SK Telecom Co., Ltd. (KRX:017670, NYSE:SKM) titled its second-quarter release, roughly translated, "solid telecom, accelerating AI DC." Management has spent two years telling investors that artificial intelligence infrastructure — data centres, cloud, the compute layer underneath Korean AI services — is what turns a no-growth carrier into something else. The market has partly believed it: the shares have nearly doubled off their 52-week low and now trade at 1.39x book, against 0.72x at KT.
Data centres are the most capital-intensive thing anyone is building right now. So the obvious place to check the story is the capital expenditure line.
It has been falling for three years.
Capital spending was ₩2,973.9bn in 2023. ₩2,487.4bn in 2024. ₩2,206.6bn in 2025. In the first half of 2026 it was ₩809.97bn, which annualises to roughly ₩1.62tn.
From 2023 to that run rate, that's a decline of about 45%.
As a share of revenue the series is starker, because revenue also fell: 16.9% in 2023, 13.9% in 2024, 12.9% in 2025, and 9.3% in the first half of 2026. A network operator spending 9.3% of revenue on capital is spending at the low end of what global carriers do in the trough between build cycles. It is not what a company constructing data centres spends.
The asset base agrees. Property, plant and equipment fell from ₩11,902.2bn at the end of December to ₩11,221.9bn at the end of June — ₩680.3bn in six months, a 5.7% decline. Intangible assets other than goodwill fell from ₩1,710.6bn to ₩1,370.2bn as spectrum licences amortise. Goodwill was unchanged at ₩2,072.5bn, so nothing was acquired.
Depreciation is running ahead of investment across the whole fixed asset base. That is the opposite of a build.
The obvious rebuttal is that modern data centre capacity often doesn't appear as capex. Operators lease space, take capacity from a developer, or build through a joint venture funded by a partner. Under IFRS 16 the leasing version shows up as a right-of-use asset and a lease liability, and the joint venture version shows up in investments accounted for using the equity method.
Neither happened.
Lease liabilities fell. Current lease liabilities went from ₩407.96bn to ₩387.91bn and non-current from ₩1,117.84bn to ₩988.32bn — ₩1,525.8bn down to ₩1,376.2bn, a decline of ₩149.6bn over the half. Investments accounted for using the equity method rose from ₩2,238.5bn to ₩2,356.3bn, up ₩117.8bn, which is small enough to be routine.
So on the consolidated balance sheet, across the three places a data centre build could appear — owned assets, leased assets, joint ventures — SK Telecom's capacity commitment went down in the first half of 2026, not up.
I want to be careful about what that does and doesn't prove. It doesn't prove the AI business isn't real: SK Telecom reports data centre revenue growing, which means capacity exists and is being sold. It doesn't prove nothing is planned; capital commitments disclosed in the notes to the half-year report would show contracted future spending, and I haven't read them. What it establishes is that as of 30 June 2026, the capital formation a large build produces had not yet started appearing.
The cash is going out rather than in, and the numbers are good.
Operating cash flow in the first half was ₩2,271.6bn against ₩810.0bn of capital spending. Free cash flow of ₩1,461.6bn in six months, which annualises to roughly ₩2.92tn — the strongest run rate in four years. For comparison: ₩1,973.3bn in 2023, ₩2,599.9bn in 2024, ₩1,717.3bn in 2025.
Almost none of that improvement came from the business. Operating cash flow in 2024 was ₩5,087.3bn and in 2025 ₩3,923.8bn; the first-half figure annualises below the 2025 level. The free cash flow improvement is the capex line falling faster than the operating line.
Where it went: financing outflows of ₩979.6bn in the half, total liabilities down from ₩17,152.5bn to ₩16,935.2bn, and a quarterly dividend held at ₩830 a share for two consecutive quarters, which annualises to about ₩713bn a year across the share count.
That is a well-run deleveraging and payout programme. It is also, precisely, the financial profile of a company harvesting a mature asset rather than building a new one.
Three arguments, and the first two are strong.
The 5G build is finished. Korea deployed 5G earlier and faster than almost anyone, and the capital cycle that funded it ended. Every large carrier in every developed market is in the same trough, and falling capital intensity is the correct behaviour, not a failure of ambition. Judging SK Telecom against a hypothetical data centre build ignores that the baseline should be falling anyway.
Second, the structures genuinely might be invisible here. Korean AI data centre projects have typically been announced as partnerships in which the carrier contributes land, power interconnection and operational capability while a financial or hyperscale partner funds construction. If SK Telecom's contribution is an existing site and a power contract, the balance sheet shows very little while the revenue and the strategic position are real. My lease and equity-method checks catch some structures and not all of them.
Third, and weakest: the spending may simply be ahead. Announced projects take two to three years from agreement to energised capacity, and a build commencing in 2027 would leave 2026 capex looking exactly like this.
The reason I still think the capex line is the right thing to watch is that all three of those explanations are testable and none of them has been tested yet. If the second is true, the revenue and margin from data centres should scale without SK Telecom's capital, which would be a genuinely excellent outcome and should be visible in segment disclosure. If the third is true, capex turns up. If the first is the whole story, then SK Telecom is a mature carrier trading at 1.39x book and 13.7x annualised earnings while KT trades at 0.72x and 8.1x, and the premium is being paid for a business that hasn't been built.
Third-quarter capital expenditure against the ₩810.0bn first-half figure. A nine-month number above ₩1.6tn would mean the run rate is turning up. Below ₩1.3tn and the full year comes in under ₩1.8tn, which would be the fourth consecutive annual decline and a fifth of revenue lower than 2023's intensity.
Second, property, plant and equipment against ₩11,221.9bn. Falling fixed assets and a growing data centre business cannot both continue indefinitely.
Third, lease liabilities against ₩1,376.2bn. This is the cleanest single indicator of off-balance-sheet capacity being taken on, and it has been going the wrong way for the AI story. A sharp increase would tell you the build is happening in a structure the capex line never shows — which would resolve the contradiction in the company's favour, and is the outcome I'd assign the highest probability to if the revenue keeps growing.
kstock reads DART every morning and writes up what moved — the contract, the buyback, the number that does not add up. The daily post and a Saturday roundup, by email.