The Balance Sheet of AI
Updated: 3 days ago
How strongly you believe in a technology is revealed not by how much you spend on it, but by what you are still willing to finance when money becomes expensive.
AI has now reached precisely that point.
According to Reuters’ reporting on Anthropic’s IPO documents, Broadcom could provide Anthropic with up to $42 billion in financing. Behind that funding lies a much larger infrastructure commitment: Anthropic’s enormous need for TPU capacity over many years. Broadcom would be more than a technology supplier; it could also become a financier and, under certain conditions, a shareholder.
This arrangement may excite the technology world. From a financial perspective, it looks somewhat more unusual.
A company sells you capacity. Then it gives you money to buy that capacity. The money it provides may later convert into equity in your company.
At that point, you cannot ask only, “How strong is demand for AI?”
You also need to ask:
How much of this demand can support its own financing?
Vendor financing is not new. It has been used for years in sectors ranging from aircraft to telecommunications equipment. But the capital intensity of AI infrastructure takes the issue to a different scale.
Chips, data centres, energy, long-term capacity contracts and financing are increasingly becoming parts of the same economic chain.
So the AI race is no longer only about building better models. Another race is underway:
Whose balance sheet will carry this enormous need for capital?
In Türkiye, the same question takes on a different meaning.
For most companies here, the issue is not how to finance $42 billion of AI infrastructure. It is a far more everyday—and much tougher—question:
For every lira I allocate to AI, what other use of my working capital am I giving up?
This difference matters.
TurkStat’s latest figures show that AI adoption in Türkiye is not declining; it is rising rapidly. Among enterprises with ten or more employees, the adoption rate increased from 7.5% to 14% in a year. Among large companies, it exceeds 37%.
But I find another figure more interesting than the adoption rate:
Around two-thirds of companies using AI are using open-source solutions.
This detail may tell us something about capital allocation as well as technological preferences.
Turkish companies are not rejecting AI. They may be experimenting with ways to incorporate it into existing work at the lowest possible initial cost.
That need not be a sign of falling behind.
Sometimes capital scarcity is a ruthless filter that weeds out poor investments.
The problem is that, if the filter becomes too severe, it eliminates good investments too.
This is where the picture for Turkish companies becomes more complicated.
In an economy with a 37% policy rate, continuing producer-price pressures and a manufacturing PMI still below the 50 threshold, the calculus of capital expenditure inevitably changes.
But the same data do not suggest that the economy has stopped. Capacity utilisation is around 74%. Economic confidence is above 100. Production continues, companies are operating and investment has not disappeared.
So saying “nobody can afford to invest” is too easy.
A more accurate statement might be:
Investment has less room for error.
The question a company asks when spending TRY 10 million today differs from the one it asked when capital was much cheaper.
As the payback period lengthens, the opportunity cost grows.
If collections are already under pressure, inventory turns are slowing, borrowing costs are high and working capital is expensive, a technology investment promising “strategic benefits in three years” is harder to defend in the boardroom than it once was.
The first AI winners in Türkiye may therefore be less spectacular projects.
It might be an application that halves the time a sales team spends preparing quotations.
It might be a system that detects collection risks earlier.
It might be a small model that identifies pricing anomalies in procurement.
It might be a tool that reduces a three-day financial consolidation task to a few hours.
None of these may bring a technology-conference audience to its feet.
But from a CFO’s perspective, they have another quality:
Their path to cash-flow impact is shorter.
Again, caution is needed.
A quick payback and lasting value are not the same thing.
If a company consistently selects only projects that repay their cost within six months, longer-term investments in data infrastructure, process standardisation, employee capabilities and enterprise technology architecture may be postponed indefinitely.
That can produce an odd result:
More people in the company use AI, but the company’s AI capability does not grow at the same pace.
This is probably one of the central risks for Türkiye.
In TurkStat’s data, the biggest obstacle for companies considering AI but not yet using it is not money: it is a lack of expertise. Legal uncertainty and concerns about data security and privacy follow closely.
This takes the issue beyond a simple “digitalisation budget”.
You can give a company the best model in the world.
If it is fed incorrect data, applied to the wrong problem or used without anyone capable of questioning its output, the result may be faster mistakes rather than greater productivity.
Perhaps this is the real financial question for AI investment in the period ahead.
Not how powerful a model we buy, but
which decisions it enables the company to make better.
For American technology companies, the challenge is increasingly about finding ways to finance data centres, chips and energy.
Türkiye faces a different optimisation problem:
Allocating scarce capital where it will have the greatest impact on decision quality and cash.
That is why the same technology need not produce the same economic outcome in the two countries.
Cheap or abundant capital gives technology scale.
Expensive capital forces technology to make choices.
AI’s real test in Türkiye over the next few years may not be how many companies use it. A more important measure will be whether companies can turn AI spending from operational curiosity into measurable economic value.
Because, in the end, the balance sheet records the outcome, not the model.
Comments