Morgan Stanley: China's Cumulative AI Investment to Reach ¥8.5 Trillion Over Five Years, Domestic Chip Shipments to Hit 11 Million Units by 2030
nashnova research
Morgan Stanley projects cumulative Chinese AI capex of RMB 8.5 trillion from 2026 to 2030, with domestic chip shipments growing tenfold to 11 million units. This means → a massive buildout funded by three distinct investor classes is now being priced.
RMB 8.5 trillion — where does it go and who pays?
Three investor classes split the bill: hyperscale cloud operators ~RMB 6.3 trillion, emerging compute-cloud firms ~RMB 1.5 trillion, telecom carriers ~RMB 653 billion. Cloud operators dominate.
Roughly RMB 2 trillion is earmarked for overseas expansion — it cannot all be counted as domestic AI procurement, nor equated with the AI chip market.
This means → when you see "8.5 trillion," unpack it first: onshore buildout, offshore footprint, and supporting infrastructure each take a share. No single segment owns the whole number.
Can the cloud giants actually fund this?
Morgan Stanley estimates annual capex for Alibaba at RMB 217–259 billion, Tencent at RMB 193–200 billion, Baidu at RMB 20–24 billion.
Tencent's cash coverage is relatively solid. Whether Alibaba's local-services spending or Baidu's search revenue stabilizes could shift the cash-sufficiency picture.
The ~RMB 2 trillion offshore spend competes with overseas debt repayment, dividends, and buybacks — and may require asset sales, bond issuance, or equity financing.
In plain terms = being able to afford it is not the same as affording it comfortably — offshore funding pressure is real.
Why would domestic chip shipments grow tenfold?
Domestic AI chip shipments are projected to rise from 1.1 million units in 2025 to 11 million by 2030, with market value expanding from RMB 94 billion to RMB 646 billion.
Domestic chips are expected to account for 70–85% of server deployments. Morgan Stanley notes that localization's economic value is first about "increasing deliverable supply," only second about improving unit cost.
This means → at this stage, domestic chips matter because they exist to be deployed — not because they are cheaper. Availability outranks cost-efficiency.
Why are deployment costs rising, not falling?
Domestic inference deployment costs — including memory, CPUs, and networking — rise from ~$19 billion per GW in 2026 to $22 billion per GW in 2027.
The drivers: supercluster configuration upgrades and memory price increases. Morgan Stanley explicitly flags that the model has not fully priced in further memory inflation; capex estimates may be revised upward.
In plain terms = localization is advancing, but component inflation is pushing costs higher. The equation "domestic substitution = savings" does not hold yet.
Three business models — how different are the returns?
Morgan Stanley benchmarks ROIC — return on invested capital, measuring how much each dollar of investment earns back — for three compute business models: GPU leasing ~13%, proprietary model services ~19%, hosting third-party models via API ~29%.
Using a next-generation domestic server (assuming inference performance at 30% of the overseas benchmark), third-party model-service ROIC drops to ~9.1% with a payback of roughly 3.5 years; if the performance ratio shifts between 20% and 40%, ROIC ranges from 0.3% to 17.9%.
This means → business model matters more than hardware choice for returns — but Morgan Stanley stresses these are project-level model outputs, not realized company returns, and upgrading the model does not guarantee higher profits.
Who does Morgan Stanley favor, and where is the risk?
Preferred names: Alibaba, Tencent, Kingsoft Cloud, VNET, plus MiniMax, Zhipu, Cambricon, Iluvatar CoreX, and Hygon Information among domestic compute players.
Morgan Stanley cautions that the ROIC figures must be weighed alongside cost of capital, competitive pressure, and current valuations — they are not standalone buy signals.
This reflects two verification checkpoints for whether this buildout converts to stable cash flow: can domestic chip performance keep improving, and can paid throughput cover upfront costs.
市场有风险，内容仅供研究参考，不构成投资建议。