top of page
Search

What Really Costs Money in AI?

Writer: AYŞEGÜL ÖZLEYEN
AYŞEGÜL ÖZLEYEN
8 hours ago
5 min read

A cybersecurity company has agreed to buy an AI startup in a deal valued at $4.5 billion. The interesting question is not just what it is buying, but what its customers might pay for next.

22 November 2026

You might expect it to become harder to make money from a technology as it becomes easier to access. Sometimes the opposite happens.

On 18 November, Palo Alto Networks announced that it had agreed to acquire AI startup Applied Compute in a deal valued at $4.5 billion.

The buyer is one of the world’s leading cybersecurity companies. The startup specialises in developing, adapting and running AI models for individual businesses.

The price is not the only thing that makes the deal interesting. Palo Alto is not buying access to models it can already use. It wants to strengthen its ability to put those models to work for its enterprise customers.

The deal has not closed. The purchase price is to be paid in Palo Alto shares, not cash. The company will not have to pay $4.5 billion out of its cash reserves, but there is still an economic cost to its existing shareholders.

So why make an investment of this size?

Getting a model is one thing. Putting it to work is another.

Businesses can now get access to powerful AI models in several ways. They can buy a ready-made service, use open-weight models or build systems tailored to their own data.

“Open-weight” is worth unpacking. Having access to a model’s weights does not mean the entire technology is free to use, unrestricted or secure.

A model that can read a contract may be useful. Connecting that same model securely to a company’s contract archive, getting it to interpret commercial terms correctly, spot missing information and flag cases for human review is a different job altogether.

In the first case, you are buying access to a technology. In the second, you are paying for a process that works.

Palo Alto’s move points to the second.

According to Palo Alto, Applied Compute has technology that helps businesses refine open-weight models after their initial training and put them into use. Palo Alto wants to combine that capability with its own security infrastructure so customers can run their AI systems more securely.

The deal brings more than software: it includes researchers, specialist skills, the capacity to develop new products and the potential to sell new services.

The company has not disclosed how much of the $4.5 billion it assigns to any of those things.

The business logic is worth a look. There may be substantial value in making models work for companies without having to build those models from scratch.

As the technology gets cheaper, the responsibility costs more

An AI application can look impressive in its first demo. In seconds it can prepare a report, analyse thousands of rows or pull risky clauses out of a long contract.

The real test begins when it becomes part of the company’s daily work.

What happens when the data is wrong? If figures in two systems do not match, which one does it trust? If a supplier’s bank details change, does it go on recommending payment or stop and ask someone to check?

These questions have less to do with how smart a model is than with how the system was designed.

The licence is also only part of the cost. Integration, data preparation, security, maintenance, specialist staff and auditing can eventually cost more than the software did at the start.

A project that is cheap to launch is not necessarily good value.

There is another possibility. Rather than joining a race to develop their own models, companies may earn a better return by putting the right model to work in the right process.

That will not hold in every industry. Building a specialised model can be a genuine competitive advantage in some fields. But the idea that every company needs a model of its own deserves a closer look.

Why the numbers can look different in Türkiye

Businesses in Türkiye are making these decisions in a different cost environment.

TurkStat’s 2026 survey found that the share of enterprises with at least 10 employees using AI rose from 7.5% to 14% in a year. Usage reached 37.1% among enterprises with at least 250 employees, compared with 12.8% among those with 10 to 49.

Use is clearly rising. Among companies that were considering AI but had not yet adopted it, many cited a lack of expertise, legal uncertainty and concerns about data security.

These are not just technical issues. Each can affect what a project costs and how long it takes to pay back.

Financing conditions also change the calculation.

The Istanbul Chamber of Industry’s manufacturing PMI stood at 47.3 in October 2026, signalling a decline in output and new orders from the previous month. TurkStat’s industrial production index was up 2.6% in September compared with a year earlier. The figures compare different periods, so they do not show a uniform slowdown. But for a business facing weak demand, the question of where to put its cash becomes more pressing.

With lira-denominated commercial lending rates still high, businesses cannot ignore the gap between the cost of financing a software investment and the return they expect from it.

Take a manufacturer considering AI for purchasing and inventory management.

Halving the time it spends on reporting would help. But the bigger economic gains might come from freeing up cash tied up in excess inventory, planning supplier payments more effectively or spotting material shortages before they interrupt production.

An hour saved does not automatically put money in the bank. Better inventory management and payment terms can have a direct effect on working capital.

A company with cheap financing and ample cash reserves is unlikely to assess the project in the same way as one facing slow sales and relying on short-term loans.

For the second company, the most advanced AI system may not be the best investment.

A smaller application that connects to existing systems and delivers measurable economic results within a few months may be worth more.



The $4.5 billion question is still open

We do not yet know whether Applied Compute will prove a good investment for Palo Alto.

We will need to see how much additional revenue the acquired technology generates, what customers will pay, what integration costs and how much cash the business generates over time.

Paying with shares does not, on its own, create value. Existing shareholders will own a smaller percentage of the company, and what the acquired business contributes has to justify that dilution.

The deal does point to a possibility worth taking seriously.

A meaningful share of the value in AI may go not only to the companies that build the strongest models, but also to those that turn existing models into systems businesses can use safely, audit and earn a return from.

If so, the big question for a company will not be which model it owns. It will be which job it can now do better.

Judge an AI investment by the extra cash it generates, the risk it reduces and the capital it ties up, not by the most impressive answer it produces. Technology can get cheaper without making a bad investment any less costly.

 
 
 

Comments


bottom of page