01 Trigger Event

Alibaba's net profit in a quarter of fiscal year 2026 plunged more than 75% year-over-year, while quarterly capital expenditure (capex) climbed to nearly $10 billion, earmarked for "holding its ground in the global AI competition" — per Bloomberg citing the company. This figure is not annual; it is quarterly.

02 What This Really Means

On the surface, it looks like "a Chinese internet company's earnings blow-up," but the real story is capex.

$10 billion per quarter, annualized that is $40 billion. This pulls Alibaba directly into the same hyperscaler capex club as Microsoft / Google / Meta / Amazon — and a year ago, this company was still pitching its cloud story using "asset-light" logic.

This tells three stories:

First, China-US hyperscalers have entered a synchronized capex cycle. US hyperscalers have collectively pushed annualized capex above $100 billion over the past four quarters, and the market has partially priced that in. Alibaba's earnings report is a wake-up shot to the market: the capex intensity on the China side is not "following" — it is of equal magnitude.

Second, the profit collapse is not an operational issue — it is a strategic choice. The 75% profit decline is not driven by Alibaba Cloud revenue falling; it is capex front-loading depreciation + interest. Management clearly believes that in the window where the Qwen generation of models is still iterating, compute absence = model absence = cloud absence. I agree with this prioritization.

Third, this means the Qwen team's training resources from H2 2026 through H1 2027 will not be throttled by compute constraints. This is a major event for the open-source ecosystem.

03 Historical Analogy

The closest analogy is not 2014 AWS, but the 1999–2000 telecoms capex cycle.

Back then, US long-distance phone companies + European telecom operators collectively laid fiber, with capex at one point hitting 30–40% of revenue. The result was the 2001–2002 global telecom industry collective write-downs, with WorldCom and Global Crossing going bankrupt. But the survivors — AT&T, Verizon, and later Level 3 — grabbed the surplus fiber assets and harvested a decade of bandwidth dividends from 2003–2008.

AI capex is most likely tracing the same curve now, but this time it will not be written off all at once — because downstream demand (token consumption) genuinely exists, unlike the pure speculation of 1999's "dig fiber and bet on future bandwidth demand."

A more recent analogy is the 2014–2015 shale oil capex cycle. Oil prices fell from $100 to $30, a batch of high-cost producers got washed out, but the capacity preserved by capex plus technological progress permanently shifted the US marginal cost curve downward.

My judgment: the "survivor dividend" of this round of AI infra capex will be reaped by developers in 2027–2029, in the form of cheaper inference, more open-weights models, and more aggressive batch pricing. The premise is that the company you are betting on survives.

04 What This Means for AI Builders

Over the next week or two, I will re-examine three things:

First, Qwen roadmap expectation revision. Previously I assumed Qwen would slow down in 2026 due to compute constraints. Now I need to reverse that assumption — at this capex intensity, Alibaba will most likely roll out dense model iterations matching Sonnet 4.5 / GPT-5 class from Q4 2026 through Q1 2027. On the builder side, self-hosted + private deployment stacks may need to prepare compatibility for new Qwen specifications in advance, especially for context window and MoE routing configurations.

Second, Alibaba Cloud inference pricing pressure. With capex this heavy, only two paths to recovery remain: API price hikes, or token volume doubling. Given Alibaba Cloud's current market share still significantly lags Huawei Cloud + Tencent Cloud, I am betting on token volume doubling — meaning from Q4 2026 through H1 2027, Alibaba Cloud's inference pricing will drop another 20–40% from current levels, with more aggressive batch / caching discounts rolled out. For middle-layer players like opcx doing model routing, this is a structural arbitrage window.

Third, GPU supply chain reverse impact. Once this $10 billion quarterly capex lands on orders for H20 / B20 / domestic substitute cards, it will further tighten global high-end inference card supply in H2 2026. I have not run specific model supply-demand models internally, but procurement at this scale will most likely push up spot prices — a direct negative for all builders relying on spot compute, including some fine-tuning service providers.

05 Counter-arguments / Risks

I could be wrong in three places.

First, the capex cycle could last longer than I think. The flip side of the 1999 analogy is that the real overcapacity cleanup took 3–4 years, and I may have underestimated the token demand curve — if agent workload genuinely takes off in 2027, Alibaba's $40 billion annualized capex may not be enough at all, rather than being excessive.

Second, I assume Alibaba will choose "use low prices to grab token volume" to recover capex, but management could alternatively choose "maintain gross margin, let profits continue to collapse." Historically, Alibaba's pricing on cloud business has been tougher than AWS, culturally not a spot-market play. If they maintain discipline this time too, my bet on pricing cuts will not materialize.

Third — and this is what worries me most — how much of this $10 billion actually became "H100/H200 clusters for Qwen training," how much became "cards running inference for the Qwen app," and how much was actually for upgrading Taobao's ad recommendation models.

If a significant portion is the latter, then the essence of this capex is internal compute reallocation at Alibaba, not external AI competitiveness improvement, and Qwen's iteration speed will not be pushed as high as I expect. I have not seen enough data in the earnings breakdown to confirm; this could be a misjudgment.

But even so, the meta-signal of this event — China-US hyperscalers synchronously entering hypercapex, and the China-side intensity being underestimated — holds up. This is something builders really need to recalibrate within the next week or two.