Alibaba Raises $10.2B in Hong Kong's Record Share Sale to Fuel AI Infrastructure

Alibaba Group completed Hong Kong's largest-ever primary follow-on equity offering on Sunday, raising HK$80 billion ($10.2 billion) through the sale of 710 million newly issued shares — and committed 100 percent of the net proceeds to its artificial intelligence infrastructure buildout. The company's stock fell 8.5 percent on Monday, its steepest single-day decline since early 2025, as investors weighed share dilution against Alibaba's increasingly urgent bet that AI cloud revenue will compound fast enough to justify a capital raise that equals what it already spent on AI compute in a single recent quarter.

A Company That Has Structurally Outgrown Its Own Cash Flow

The equity placement was structured as a Regulation S offshore transaction, excluding US investors by design. China International Capital Corp., HSBC, Morgan Stanley, and UBS acted as joint bookrunners. Institutional demand reached roughly three times the offering size, according to people familiar with the matter who asked not to be identified, with sovereign wealth funds among the confirmed participants. Alibaba increased the size of the offering after the deal was oversubscribed. The offering priced at HK$112.70 per share — a 3.6 percent discount to the Friday closing price of Alibaba's US-traded American depositary receipts and an 8.4 percent discount to Friday's Hong Kong close — and is expected to settle on August 26.

The deal surpasses by a significant margin Hong Kong's previous largest follow-on offering and represents the city's biggest share sale since Prosus NV sold $14.7 billion worth of Tencent Holdings stock in 2021, according to Bloomberg data. Globally, Alibaba's offering ranks third among primary follow-on share sales in 2026, behind offerings from Alphabet and Intel.

What makes the timing notable is the arithmetic behind it. In the quarter that ended June 30, 2026, Alibaba's capital expenditures surged 75 percent year-over-year to RMB 67.7 billion — roughly $10 billion, nearly matching the entire equity raise in a single quarter. Operating cash flow for that period was positive at RMB 22.9 billion, but free cash flow turned deeply negative at -RMB 44.7 billion, the consequence of an investment pace that Alibaba's operations alone are not generating enough cash to sustain at the desired speed. The company ended the quarter with $69.9 billion in cash and liquid investments — a substantial buffer — but that figure was declining rapidly. The equity raise does not resolve the funding gap structurally; it extends the runway by roughly one quarter of capital at the current pace.

Chairman Joseph Tsai purchased approximately HK$81 million worth of shares in Hong Kong on Monday, and Chief Executive Officer Eddie Wu acquired approximately HK$39 million, according to Hong Kong Stock Exchange disclosure filings. The two combined purchases totaling roughly $15.3 million in US dollar terms represent a relatively small fraction of each executive's overall stake, but as a visible public signal made into a falling share price, they carry a clear intent: management is backing the thesis with its own capital.

Alibaba also disclosed during the share offering announcement that it had already deployed approximately half of its previously announced CNY 380 billion (~$56 billion) three-year AI investment plan in the roughly eighteen months since that plan was announced in February 2025. With the June quarter representing the halfway mark on cumulative spending, the equity raise adds external capital on top of what operations and existing cash are contributing — a signal that the investment plan is proceeding faster than originally modeled or that the desired scope has expanded.

What "Full-Stack AI" Actually Means in Practice

Alibaba's SEC filing described the use of proceeds in a single phrase: "invest in its full stack AI capabilities, including to expand and enhance its AI infrastructure." That language is Alibaba's chosen shorthand for a vertical integration strategy that spans five distinct layers of the AI technology supply chain — each of which is now being explicitly funded by this raise.

The foundation is the T-Head semiconductor subsidiary, which designs Alibaba's proprietary AI accelerator chips. T-Head unveiled its Zhenwu 810E chip in January 2026, a fully in-house-designed parallel processing unit delivering performance comparable to Nvidia's H20 — the GPU that Nvidia engineered specifically for the Chinese market to comply with US export restrictions. In May 2026, T-Head followed with the Zhenwu M890, which delivers three times the throughput of its predecessor, carries 144 gigabytes of GPU memory, and achieves 800 gigabytes per second of interchip bandwidth — purpose-built for the emerging generation of AI agents that must retain long context windows and coordinate with other models in real time. T-Head has shipped more than 560,000 Zhenwu units to date across more than 400 enterprise customers. Alongside the M890, Alibaba introduced the Panjiu AL128 server, which integrates 128 M890 accelerators into a single rack — the kind of high-density configuration required for large-scale agentic AI workloads.

The strategic logic behind T-Head is not purely about performance. CEO Eddie Wu said on Alibaba's earnings call that as the company ramps deployment of its own chips in its data centers to replace commercially procured silicon, it expects to see "substantially higher gross margin as well as profitability." The current capex surge includes significant procurement of CPU chips because demand for AI agent compute is outpacing supply — and some of that demand is being met with external components at margin-compressing commercial prices. Zhenwu chips, once deployed at scale, are expected to reduce that margin drag. T-Head has a chip roadmap extending through 2028, with a successor to the M890 (the V900) planned for the third quarter of 2027 and a further generation (the J900) planned for the third quarter of 2028. Alibaba has also disclosed plans to consider a public listing of T-Head, with restructuring underway.

Above the silicon layer sits Alibaba Cloud's compute infrastructure, which reorganized its reporting this quarter to merge the Cloud Intelligence Group with T-Head into a single "AI Cloud and Compute Services" segment — making the chip-to-cloud integration not just a technical claim but a structural organizational fact. The cloud platform includes Platform for AI (PAI), which provides end-to-end ML services spanning training, fine-tuning, and inference deployment, and the Lingjun Intelligent Computing Service, which manages large-scale cluster operations with near-linear acceleration for model training at scale. The cloud business already serves approximately 67 percent of A-share listed companies in China. In the June 2026 quarter, it posted external revenue growth of 45 percent — the fastest growth in 22 quarters — while simultaneously expanding its adjusted EBITA margin to 12 percent, up from approximately 7 percent a year earlier.

The model layer is anchored by the Qwen family. Alibaba's flagship model, Qwen3.8-Max, is a 2.4-trillion-parameter sparse mixture-of-experts system that activates approximately 95 billion parameters per inference pass, processes up to one million tokens of context, and handles text, image, and video inputs simultaneously through the API. The mixture-of-experts design matters economically: a gating network routes each incoming token to a small subset of specialized expert sub-networks rather than activating every parameter. At a 95-billion active parameter load, the inference compute cost is comparable to running a much smaller dense model, not a frontier-scale system — which is what makes Alibaba's API pricing of $2.00 per million input tokens and $6.00 per million output tokens viable at the scale Alibaba needs. The release of Qwen3.8-Max open weights on August 12 and the accompanying Qwen3.8-27B dense model on August 14, both available on Hugging Face and ModelScope, extended Alibaba's ecosystem strategy: developers who deploy Qwen locally or build derivative models become part of a funnel that generates Alibaba Cloud dependency over time. The Qwen model series has accumulated more than 3 billion downloads globally and spawned more than 300,000 derivative models, according to Alibaba's own reporting.

Read more: Alibaba's Qwen models and the global AI infrastructure consolidation wave

At the enterprise application layer, QwenWork — an AI-native workforce agent platform that consolidates coding workflows, task automation, and knowledge management — launched in public beta on August 3 and began paid subscription tiers on August 24. At the consumer layer, the Qwen App has reached 250 million users, according to the company, with paid subscriptions for productivity features now active.

The Return Thesis — and Its Key Uncertainties

Alibaba's management laid out the investment economics on the June quarter earnings call: servers have a five-year operational lifespan, AI workloads recover their full capital cost within three years under current margins, and years four and five become a period of "pure free cash flow" from hardware that is already depreciated. Management cited operational data to support this: Nvidia V100 chips purchased in 2018 and A100 chips purchased in 2020 are still running at near-full capacity for customers today. CEO Wu stated that the payback period for AI compute assets is currently on track to fall from three years to approximately 2.5 years as AI-related product gross margins improve — and that he expects it could potentially reach "around two years."

That improvement trajectory, if it holds, implies that the capital Alibaba is deploying today compounds substantially faster than its original investment plan assumed. Nomura analysts wrote after the June quarter results that cloud growth had not yet peaked and that accelerating growth was occurring alongside meaningful margin expansion rather than at the expense of profitability — a significant distinction from the more typical pattern in which infrastructure investment drives margin compression without near-term offset.

The uncertainties are real and should not be papered over by the management narrative. Alibaba's AI Labs and Applications segment — the unit housing the Qwen model development and consumer AI products — reported an adjusted EBITA loss of RMB 13.9 billion in the June quarter, up from a loss of RMB 3.2 billion a year earlier. That is the cost of model training and application development before those products generate sufficient cloud API revenue to cover their own expenses. Net income for the quarter overall fell 75 percent year-over-year. The company also cut share buybacks by approximately 80 percent in the quarter — from $815 million in the comparable quarter of 2025 to $162 million — a clear signal that capital earmarked for shareholder return is being redirected to the AI buildout.

The broader context of the capital allocation is also relevant. China's internet sector spent 81.8 percent more on equipment in the first seven months of 2026 than in the same period a year earlier, according to the National Bureau of Statistics — the most authoritative confirmation yet that China's AI infrastructure buildout has converted from announced intention to actual hardware purchase. US hyperscalers are forecast to spend approximately $725 billion in aggregate capital expenditures in 2026, much of it AI-related. Alibaba's $10.2 billion equity raise and its total three-year AI plan of roughly $56 billion put it in a different order of magnitude from AWS, Azure, or Google Cloud individually — but structurally, the trajectory is the same: sustained heavy capital commitment based on the conviction that AI demand will grow into and beyond current infrastructure capacity.

Alibaba's Position in the Global AI Race — and Its Complicating Factors

Alibaba faces a different competitive picture depending on which axis you examine. Domestically, its cloud business leads China in both market share and growth velocity: AI-related product revenue has grown at triple-digit year-over-year rates for twelve consecutive quarters. Qwen ranks first in enterprise model invocation market share among Chinese companies according to Alibaba's own data. ByteDance is spending aggressively on its own cloud AI infrastructure and Doubao consumer AI application, but Alibaba's infrastructure advantage — particularly the PAI platform and the Zhenwu chip program — gives it structural depth that pure-play model competitors cannot easily replicate.

Internationally, the picture is more complicated. The US Department of Defense added Alibaba to its Section 1260H Chinese Military Companies list on June 8, 2026, citing military-civil fusion ties through Alibaba's indirect affiliations with Chinese state bodies. Alibaba filed a federal lawsuit in San Jose challenging the designation on June 24, calling it "arbitrary and capricious." A 1260H listing does not freeze assets or bar US commercial activity, but it prevents the DoD from contracting with Alibaba and — effective June 30, 2027 — restricts DoD procurement through Alibaba's supply chain. Enterprise buyers with government or defense-adjacent relationships must now factor that constraint into their Alibaba Cloud and Qwen deployment decisions.

The legal structure of Alibaba Cloud's operations also matters for international enterprise adoption in a way that cannot be fully resolved by model performance. Alibaba, while incorporated in the Cayman Islands and headquartered for HKEX purposes in Hong Kong, operates its principal cloud infrastructure and model development from Hangzhou, China. China's National Intelligence Law (Article 7) requires Chinese organizations to cooperate with state intelligence requests. China's Cybersecurity Law, amended effective January 1, 2026, extends AI governance and access obligations explicitly to AI systems and network operators. These obligations apply regardless of where an Alibaba Cloud server is physically located at inference time. For enterprises deploying QwenWork — Alibaba's enterprise knowledge-base and workflow automation platform — or using Qwen models through Alibaba's cloud API, the data-routing question requires assessment rather than assumption.

The Qwen open-weight release strategy partially mitigates this for the developer community: a self-hosted deployment of Qwen3.8-Max on enterprise-owned infrastructure eliminates active data transmission to Alibaba's servers, though it does not eliminate the ongoing dependency on Alibaba for future model releases. The rapid global uptake of Qwen weights — 300,000-plus derivative models, adoption by enterprises including Airbnb for customer service workflows — suggests that the model quality argument is winning for many international users, legal considerations notwithstanding. Whether that dynamic holds at the enterprise scale QwenWork is targeting is a separate and unresolved question.

What the Capital Will Be Deployed Against

Alibaba's official use of proceeds statement does not break down allocations by layer — how much goes to Zhenwu chip production and procurement versus compute infrastructure buildout versus model training versus application development is not disclosed. The "full-stack" framing encompasses all of these simultaneously. What management has signaled is that the immediate demand constraint is compute capacity: CEO Wu said on the earnings call that demand for AI compute currently exceeds supply, which explains why a company sitting on nearly $70 billion in cash and liquid investments is nonetheless issuing new equity. The issue is not total resources; it is the pace at which those resources can be deployed productively without creating idle capacity.

The chip roadmap gives some structural indication. T-Head's publicly disclosed roadmap targets the V900 for Q3 2027 and the J900 for Q3 2028. Each generation is expected to expand Alibaba's ability to replace commercially procured AI processors with in-house silicon, improving gross margins on AI cloud services over the 2027–2028 window. If the payback period genuinely shrinks toward 2.5 years as Zhenwu chips scale in deployment, the free cash flow inflection point — when today's capital expenditures begin to generate net cash — would arrive somewhere in the 2028–2029 range for infrastructure committed now. That is the timeline Alibaba is essentially asking investors to fund through today's dilutive offering.

Morgan Stanley selected Alibaba as a Top Pick and assessed that higher capex should fuel better medium-term returns. Of 40 analyst ratings tracked by one data provider, 38 are Buy or Strong Buy. The share purchases by Tsai and Wu today — bought directly into the 8.5 percent stock decline — suggest that management agrees with that conviction, at whatever level of personal capital they chose to commit.

Competitive Asymmetries That the Capex Number Obscures

The raw comparison of $10.2 billion raised against the US hyperscalers' $725 billion aggregate 2026 capex budget is arithmetic, not analysis. The more useful framing is structural: where Alibaba has genuine competitive leverage and where it is structurally constrained.

The clearest advantage is the Chinese domestic cloud market. Alibaba Cloud serves approximately 67 percent of A-share listed Chinese companies and holds the leading position in enterprise large-model invocations according to its own reporting. That captive installed base — enterprise customers who built their cloud architecture on Alibaba infrastructure, often years before the current AI cycle — creates a natural upsell runway for AI cloud services that a foreign competitor cannot easily replicate regardless of model performance. When Alibaba Cloud's EBITA margin expanded to 12 percent while growing 45 percent, that combination indicated genuine operating leverage, not simply scale buying market share.

The international market presents a different picture. Alibaba Cloud's global footprint is meaningful — data centers in Asia Pacific, Europe, the Middle East, and the Americas — but international AI enterprise adoption faces headwinds that domestic metrics do not capture: the DoD 1260H designation, the state-access legal framework, and the practical reality that enterprises in regulated industries (financial services, healthcare, defense supply chain) typically cannot use Chinese-operated AI infrastructure regardless of model quality. Qwen's open-weight strategy partially addresses this by enabling self-hosted deployments that remove Alibaba from the data path, but it also means that international adoption via self-hosting does not necessarily translate into Alibaba Cloud API revenue. The company gets ecosystem influence without direct monetization in those cases.

The relationship with Moonshot AI — in which Alibaba holds approximately a 36 percent equity stake — gives it partial financial exposure to the Chinese AI model race even where Qwen is not the developer's first choice. If Kimi K3 or its successors win market share from Western models in Asia Pacific markets, Alibaba benefits financially regardless. That portfolio positioning is an underappreciated hedge in Alibaba's competitive architecture.

Risks That the Investment Thesis Must Navigate

Alibaba's core e-commerce segment reported an 8 percent year-over-year revenue decline in the June quarter from its Chinese marketplace business, a consequence of both Pinduoduo's sustained domestic competition and Alibaba's own deliberate pullback from direct-sales businesses. The AI business — specifically, cloud revenue and the new AI Labs and Applications segment — must generate enough growth to compensate for a stagnating and in some segments shrinking commerce core. That is not an unusual structural position for a large technology company undergoing platform transition; it is, however, a transition that has a real time cost.

Independent verification of Alibaba's AI performance claims also remains incomplete in ways that matter for enterprise buyers. Qwen3.8-Max, released in early August, carries a vendor-published benchmark table showing strong results on SWE-bench Pro, GPQA Diamond, and multimodal evaluations. As of this writing, the model has not yet appeared on Artificial Analysis's third-party intelligence index or the Hugging Face Open LLM Leaderboard — the two most widely referenced independent benchmarking platforms in the enterprise AI evaluation community. Arena.AI's crowdsourced rankings place Qwen3.8-Max second globally on multimodal tasks, per Alibaba's own reporting of those rankings, though Arena.AI rankings reflect human preference votes rather than controlled evaluations. The model's performance on knowledge-retrieval hallucination tasks — a key failure mode for enterprise deployment — has not been independently measured.

The question that the $10.2 billion raise cannot answer is whether the AI revenue line — cloud API consumption, QwenWork enterprise subscriptions, Qwen App paid tiers, and downstream e-commerce AI integration — grows fast enough over 2027 and 2028 to validate the capital commitment before the next funding decision becomes necessary. Alibaba's e-commerce core, once the engine funding everything else, faces sustained domestic pressure from Pinduoduo and cross-border headwinds from TikTok Shop's continued international expansion. The AI business must compound, not merely grow, to carry the financial structure that this buildout is creating. Whether the payback period actually compresses toward two years — or whether the model assumptions behind that figure prove too optimistic as the broader AI infrastructure cycle matures — is the empirical question that the next four quarters of Alibaba's cloud margin data will begin to answer. The raise is funded. The thesis is running.