I'm Earning 6 per cent Yield to Own China's AI Infrastructure

Three Chinese state telcos. AI infrastructure inside a Five-Year Plan chapter. And one filing from 2009 that tells me when to worry.

Beijing, China. April 2024.

I own three Chinese state telcos. They are paying me roughly 6 per cent a year while I wait to find out whether the state’s AI infrastructure build shows up in the earnings — or eats them. That is the entire trade. Everything else in this piece is the work I did to convince myself the wait is worth it.

In the United States, a telco is a pipe. Traffic flows through it, built by others. In China, the operator owns the build site, holds the state procurement, and writes the cheque. That is not a nuance. It is a different business. China Mobile (CM), China Telecom (CT) and China Unicom (CU) are not pipes for AI built elsewhere. They are the licensed infrastructure inside which the state’s computing build is being capitalised.

The last time these three were asked to build national infrastructure at scale, shareholders got less cash for six years. The 2008 to 2013 3G build-out is the analogue. The state has done the naming again. That is the risk I am underwriting against, and it has not gone away. The market has stopped treating these three like dead phone utilities; it has not yet decided whether they belong inside China’s AI+ infrastructure stack. That undecided gap is the trade.

All three are dual-listed: H-shares in Hong Kong, A-shares in Shanghai. Same underlying company, same annual report, same dividend cheque. Different price. A-shares trade at a premium that implies roughly 4 to 5 per cent yield for mainland investors. H-shares imply 5.6 to 6.5 per cent for Hong Kong investors on the same dividend. I hold H-shares for that reason.

I started buying all three in 2023 and 2024, and added again on sell-offs in 2025 and on the 2026 VAT reclassification. The reasoning then was simple: depressed prices, attractive yields, three state-owned enterprise (SOE) backbones inside the national network. A value investor did not need to believe anything else about the future to find that combination interesting. One of my mentors had bought a year earlier, at the 2022 low; his work was done. Mine was not, and I waited.

Why a Chinese telco is not a Western telco

A Chinese state telco is a licensed builder of national infrastructure under a Five-Year Plan mandate, sitting beneath a single shareholder — the State-owned Assets Supervision and Administration Commission (SASAC) — with state procurement directed to domestic providers by both regulation and incentive. A Western telco is a pipe for compute built elsewhere. Same word on the stock screen, different business behind it.

Verizon’s AI infrastructure is a low-latency fibre product for third-party hyperscalers. British Telecom’s (BT) AI thesis is cost-out inside a regulated incumbent with pension obligations. The compute is somebody else’s. Deutsche Telekom (DT) is the closest Western analogue: sovereign-compute pivot, regulated domestic procurement, premium multiple for the digital-sovereignty story. Even DT is buying capacity that others operate.

In China, the 15th FYP names the National Unified Computing Power Network as core infrastructure and writes the operators into the chapter. SASAC owns the controlling stake in all three. The shareholder, the regulator and the procurement counterparty are the same balance sheet.

The 15th FYP screen, applied to the three

China Telecom’s 2025 Annual Results Presentation guides total capex down 9.2 per cent to RMB 73 billion (~US$10.7 billion), and inside that smaller envelope computing infrastructure rises from 25 to 35 per cent of mix while network infrastructure falls from 51 to 41 per cent. Cash reallocated to where the plan says the build will happen, not added on top of where it was already going. That single slide is what Five-Year Plan execution looks like inside a corporate budget.

I have a rule on Chinese equities: I will not own one unless I can see how it rides the current Five-Year Plan. The transmission from chapter language to dividend cheque runs in five steps. The state names a priority. The Ministry of Industry and Information Technology (MIIT) defines the network architecture. The telco moves capex. The segment line shows revenue. The dividend tells me whether shareholders are still being paid while that happens. Break any one and the trade weakens.

The chapter language is unambiguous. Chapter 7 of the 15th FYP mandates the National Unified Computing Power Network: ultra-large-scale clusters, compute-energy coordination, allocation scheduled across regions. The AI Plus action lists model, chip, cloud and application as a coordinated stack. The 2026 Government Work Report names both as priorities. CT’s capex slide is what step three of the transmission looks like in a filing.

Stage three: the growth gets rails

In late April and early May 2026, all three operators stood up within a fortnight and named what they were building: a Token factory. The phrase each used varied. The structure was identical — IaaS compute at the bottom, a scheduling and model layer in the middle, Token billing at the top. Stage three of this thesis is the moment that build went from inferred to publicly committed.

For two years before that, the growth half of this thesis ran on inference. The FYP chapters were clear. The capex mix was shifting. The segment lines were appearing. But no operator had stood in front of an audience and named what they were building.

At the 9th Digital China Summit, China Telecom’s cloud subsidiary Tianyi Cloud launched what it calls a one-stop Token service system, structured around serving AI Token demand at scale. The IaaS layer is already at 91 EFLOPS of self-owned and accessed compute; the stated ambition is a full-stack national infrastructure for Token production. China Mobile followed on 8 May at its annual cloud conference in Suzhou. The chairman announced a Trillion-Token Service Trial Package. The company launched MoMA, a model management platform integrating over 300 AI models with real-time Token billing. China Mobile’s management said explicitly that Token operations cannot yet replace mobile data revenue. I found that last disclosure more useful than the headline number. It tells you the growth runway is still early, and management knows it.

China Unicom announced Token-based computing packages and said it will drive Token volume growth through computing power sales and model access.

The state has provided the framing. In May 2026, Xinhua and CCTV both described the national computing power network as the “computing version of the state grid.” Daily Token calls in China exceeded 140 trillion in March, more than 1,000 times the level at the start of 2024. Computing power has been classified as one of China’s “six networks,” alongside water, power, and logistics. This is what the 15th FYP chapters look like when they are being executed in real time.

Yield in 2023. Yield plus inferred growth in 2024 and 2025. Yield plus named, capitalised, publicly committed growth in 2026. The Token is the unit of account the industry has chosen, and the state has ratified it. Whether this translates to a sustained earnings upgrade is what the August interims will begin to answer. I do not know the answer. What follows is the sizing the arc earned.

The sector bet, in order: CM > CT > CU

I size CM largest, CT second, CU smallest. CM is the cohort’s cleanest income position — largest stack, 77 per cent payout, the operator the state would call on first if shareholder cash were needed. CT is the cloud and government-enterprise growth play, where the August 2026 interim results — the first H1 earnings print since the Token factory was named in April — are the first read on whether the Token scheduling layer carries margin. CU is the capital-discipline tail — the thinnest growth disclosure, the most room to disappoint, and the cleanest payout-convergence optionality. I own all three. The differences inside the cohort earn the sizing.

China Mobile is the scale anchor. Backbone operator, 92.5 EFLOPS of computing capacity at FP16, and the largest balance sheet of the three. Payout 77 per cent in 2025, up from 71 in 2023, with 2026 guidance flagged as stable-to-rising at the Capital Markets Day. Its growth play is what the chairman named in Suzhou: the Trillion-Token Service Package and the MoMA platform, sold on the largest stack in the cohort. If the income thesis breaks anywhere in the cohort, it breaks here last. The live alternative is not a revenue miss — it is a payout decision: CM is the operator the state would call on first if the build required shareholder cash, and it has done exactly that before.

China Telecom is the cloud and government-enterprise position. Cloud revenue RMB 120.7 billion (~US$17.8 billion) in 2025, the cohort’s first RMB 100 billion milestone. AIDC RMB 34.5 billion (~US$5.1 billion). Security revenue RMB 16.6 billion (~US$2.4 billion) as a standalone line. Its growth play is the Tianyi Cloud Token service named at the Digital China Summit: five layers from compute to application. The scheduling layer is the structural differentiator: guaranteed Token throughput sold with SLA-level commitments, not raw capacity. Whether the margin model follows is the August question. The disclosure says the model exists, not that it is profitable. CT is the cohort name where the story is easiest to believe. That is why I am hardest on it.

China Unicom is the capital-discipline tail. Smallest position. Free cash flow RMB 36.0 billion (~US$5.3 billion) in 2025, up 28.5 per cent on net profit growth of only 1 per cent. I read this as capex restraint rather than revenue momentum; the live alternative is that CU cannot find shovel-ready AI spend at its scale, which would make restraint a tell. The 2027 budget separates the two. Payout sits at 63 per cent against 77 per cent at the other two. That is where the re-rating optionality sits: if CU’s payout drifts upward toward 77 per cent, the dividend grows, and the price has to re-rate to hold the yield in line with CM and CT. Its growth play is the Token-based computing packages announced alongside the others, sold for computing power and model access, with smaller scale framed as a flexibility advantage. The growth story has the most room to disappoint and the disclosure quality is the thinnest.

I am not picking the telco that wins the computing build. I own the listed SOE layer through which part of the build flows: income prices in 2023, growth prices in 2026.

The valuation anchor: CM 11.95×, CT 13.70×, against Deutsche Telekom at 14.02×. The income re-rating from the 2021 trough — CM from 6.85× — is largely in. The next leg of return, if it comes, depends on whether the growth named in April converts to segment-line earnings inside the August interims and the FY2026 Capital Markets Days. CU sits inside the same valuation logic but trades on a free-cash-flow story rather than a clean trailing multiple, which is part of why it is the smallest of the three weightings. That is the valuation case. The reason I will not pay for it naively is what happened the last time these three were asked to build.

Why 2009 still sits in the middle of my thinking

Across the 2008 to 2013 3G build-out, China Mobile held its payout ratio at 43 per cent for six consecutive years; China Telecom held DPS flat through 2008 and 2009 at an inferred 30 to 35 per cent payout; China Unicom’s 2012 payout sat at 25 to 30 per cent against RMB 99.79 billion (~US$14.7 billion) of capex. Payout suppression of 25 to 40 percentage points lasted six years. The only thing that historically broke the income on these three was the state asking them to build something, and these are the numbers from the last time it happened.

I went to the HKEX primary filings to read what that looked like underneath the headline. CT’s DPS sat flat at HK$0.085 across 2008 and 2009 while the C-network integration ran. CU’s 2012 DPS was RMB 0.12 (~US$0.02) against that RMB 99.79 billion capex envelope. Board language across all three referenced “investment needs” and “long-term sustainable development”. Calm. Administrative. The payout ratio did the talking.

No SASAC or MIIT directive appears in any board disclosure I retrieved. The instruction may have existed. It was not in the public text.

BT cut its dividend in 2020 while funding a national fibre roll-out the UK government had defined as strategic infrastructure. When a company is the national network, shareholder cash becomes the balancing item when the build requires it. China is more direct about that balance.

That is the risk control argument for not paying 12× earnings naively.

Why 2026 is not 2009

Three things have changed.

Capex direction. In 2008, the operators were ramping into a build. In 2026, total capex is flat to declining across the cohort. CM cut 2025 capex 8 per cent to RMB 150.9 billion (~US$22.2 billion). CT guided 2026 down 9.2 per cent to RMB 73 billion. CU’s free cash flow rose 28.5 per cent on net profit growth of only 1 per cent, with the capex burden falling. The mix inside the smaller envelope is shifting to AI infrastructure, but the envelope itself is not exploding.

Revenue visibility. The 3G build created the network first and the monetisation later. The AI+ build is already showing up in cloud, AIDC, computing services and security revenue, year by year, in the segment lines. That is a different sequence inside the income statement.

SOE reform direction. The SASAC value-creation framework, often referenced as Document 79, and the later SASAC market-value management push, are distinct documents and I treat them so. The first is a profitability and value-creation campaign for central SOE managers. The second pushes listed central SOEs to manage market value, capital returns and shareholder communication. CM’s payout moved from 71 per cent in 2023 to 77 per cent in 2025 inside that framework. The trajectory tracks the framework dates, CT and CU have moved in the same direction over the same window, and management cites shareholder-return language at the Capital Markets Day. Declining capex is the live alternative driver. I read the SASAC direction as the most plausible cause; I cannot separate it cleanly from the cash-flow effect on the public disclosure.

That is not a guarantee. It is the changed incentive set I am underwriting on top of the income floor.

What I am watching for, on a five-year horizon

I am watching three things across three release windows: capex mix at the August 2026 interims and the FY2026 Capital Markets Days; payout trajectory across the same prints; the 2027 budget disclosures, where guided intent either becomes durable or does not. If the capex mix and payout trajectory hold through that window, the cohort thesis is intact. The horizon is the FYP duration, not a single earnings print.

I am a Bayesian investor on long positions. I update on what the data tell me, in proportion to how informative each release actually is. The August interims are one input, not the picture.

On CM, the indicator is the disaggregation of the AI services line. If management starts separating infrastructure services (computing, cloud, security) from bundled content (MIGU, e-commerce), the disclosure quality improves and the underwritten thesis gets cleaner. If the bundle hardens, the underwritten thesis weakens whether or not the headline number grows.

On CT, the indicator is the capex mix continuing to shift toward computing infrastructure inside a flat-to-declining envelope. The 26 per cent to 35 per cent guided shift for 2026 is the signal of intent. The 2027 budget will tell me whether the intent is durable.

On CU, the indicator is payout convergence. If the payout drifts from 63 per cent toward 70 per cent, the yield compresses through price and the re-rating optionality earns out. If the payout stays low while capex obligations rise, the income story weakens before the growth story has had a chance.

Across all three, the indicator that would change my view of the cohort, not the timing, is SASAC’s central-SOE performance assessment. CU’s 2025 Sustainability Report ties performance assessment for central enterprise managers to AI computing power targets. If a future round of criteria encodes computing-power floors as a hard performance measure, the cohort-wide capex discipline is at risk and I would revisit the sector weight before any individual name. That is the single most important unresolved data point in the position.

The trade I am still in

Adding now is harder than buying in 2023. The 2023 buy was easy. Price was low, yield was high, and the underwriting required nothing more than valuation discipline. The 2026 hold is harder. The easy part has played through, and the remaining return needs a state-mandated growth story to deliver where chapter language and filing disclosure currently disagree.

That is the trade I am still in. The work bought me the position. The work is what tells me whether to keep it. The 2009 filing stays on the desk for the same reason: not because history repeats cleanly, but because it tells me what to watch when a national infrastructure cycle reaches the shareholder.

---

As of the date of publication, I hold positions in China Mobile (HKEX: 941), China Telecom (HKEX: 728), and China Unicom (HKEX: 762). Positions may change after publication without notice. Cohong Lane is a periodical publication made generally available to the public; this is disclosure of my positions, not a recommendation to buy, sell, or hold any securities. Full disclaimer · About Philip.