The recent launch of Kimi K3, boasting 2.8 trillion parameters and claiming the title of the world's largest open-source model, has sent ripples through the AI community. Elon Musk succinctly commented "Impressive." In the Artificial Analysis comprehensive evaluation, K3 scored 57 points, ranking third globally, trailing only Claude Fable 5 and GPT-5.6 Sol, and surpassing Claude Opus 4.8 and GPT-5.5. In the Code Arena front-end programming blind test, K3 achieved a top global score of 1679, directly exceeding Claude Fable 5's 1631 and GPT-5.6 Sol's 1618. Just three days post-launch, surging user demand pushed computational resources to their limit, prompting Kimi to temporarily halt new consumer subscriptions. Concurrently, the day K3 was released saw the highest single-day increase in Annual Recurring Revenue (ARR) in the company's history. Wall Street is now discussing the "Kimi moment," drawing parallels to the "DeepSeek moment" of early 2025. Bernstein described K3 as a "home run," while Morgan Stanley noted China's AI capability to continuously track the world's most advanced levels and gradually capture more market share over time. While these developments are undoubtedly exciting, for investors, a more critical question arises: as domestic models increasingly close the gap with their overseas counterparts, what does this mean for AI investment? Is it a positive or negative signal, and for whom?
Shrinking Performance Gaps Shift the Competitive Landscape
A fundamental business principle applies here. If Claude scores 100, Kimi 98, DeepSeek 97, and GLM 95, for most enterprises, all scores are sufficient. For over 80% of enterprise AI use cases—translation, summarization, code completion, document analysis, customer service Q&A—the difference between 97 and 100 is imperceptible to the end-user. When performance differences become marginal, corporate focus shifts from which model is smarter to three key factors: cost, stability, and ecosystem integration.
First, cost. Goldman Sachs data shows prices per million tokens for mainstream large language models have fallen over 90% since early 2022. DeepSeek V4-Pro's output price is as low as $0.87 per million tokens, 34 times cheaper than GPT-5.5. Xiaomi's MiMo V2.5 Pro has standardized the output price for 1M-context long texts at $3, eliminating all tiered pricing. The entire industry is transitioning from "capability-based pricing" to "cost-based pricing."
Stability and ecosystem integration are the true differentiators. When a company deploys an AI Agent, it needs not just a conversational model, but an execution body capable of accessing data, sending emails, operating ERP systems, and managing approval workflows. The model is the brain, but a brain without limbs accomplishes nothing. As model performance converges, the advantages of established players like TENCENT and Alibaba may become more pronounced due to their extensive enterprise ecosystems and access points. Take WorkBuddy as an example. It employs a model-agnostic design, allowing seamless switching between five major models—Hunyuan, DeepSeek, GLM, Kimi, and MiniMax—and supports custom integration of any OpenAI-compatible API. Users can switch models mid-conversation or enable an Auto mode where the system routes tasks based on complexity, using the cheapest DeepSeek Flash for simple translation and Kimi K3 for complex reasoning. WorkBuddy absorbs the switching costs between models.
More importantly, WorkBuddy is backed by the entire TENCENT ecosystem. After WeChat authorization, users can issue commands via mobile with automatic execution on PC. It integrates with Tencent Docs, Tencent Meeting, WeChat Work, TAPD, and connects to over 30 external tools like Feishu, Kingsoft, and Qichacha via the MCP protocol. A command like "create a PPT from last week's sales data and send it to meeting attendees" is executed autonomously. This is not a chatbot; it's an "Agent that gets work done." In the Agent era, the model is merely the engine, while connectivity forms the chassis.
Pure Model Companies Face a Challenging Outlook
As model performance converges, a price war becomes inevitable. The impact of such a war varies drastically across companies. For tech giants like Alibaba, ByteDance, and Xiaomi, model APIs essentially serve as customer acquisition channels for their broader commercial empires—Qianwen ties to Alibaba Cloud, Doubao to Volcano Engine, and MiMo to Xiaomi's device ecosystem. These APIs don't need to be profitable; they can even sustain minor losses long-term if they drive cloud service subscriptions, compute consumption, and hardware sales. This is the "ecosystem subsidy" logic.
However, pure model companies like Zhipu AI and Moonshot AI lack such a subsidy pool. They must rely on API revenue to cover R&D and compute costs. In an era of exponentially growing Agent-driven token consumption, maintaining low prices means losses increase with sales volume. The market is already voting with its feet. On July 17, the day K3 launched, shares of several leading domestic model companies plunged significantly more than other AI leaders. K3 was the catalyst for two reasons. First, it proved "Chinese open-source models can catch up to top-tier closed-source models." Second, Moonshot AI is reportedly negotiating a new funding round of up to $20 billion, targeting a $30 billion valuation and accelerating its Hong Kong IPO plans. When a competitor both open-sources for free and heads towards a $30 billion listing, the "scarce AI investment ticket" becomes less scarce.
A deeper issue exists: in products like WorkBuddy, where users can freely choose between models with minimal performance differences, both enterprises and individual consumers will opt for the cheaper option. Pure model companies are caught in the middle—squeezed from above by open-source competitors with only a 2-3 point performance gap and from below by price wars backed by ecosystem subsidies from giants. Pricing power is eroding.
Complex Implications for Upstream Silicon Hardware
The impact of falling model prices on upstream silicon hardware is complex and requires the dust to settle. On one hand, it appears negative. A relentless token price war suggests model company profits will struggle to improve. Without sufficient profits and cash flow, their willingness to increase capital expenditure is likely weak.
On the other hand, a different narrative emerges. Cheaper tokens incentivize enterprises to integrate AI Agents into every business process. A reimbursement approval that previously took two hours of manual work can now be handled by an Agent for mere pennies. Companies will automate every automatable process. Data supports this: in Q2 2026, weekly token calls on major platforms surged from approximately 21 trillion in early April to 46.66 trillion by June 15, doubling in a quarter. CMB International predicts that by 2030, global monthly token consumption for Agent scenarios will exceed 21 quintillion, with a compound annual growth rate of 165%. This reflects the Jevons paradox in economics: increased efficiency leads to greater total consumption. Cheaper coal didn't reduce its use; it led to more applications for coal. Similarly, cheaper tokens won't reduce compute demand; they will lead to more AI adoption across scenarios. From this perspective, demand for computing power remains robust, supporting the upstream silicon narrative.
This divergence currently lacks a clear resolution. If it were clear, current silicon stock price volatility wouldn't be so pronounced. Compounding this uncertainty, the upstream silicon narrative's sustainability faces some doubt just as the Korean stock market grapples with a deleveraging crisis. The KOSPI has retreated 26% from its June 22 high, entering a technical bear market. All 14 leveraged ETFs have fallen below their issue price. Korean financial regulators urgently raised the margin requirement for single-stock leveraged ETFs from 10 million won to 30 million won and suspended new product listings. While financing balances have only dropped from 38.6 trillion won to 35.6 trillion won (less than 8%), deleveraging is far from over.
Therefore, assessing silicon hardware is currently in an awkward window. The long-term direction seems clear—cheaper tokens drive Agent adoption and exponential compute demand growth—yet is clouded by uncertainty. Simultaneously, the short-term deleveraging shock hasn't fully played out, with selling pressure from Korea potentially spilling over again. Clarity may emerge once Korean leverage unwinds and trading dynamics for silicon stocks stabilize.
Where is Certainty Found?
Relative certainty lies with companies like TENCENT and Alibaba, and their potential to drive the Hang Seng Tech Index. The logic is clear. As models converge, competition shifts from "whose model is smarter" to "whose ecosystem is stronger." TENCENT boasts WeChat with 1.43 billion monthly active users, WeChat Work, Tencent Docs, Tencent Meeting, TAPD, and the WorkBuddy Agent portal with 13 million daily active users. Alibaba has DingTalk, Alipay, Alibaba Cloud, Cainiao, Taobao, and its Qianwen model is the backend for Apple Intelligence in China. These ecosystems were not built overnight and cannot be overturned by a single model. When model performance differences are only 2-3 points, ecosystem barriers become the deepest moat.
On July 20, the Hang Seng Tech Index rose over 3.5% intraday. Alibaba surged over 4%, TENCENT gained over 3%, and Meituan climbed over 3%. Beyond the ecosystem revaluation logic driven by model convergence, an incremental catalyst emerged: the U.S. did not renew the national emergency concerning Hong Kong, allowing Executive Order 13936 to expire. Valuation pressures on Hong Kong's tech sector are being removed one by one.
From a longer-term perspective, the significance of K3 isn't whether it's 2 points stronger or weaker than Claude. Its significance, alongside GLM-5.2 and DeepSeek V4, lies in proving that model capability convergence is an irreversible trend. As models converge, value migrates from the model layer to the application and ecosystem layers. A statement by Professor Li Hongbing of Beijing University of Posts and Telecommunications aptly summarizes this: "Digital economy competition is evolving from model competition to scenario competition, ecosystem competition, and platform competition. Enterprises with industry data, high-frequency scenarios, and ecosystem synergy capabilities will hold greater advantages." Translated into investment language: pure model companies sell "brains," which are easily replaceable; companies with ecosystems sell "nervous systems," which are far harder to replace.
Comments