Tencent capex turns AI into an investment test

Tencent's Hunyuan and WorkBuddy now sit inside a more expensive bargain: the company's capital expenditure reached 52.8 billion yuan in the second quarter, up 176% year on year. According to woshipm, the same earnings cycle sent shares lower for Tencent, Alibaba, Baidu, Kuaishou, and NetEase, with Xiaomi the exception among the companies it discussed.

That market reaction matters because the AI story has moved beyond tool adoption. Alibaba has Qwen and Alibaba Cloud, Baidu has Kunlun chips and AI Cloud, Kuaishou has Kling, Xiaomi has MiMo, and NetEase uses AI widely in game development, woshipm notes. The question is what these systems can repay.

The bill is already visible. Alibaba's capital expenditure reached 67.7 billion yuan in a single quarter, up 75% year on year, while its AI cloud and computing power service revenue grew 45%. Baidu Core's AI business revenue reached 12.5 billion yuan in the second quarter, up 25% year on year. AI is becoming an investment test before it becomes a premium.

The AI bill now arrives before the AI premium

The first visible change in China's AI cycle is that the bill now shows up before the premium. According to woshipm, after the latest earnings disclosures, shares of Tencent, Alibaba, Baidu, Kuaishou, and NetEase fell, while Xiaomi was the exception among the six companies discussed. That reaction matters because these are not companies without AI stories: Baidu has Kunlun chips and AI Cloud, Kuaishou has Kling, Alibaba has Qwen and Alibaba Cloud, Tencent has Hunyuan and WorkBuddy, Xiaomi has MiMo, and NetEase has widely used AI in game development.

The market is no longer paying simply for participation.

Tencent is the clearest marker of the shift. Woshipm put Tencent's capital expenditure at 52.8 billion yuan in the second quarter, up 176% year on year. 36kr gave the figure as 52.78 billion yuan in 2026Q2, up 176% year over year and up 65% quarter over quarter, and said Tencent's free cash flow turned negative once that spending was included. An AI strategy can look credible and still turn into a cash-flow test.

Alibaba and Baidu show the same pressure from different angles. Alibaba's AI cloud and computing power service revenue grew 45% in the latest quarter, according to woshipm, but its capital expenditure reached 67.7 billion yuan in a single quarter, up 75% year on year. Baidu Core's AI business revenue reached 12.5 billion yuan in the second quarter, up 25% year on year, while Baidu's online marketing revenue fell 19% and total revenue fell 4% in the same period.

Growth is real. Replacement is harder.

Kuaishou adds another version of the investment test. Kling AI's revenue exceeded 850 million yuan in the second quarter, up more than 200% year on year, yet woshipm also noted that Kuaishou's adjusted net profit declined significantly in the period discussed. Wu Duidui's argument is that stronger consensus around AI makes an AI valuation premium harder to obtain. AI has not reset Chinese internet companies to the same starting line; it has amplified cash flow, legacy businesses, organizational capability, historical burdens, and real competitive advantages.

Tencent makes the case for spending, and the warning against copying it

Tencent's numbers show the cleanest version of the pro-spending case. In 2026Q2, according to 36Kr, it had 204.785 billion yuan in revenue, up 11% year over year, and 67.276 billion yuan in operating profit, up 12%. That scale matters because internal AI infrastructure is easiest to justify when the buyer already has large cash-generating businesses, many products that can absorb model capability, and enough traffic to turn inference capacity into repeated use rather than idle prestige hardware.

The bill still changed the shape of the quarter. Tencent's research and development spending reached 27.28 billion yuan in 2026Q2, up 35% year over year. Capital expenditure reached 52.78 billion yuan, up 176% year over year and 65% quarter over quarter, and free cash flow turned negative after capital expenditure was included. Management said on the earnings call that Tencent substantially increased computing-power procurement.

This is not casual experimentation.

36Kr says the spending went into infrastructure construction, PAPI and coPAPI inference demand for upgrades to Tencent's self-developed Hunyuan model, and AI capabilities inside products and services. The named product list matters: Hy, Yuanbao, CodeBuddy, WorkBuddy and Xiaowei are not a single chatbot wrapper. They are attempts to distribute AI across consumer use, coding, work assistance, and device interaction. After excluding contributions from new AI products including those services, adjusted Non-IFRS operating profit was 86.1 billion yuan, up 19% year over year.

The warning is that Tencent's logic does not automatically travel. The 36Kr author frames the move as using core-business cash flow as a cash cow to build AI and seek a new growth engine, while accepting negative book free-cash-flow pressure to secure computing power and stockpile storage. woshipm's author reaches a different prescription for large companies: investment and financing may be wiser than self-operating every AI bet. Kuaishou's Kling shows that path, with an external capital increase capped at $3 billion, nearly $2.8 billion in agreements, and a post-money valuation as high as $18 billion. Baidu's Kunlun chip business has also started an independent listing plan.

Storage shortages turn workflow choices into infrastructure bets

A workflow tool looks light until its usage pattern starts reserving machines and memory on someone else's balance sheet. According to 36Kr's author, Tencent is spending to secure computing power and stockpile storage even as negative book free cash flow pressure shows up. The same author argues that Tencent's entry reinforces the case that AI capital expenditure is feeding storage-company growth, with prices still strong because supply is extremely tight.

That turns a product decision into an infrastructure bet.

The clearest market version came through Situational Awareness. Ifanr reported that Leopold Aschenbrenner, a 24-year-old former OpenAI researcher, published the 165-page essay "Situational Awareness: The Decade Ahead" in June 2024. The fund that carried the same name grew from about $1.5 billion to more than $45 billion in less than two years, then reached a peak size close to $100 billion after debt was added. Its usual structure borrowed an additional $3 for every $1 of principal it held.

Its portfolio made the thesis explicit. Ifanr reported that Situational Awareness bought AI infrastructure companies including SanDisk, Micron, Bloom Energy, and Nebius, while shorting software companies including Adobe, AppLovin, and Figma. As of the end of June, it held about $5.7 billion in SanDisk and $5.6 billion in Micron, together about 56% of its U.S. stock portfolio.

The exposure cut both ways. Situational Awareness lost about 67% in July and erased nearly $30 billion. SanDisk fell about 47% in July, and Micron fell about 29% in July; from July 24 to July 28, Nebius, Bloom Energy, SanDisk, and Core Scientific fell by about 9% to 24%. On July 29, the fund offered to sell Anthropic shares at about a 20% discount and gave potential buyers about 12 hours to decide. Jane Street's about $2.5 billion investment once approached $10 billion, then fell to about $3 billion to $3.5 billion after July.

For product teams, the lesson is not that every AI workflow must own chips or disks. It is that storage shortages, vendor pricing, and financing structure can migrate into the unit economics of a feature that users experience as a simple button.

Aureka shows what an AI asset looks like

Aureka is useful here because it makes the investment test visible. woshipm's author argues that old Internet-era paths are a poor way to compete in AI, and that path dependence is an ultimate trap for giants. The same argument says some major Internet companies will be eliminated regardless of how hard they struggle, while AI is still far from its endgame and companies must survive long enough to see it.

That view sounds bleak for platform competition. Aureka shows the narrower case where AI spend can become an asset rather than a feature.

Founded by Zhao Wei'an in 2023, Aureka has built a generative AI platform connected to a high-throughput digital biology platform, according to 36kr. The point is the loop: AI designs molecules, wet-lab validation tests them, and the results feed the system again. That is different from shallow automation because the company is not only trying to make an existing task cheaper. It is building reusable model infrastructure around biological feedback.

The bill comes first. Aureka invested in a kilocard-scale computing cluster at the end of last year to develop the biological foundation model AuraIDE. Zhao argues that AI is not fundamentally a cost-reduction and efficiency-improvement game, and may increase costs and reduce efficiency in the short term. That claim matters because it matches the shape of the work: compute, lab capacity, model iteration, and pipeline selection all sit before the payoff.

There is evidence of payoff, not just promise. 36kr says Aureka's AI-developed antibody molecules have generated tens of millions of dollars in licensing revenue. The company has raised $100 million in Series B financing, released OpenDDE in July, and produced more than 10 candidate pipelines across cardiovascular metabolism, autoimmune diseases, and central nervous system diseases.

The strongest signal is hit rate. Zhao says earlier work using AF open-source models and internal models from 2023 to 2024 usually produced only five or six active sequences out of 50 for an autoimmune disease target. After OpenDDE entered the design process this year, more than 30 of 50 showed high biological activity, a 60% hit rate.

That is what durable AI capability looks like: a model tied to validation, licensing income, and a pipeline moving toward an IND application in the first quarter of next year.

When AI spend is worth it-and when it is just expensive motion

• A company has strong cash flow from existing businesses and is trying to build a new AI growth engine. Treat AI as a reinvestment decision, not a simple productivity program. Tencent is the clearest example: its 2026Q2 capital expenditure was 52.78 billion yuan, up 176% year over year, and free cash flow turned negative after capital expenditure was included, even as revenue and operating profit still grew. The practical test is whether the spend creates infrastructure, model capability, or product distribution that can become a defensible asset.
• AI revenue is growing, but the legacy business it is meant to replace is declining. Use AI only if the new business can realistically offset the old one, not merely produce a good growth headline. Baidu Core's AI business revenue reached 12.5 billion yuan in the second quarter, up 25% year on year, while Baidu's online marketing revenue fell 19% and total revenue fell 4% in the same period. This is a replacement problem, not just an adoption problem.
• A team or investor is betting on AI infrastructure as the obvious bottleneck. Be wary of leverage and crowding. Situational Awareness bought AI infrastructure companies including SanDisk, Micron, Bloom Energy, and Nebius, while typically borrowing an additional $3 for every $1 of principal it held. It grew from about $1.5 billion to more than $45 billion in less than two years, but lost about 67% in July and erased nearly $30 billion. The AI capex thesis may be right, but the financing structure can still break first.
• A company can connect AI model output to a domain-specific validation loop. This is the strongest case for higher AI spending. Aureka built a generative AI platform and a high-throughput digital biology platform to connect AI design with wet-lab validation in a closed loop. Its AI-developed antibody molecules have generated tens of millions of dollars in revenue through licensing, and the company invested a kilocard-scale computing cluster to develop AuraIDE. Here, the spend aims to create a scientific and workflow capability, not just automate office work.
• An individual worker is deciding whether to leave a large internet company for an AI startup or an independent AI product. Price the opportunity as a career-risk tradeoff, not only a salary upgrade. One cited algorithm engineer joined an AI startup established less than two years earlier for double his previous pay, a title one level higher, and stock options of uncertain future value. Another independent developer's AI writing tool income reached 1.5 times her major-tech-company salary in good months and almost zero in bad months. The source's own career formula is "growth net value = capability improvement speed - energy consumption speed"; positions with growth net value greater than 0 are worth taking, while those below 0 should be approached cautiously.

Workers need their own AI investment memo

The capex question has a personal version: which AI bet turns time, stress, and tool fluency into an asset that travels with the worker? Woshipm frames Chen Mo as a 27-year-old algorithm engineer who left a major tech company for an AI startup founded less than two years earlier. The package looked like classic frontier compensation: double his previous pay, a title one level higher, and stock options whose value could not yet be known.

That is not just a raise. It is a risk budget.

The labor market is pushing people toward that budget. Woshipm cites Zhaopin data showing resume submissions for AI-related jobs up 210% year on year in the first half of 2026, while employer demand for AI-related roles grew 450%. The same article says large-model training experience can lift startup algorithm-engineer starting salaries by 30% to 50% over major tech companies. Pay is the visible signal. The harder question is whether the role creates transferable judgment, not only a hotter line on a resume.

Woshipm's useful test is blunt: growth net value = capability improvement speed - energy consumption speed. A role is attractive when that value is greater than 0; it needs caution when it is less than 0. One employee in an AI startup department of an unnamed cloud platform said three months there brought business complexity equivalent to two years at a major tech company. That is high-density learning, if the worker can absorb it.

The independent path has the same math with less padding. Lin Yi, a 32-year-old former product manager, left a major tech company three years earlier and became a one-person company this year. Her AI writing tool brought 1.5 times her old salary in good months and almost zero in bad months. Woshipm's author recommends one year of living reserves, or keeping a major-tech-company job until side income reaches 50% of salary.

The career risk is not only startup failure. The author argues that AI can complete 60% of middle-manager information-transfer work, making the remaining 40% harder to identify and reward. The worker's memo should ask one thing first: does this move make me someone who controls AI, or someone managed by it?

AI spend versus claimed payoff across Chinese tech, startups, investors, and workers

Watch the ratio that matters, not the demo. In Aureka's drug-design case, Zhao Wei'an told 36kr that earlier AF open-source models and internal models produced five or six active sequences out of 50; after OpenDDE entered the same autoimmune target process, more than 30 of 50 showed high biological activity. That is the kind of before-and-after worth funding: a workflow changes the yield of scarce experiments, not just the speed of drafting slides.

Make your own AI budget pass the same test.

Ask what durable asset appears after the spend. Aureka is building more than 10 candidate pipelines, preparing an IND application in the first quarter of next year, and working with Nvidia on molecule inference acceleration, according to 36kr. Those are harder assets than a prompt habit. If your project mainly creates tool bills, extra review meetings, or dependence on infrastructure nobody on the team controls, treat that as a warning, not modernization.

For readers outside China

• Availability: The source material covers Chinese internet companies, Chinese AI startups, and Chinese career cases, but does not provide product availability details for readers outside China. It names products and businesses including Hunyuan, WorkBuddy, Qwen, Alibaba Cloud, Kunlun chips, AI Cloud, Kling, MiMo, Doubao, DeepSeek, and Aureka's OpenDDE, but cross-border access, English-language support, enterprise procurement routes, and API availability are not disclosed in sources.
• Pricing: The sources give company-level financial figures and financing data, not user-facing subscription prices. Reported figures include Tencent capital expenditure of 52.78 billion yuan in 2026Q2, Alibaba capital expenditure of 67.7 billion yuan in a single quarter, Kling AI revenue exceeding 850 million yuan in the second quarter, Kling external capital increase agreements of nearly $2.8 billion, Aureka Series B financing of $100 million, and Aureka's AI-developed antibody licensing revenue of tens of millions of dollars. Specific product prices, API rates, seat pricing, cloud-unit pricing, and consumer subscription fees are not disclosed in sources.
• Closest Western equivalents: Tencent Hunyuan, Alibaba Qwen, Baidu AI products, Doubao, and DeepSeek are closest to large-model platforms such as OpenAI, Anthropic, Google Gemini, and Meta Llama, though the sources do not provide feature-by-feature comparisons.; Kuaishou Kling is closest to generative video tools such as OpenAI Sora, Runway, and Pika, but availability and pricing outside China are not disclosed in sources.; Tencent CodeBuddy and WorkBuddy are closest to AI coding and workplace assistants such as GitHub Copilot, Microsoft Copilot, and Google Workspace AI tools, but the sources do not compare capabilities directly.; Aureka's OpenDDE and AuraIDE sit closer to AI-for-drug-discovery and biological foundation-model efforts such as AlphaFold3 and Boltz; the source claims OpenDDE's training dataset and data-splitting rules are consistent with those models.
• Data residency: The source material does not cover data residency, hosting regions, cross-border data transfer, model-training data retention, enterprise compliance terms, or whether foreign users' data is stored inside or outside China. For any workflow involving proprietary code, documents, customer data, or biomedical data, those terms would need to be verified directly with the vendor.

Sources

• woshipm 一定会有大厂被AI淘汰，无论它们如何挣扎 https://woshipm.com/share/6433184.html
• 36kr 腾讯能不能为存储续命？ https://36kr.com/p/3937749806464389
• woshipm AI已经救不了大厂的股价了 https://woshipm.com/ai/6452551.html
• woshipm AI时代的职业十字路口：大厂、创业、单干，还是深耕? https://woshipm.com/chuangye/6432197.html
• ifanr 牛来了又走了，AI 股神暴亏 300 亿内幕曝光：成也 AI，败也 AI https://ifanr.com/1675462
• 36kr 又融了1亿美金的AI公司，说AlphaFold不是生物世界天花板 https://36kr.com/p/3933494167108738