Chinese AI platform executives face pressure to secure productive use cases ('find jobs') for 20GW of existing compute capacity as utilization challenges emerge.

At the Cloud Summit on September 22, Alibaba Group CEO Wu Yongming announced a target: by 2032, Alibaba Cloud will operate global data centers exceeding 20GW in scale. Pingtouge's next-generation AI chip, Wenyu V900, will deliver three times the single-chip computing power of its predecessor M890, with mass production scheduled for Q1 2027. Qwen's subsequent model parameters will expand to between 50 trillion and 100 trillion.

Parameter scale determines what a model can do; cluster scale determines how much Alibaba can supply; together, these two factors determine who pays the electricity bill and depreciation.

According to forecasts from Goldman Sachs and other institutions, Alibaba Cloud currently operates about 4-6GW of computing power. Bloomberg reported that Microsoft plans to expand its data center capacity to 38GW by 2032, while AWS previously disclosed a pace of adding 3.8GW annually.

Alibaba needs to quadruple its computing capacity in six years, requiring annual growth rates close to the pace of the two leading global cloud providers. Wu Yongming himself acknowledged in the same speech that global supply chain shortages in AI data centers continue to constrain computing capacity growth.

Scale is only half the story. After chips are purchased, data centers built, and power and networks connected, depreciation, maintenance, and energy costs accrue by the second regardless of whether the computing power is being used. What truly matters is how much time these server racks spend running for external customers.

Wu Yongming compared future machine intelligence to electricity—plug in and use. This metaphor perfectly captures another side of this business: the value of an electrical grid has never depended on how advanced the grid is, but rather on how many people are willing to pay for electricity.

This calculation is not unique to Alibaba. A week earlier, Huawei's Connect conference released the Ascend 960 supernode, with highly consistent technical answers: constrained by manufacturing processes, they don't pursue single-chip performance, instead relying on interconnection protocols and system-level innovation to turn clusters into "one computer."

The difference lies in sales strategy. Huawei explicitly pursues "hardware monetization"—the supernode itself is a commodity that customers purchase, with utilization being the customer's responsibility. Pingtouge doesn't pursue independent chip profits; computing capacity is an asset Alibaba owns itself, and utilization becomes Alibaba's own profit-and-loss matter.

**20GW First Enters the Cash Flow**

In February 2025, Alibaba announced it would invest at least 380 billion yuan over the next three years in AI and cloud infrastructure. By the first quarter of fiscal 2027 (ending June 30, 2026), 190 billion yuan had already been invested. With less than half the time elapsed, half the budget has been spent.

380 billion yuan is likely just the first check. UBS calculates that each additional gigawatt of AI data center requires approximately 100 billion yuan in IT equipment capital expenditure alone. From current scale to 20GW represents roughly 15GW of new additions, corresponding to trillion-level investment.

Investment banks have reverse-engineered revenue targets based on this goal. Citi believes 20GW corresponds to about $160 billion in external cloud revenue by fiscal 2033, while UBS estimates about $170 billion by 2032—both methodologies imply roughly 40% annual compound growth. Based on Alibaba's current quarterly cloud revenue annualized at about $27 billion, this means cloud revenue must grow nearly sixfold over six years.

Quarterly data shows steeper trends. Alibaba's Q2 capital expenditure this year was 67.678 billion yuan, up 75% year-over-year, far exceeding market expectations of 36 billion yuan. In the same quarter, AI cloud and computing power service revenue totaled 48.437 billion yuan. Capital expenditure was 1.4 times this business segment's revenue.

Pressure is already visible in cash flow. This quarter's free cash flow net outflow was 44.67 billion yuan, compared to 18.815 billion yuan in the prior year. The group's operating profit declined 57% year-over-year to 15.161 billion yuan.

In August, Alibaba completed 80 billion Hong Kong dollars in rights offerings, at an 8.4% discount to closing price—the first discount since its 2019 Hong Kong listing. After the announcement, Hong Kong stock fell 8.54%. Cai Chongxin and Wu Yongming subsequently increased their holdings by approximately 80 million and 40 million Hong Kong dollars respectively.

Money is being spent faster than planned both because demand is indeed growing and because chips are appreciating in price. Management attributed this quarter's elevated capital expenditure to hardware delivery schedules, increased CPU demand, and chip price increases.

Alibaba has its own calculation for returns on this spending.

Wu Yongming's judgment in the earnings call was: industry consensus holds that no demand ceiling for AI computing capacity is visible before 2030; based on current average gross margins of AI products, computing capacity investments can be recovered in three years, and with improved gross margins, could potentially be shortened to 2.5 years or even 2 years.

This statement transforms 20GW from aspiration into a time-bound financial commitment. Three-year payback requires two premises: AI product gross margins remain at current levels and computing capacity stays fully utilized. But the problem is that neither premise is entirely within Alibaba's control.

**Pingtouge Must Lower Per-Unit Computing Costs**

The role of Wenyu V900 in this system is more concrete than the title of "China's Most Powerful AI Chip."

According to Alibaba's disclosed data, V900 features 216GB of video memory, inter-chip interconnect bandwidth of 1200GB/s, and a single cluster built on this chip can scale up to 500,000 cards.

216GB of video memory already exceeds NVIDIA's B200 at 192GB and Huawei's Ascend 910C at 128GB, far higher than Cambrian's MLU590 at 64GB and NVIDIA's H20 at 96GB. For the first time, a domestically-produced AI chip surpassed NVIDIA's current products in this core specification.

But video memory is not computing power. The launch event did not disclose complete absolute computing performance, power consumption, manufacturing process, yield rate, or third-party testing—"most powerful" remains Alibaba's own assessment for now.

The iteration pace reveals another layer of information. M890 just debuted in May this year; four months later V900 takes the stage. The previous generation hasn't had time for sufficient market validation, yet the next generation is already in position. This pace was, in a sense, forced by competitors.

Five days before, at Huawei's Connect conference, the Ascend 960 supernode was released—4,096 NPUs per node connected via self-developed Lingqu protocol. Huawei's Ascend chips follow a clear "one generation per year" cadence; 960DT arrived three quarters ahead of schedule in Q1 2027, nearly coinciding with V900's production timeline. Both sidestep manufacturing process constraints and bet on interconnection protocols—Huawei calls theirs Lingqu, Alibaba calls theirs ICN. Next comes competition over which ecosystem follows each protocol.

One partially verifiable signal: the Lingxun supernode based on M890 is already the first domestically-produced 100,000-card cloud cluster, having successfully run training and inference of models exceeding 2 trillion parameters including Qwen 3.8 and Kimi K3. It will launch new services in Ningxia Zhongwei in Q4 this year. The Wenyu series has cumulatively served over 650 enterprise customers, but "workable" doesn't directly equal "cost-effective."

At the same launch event, another set of numbers came closer to "utilization." The new-generation high-performance storage CPFS reduced average model startup latency by 50%, improved peak computing power utilization by 30%, and lowered AI storage costs by 69% in actual model training. On the inference side, Tair KVCM cache scheduling system achieved 99% cache hit rates and reduced per-token cost by 50%. The less time GPUs wait for data, the higher the effective output per unit computing power: chip peak computing power sets the ceiling, utilization determines the cost line.

Wu Yongming himself said in the earnings call that Alibaba Cloud's servers have virtually no idle cards. The other side of that statement is: he must say it, and he must deliver it.

Organizational boundaries have been redrawn according to this logic.

In June this year, Alibaba merged its Cloud Intelligence Group and Pingtouge into the "AI Cloud and Computing Power Services" segment, with the Model Laboratory, Qwen consumer business group, and QwenWork moved into a separate "AI Laboratory and Applications" segment.

This arrangement allows Pingtouge's value to flow more directly into the cloud business: a self-developed chip, even without generating independent profit through external sales, creates value as long as it reduces per-hour computing costs and improves Alibaba Cloud's gross margins. Management mentioned in the earnings call that cloud business actual profit margins are even higher when excluding the impact of consolidating AI chip accounting.

Goldman Sachs provided a quantified reference point in its post-Cloud Summit research report: Pingtouge-supplied computing power currently represents about 10% of Alibaba Cloud, with a medium-term target of 50%. If this ratio materializes, Alibaba Cloud's cost structure will be rewritten—half of computing capacity will be priced at internal cost, half procured at market prices, and cloud business gross margin sensitivity to external chip price increases will decline notably. This is the first time "self-developed chips reduce costs" has a trackable progress indicator.

Whether V900 succeeds, chip market share is a secondary metric. How much Alibaba Cloud's per-token cost declines is what will faster enter the profit-and-loss statement.

A curious detail: the exhibitor list for the Cloud Summit's computing power pavilion includes NVIDIA, Intel, and AMD. Alibaba, presenting full-stack self-development achievements at the same conference, still invited the suppliers it is replacing into the pavilion. Full-stack does not mean fully closed; self-development and external procurement will run in parallel for considerable time.

**AgentCore Turns Invocations into Long Bills**

Alibaba's launch of Agentic Cloud and AgentCore at the conference carries commercial significance potentially exceeding another large model release.

Li Feifei, Alibaba Cloud's CTO, systematically articulated this strategy. She breaks down Agentic Cloud into three layers: Model, Harness, and Context—models provide capabilities, Harness lets intelligent agents run stably in production, Context enables agents to maintain continuously updated, domain-wide cognition. Alibaba doesn't plan to compete solely on model capabilities; it wants to turn the entire chain from agent invocation to execution into its own product.

Traditional chatbots receive questions and generate answers, with billing primarily centered on tokens. After an agent receives a task, it might run continuously for minutes or even hours, needing to maintain context, call external tools, access databases, execute code, manage identity and permissions, and retry after failures. A single invocation thus expands into combined consumption of computing power, memory, storage, gateways, networks, and databases.

The billing structure directly reflects this change: beyond model fees, Alibaba Cloud separately charges for computing CUs, gateway calls, storage, exclusive control planes, and public network traffic, with streaming responses, MCP long connections, and agent long sessions calculated as invocation counts based on connection duration. The revenue Alibaba seeks extends beyond model inference, stretching across every run, wait, memory, and tool call by agents.

Alibaba's official position is that AgentCore can improve task completion rates to 99% and total cost of ownership by 70%. The accompanying Agent Sandbox can create 100,000 sandboxes per minute, awakening from deep hibernation in under 600 milliseconds.

However, like Wenyu chip's "most powerful" claim, these figures currently represent Alibaba's own assessment, without independent customer cases or third-party testing available for cross-validation.

Capital expenditure disclosures already reflect this change: one reason for the dramatic quarterly capital expenditure increase is the company's expectation that customers will adopt more AI Agents, so it's preemptively increasing CPU computing capacity. Large models primarily consume GPUs; agents must run substantial tools and business logic beyond calling models, and CPUs are re-entering AI infrastructure procurement lists.

AgentCore allows enterprises to connect multiple model suppliers. Alibaba didn't make Qwen the sole option because models don't establish the most durable customer relationships. Enterprises can switch models, but migrating an entire ecosystem of identity, permissions, credentials, tools, memory, and monitoring is far more difficult.

In the cloud computing era, vendors retain customers through compute, storage, and databases. In the Agent era, whoever manages what data agents can see, what tools they can call, and under whose identity they act, controls the most difficult-to-migrate layer of enterprise AI.

Qwen brings customers in; AgentCore keeps the bills coming.

**Qwen's Losses Are Cloud's Customer Acquisition Costs**

Alibaba's latest segment data illustrates this strategy's cost allocation.

This quarter, AI cloud and computing power service revenue totaled 48.437 billion yuan, with adjusted EBITA of 5.628 billion yuan, up 133% year-over-year, with profit margin rising to about 12%. AI Laboratory and Application revenue was 3.338 billion yuan, with adjusted EBITA loss of 13.861 billion yuan—a loss roughly 2.5 times the cloud and computing power segment's profit. The two segments offsetting each other, Alibaba's AI segment recorded net losses of about 8.2 billion yuan this quarter, or approximately 32.9 billion yuan annualized, equivalent to more than half of the group's operating profit.

Alibaba's AI machine currently operates on three sets of accounts: e-commerce generates cash, Qwen burns cash, and cloud business expands production and achieves profits. Qwen currently resembles demand subsidy for Alibaba Cloud, using group profits to purchase user habits and inference invocations, seeking loads for pre-built computing capacity.

Wu Yongming offered a historical parallel in his speech: in 1882, Edison's Pearl Street Station lit approximately 400 nearby lamps—initially selling light bulbs with electricity, much like today selling tokens with Agent software.

On one side, building computing capacity and achieving profits; on the other, deploying applications to consume capacity and bearing customer acquisition costs, form Alibaba's current AI cycle. Qwen App integrates with Taobao, Tmall, and instant retail; QwenWork integrates with DingTalk and enterprise workflows; Qwen Intelligence enters the Honor Magic9 launching September 28. For the applications team, this is user competition; for Alibaba Cloud, these entry points continuously generate inference, storage, and tool calls.

Measurable consumption has already emerged on the customer side. China Pacific Insurance disclosed that daily token consumption since partnering with Alibaba exceeds 300 billion. But pulling back the view to the national level reveals penetration in this phase: according to the National Data Bureau, as of March this year, China's daily word token invocation exceeded 1.4 quadrillion; Wang Tao's figure at Huawei Connect in September was about 5 quadrillion. Pacific Insurance's 300 billion represents only about 0.006% of the national pool. 20GW doesn't correspond to existing demand; it corresponds to demand not yet created.

Alibaba itself offered a demand-side slope at the Cloud Summit main forum: Liu Dayiheng, head of Token Foundry, disclosed that two months after Qwen 3.8-Max's release, real user revenue growth was 8.5 times, token consumption increased 12 times. The slope is steep enough, but the base wasn't disclosed.

Alibaba management disclosed that last quarter's AI-related product annual recurring revenue (ARR) broke through 49.5 billion yuan (approximately $7.3 billion), expected to approach $10 billion next quarter. This is currently the closest-to-"sustained external customer payment" public metric.

But placing it back in the accounts, the proportions remain stark. 49.5 billion yuan in ARR equals only about 18% of annualized quarterly capital expenditure of 270.7 billion yuan. Single-quarter AI-related product revenue of 12.376 billion yuan represents only about one-quarter of AI cloud and computing power service revenue.

Alibaba's AI can already drive cloud business growth, but existing AI revenue falls far short of consuming the 2032 20GW supply. Internal calls can improve server utilization, yet generate substantial inference costs. Only when external enterprises sustain payment, or consumer agents generate traceable transaction and service revenue, will these costs transform from internal group circulation into new revenue.

**20GW Awaits External Customers to Pay**

The Cloud Summit discussed the technological path toward super artificial intelligence; the earnings report discusses depreciation and cash flow. Alibaba Group Board Chair Cai Chongxin offered a more macro judgment at the same conference: the AI industry is transitioning from isolated breakthroughs in frontier technology toward scaling technology to generate value in real scenarios. He also said, "we need open cooperation more than ever," and reading this statement alongside the five-layer full-stack self-development system, the tension becomes clear. It places Alibaba in a position that must deliver.

The difficulty in delivery lies in timing mismatch. Internet companies previously could adjust server procurement based on traffic growth; 20GW pushes Alibaba toward a different operational discipline: build capacity first, then improve utilization, finally achieve cost absorption through scale. The larger the capacity, the more Alibaba cannot afford server idle time.

The most concerning scenario isn't lack of AI demand, but demand growing slower than capacity expansion. Once computing supply becomes cyclically oversupplied, cloud providers must reduce prices to improve utilization. Price declines then extend the capital recovery cycle, pushing Wu Yongming's stated three-year payback further out.

Alibaba holds cards. It simultaneously owns cloud, models, office software, e-commerce transactions, and numerous consumer touchpoints—internal demand can fill part of the server racks first, AgentCore can make the operation layer a difficult-to-migrate foundation, and Pingtouge can push per-unit computing costs down.

External cloud revenue growing 45% year-over-year and AI-related product ARR nearing 50 billion yuan indicates demand isn't hypothetical. The 13.861 billion yuan in application losses, meanwhile, shows model and application investment costs currently far exceed cloud segment profits.

As long as application losses grow faster than cloud profit and e-commerce increments, so-called full-stack advantages amount only to cost shuffling within the group.

This also distinguishes this AI investment round from any prior Alibaba subsidy. Taobao back then burned traffic costs—expenditure ended when activities concluded. This time it's burning depreciation; servers, once powered on, accrue costs by the second, irretrievably. 20GW first enters capital expenditure and cash flow; whether it becomes AI revenue depends on how many enterprises will let agents run continuously in their operations long-term.