the same wall: organizations sitting on enormous volumes of sensitive data wanted the power of AI but had no workable way to use it without handing control of that data to an outside provider.
Todorov was well-positioned to see the problem early. Forbes describes how he had already led CleverPine’s transformation into an AI-first company, and understood almost from ChatGPT’s release that the technology would reshape entire industries. After roughly a year and a half spent building a MVP, he spun the effort out into its own company: DiscreteStack.
What Is the AI Cost Paradox?
The AI cost paradox is the pattern Forbes names directly: costs that rise unexpectedly, driven not by one factor but by several stacking at once. It is not only that AI agents consume large volumes of tokens – it is that compute power, maintenance, and data processing all pile on top, so the bill grows faster and less predictably than any single metric would suggest.
As Forbes describes it, this creates situations that are almost absurd. A company can open an invoice from its AI provider – one of the hyperscalers – for hundreds of thousands of dollars, with no clear explanation of how the figure was generated. Or the tokens included in a team’s subscription plan run out just two hours into the workday, cutting engineers off from the very tool they were promised.
The problem is not confined to any one budget. Gartner expects worldwide AI spending to reach $2.52 trillion in 2026, up 44% in a single year.¹
That trajectory only makes sense once you stop thinking of AI as a piece of software and start thinking of it as something more basic. It is the comparison Todorov keeps coming back to, in his own words to Forbes:
“Artificial intelligence looks like it’s becoming the new essential commodity, the way energy is.”
Forbes places that line next to a similar one from a very different source. OpenAI CEO Sam Altman recently described his company’s own ambitions in almost identical terms: “We see artificial intelligence as a utility, like electricity or water supply. And people will buy it from us for a fee.” Forbes notes the remark drew sharp pushback and pulled the AI-oligopoly question back into public view – the same oligopoly Todorov built DiscreteStack to challenge, and the same tension this article returns to below.
Treated as a commodity you rent rather than a utility you own, the rental model starts to look less like convenience and more like a meter you do not control, running against a resource you cannot do without.
How Forbes Describes DiscreteStack’s Approach
As presented in the Forbes profile, DiscreteStack does not train a proprietary foundation model or require customers to depend on one. It works with open large language models, optimizes their inference performance, and deploys them in an environment controlled by the customer. According to the company’s figures cited by Forbes, this can reduce AI expenditure by as much as 90%.
The company’s method is deliberately unlike the hyperscaler playbook. It begins with open large language models and concentrates its engineering on making them run lean. Rather than spending tens or hundreds of millions to train a model from scratch, it puts that effort into inference: the phase where the model actually has to run, and has to run efficiently, in a production environment.
The payoff is efficiency. By DiscreteStack’s account in the Forbes feature, the optimized system does the work of the original model-software-hardware stack at a fraction of the resource cost, enough to cut AI spend by as much as 90% without surrendering autonomy or putting data at risk.
Lower cost follows from that efficiency rather than from a discount. One of the first companies to trial the system,as Forbes states, was a U.S. company – and it chose DiscreteStack for exactly that reason: it ran far cheaper than renting equivalent capability from a hyperscaler. Under the right conditions, Todorov told Forbes, the cost can fall to a tenth of what the same work would run in metered tokens, largely because the system squeezes so much more out of the hardware it sits on. Ownership adds a second advantage that metering cannot. Because the capacity is bought outright rather than billed per request, AI can keep working through the night at essentially no additional cost.
The contrast drawn throughout the Forbes profile can be summarized as follows:
| Rented AI (Hyperscaler) | Customer-controlled deployment | |
| Where data is processed | Outside the company perimeter | Inside the company perimeter |
| Pricing | Metered per token and per agent step | Fixed annual license per execution node |
| Cost as usage grows | Climbs with every prompt | Tracks the size of the business |
| Control of the model | Provider can change or retire it | Customer controls the stack |
| Off-hours use | Billed like any other request | No per-request charge within available capacity |
The result, as Forbes describes it, is a deployment model built around customer control, open-weight models, and predictable licensing rather than consumption-based billing. DiscreteStack provides a more detailed explanation of its commercial model on its website.
Why AI Ownership Matters Now
The Forbes argument is that two long-standing reasons for renting AI have weakened: capable models are no longer exclusively closed, and operating open-weight models has become more practical than it was several years ago. DiscreteStack is presented as one example of the infrastructure layer intended to make that transition workable for enterprises.
For European companies there is a second force in play, beyond cost and control. Regulation already nudges them toward keeping data processing local and reducing reliance on providers outside the bloc, and Todorov told Forbes he reads the policy direction as the right instinct:
“European institutions have very rightly recognized that they shouldn’t be dependent in this respect. Or at least not this dependent.”
A system that is owned and on-premises answers both demands at once: it keeps the data home and takes the dependency off the table.
The Strategic Choice Ahead
Not every workload needs to be owned, but where data is sensitive, budgets cannot run unpredictably, and compliance cannot be outsourced, the case for owning AI strengthens with every quarter that spending climbs.
Plenty of workloads will keep running on rented AI, and for many of them that is the sensible choice. But the calculus changes where the data is sensitive, where the budget cannot be left to chance, and where compliance cannot be handed to a third party.
That is the decision the Forbes feature ultimately sets in front of enterprise leadership, and it is the one Todorov means to force, in the interview’s closing line:
“What we want is to challenge this oligopoly. To have an alternative. Not to rent AI, but to own it and have control over it.”
The question is no longer which AI vendor to pick. It is whether AI is something the company rents or something the company holds.
This article draws on ideas and quotes from the Forbes Bulgaria feature “Дигитален суверенитет” (“Digital Sovereignty”), published in Forbes Insights, and featuring DiscreteStack CEO Hristo Todorov. Read the original feature on Forbes Bulgaria. The original is in Bulgarian; quotes have been translated into English for this piece.
Continue exploring the topic: read DiscreteStack’s ownership-versus-renting analysis, or learn more about the company featured in the Forbes Bulgaria profile.