Alibaba is rebuilding its operating backbone around an in-house AI compute supply chain and a 10x data center expansion, changing how capacity, cost, and risk are managed across its networks.
In Brief
- Alibaba is treating AI infrastructure as a core production asset, not a support function, with a planned tenfold increase in data center capacity by 2033.
- Ownership of proprietary AI chips and flexible capacity models is reshaping sourcing strategy, pricing power, and margin structure under severe hardware scarcity.
- Quick commerce and cross-border logistics investments are being tuned for unit economics and network synergies, funded by higher-margin AI and cloud services.
The Strategic Break: Compute as a Primary Supply Chain
Alibaba has drawn a clear line between its pre‑AI infrastructure model and the one it is now building. Leadership describes two new ‘factories’ for AI training and AI inference, both powered by a rapidly expanding estate of data centers. The company expects to need about ten times its 2022 data center capacity by 2033 and signals that earlier capital expenditure guidance of RMB 380 billion is likely to be exceeded.
This is not presented as an IT upgrade. It is framed as the construction of core production capacity, equivalent to adding a new tier of plants or hubs into a physical network. AI and cloud products already account for 30% of external cloud revenue, with AI revenue of RMB 9 billion in the recent quarter and an annual run rate of RMB 36 billion. Model and application services are on track for annual recurring revenue of RMB 30 billion by year‑end, from more than RMB 8 billion already.
At the same time, Alibaba has moved to secure control over the most constrained element in this digital supply chain: advanced compute hardware. Through its T‑Head unit, it has proprietary AI chips in scaled production, with over 60% of its compute capacity already serving external customers. Management positions this as autonomy over the compute supply chain in an environment where the cost of deploying a new server has more than doubled in a year and where constraints on chips and memory are expected to last three to five years.
These decisions mark a structural shift. Compute capacity moves from being a purchased utility to being a strategically controlled input, with implications for how capacity is planned, financed, priced, and allocated across the wider operating network.
How Alibaba Is Redesigning Capacity and Sourcing Logic
In operational terms, this kind of shift requires a new approach to capacity planning and sourcing governance.
On capacity, Alibaba is committing to sustained, high utilisation of its AI assets. Management notes that there is not a single GPU card sitting idle and that the return on this investment is expected over a three‑ to five‑year horizon. This implies a planning cadence closer to long‑lead manufacturing or large hub investments than to incremental IT provisioning. Demand from external clients and internal workloads must be forecasted and buffered at the portfolio level, not server by server.
On sourcing, Alibaba is diversifying how it acquires compute capacity. Leadership explains that capacity can come through capital expenditure, operating expenditure, and co‑build models. Some AI servers can be sold to other computing centres; some data centres can be co‑developed with partners. This mirrors asset‑heavy and asset‑light mixes in transport and warehousing, where owned, leased, and joint‑venture assets are balanced to manage risk, utilisation, and balance sheet impact.
T‑Head chips sit at the centre of this strategy. Domestically produced semiconductors in China are acknowledged to lag leading overseas chips on energy efficiency, and current penetration in Alibaba’s infrastructure is still relatively low due to production constraints. However, global AI chip vendors are seen earning gross margins of 60% to 80%. As domestic chip production ramps and performance improves, Alibaba expects to substitute away from imported margin and lower its own unit compute cost, while retaining pricing power created by global scarcity and higher replacement costs.
Pricing logic is already changing. With the cost of a new server more than 100% higher than a year ago, Alibaba reports having pricing power with both new and existing customers, particularly as demand for AI inference grows. Management links higher replacement costs, rising inference volumes, and expanding deployment of proprietary chips to expected gross margin improvement over the next two to three years.
For any large operational network, this model points to a few practical shifts:
- Capacity is staged in larger, more centralised units, backed by multi‑year utilisation plans rather than short‑term scaling.
- Critical components are moved under partial or full control, even if performance is initially lower, to avoid imported margin stacks and supply risk.
- Commercial pricing is explicitly tied to replacement cost and capacity tightness, not just historical rates or commodity indices.
AI Infrastructure as The Funding Engine For Network Expansion
The AI and cloud business is not only a user of capital; it is positioned as the main funding engine for broader network investments. External cloud revenue grew 40% year‑on‑year, with a relatively stable adjusted EBITA margin of 9.1%. AI‑related products have delivered eleven consecutive quarters of triple‑digit revenue growth, and token usage for models has increased more than tenfold from late 2025 to mid‑2026.
Management presents a flywheel: ongoing investment in cloud infrastructure increases revenue potential in AI and cloud offerings; as those offerings scale, their gross margins rise; the resulting cash flow is then reinvested into further infrastructure. This is similar to a centralised logistics hub that becomes more efficient as throughput increases, releasing capital to build additional hubs or upgrade automation.
The balance sheet is configured to support this. Operating cash flow in the last financial year was an inflow of RMB 9.4 billion, while free cash flow was an outflow of RMB 17.3 billion, primarily due to AI investments. The company holds a net cash position of about USD 38 billion, or approximately USD 59 billion excluding long‑maturity debt, and plans to sustain elevated investment levels in AI and infrastructure over the next two years.
This structure allows Alibaba to carry other parts of the network through heavy investment phases. Within China e‑commerce, adjusted EBITA fell 40% to RMB 24 billion, primarily due to investment in quick commerce, user experience, and technology. Excluding quick commerce, EBITA would have been stable year‑on‑year.
Quick Commerce and Cross‑border Logistics as Testbeds For Unit Economics
On the physical side of the network, Alibaba is applying similar discipline to quick commerce and cross‑border operations.
Quick commerce revenue reached RMB 20 billion in the quarter, up 57% year‑on‑year. Overall order volume was 2.7 times the prior year, with non‑food orders at three times. At the same time, average order value increased quarter‑on‑quarter and unit economics improved, driven by order mix optimisation. The company links this directly to better fulfilment logistics efficiency and category focus, particularly in food, fresh produce, and healthcare, which in turn support growth in Freshippo and Tmall Supermarket.
The target is explicit: unit economics in quick commerce are expected to turn positive by the end of fiscal year 2027, even as order volumes are maintained. Alibaba frames quick commerce not just as a standalone business, but as a source of synergies for its conventional e‑commerce operations. It is used to drive customer acquisition, increase engagement, fulfil diverse demand, and support logistics infrastructure. In network terms, same‑day and near‑store flows are being designed to share assets and data with standard e‑commerce flows, aiming for economies of scope in last‑mile and local inventory.
Cross‑border logistics is on a similar path. The AIDC segment, which includes AliExpress Choice, grew revenue by 6% and significantly narrowed its adjusted EBITA loss, approaching breakeven. Management credits logistics optimisation and operating efficiency, and notes that unit economics for AliExpress Choice have improved substantially on a sequential basis. This implies active work on route design, consolidation, mode mix, and warehouse operations to reduce cross‑border cost‑to‑serve and compress lead time variability.
In both quick commerce and cross‑border, Alibaba is trading near‑term profit for network position and learning. The AI and cloud businesses provide cash and margins; quick commerce and AIDC consume capital but are being tuned toward defined unit‑economics targets and integrated more tightly into the wider fulfilment structure.
Merchant Economics and Demand Shaping as Part Of The Operating Model
Alibaba’s disclosures also show changes in how merchant behaviour is shaped and how that ties back into operations. Customer management revenue (CMR) grew 1% as reported but would have been up 8% on a like‑for‑like basis, once the impact of a new subsidy accounting treatment is removed. The company has upgraded its business development program for select merchants so that platform subsidies are directly tied to marketing spend on its properties. Those subsidies are now recorded as a contra revenue item rather than as sales and marketing expense.
This design influences where and how merchants invest in promotions and assortment, which in turn affects order flow, category mix, and load on fulfilment assets. When combined with quick commerce, where order mix optimisation is central to unit‑economics improvement, and with AI‑driven demand capture through the Qwen assistant integrated into core consumer applications, the picture is of a platform that is using spend incentives and AI front ends to steer demand into flows that the network can serve more efficiently.
For any large operation, this highlights a simple point: commercial levers and logistics design cannot be separated. Subsidy rules, marketing programs, and AI‑based recommendations are part of the operating model, not ancillary functions.
Constraints and Trade‑offs That Will Shape Execution
Several hard constraints underpin this strategy.
First, hardware remains physically constrained. Management expects production limits for chips and memory to last three to five years, and notes that server deployment costs have already more than doubled year‑on‑year. This caps the speed at which compute capacity can grow and raises the financial impact of mis‑forecasting demand.
Second, domestic chip production is itself constrained. T‑Head chips represent a relatively low proportion of total infrastructure today, in part due to Chinese production limits and performance gaps versus global leaders. The strategy assumes that domestic capacity and capability will improve fast enough to underpin gross margin expansion as reliance on high‑margin foreign chips is reduced.
Third, the company is prioritising AI cloud growth and token consumption over near‑term margins. Leadership states that margin is secondary for the next three to five years as AI penetrates more industries. This mirrors earlier periods in logistics where networks were built for scale first and optimised for profitability later. It places more pressure on other parts of the portfolio to manage working capital, service levels, and cost discipline while the AI engine scales.
What This Operating Model Now Enables
Alibaba has shifted from buying compute as a utility to running compute as a core, vertically integrated supply chain, backed by long‑horizon capital commitments and explicit utilisation and pricing logic. The same mindset is being applied to fast fulfilment and cross‑border logistics, where volume growth is coupled with defined unit‑economics targets and network synergies rather than isolated growth metrics.
The result is an operating model that links AI infrastructure, merchant incentives, and fulfilment design into a single capacity system. It enables more direct control over bottleneck inputs, greater scope to steer demand toward efficient flows, and a clearer path to margin expansion once current hardware and unit‑economics constraints ease. It also raises the demands on planning discipline, sourcing governance, and capital deployment, since errors compound across both digital and physical networks rather than in one domain alone.
