Everyone is talking about Saudi AI. Nobody is asking who pays for it.

Saudi Arabia’s AI ambitions are becoming difficult to ignore.

PIF created HUMAIN. Partnerships have been announced with Nvidia, AMD, xAI and some of the world’s largest technology and infrastructure companies. HUMAIN is targeting more than 6GW of data-centre capacity over the coming decade.

Most of the coverage understandably focuses on the scale of that ambition.

I am more interested in the balance sheet.

Six gigawatts of data-centre capacity has to be financed, built, connected to the grid, filled with equipment and ultimately used by customers willing to pay enough for the infrastructure to generate an acceptable return.

That makes the Saudi AI story as much a capital-allocation story as a technology story.

And for a Tadawul investor, that distinction matters.

Saudi Arabia already has a growing data-centre industry, but HUMAIN’s plans would take the market to an entirely different scale.

The company is targeting approximately 6GW of capacity over the next decade and has secured 211 plots of land across Saudi Arabia for its expansion. Facilities in Riyadh and Dammam were initially expected to provide up to 100MW each.

There is also a much larger infrastructure plan emerging around Riyadh. In May, HUMAIN tendered infrastructure works for a proposed 6GW hyperscale AI data-centre campus on a 24-square-kilometre site in Al-Saad, east of Riyadh. The plans include major electricity distribution infrastructure, substations and a 2,000MVA bulk supply point.

The announcements are therefore becoming increasingly tangible.

But announced capacity and operational capacity are not the same thing.

Madar Strategy recently audited 13 Saudi and Emirati AI infrastructure projects, programmes and cloud regions representing roughly 4,000MW of headline capacity. Its conclusion was not that none of those projects exists or will be built. It was that the public evidence does not yet allow investors to convert those headline megawatts into a reliable number for delivered compute.

Of the cases examined, only one explicitly defined its capacity as IT load. None publicly disclosed firm connection megawatts, and Madar found no project-specific public energisation record across the audited cases at its July 22 evidence cut-off.

That is an important distinction.

A press release tells you what someone intends to build.

An energised data centre tells you what has actually been delivered.

The gap between those two things is where execution risk lives.

This is where the story becomes more interesting to me.

AI infrastructure is extraordinarily capital intensive. The cost is not simply GPUs. Data centres require land, substations, transmission infrastructure, cooling systems, fibre, construction and enormous amounts of power before a single customer starts paying for compute.

According to Alvarez & Marsal estimates reported by Semafor, financing even half of Saudi Arabia’s announced data-centre pipeline could require as much as $32 billion of debt. That is more than Saudi banks are likely to want to assemble on their own.

We are already beginning to see what that financing architecture might look like.

Reuters reported in May, citing two people familiar with the matter, that HUMAIN had selected Goldman Sachs to advise on a financing package worth at least SAR 20 billion for data centres representing approximately 2GW of capacity. HUMAIN and Goldman Sachs declined to comment.

That matters because Saudi banks are not starting from empty balance sheets.

They are already financing mortgages, corporate expansion, infrastructure and the broader Vision 2030 investment programme. AI infrastructure now arrives as another large category of credit demand competing for capital.

International banks, infrastructure funds, private credit and potentially institutional investors will therefore almost certainly have roles to play alongside domestic lenders.

And they will not provide that capital simply because AI is strategically important.

They will want a return.

This is the question I think gets lost whenever another gigawatt announcement appears.

Building a data centre is not the same thing as creating an economically attractive data centre.

Someone has to pay for the land, construction, grid connection, equipment and financing costs. Once operational, the facility needs enough utilisation, at sufficiently attractive pricing, for the return on invested capital to exceed the cost of financing it.

That hurdle has become more demanding.

Long-term US government bond yields have recently been above 5%. That matters because global infrastructure investors are constantly comparing returns across opportunities. If relatively low-risk assets offer attractive yields, a project carrying construction risk, technology risk, utilisation risk and geopolitical risk has to compensate investors accordingly.

There is another complication specific to AI: technology moves extraordinarily quickly.

A toll road built today can still be a toll road in 30 years. A data centre filled with today’s most advanced hardware faces a very different technology cycle. GPUs become obsolete. Cooling requirements change. Compute density increases. Customers shift between models and providers.

That does not make AI infrastructure a bad investment. It means the price of capital and the structure of customer contracts matter enormously.

A facility backed by a 10- or 15-year hyperscaler agreement is a very different credit proposition from one being built in anticipation of future demand.

Which brings me back to a question I raised previously about STC’s Center3 strategy.

The gigawatt headline is not the number that ultimately matters. The ROIC is.

How much capital is invested? How much debt sits against it? What utilisation does the facility achieve? What are customers contracted to pay? And what return remains for shareholders after financing costs?

Until we have those answers, I am treating capacity announcements as evidence of ambition rather than evidence of returns.

This is perhaps the most interesting part of the Saudi AI story.

The obvious investment is not necessarily the best investment.

AI demand creates demand for compute. Compute creates demand for data centres. Data centres create demand for electricity, grid infrastructure, cooling, fibre, construction and financing.

The further down that chain you go, the less glamorous the businesses become.

They may also become more investable.

I already own three businesses that sit somewhere along this chain, although none of them was bought specifically as an AI investment.

Al Rajhi and SNB

Saudi banks are obvious potential beneficiaries of the financing requirement. Large infrastructure programmes create loan growth, fee income and opportunities to participate in syndicated financing.

But there is a second side to that trade.

More lending is only attractive if it is appropriately priced. Banks still have to manage funding costs, capital ratios and concentration risk. I therefore would not automatically treat every SAR 20 billion infrastructure announcement as positive for bank shareholders.

The question is whether they earn an adequate spread for the risk they are taking.

STC

STC is closer to the physical infrastructure.

Connectivity sits underneath every data centre, while Center3 gives STC exposure to data-centre and subsea-cable infrastructure. That makes STC one of the more direct listed ways I currently have of participating in Saudi Arabia’s digital infrastructure buildout.

It also makes capital allocation more important.

The more STC moves from a traditional telecom operator into capital-intensive digital infrastructure, the more closely I want to watch the returns generated by those investments.

Aramco

Aramco is much further removed, and I would not describe my Aramco position as an AI investment.

But Saudi Arabia’s ability to offer abundant and relatively low-cost energy is part of the country’s pitch as a data-centre location. AI may be digital at the customer end, but the infrastructure underneath it is extremely physical and extremely energy intensive.

Aramco sits somewhere underneath that energy system.

That is enabling exposure, not an AI thesis.

And I think keeping those distinctions clear matters.

I believe Saudi Arabia will build a significant AI infrastructure industry.

The government commitment is real. PIF’s capital is real. The partnerships are real. Land is being secured, infrastructure is being tendered and financing discussions are moving forward.

But that is only the first half of the investment case.

The second half is much less exciting to put in a press release: cost of capital, construction timelines, power connections, utilisation, customer contracts, debt service and return on invested capital.

Those numbers will eventually tell us whether Saudi Arabia has simply built a lot of AI infrastructure or built a profitable AI infrastructure industry.

There is a big difference between the two.

For now, I am less interested in counting announced gigawatts and more interested in watching how those gigawatts move from announcement, to financing, to construction, to energisation and finally to cash flow.

That is where I think the real Saudi AI investment story will be found.

Sources: HUMAIN and company announcements; Reuters; Semafor/Alvarez & Marsal; Madar Strategy AI infrastructure audit; Data Center Dynamics; MEED. Company information from Saudi Exchange disclosures.

The Quiet Compounder is for educational purposes only and is not financial advice. Always do your own research or consult a licensed advisor before making investment decisions.