Equity volatility

BABA Is Priced as Commerce. Its Long-Dated Options Ignore the AI Rewrite.

Alibaba's commerce engine funds Cloud, Qwen, and T-Head. The stock carries a China discount; long-dated options barely price the chance that the company changes category.

AIEquitiesVolatility

Alibaba is a Chinese commerce cash machine financing a cloud-and-AI stack. The stock prices the cash machine. The long-dated options barely price the chance that the rest changes what the company is.

For an American reader, Alibaba starts with two apps. is the open bazaar: a vast assortment sold by merchants and individuals. is the mall: official storefronts for domestic and international brands. Sellers own most of the inventory. Alibaba sells attention, placement, transaction services, and fulfillment.

That commerce engine sits beside a much larger collection. , , and serve Chinese consumers. , , , and make up most of its international commerce business.

Then there is , , , and . Cloud, Qwen, and T-Head form the technology stack at the center of this article.

On paper, the collection resembles Amazon's commerce-and-cloud machine crossed with Alphabet's ads, models, and custom silicon. The public-market record is the opposite.

THE ACTUAL TRACK RECORD

Same ingredients, completely different outcome

Total return index · log scale
Alibaba, Amazon, and Alphabet total returns since September 2014Alibaba returned 49%, Alphabet 1,079%, and Amazon 1,516%.1002004008001600Sep 201420162018202020222024Aug 2026OPENAI LAUNCHTRANSFORMER PAPERBABA+49%AMZN+1,516%GOOGL+1,079%
$100 invested at the end of Alibaba's first public month became about $149. The same $100 became $1,179 in Alphabet and $1,616 in Amazon.

Measured from Alibaba's first full month as a public company in September 2014 through August 17, 2026, the same window for all three stocks, BABA returned 49%, Amazon returned 1,516%, and Alphabet returned 1,079%. The chart's dashed lines mark OpenAI's 2015 launch and the 2017 Transformer paper. The U.S. hyperscalers rode the infrastructure trade; Alibaba spent most of the period below its 2020 peak.

The article's thesis is narrower than “BABA should trade like GOOGL.” China explains a stock discount. It does not explain why BABA's long-dated options trade near realized volatility while Cloud, Qwen, and T-Head are trying to change the market's classification of the company.

What Alibaba is

The closest U.S. shorthand for Taobao and Tmall is Amazon Marketplace with a Google Ads auction laid over it. Merchants bring the products and most of the markdown risk; Alibaba charges for discovery, advertising, commissions, and services. This marketplace, not Cloud, still pays for the group.

HOW THE MONEY ACTUALLY WORKS

One marketplace pays for everything else

March quarter 2026 · adjusted EBITA
WHAT IT IS NOTA retailer holding the inventory

Buy goods, store them, mark them up, and absorb what does not sell. That is Amazon's first-party retail business.

WHAT IT ISA toll booth on other people's inventory

Merchants bring the products and risk. Alibaba sells them consumer attention and takes a cut of the transaction.

01Merchants bring the inventory

Millions of sellers list goods on Taobao and Tmall. They own the stock and carry most markdown risk.

02Attention is auctioned

Merchants bid for placement, traffic, and conversion. This is closer to Google Ads than Walmart.

03Alibaba keeps the toll

Advertising, commissions, delivery, and service fees turn someone else's inventory into Alibaba revenue.

$17.72Bquarterly core revenue$3.48Badjusted EBITA from it20%segment margin on the toll
WHERE THE $3.48B GOES · SAME QUARTER
China e-commerceTaobao, Tmall, Ele.me, Fliggy
+$3.48B
Cloud Intelligencecompute, databases, model APIs
+$0.55B
InternationalAliExpress, Lazada, Trendyol
$0.02B
All othersCainiao, Amap, Qwen app, DingTalk
$3.07B
Group totalThe core earns $3.48B. Cloud adds profit; international and the investment portfolio consume almost all of it.+$0.94B
Segment revenue includes inter-segment activity. The 20% figure is China e-commerce adjusted EBITA divided by segment revenue, not a consolidated margin. Commerce funds the group; Cloud and the other businesses spend it.

China e-commerce produced $17.72B of quarterly revenue and $3.48B of adjusted EBITA. Cloud added $550M. International commerce was roughly break-even. “All others” lost $3.07B as Alibaba funded technology, quick commerce, and consumer AI.

Adjusted EBITA is not cash flow, and segment revenue includes eliminations. The direction is still plain: the marketplace produced the profit; Cloud contributed a little; the rest spent most of it.

What Alibaba is building with the cash

The stack has three parts. Cloud sells the compute. Qwen creates demand for it. T-Head lowers the cost and reduces dependence on foreign chips.

HOW QWEN MAKES MONEY

Usage, Cloud invoices, and T-Head compute

MODELQwen

A family of Alibaba-built language, vision, and reasoning models. Some weights can be downloaded; larger managed versions are sold as APIs.

Closest U.S. comparisonGemini is owned by Google. Claude is built by Anthropic, where Amazon and Google are minority investors.
DISTRIBUTIONAlibaba Cloud

The metered enterprise layer: hosted models, compute, storage, databases, networking, and security.

Closest U.S. comparisonGoogle sells Gemini and Claude through Vertex AI. AWS sells Amazon and third-party models through Bedrock.
COMPUTET-Head Zhenwu

Alibaba-designed processors for training and running AI models. Alibaba reported more than 100,000 units deployed on its public cloud by March 2026.

Closest U.S. comparisonGoogle calls its AI accelerators TPUs. Amazon calls its training chips Trainium and inference chips Inferentia.
Better models create usageCloud meters the usageOwned chips can lower cost
Alibaba owns all three layers. The model matters only if usage turns into Cloud revenue and lower compute cost.

Qwen works like GPT or Gemini as a brand: it names a family of models, not one chatbot. Alibaba trains the family, runs consumer products on it, releases downloadable weights, and sells managed access through Cloud.

DeepSeek is not part of Alibaba. Alibaba publicly denied the rumored investment. Hosting a DeepSeek model creates Cloud revenue; using Qwen also gives Alibaba the model economics.

QWEN IN CONTEXT

Qwen against the current model field

FOUNDATION MODEL

The general engine behind chat, coding, search, and enterprise agents.

MODEL FAMILY

Several sizes and products sharing a name, from cheap local models to hosted flagships.

OPEN WEIGHTS

Developers can download the trained parameters and run them away from the maker's cloud.

INFERENCE

The repeated compute used every time a trained model answers a request.

Model / ownerIndependent capability indexAccess and distribution
Claude Opus 5Anthropicindependent lab
61
API onlyAnthropic · AWS · Google Cloud
GPT-5.6 SolOpenAIMicrosoft partner
59
API onlyOpenAI · Azure · AWS
Gemini 3.5 FlashGoogleGoogle-owned
45
API onlyGoogle AI · Vertex AI
Qwen3.7 MaxAlibabaAlibaba owns it
46
API flagship + open-weight familyAlibaba Cloud · Qwen
DeepSeek V4 ProDeepSeekindependent of Alibaba
44
Open weights + APIDeepSeek · third-party clouds
Qwen is not DeepSeek.Alibaba trains and owns Qwen. DeepSeek is an independent lab whose models Alibaba Cloud can host, just as AWS and Azure host models they do not own.
Artificial Analysis snapshot, August 2026. Highest tested model shown for each family; configurations and effort settings differ. The index combines reasoning, knowledge, coding, and agentic evaluations.

On the August 2026 Artificial Analysis leaderboard, Qwen3.7 Max trails the highest-scoring Claude and OpenAI configurations while sitting near the tested Google and DeepSeek flagships. Click View more for cost, speed, and latency. The commercial question is whether developers and Chinese enterprises run useful workloads through Alibaba's meter, not whether Qwen wins every benchmark.

Open weights help Qwen travel. Developers can download variants and run them without an Alibaba account. When a company wants managed hosting, security, databases, or support, Alibaba Model Studio sends the bill.

Every request consumes inference. T-Head designs the Zhenwu processors that run some of it. The deployment figure comes from Alibaba; it is not an independent chip benchmark.

Cloud rents the compute. SanDisk sells the storage that keeps the same AI workloads fed. The two businesses sit at opposite ends of the same capex cycle.

Cloud margin is the fulcrum

Alibaba Cloud's March-quarter revenue grew 38% to $6.04B. External-customer revenue grew 40%. AI products reached 30% of external Cloud revenue. Adjusted EBITA grew 57% to $550M.

Cloud's adjusted EBITA margin was about 9%. AWS produced a 39.4% operating margin last quarter; Google Cloud produced about 35.6%. Those measures are not accounting-identical, but they describe very different levels of maturity.

CLOUD SENSITIVITY

What a 20% margin would do to earnings

Quarterly · USD
China e-commerce now$17.7Brevenue
$3.48Badjusted EBITA
Alibaba Cloud nowrevenue · 38% growth
$0.55Badjusted EBITA · 9% margin
Three-year scenario$13.3B30% annual growth
$2.65B20% margin · 76% of commerce profit
This is a sensitivity, not company guidance. It holds commerce profit flat and applies 30% annual Cloud growth with a 20% terminal adjusted EBITA margin.

Hold Cloud growth at 30% for three years, below the latest print, and quarterly revenue reaches roughly $13.3B. At a 20% adjusted EBITA margin, barely half AWS's current margin, quarterly profit reaches $2.65B, equal to 76% of what China e-commerce earns today. Failure is legible too: if margin stays near 9%, commerce keeps subsidizing the second act.

Put it next to the U.S. hyperscalers

With the Alibaba pieces defined, the comparison becomes easier to read. Amazon, Alphabet, Microsoft, and Alibaba each combine a cash-generating distribution surface, public cloud, first-party models, and custom silicon. The ownership and economics differ at every layer.

FOUR COMPANIES, SIX LAYERS

Alibaba against the U.S. hyperscalers

Download as GIF ↓
Choose a layer04 / 06
04

Custom silicon

controls cost and supply

OWNED

100K+ processors deployed

commerce + ads + an owned AI stack
OWNED +

separate first-party chips for training and inference

retail + ads fund AWS
OWNED 8t + 8i

separate first-party accelerators for training and inference

advertising funds cloud + Gemini
OWNED

first-party inference accelerator live in Azure

software funds Azure + AI partnerships
Published per-chip specificationsn/d = the vendor does not disclose a comparable figure
Zhenwu M890
FP8n/d
GB144
TB/sn/d
Trainium3
FP82.52
GB144
TB/s4.9
TPU v7 Ironwood
FP84.61
GB192
TB/s7.37
Maia 200
FP8>5.0
GB216
TB/s7.0
READFor Alibaba, custom silicon is both an efficiency project and insurance against losing access to leading U.S. chips.
Alibaba owns its model, cloud, and silicon.BABA adds an ADR and VIE between those assets and the shareholder.
Vendor-published specifications for current named accelerators. FP8 compute, memory capacity, and memory bandwidth are different dimensions, not a composite benchmark. Alibaba discloses 144GB of HBM3 and 800GB/s inter-chip bandwidth for Zhenwu M890, but not comparable FP8 compute or HBM bandwidth.

Amazon owns Nova and Trainium, and it is a minority investor in Anthropic. Alphabet owns Gemini and Tensor Processing Units, and it also holds a minority Anthropic stake. Microsoft owns Azure, Phi, MAI, and Maia 200. Its OpenAI relationship includes an economic interest, licensing, and revenue rights; OpenAI is not a Microsoft subsidiary.

Alibaba's architecture belongs in the comparison. Its scale, margins, governance, and shareholder claim do not.

Why the discount is justified

China's technology crackdown changed the rules on Ant and the marketplace. Weak domestic demand hit merchants and ad budgets. U.S. chip controls limited the hardware Alibaba could buy. The variable interest entity structure adds contractual risk between some Chinese operations and foreign holders. All four sit between Alibaba's products and a U.S. investor's claim on them.

WHY THE DISCOUNT EXISTS

Four risks between the assets and the ADR

01RegulationMonetization and capital allocation

Ant IPO cancellation and Alibaba antitrust fine

02Domestic demandCommerce growth and merchant spending

Weak consumption and a shrinking population

03Export controlsAI capacity and input cost

Advanced chips require shifting U.S. licenses

04VIE structureOwnership and capital mobility

ADR holders rely partly on contractual control

These are not generic country-risk labels. Each one interrupts a specific link between Alibaba's assets and the value an outside shareholder can realize.

Regulation. Ant Group was days from a record IPO when the Shanghai Stock Exchange suspended it. The overhaul ended with RMB7.123B ($984M) in penalties and a much smaller private valuation. Alibaba still owns 33% and received a $477M Ant dividend in fiscal 2026, but Ant is separately governed and privately marked.

Demand. China's population fell by 3.39M in 2025, with 23% of the population age 60 or older. Retail sales grew only 2.7% in the first half of 2026. That slows the merchant and advertising engine just as Alibaba spends on quick commerce and AI.

Chips. Access to advanced processors turns on U.S. licensing policy. T-Head can replace some foreign accelerators, but fabrication, memory, and semiconductor equipment still cross the trade-policy boundary.

Ownership. A BABA American Depositary Receipt represents shares in Alibaba's Cayman holding company. For businesses held through variable interest entities, the group relies on contracts rather than direct equity ownership. Alibaba's annual filing also notes that the significant majority of its revenue, profit, and operating cash flow sits in directly owned subsidiaries. Not every dollar crosses the contractual seam.

Estimating fair market value without China Risk?

There is no neutral “American-company multiple.” Two May 2026 sum-of-the-parts reports offer a live range of assumptions: BOCI put BABA at $187 per American Depositary Share; DBS used $204. Those targets imply roughly 50% and 64% upside from the August 17 close, respectively. Put another way, BABA traded at 33% and 39% discounts to the two targets. Each estimate depended on forward assumptions for commerce and Cloud.

Reverse the current price instead. BABA closed at $124.71 on August 17 for a market value near $301.4B. Alibaba had about $37.8B of net cash. Fiscal 2026 China commerce adjusted EBITA was $15.586B. The slider below shows what remains after changing the multiple on that profit stream.

REVERSE VALUATION

What does the stock leave for Cloud?

Approximate equity value · USD
Try a multiple
July 30–31, 2026 snapshots. Different business mixes and a different profit measure from Alibaba adjusted EBITA.
Equity value$301.4BBABA at $124.71
Net cash$37.8Bcash and liquid investments less debt
China commerce$218.2B14.0× FY26 adjusted EBITA
=
Everything else$45.4BCloud, Ant, international, logistics and other
Less of $25.9B$19.5Bleft for Cloud, international commerce, logistics, Qwen, T-Head and other businesses
Two May 2026 sum-of-the-parts reports put BABA at$187 BOCI$204 DBS50% and 64% upside from the August 17 close, equivalent to 33% and 39% discounts to the targets. Both rely on their own forward estimates and multiples.
Uses FY2026 China commerce adjusted EBITA of $15.586B. The residual omits tax, liquidity, VIE and capital-allocation haircuts and includes current losses in the non-commerce businesses.

At 14× commerce, the residual is $45.4B. Mark Alibaba's 33% of Ant at the company's last disclosed 2023 transaction valuation and roughly $19.5B remains for Cloud, international commerce, logistics, Qwen, T-Head, and the losses elsewhere. Move commerce to 10× and the residual expands to $107.8B. The cheap reading depends on the multiple; now the dependency is visible.

That is the stock half of the thesis: Alibaba can deserve a China discount without Cloud, Qwen, and T-Head being worth almost nothing.

The options market asks a stranger question

Alibaba reports June-quarter results before the U.S. market opens on August 20. The August 21 options should be expensive, and they are: at-the-money implied volatility was 80.8% in Cboe's delayed chain on August 17, against about 30% for Amazon and Alphabet.

BABA TERM STRUCTURE

Earnings is expensive. The long end is not.

BABAAMZNGOOGL
BABA, AMZN, and GOOGL implied volatility term structuresToggle between headline at-the-money implied volatility and forward implied volatility with the earnings event isolated.20%30%40%50%Aug 21→28Aug 28→Sep 18Sep 18→Oct 16Oct 16→Dec 18Dec 18→Jun 17BABA Aug 21→28: 39.4% IVBABA Aug 28→Sep 18: 37.7% IVBABA Sep 18→Oct 16: 39.9% IVBABA Oct 16→Dec 18: 42.3% IVBABA Dec 18→Jun 17: 44.7% IVAMZN Aug 21→28: 27.7% IVAMZN Aug 28→Sep 18: 27.9% IVAMZN Sep 18→Oct 16: 30.5% IVAMZN Oct 16→Dec 18: 37.1% IVAMZN Dec 18→Jun 17: 36.3% IVGOOGL Aug 21→28: 27.1% IVGOOGL Aug 28→Sep 18: 27.7% IVGOOGL Sep 18→Oct 16: 30.9% IVGOOGL Oct 16→Dec 18: 35.1% IVGOOGL Dec 18→Jun 17: 35.7% IV
39.4% first post-event forward44.7% four-to-ten-month forward44.4% trailing one-year realized vol
Cboe delayed option chains, frozen August 17, 2026 at approximately 19:13 UTC. Forward volatility is calculated in variance space between adjacent expiries.

The chart opens on forward volatility, which strips out variance already packed into the earlier expiry. The August 21-28 window is 39.4%. The next two are 37.7% and 39.9%. Click Headline IV to put the 80.8% earnings spike back on the chart.

The odd price is at the other end. The December-to-June forward is 44.7%. BABA realized 44.4% over the past year. A seller of variance across next winter and spring gets almost no premium over what the stock just delivered.

Amazon and Alphabet pay one. Their corresponding forwards are 36.3% and 35.7%, against realized volatility of 33.8% and 32.4%.

LONG-DATED VOLATILITY

BABA pays almost nothing over realized

Annualized volatility · Aug 17, 2026
Forward implied volatility compared with trailing realized volatilityBABA has a 0.3-point premium, compared with 2.5 points for Amazon and 3.3 points for Alphabet.30%35%40%45%50%BABA+0.3 pts44.4 RV → 44.7 FWDAMZN+2.5 pts33.8 RV → 36.3 FWDGOOGL+3.3 pts32.4 RV → 35.7 FWD
The hollow dot is trailing one-year realized volatility. The solid dot is the December-to-June forward implied volatility. BABA's gap is effectively a hairline.
THE DURATION MISMATCH

Four days of fear, ten months at cost

Block width = calendar days
BABA implied volatility windows drawn at their true calendar durationThe four-day earnings block is 80.8% implied volatility. The 181-day December-to-June forward is 44.7%, nearly level with 44.4% trailing realized volatility.20%40%60%80%Aug 17→21: 80.8% IV across 4 calendar daysAug 21→28: 39.4% IV across 7 calendar daysAug 28→Sep 18: 37.7% IV across 21 calendar daysSep 18→Oct 16: 39.9% IV across 28 calendar daysOct 16→Dec 18: 42.3% IV across 63 calendar daysDec 18→Jun 17: 44.7% IV across 181 calendar days44.4% TRAILING REALIZED80.8% · 4-DAY EVENT44.7% · 181 DAYSSEP-Q PRINTCLOUD MARGINDEC-Q PRINTAI REVENUE SHAREMAR-Q PRINTT-HEAD EXTERNALAUG 17, 2026JUN 17, 2027 · 304 DAYS
The first block runs from the snapshot to the first expiry; the rest are forwards between listed expiries. Every block is drawn at its true calendar width. Height is implied volatility; the white dashed line is trailing one-year realized volatility. Alibaba print markers are expected quarterly windows, not company guidance.

The curve is quiet around other scheduled shocks too. The August 21-28 forward, which contains NVIDIA's August 26 results, is 39.4%. The next window, which contains China's August 31 PMI release, is 37.7%. Neither event commands much extra premium.

The curve charges a lot for four days and almost nothing extra for ten months. Export controls, variable-interest-entity governance, consumption, and domestic competition will not resolve on August 20. Neither will Cloud margin, paid AI demand, or T-Head utilization.

💡
The mismatch

At 30% annual growth and a 20% adjusted EBITA margin, Cloud would earn quarterly profit equal to 76% of today's China-commerce profit. At a 14× commerce multiple, the stock leaves $45.4B for everything else before separately marking Ant. The December-to-June forward trades at 44.7% against 44.4% realized.

What would prove this wrong

The August 20 report will move the inputs. It cannot settle a three-year cloud buildout, but it can make the cheap reading look better or worse.

WHAT TO WATCH

Five numbers that change the valuation

Latest reported
01
Cloud keeps outgrowing commerceMAR-Q RELEASE
38%Cloud revenue growth against 6% China e-commerce growth
02
AI is paid external demandMAR-Q RELEASE
30%AI-related products as a share of external Cloud revenue
03
Cloud margin reaches the high teensMAR-Q EXHIBIT 99.1
9%Adjusted EBITA margin, against 39% at AWS and 36% at Google Cloud
04
T-Head reaches paying customersMAY-26 RELEASE
400+Alibaba-reported external customers across 20 industries by May 2026
05
Assets produce holder-accessible cashFY26 20-F
$477MAnt Group dividend received in fiscal 2026
These thresholds are our framing, not Alibaba guidance. Update them with each quarterly result.

Cloud has to keep outgrowing commerce. AI revenue has to come from paying companies, not consumer-app downloads. Margin has to leave single digits without quick-commerce losses swallowing the improvement. T-Head has to prove useful beyond Alibaba's own infrastructure. Some of the value inside Ant and the rest of the structure has to reach the American Depositary Receipt holder.

If Cloud slows back toward commerce growth, margin stays near 9%, and the rest of the portfolio keeps consuming the core's profit, the commerce classification survives. If growth holds and margin moves into the high teens, the earnings mix changes with it.

The stock prices Alibaba as Chinese e-commerce. The long-dated options barely price the chance that the market changes its mind, in either direction.