As global AI frontrunners, including Anthropic, OpenAI, and Elon Musk's SpaceXAI, collectively urge a slowdown in frontier AI model development, global equity markets are simultaneously reassessing growth expectations for AI compute infrastructure investments. This, combined with rising oil prices and renewed Federal Reserve rate hike risks, has left AI-themed stocks broadly weak. SK Hynix fell over 5% in early Korean trading, Samsung Electronics dropped more than 3%, and the KOSPI index, often dubbed the "AI compute bellwether," closed Monday down over 3%. However, many veteran Wall Street analysts indicate these latest developments will not have a lasting impact on the industry and are unlikely to derail the long-term bull case for AI compute investments.

Amid the global semiconductor sell-off triggered by "AI slowdown" concerns, Morgan Stanley's latest research report, titled "China AI Pathway: From 8.5 Trillion Yuan Capex to Triple Compute Capacity," reveals that China's AI compute industry is poised to enter a new phase driven by a massive expansion of domestic chip supply, cloud AI capital expenditure, and the commercialization of AI inference. The report projects cumulative investments of approximately 8.5 trillion yuan (around $1.3 trillion) from 2026 to 2030 by hyperscale cloud vendors, emerging AI compute cloud firms, major telecom operators, and other new AI entities. This spending is expected to lift China's IT power capacity from 26GW in 2025 to 81GW by 2030. The "triple AI infrastructure spending" referenced by Morgan Stanley precisely implies capacity reaching roughly 3.1 times current levels; GW measures IT-side equipment power capacity, with actual total AI compute cluster capability also depending on chip performance, memory bandwidth, high-speed optical interconnect efficiency within clusters, and overall power system energy utilization.

The report indicates that this investment covers both domestic construction and approximately 2 trillion yuan in overseas expansion, so it cannot all be considered domestic AI compute infrastructure procurement orders. Morgan Stanley's investment thesis centers on three interconnected links: improved supply of domestic GPUs and ASICs, which allows previously chip-constrained demand to materialize; major cloud vendors supporting the supply chain through capex, prepayments, long-term contracts, and equity investments; and cloud platforms converting equipment investments into recurring revenue via GPU leasing, model services, and third-party model hosting. Additionally, the analyst team notes that business models determine return differences—the benchmark ROIC for China's compute infrastructure solutions rises from approximately 13% for self-owned GPU leasing to about 19% for proprietary model services and roughly 29% for third-party model hosting. A next-generation domestic-substitute chip hardware solution yields around 9.1%, still dependent on performance and cost improvements.

The report's most valuable contribution is connecting "capital investment—available compute—paid inference—capital returns" into a testable investment framework for China's AI supply chain. Specifically, Morgan Stanley calculates benchmark ROICs under different business and hardware assumptions for the Chinese market at approximately 13% for self-owned GPU/ASIC compute leasing (enterprises buying servers and renting GPU capacity, i.e., IaaS on owned infrastructure), 19% for running proprietary models on owned compute (enterprises owning infrastructure and operating their own models, charging via model APIs), and 29% for running third-party AI models on owned compute (operating third-party models on owned infrastructure, providing API services with revenue sharing to model providers). If the third-party model service adopts the report's assumed next-generation domestic servers (replacing core AI hardware with hypothetically next-gen domestic chips), benchmark ROIC drops to approximately 9.1%. These figures reflect project-level unit economics and should not be interpreted as realized company returns or a guarantee that business model upgrades necessarily boost profitability.


Where the 8.5 Trillion Yuan Goes: Domestic Chips Open Expansion Space, Real Cash Flow Determines Construction Pace

Internet giants are the absolute main force in capex, but three types of investors shoulder different roles. Morgan Stanley projects hyperscale cloud vendors and major internet companies will invest approximately 6.3 trillion yuan cumulatively from 2026 to 2030, reaching 995 billion yuan in 2026, up 121% year-over-year and higher than the 41% growth forecast for US peers. Spending rises a further 23% to about 1.2 trillion yuan in 2027, reaching roughly 1.4 trillion yuan by 2030. At the company level, Morgan Stanley estimates Alibaba's annual capex from 2026 to 2030 will range between 217 billion and 259 billion yuan, Tencent between 193 billion and 200 billion yuan, and Baidu between 20 billion and 24 billion yuan. Alibaba's more aggressive spending reflects revenue visibility from AI GPU compute infrastructure services and model services; the report notes its June quarter MaaS annualized revenue run-rate has exceeded 10 billion yuan, with expectations of reaching 30 billion yuan by year-end. Morgan Stanley clarifies this figure is an annualized run-rate, not full-year realized revenue.

The second pillar comprises emerging compute cloud firms, with five-year investments of approximately 1.5 trillion yuan, scaling from 234 billion yuan in 2026 to 363 billion yuan by 2030. Over 95% of this spending is expected to go toward high-performance AI server cluster procurement centered on AI GPUs/ASICs, serving as supplemental high-end compute capacity and leasing supply. The third pillar is telecom operators, with five-year compute-related capex of approximately 653 billion yuan, rising from about 81 billion yuan annually to 179 billion yuan, primarily serving government, state-owned enterprise, and sovereign AI demand. Operators allocate roughly half their compute budgets to supporting infrastructure like data center facilities and networks, with over 75% of remaining equipment spending directed at GPUs or ASICs. Thus, the 8.5 trillion yuan encompasses servers, storage, networking, data centers, and overseas construction—directly equating it to AI chip market size would significantly overstate chip revenue.

The economic value of localization first manifests as increased deliverable supply, then as improved unit compute costs. Morgan Stanley forecasts domestic AI chip shipments rising from 1.1 million units in 2025 to 2.4 million in 2026, 4.8 million in 2027, and 11 million by 2030. Corresponding market size expands from 94 billion yuan to 646 billion yuan, with domestic chips expected to account for 70%–85% of server deployments during 2025–2030. On the capacity front, of the new 55GW, leading cloud vendors contribute 34GW, operators' internal and government/enterprise demand adds 9GW, and other internet companies and AI labs contribute 12GW; China's top cloud vendors also plan approximately 13GW of new overseas capacity, bringing global additions to 47GW. The report shows third-party data center annual orders rising from 6.1GW in 2026 to 11GW by 2030, with about 70% of incremental orders landing in western regions like Inner Mongolia and Ningxia. Engineering-wise, this creates shared expansion opportunities for domestic GPUs, ASICs, servers, optical interconnects, power distribution, and liquid cooling, but "cheap electricity and facilities" cannot automatically offset the costs of less efficient computing. The report estimates domestic inference equipment deployment costs rise from about $19 billion per GW in 2026 to $22 billion per GW in 2027, driven by supercluster configurations and memory price increases; this metric includes memory, CPU, and networking components, not just chip prices. The report also explicitly states that models have not yet fully priced in subsequent memory price hikes, potentially leading to upward capex revisions.

For Chinese AI companies, the real optimization target is the cost per million effective tokens after meeting latency and reliability requirements, plus the paid throughput generated per watt and per yuan of capital. Chip counts, power capacity, and commercial output must be measured separately. Financing capability will determine which companies can convert expansion plans into stable operating assets. Morgan Stanley believes Alibaba, Tencent, and Baidu's domestic investments can generally be supported by operating cash flow and cash reserves, with Tencent being relatively stable; Alibaba's local services competition spending and Baidu's search business stabilization could still alter cash coverage capabilities. The approximately 2 trillion yuan in overseas capex must compete with offshore debt repayment, dividends, and buybacks, potentially requiring portfolio sales, bond issuance, or equity financing. Emerging compute cloud firms show significantly higher debt sensitivity, with finance leases typically covering 30%–40% of investments at around 5% lease rates, bank loans at 3%–4%, and customer prepayments alleviating construction-period pressure—though these should not be mistaken for recurring operating profits. The report projects independent data center operators' five-year construction capex at approximately 569 billion yuan, with project loans typically covering 60%–80% of costs, operating cash flow covering only about 10%–15% of capex, and the remainder relying on REITs, asset securitization, and equity. This constitutes a separate layer of infrastructure estimation that should not be simply added to the 8.5 trillion yuan without addressing overlapping scopes.

Morgan Stanley's forecasts for North America's Big Four tech giants provide a useful comparison: their combined operating cash flow is projected to grow from $739 billion in 2026 to $1.23 trillion by 2028, while incremental debt needs decline from $238 billion to $90 billion. However, these cash flow expectations include advertising, e-commerce, and traditional cloud businesses and cannot be entirely attributed to AI. Meanwhile, Big Four data center capex reaches $1.47 trillion and $1.64 trillion in 2027 and 2028 respectively, making the more accurate conclusion that 2028 investment growth slows to about 12%, with total spending still increasing by roughly $170 billion—demonstrating that investment growth deceleration can coexist with rising compute usage. The key lies in whether built assets can be promptly activated, attract customers, and generate sustained cash.


From "Buying Compute" to "Selling Outcomes": How the Astra Model Reshapes Storage Chip Demand and AI Investment Return Pricing

OpenAI's recently launched GPT-6 Astra model and the RSI technical pathway emphasized by AI leaders are expected to become two core drivers of exponential AI compute demand growth. Stronger AI models, broader adoption of AI application tools, and next-generation training paths with more robust compute requirements are strengthening the case for sustained AI infrastructure demand growth. Astra's investment significance lies in improving complex task success rates and economic viability, encouraging enterprises to deploy more agents and handle more specialized tasks. Morgan Stanley's recent emphasis on "shifting from demand debates back to the physical supply constraints of the AI theme" captures exactly how the frontier Astra model generates a new round of AI compute resource demand expansion. OpenAI's product head stating that demand is so unprecedented the company may pause new Pro subscriptions serves as a critical signal of AI compute service capacity strain.

As AI evolves from single-turn Q&A to programming, research, office work, and cross-software operations, a single paid task may involve planning, multiple model calls, tool execution, result verification, and persistent state saving. Broader enterprise adoption also increases the number of concurrently running tasks. However, stronger models may accomplish the same task with fewer tokens and retries—OpenAI's Astra release materials disclosed some efficiency improvements—so "stronger models" cannot directly translate to "every task consumes more compute." A more reasonable growth mechanism is: declining task costs and higher success rates make previously uneconomical work automatable; total compute demand accelerates only when new task demand exceeds per-task resource consumption reductions. As noted in Goldman Sachs' meeting notes from the SanDisk Technology Conference, long-term agreements cover approximately 50% and 67% of planned shipments for fiscal 2027 and 2028, reflecting order visibility; whether these convert to higher profits depends on product certification, customer share, pricing, and per-bit costs.

For leaders in China's AI compute supply chain, the same mechanism means domestic GPU and ASIC expansion will drive demand for complementary memory, SSDs, high-performance Ethernet networking infrastructure, data center CPUs, optical interconnect chips, and software optimization. However, which suppliers secure these orders still depends on specific procurement systems and product competitiveness. Morgan Stanley's analyst team states that whether massive AI investment scales into strong profitability trajectories ultimately depends on the combined effects of paid throughput, unit pricing, and capital efficiency. Morgan Stanley's benchmark scenarios are: approximately 13.2% ROIC for self-owned GPU leasing, about 19.4% for proprietary model services, and roughly 29% for hosting third-party models on owned compute with API services; the third option includes a 10% model revenue share and assumes performance discounts for running third-party models, so it should not be interpreted as simply reselling external APIs. The relatively higher upfront costs associated with overall high-end AI server cluster procurement, performance ratios, token throughput, and utilization-related expenses are the primary reasons GPU leasing returns trail US benchmarks. The next-generation domestic server benchmark assumes a price of 4 million yuan with inference performance at 30% of overseas comparison solutions, yielding 9.1% ROIC and a roughly 3.6-year payback period; when the performance ratio varies between 20%–40%, ROIC ranges from approximately 0.3% to 17.9%. This highlights that the value of HBM3e, software-hardware co-design, and cluster optimization lies in increasing billable throughput. Prefill/decode separation may be more effective for long-input tasks, but actual benefits are also affected by first-token latency, output latency, and KV cache transfer—so it cannot yet be concluded that domestic projects universally "cross the profitability threshold."

In terms of investment strategy, Morgan Stanley advises investors to simultaneously examine hardware suppliers' order quality and capital discipline, cloud platforms' paid utilization and unit gross margins, and model companies' ability to have revenue growth cover training and inference costs. Accordingly, Morgan Stanley prefers Alibaba, Tencent, Kingsoft Cloud, and GDS Holdings, along with domestic compute companies including MiniMax, Zhipu AI, Cambricon, Tianshu Zhixin, and Hygon Information Technology. These represent the latest domestic compute stock selections from Morgan Stanley's research, while ROIC must still be evaluated alongside cost of capital, competitive pressures, and current valuations.