Alibaba Cloud AI Pricing Power and Huawei Ascend Competition

Can hyperscale cloud providers exercise pricing power in AI services amidst competition from integrated hardware rivals?

By Austin Kline

Founder & Lead Analyst · Research methodology

Published 2026-11-14 · Last updated 2026-11-14

Drafted with AI assistance from company filings and reported figures; published under Austin Kline's research rules. How reports are made

Executive summary

Alibaba Cloud holds pricing leverage at the Model-as-a-Service and enterprise software layer through its proprietary Qwen foundation model ecosystem, where high switching costs deter migration. Conversely, in raw compute infrastructure, modest 9.0% adjusted EBITA margins and aggressive competition from state telecoms and Huawei Cloud constrain pricing flexibility. Alibaba's heavy capital expenditures require a successful customer transition toward higher-margin model invocation rather than commoditized IaaS.

Alibaba by the numbers

Historical · FY2022–FY2026

Source: Reported annual filings (10-K / 20-F) and provider data. Fiscal years as reported; margins are calculated from the reported figures shown.

Software Differentiation Versus Infrastructure Commoditization

Added 2026-11-14

To evaluate whether Alibaba Cloud can raise its artificial intelligence prices without losing market share to Huawei’s Ascend ecosystem, one must dissect the AI value chain. The answer depends entirely on which layer of the stack Alibaba is attempting to price: raw compute infrastructure (IaaS) versus proprietary models, software, and agentic workflows (Model-as-a-Service and PaaS).

At the raw compute layer, Alibaba has limited unilateral pricing power. At the model and application layer, however, Alibaba possesses structural advantages that allow it to increase monetization without ceding customer accounts to Ascend.


1. The Distinction Between Silicon and Cloud Services

Comparing Alibaba Cloud directly to Huawei’s Ascend requires separating the hardware vendor from the cloud service provider. Ascend is Huawei’s proprietary AI accelerator family (such as the Ascend 950DT and upcoming 960DT). Huawei sells this silicon directly to enterprises, state entities, and third-party data center developers, while also using it to power its own competing Huawei Cloud.

In the hardware market, Huawei has already demonstrated aggressive pricing behavior:

  • Due to soaring demand amid tightening US semiconductor curbs and acute shortages of high-bandwidth memory (HBM), Huawei has pushed indicated prices for its flagship Ascend 950DT accelerator up by 20% to 60%, taking individual card prices above RMB 250,000.
  • Cambricon and other domestic chipmakers have instituted similar double-digit price increases.

Because the underlying domestic hardware benchmark is becoming substantially more expensive, the total cost of ownership for any enterprise attempting to build or rent an on-premise Ascend cluster is surging. When industry-wide hardware replacement costs rise, Alibaba Cloud can adjust its baseline compute rental pricing upward to reflect input-cost inflation without immediately appearing uncompetitive relative to self-hosted Ascend alternatives.


2. The Commodity Trap in Raw Compute (IaaS)

Where Alibaba Cloud cannot easily hike prices is in raw, generic compute hosting.

According to Alibaba Group Holding Limited’s Form 20-F for the fiscal year ended March 31, 2026 (filed with the SEC on May 20, 2026):

  • Cloud Intelligence Group generated revenue of RMB 158,132 million ($22,924 million), an increase of 34% year-over-year.
  • Cloud Intelligence Group adjusted EBITA grew 35% to RMB 14,265 million ($2,068 million), yielding an adjusted EBITA margin of approximately 9.0%.
  • External cloud revenue accelerated to 40% growth in the fourth quarter of fiscal 2026, with AI-related products accounting for 30% of that volume.

While a 34% revenue expansion is solid, a 9.0% operating margin demonstrates that Alibaba Cloud does not enjoy monopolistic pricing power in basic infrastructure. In standard compute and storage, Alibaba competes not just with Huawei Cloud, but with aggressive state-owned telecom cloud providers (China Telecom, China Mobile) who operate with lower return-on-capital hurdles.

Furthermore, Alibaba’s capital intensity has escalated rapidly. Consolidated capital expenditures for fiscal 2026 surged to RMB 126.94 billion, depressing consolidated operating cash flow to RMB 76.21 billion and plunging free cash flow to negative RMB 50.72 billion. The company transitioned to a net-debt position of RMB 86.11 billion.

Alibaba is caught in a capital-expenditure cycle where it must spend heavily to secure infrastructure. If Alibaba hikes raw IaaS compute prices significantly above the cost of operating on alternative clouds, price-sensitive enterprise workloads can and will migrate to Huawei Cloud or state-backed infrastructure.


3. The Proprietary Moat: Qwen, Model-as-a-Service, and Integration

The picture changes dramatically once a customer moves up the stack into Model-as-a-Service (MaaS) and enterprise software integration. This is where Alibaba has defensible pricing power.

As disclosed in Alibaba's fiscal 2026 Form 20-F:

  • Alibaba Cloud held China's largest public cloud market share by revenue in 2025 across IaaS and PaaS (according to IDC data).
  • Alibaba Cloud captured a 35.8% market share in China’s AI cloud market in the first half of 2025 (according to Omdia).
  • The proprietary Qwen foundation model family ranked first with a 32% market share in enterprise-level large model invocation in China in the second half of 2025 (according to Frost & Sullivan).
  • At the consumer level, the Qwen application surpassed 295 million users by March 2026.

Enterprises building production applications around Qwen—utilizing Alibaba’s Platform for AI (PAI), the Lingjun Intelligent Computing Service, and the Wukong enterprise agent platform—cannot easily swap out their infrastructure for an Ascend cluster. Migrating complex agentic workflows requires rewriting API integrations, fine-tuning new model weights, and re-architecting software pipelines.

Alibaba has already begun exercising this pricing power:

  • It has introduced commercial revenue-sharing licenses on large open-weight releases (such as Qwen3.8-Max) for high-revenue commercial users.
  • It has introduced tiered subscription plans on its Qwen office assistant tools, ranging from RMB 200 to nearly RMB 1,500 annually.

When an enterprise pays for token consumption on Qwen, it is not buying raw server cycles; it is buying reasoning capability, accuracy, and enterprise tool integration. A price increase on proprietary model invocations does not drive clients to Huawei’s hardware, because Ascend is a chip architecture, not a turnkey replacement for Qwen's software ecosystem.


4. The Critical Vulnerability: The Open-Weight Ecosystem on Ascend

The primary scenario where Alibaba Cloud would lose share to the Ascend ecosystem during a price hike involves third-party open-source models optimized specifically for Huawei hardware.

This dynamic is already visible in the domestic market:

  • Frontier domestic AI labs like DeepSeek have actively stationed engineering teams with Huawei to optimize model training and inference natively on Ascend clusters (such as the planned 160,000-chip deployment in Inner Mongolia).
  • If high-performing open-source architectures (like DeepSeek's Flash variants) achieve parity with Qwen while running with higher efficiency on Ascend hardware, cost-conscious enterprises could deploy those models directly on Huawei Cloud or private Ascend infrastructure.

If Alibaba attempts to extract excessive rents on token pricing while open alternatives running on Ascend deliver superior price-performance per token, developer inertia will shift toward the Ascend-native open ecosystem. Alibaba is attempting to mitigate this internally through its own chip subsidiary, T-Head, which has scaled proprietary GPUs to lower internal inference costs, but it remains in an active engineering race against Huawei’s hardware-software co-design.


The Fundamental Verdict

Alibaba Cloud can selectively raise prices on its software layer—proprietary Qwen model APIs, enterprise agent workflows, and ecosystem-integrated business tools—without material defection to Huawei Ascend. The switching costs and software tooling provide adequate protection.

However, Alibaba Cloud cannot unilaterally raise prices on commoditized infrastructure rental without losing share to Huawei Cloud and state telecom operators. With the Cloud division earning a modest 9.0% adjusted EBITA margin and the broader group absorbing RMB 126.94 billion in capital spending, Alibaba’s priority must be migrating customers away from low-margin raw compute toward high-margin proprietary token invocation. Only through that software-layer transition can Alibaba turn its massive AI buildout into genuine economic returns on invested capital.

Hardware Replacement Costs and Model-as-a-Service Transition

Added 2026-11-14

The fundamental reality of Alibaba’s artificial intelligence cloud strategy is that at the raw infrastructure layer (IaaS), operating margins cannot be sustained at current levels once stockpiled high-performance foreign silicon runs out.

While Alibaba is attempting to offset this drag by moving up the value stack into Model-as-a-Service (MaaS), enterprise workflows, and proprietary software, the economic and thermodynamic penalties of relying on domestic substitute silicon will compress infrastructure-level unit economics.

To evaluate whether Alibaba Cloud can sustain its profitability, the business must be separated into its audited reported performance, the unit economics of computing hardware, and the pricing dynamics of the software built on top of it.


The Reported Reality: High Growth on Thin Margins

According to Alibaba Group Holding Limited’s Form 20-F for the fiscal year ended March 31, 2026 (filed with the SEC on May 20, 2026), the Cloud Intelligence Group delivered:

  • Segment Revenue: RMB 158,132 million ($22,924 million), an increase of 34% year-over-year from RMB 118,028 million in fiscal 2025. In the final quarter of the fiscal year, external cloud revenue accelerated to 40% growth, with AI-related products accounting for 30% of that external volume.
  • Segment Adjusted EBITA: RMB 14,265 million ($2,068 million), up 35% from RMB 10,556 million in fiscal 2025.

While these figures demonstrate strong commercial traction, the underlying margin profile is modest. Adjusted EBITA margin for Cloud Intelligence was approximately 9.0% in fiscal 2026.

Crucially, this growth came at an immense cash cost. Consolidated capital expenditures across Alibaba jumped from RMB 86.66 billion in fiscal 2025 to RMB 126.94 billion in fiscal 2026, driven largely by cloud and AI infrastructure commitments. Operating cash flow fell from RMB 164.82 billion to RMB 76.21 billion, pushing group free cash flow from positive RMB 78.16 billion deeply into the red at negative RMB 50.72 billion.

Alibaba is already operating its cloud infrastructure on narrow cash margins while bearing unprecedented capital reinvestment requirements.


The Hardware Dilemma: The Capital Efficiency Penalty

Stockpiled accelerators—whether compliant iterations like the RTX PRO series or earlier-generation enterprise GPUs—face inevitable physical wear and rapid technological obsolescence. Data center silicon under sustained training and inference workloads has an economic lifespan of three to four years. When these stockpiles exhaust their useful lives or become non-competitive, Alibaba must transition entirely to domestic merchant alternatives (such as Huawei’s Ascend series) or internal custom silicon developed by its semiconductor design subsidiary, T-Head.

This transition imposes a severe structural cost penalty on compute unit economics:

  1. Lower Density per Dollar and Watt: Domestic silicon currently lags leading global architectures in compute density and memory bandwidth (particularly High-Bandwidth Memory, or HBM). To achieve the equivalent FLOPS (floating-point operations per second) required to train and run massive frontier models, Alibaba must cluster significantly more chips. A larger cluster requires more physical racks, more complex networking interconnects, substantially higher electricity consumption, and increased cooling capacity.
  2. Rising Upstream Hardware Costs: Due to semiconductor export curbs and limited domestic advanced fabrication capacity, high-end domestic accelerators are in short supply. Merchant suppliers like Huawei have pushed pricing on flagship accelerators like the Ascend series up sharply. Alibaba Cloud will not have the unilateral buyer power to extract low hardware acquisition costs.
  3. Depreciation Drag: Because Alibaba must spend more capital to deploy equivalent effective compute, depreciation expense per unit of revenue will rise. In a cloud business where depreciation already represents a major operating cost, this creates automatic gross margin compression at the IaaS tier.

If Alibaba were purely a landlord renting out servers and raw compute time, its margins would decline precipitously as its existing chip inventory ages out.


The Mitigation: Moving from IaaS to Model-as-a-Service

Alibaba’s leadership is aware of this hardware ceiling. In their letter to shareholders in the fiscal 2026 Form 20-F, management highlighted that the Cloud Intelligence Group is pivoting away from traditional compute and storage toward models, AI compute, and agent services. The company established the Token Hub Business Group and introduced its proprietary Qwen foundation models and Bailian (Model Studio) platform to capture revenue higher up the stack.

The bull case for sustaining margins rests on three non-hardware advantages:

  • T-Head Custom Silicon Optimization: Form 20-F notes that T-Head has brought its proprietary GPU and specialized processors into production at scale. Tailoring custom Application-Specific Integrated Circuits (ASICs) directly to the Qwen model architecture allows Alibaba to optimize inference workloads, lowering power consumption and hardware costs per token generated compared to renting third-party merchant chips.
  • Software-Like Switching Costs (MaaS): If enterprise clients merely rent raw virtual machines (IaaS), they can easily migrate workloads to whichever cloud offers the cheapest per-hour compute. However, if clients build their internal workflows, proprietary databases, and enterprise agents on top of Alibaba’s Qwen foundation models and Bailian developer tools, switching costs rise dramatically. Software and API consumption (Model-as-a-Service) historically carry software gross margins (60% to 75%), which can subsidize lower infrastructure margins.
  • Ecosystem Synergy: Alibaba is embedding Qwen into its core domestic e-commerce operations (Taobao and Tmall), quick commerce, and consumer applications (surpassing 295 million users on consumer Qwen by March 2026). Internal consumption drives massive scale, allowing Alibaba Cloud to amortize model development costs across an enormous proprietary base.

The Fundamental Obstacles to Margin Expansion

Despite Alibaba's software and model strengths, three disciplined considerations suggest AI cloud margins will remain constrained:

  1. Domestic AI Price Competition: China’s enterprise AI landscape is marked by fierce price competition. Leading domestic model developers and cloud competitors (including ByteDance’s Doubao, DeepSeek, Tencent, and state-backed telecom clouds) have engaged in aggressive price-cutting on token calls to gain market share. This limits Alibaba’s ability to charge premium prices for proprietary model access.
  2. The "Treadmill Capex" Dynamic: AI infrastructure is not a one-time buildout that yields decades of recurring free cash flow like a pipeline or a toll road. The rapid evolution of model architectures requires continuous, high-intensity capital reinvestment simply to remain state-of-the-art. When a business must spend tens of billions of dollars annually merely to maintain its competitive position, high reported accounting margins can be an illusion that masks poor cash generation.
  3. Sub-Par Return on Capital: As of fiscal 2026, Alibaba’s return on invested capital (ROIC) stands at roughly 2.0% to 2.8%, well below any reasonable 10% to 12% cost of capital. A business segment earning a ~9% operating margin while free cash flow turns negative is not currently generating excess economic returns.

The Verdict

Alibaba cannot sustain its current AI infrastructure margins on raw compute once its stockpiled foreign chips are fully depleted. The hardware cost penalty of clustered domestic silicon and higher power requirements will inevitably compress IaaS profitability.

The company can only defend its overall Cloud Intelligence margin (currently near 9%) if it aggressively shifts its revenue mix away from commoditized compute rental toward high-margin, proprietary Model-as-a-Service and enterprise workflow software.

For an investor, the critical metrics to track are not top-line cloud growth or strategic commentary about three-year payback periods, but Cloud segment EBITA expansion and a tangible recovery in consolidated free cash flow. Until incremental capital expenditure begins translating into cash returns comfortably exceeding the cost of capital, Alibaba's AI cloud expansion must be evaluated as an expensive, capital-consumptive transition rather than a high-margin compounding engine.

Research for information only, not investment advice. Figures reflect the dates shown; earlier sections are kept as written for the historical record.

Keep exploring

company · $BABA $9988.HK

Alibaba Group: Capital Reinvestment, Cash Flow Deterioration, and Valuation

In fiscal 2026, Alibaba consolidated revenue grew 2.7% to RMB 1,023.67 billion while operating cash flow dropped 53.8% to RMB 76.21 billion and capital expenditures expanded to RMB 126.94 billion, driving free cash flow to negative RMB 50.72 billion. Operating margins compressed to 5.8% and the balance sheet transitioned to a net-debt position of RMB 86.11 billion as the firm elevated spending on AI infrastructure and quick commerce delivery networks. Valuation models across conservative, base, and optimistic scenarios indicate intrinsic values between $62 and $148 per ADS, reflecting acute capital-reinvestment risks.

industry · $BABA $TCEHY

Chinese Hyperscalers and Silicon Independence: Alibaba, Tencent, and the CANN Defense

Alibaba and Tencent are deploying compiler abstraction layers as a defensive mechanism to bypass Huawei's proprietary CANN software stack and prevent vendor lock-in across mixed accelerator fleets. These compilers aim to commoditize underlying silicon rather than establish high-margin standalone software moats. True economic value in the AI ecosystem continues to pool in capital-intensive physical infrastructure and higher-margin application layers such as Model-as-a-Service and proprietary enterprise workflows.

industry

Silicon Dependency and Pricing Power: Analyzing Huawei Ascend in Cloud Computing

Relying on a direct competitor's hardware creates severe margin compression at the raw infrastructure layer, where scarce accelerator supply allows the supplier to dictate capital costs while favoring its own downstream cloud operations. However, cloud platforms retain pricing insulation higher in the stack through proprietary models, high switching costs, and developer workflow lock-in. To avoid becoming captive distribution channels, operators must pursue multi-architecture software abstraction layers and internal ASIC development.

company · $MTN

Vail Resorts: Forensic Investigation of Epic Pass Economics, Leverage, and Valuation

Vail Resorts experienced a net income decline from $347.9 million in fiscal 2022 to $147.5 million in fiscal 2026, accompanied by a drop in operating cash flow to $479.6 million. Total debt-to-EBITDA increased to 4.70x while the dividend payout ratio rose to 163.5% of fiscal 2026 net earnings as cash reserves declined. Normalized owner earnings of approximately $260 million to $270 million suggest current trading levels exceed base-case intrinsic value estimates.

company · $IONQ

IonQ Valuation and Commercialization: Financial Reality Versus Market Expectations

IonQ reported fiscal 2025 revenue of $130.0 million against a net loss of $(510.4) million and free cash flow burn of $(299.6) million. Although the company holds $1.03 billion in cash reserves to support approximately three years of operations with minimal debt, its $10.74 billion market capitalization reflects a trailing price-to-sales multiple exceeding 50x. Commercial viability and fault-tolerant quantum computing remain an estimated five to eight years away, posing significant competitive and dilution risks.

company · $6758.T $SONY

Sony Group: Sum-of-the-Parts Valuation and Conglomerate Transformation

Sony Group generates ¥12.48 trillion in net revenue and ¥1.49 trillion in free cash flow, operating across gaming, music publishing, film, and image sensors. A normalized sum-of-the-parts assessment yields an estimated equity value of ¥21.12 trillion ($33.50 to $37.50 per ADR), indicating a 20% to 28% discount to private market asset values. The planned spin-off of Sony Financial Group and high-margin recurring royalties from music and digital gaming provide potential catalysts to unwind this conglomerate discount.

company · $SYK

Stryker's ROIC Dynamics: The Balance Sheet Drag of Serial M&A

Despite generating a 64.0% gross margin and $4.28 billion in free cash flow in fiscal 2025, Stryker recorded a Return on Invested Capital of 9.6%. Decades of serial acquisitions have capitalized substantial goodwill and intangibles onto the balance sheet, permanently expanding the invested capital denominator. This dynamic narrows Stryker's spread over its estimated weighted average cost of capital to roughly 100 to 200 basis points.

company · $SYK $JNJ $BSX

Stryker Corporation: Fundamental Moat Analysis and Intrinsic Valuation

Stryker demonstrates strong competitive advantages through high switching costs and its Mako robotic platform, generating $4.28 billion in free cash flow with gross margins of 64.0%. However, at a trailing price-to-earnings multiple of 33.6x and a share price around $325, the equity trades slightly above its estimated baseline intrinsic value of $290 to $315 per share. With net debt at 1.95x EBITDA and substantial insider selling, the stock currently offers limited margin of safety for fresh capital.

company · $Z $ZG

Zillow Group: Platform Dominance Versus Housing Market Headwinds

Zillow operates an asset-light portal with 9.6 billion annual visits and gross margins of 74.1%, but its fiscal 2025 operating margin was -1.2% and ROIC stood at -0.6%. While rentals and mortgage segments grew 39% and 37% respectively, depressed residential transaction volumes and commission compression limit Premier Agent monetization. Intrinsic value modeling yields a base-case estimate of $32.00 per share against a market price of $29.57, offering minimal margin of safety given cyclical rate sensitivities.

Browse all research →