Kimi K3 Reassesses AI Narrative: Will Cheap Models Disrupt the Computing Power Surge?

TL;DR
· After the release of Kimi K3, AI-related stocks fell last Friday, as the market revisited concerns over computing expenditures related to DeepSeek R1.
· The approximately $700 billion AI infrastructure plan by U.S. cloud giants this year is a key figure impacting the valuation of open-weight models.
· Open models may lower the prices of closed-source models, but they could also expand AI usage, which does not equate to a peak in computing demand.

Deutsche Bank noted in a report aimed at retail investors that after the release of Moonshot AI's Kimi K3, AI stocks experienced a 'DeepSeek moment' last Friday: the market began to question whether it is still reasonable for U.S. tech giants to spend hundreds of billions of dollars on AI infrastructure if Chinese open-weight models can approach cutting-edge closed-source models at a lower cost.

This is not merely a technical debate. The report mentions that last Friday, the 'Tech Seven' fell by about 1.8%, and the Philadelphia Semiconductor Index dropped by about 1.6%, bringing its cumulative decline for the week to over 10%. Alphabet's Google, Microsoft, Amazon AWS, Meta, and Oracle are expected to invest around $700 billion this year to build AI capabilities, an increase of about 70% from last year.

The reason for investors' anxiety is straightforward: the AI bull market over the past few years has been built on the assumption that stronger models require more chips, more data centers, more power, and higher capital expenditures. However, after DeepSeek R1, Kimi K3 once again reminds the market that cheap, lightweight, downloadable, and modifiable models are catching up to closed-source giants.

Why does the release of one model impact NVIDIA and cloud vendors?

The shock triggered by DeepSeek R1 in January 2025 is still part of the market's memory. The report recalls that at that time, the market was concerned that Chinese teams could create 'good enough' models using older chips and at lower training costs, leading to NVIDIA's market value evaporating by about $600 billion in a single day.

The questions raised by Kimi K3 are similar: if open-weight models can achieve usable levels in many application scenarios, do enterprises still need to pay top dollar for the most expensive closed-source models? Do cloud vendors still need to build data centers at the current pace? Will chip demand shift from 'the more, the better' to 'cheaper, more efficient, and more decentralized'?

The key issue here is not whether Kimi K3 has fully surpassed OpenAI or Anthropic, but whether it is enough to shake the market's default assumptions about AI business models. As long as 'good enough' models become cheaper, the pricing power of closed-source model vendors will be compressed, and the growth stories of cloud vendors and chip companies will also be forced to undergo more scrutiny.

This is why the release of a model can resonate in the stock market. Among the Tech Seven, many companies are both providers of AI models and applications and the world's largest buyers of data centers; the semiconductor index directly bears the market's expectations for changes in demand for GPUs, networking equipment, and AI servers.

What truly impacts valuations is not the term 'open source'

Many AI models referred to as 'open source' by the market are more accurately described as 'open weight'.

Closed-source models resemble 'plug-and-play' services. Flagship models like OpenAI's GPT and Anthropic's Claude are typically controlled by developers in terms of model weights, training methods, pricing, updates, and security restrictions, with users accessing them through applications, APIs, or cloud services. Their advantages are convenience, stability, and high integration; however, the downside is that users find it difficult to gain control over the underlying mechanisms and must accept the prices and rules set by service providers.

Open-weight models, on the other hand, make the trained parameters available, allowing users to download, deploy, fine-tune, and even technically modify certain security settings. Models like DeepSeek R1, Meta's Llama, and several from Mistral are closer to this category. They do not necessarily disclose complete training data, training code, and replication details, so they are not equivalent to strictly defined open-source systems.

This distinction is important for enterprise clients. Open-weight models mean greater control: enterprises can run models on their own servers, reducing reliance on external APIs; they can also customize for specific industries, languages, or tasks; for institutions in finance, healthcare, and government where data cannot leave local environments, local deployment is more attractive.

However, 'downloadable' does not equate to 'free'. Enterprises still have to pay for computing power, electricity, engineering, monitoring, maintenance, and security costs. Open models may not possess the same general capabilities, product experience, and service commitments as closed-source flagship models. In other words, their impact does not lie in replacing closed-source models overnight but in providing customers with another option.

Four Areas Where AI Valuation is Questioned

Models like Kimi K3 primarily challenge 'scarcity'.

If strong models can be replicated, modified, and hosted by more developers, AI models themselves may increasingly resemble foundational software infrastructure rather than high-margin products enjoyed by a few companies. Closed-source model vendors can still profit from products, data, security, and ecosystems, but charging high prices solely based on 'model capability superiority' will become more difficult.

The second area impacted is pricing power. As AI agents and enterprise employees use more tokens, the costs of model calls are shifting from small bills during pilot phases to significant items in enterprise budgets. When AI subsidies decrease and leading model vendors begin to charge more explicitly per token, cost-sensitive customers will be more willing to try open models.

The third area is capital expenditures. Currently, U.S. tech giants' AI investment scale is already substantial, with Google, Microsoft, AWS, Meta, and Oracle planning around $700 billion for AI capability construction this year, one of the most closely watched figures in the market. If enterprises can solve most tasks with smaller, cheaper, and more specialized models, investors will naturally question whether the growth in spending on chips, networks, electricity, and data centers is too rapid.

The fourth area is the U.S. technological moat. The continuous release of Chinese open-weight models has weakened the market's perception of the 'unshakeable' nature of U.S. AI leadership. The report mentions that this year, the number of tokens processed by Chinese models on the OpenRouter developer platform has already surpassed that of U.S. models. This does not mean that the U.S. AI advantage has disappeared, but it indicates that developer usage and ecosystem diffusion are becoming more decentralized.

These impacts collectively point to one result: closed-source models remain the benchmark for cutting-edge capabilities, but the narrative that 'closed-source models monopolize everything' has weakened.

Open Models Do Not Only Bear Bad News for AI; They May Also Amplify Demand

The market is most likely to interpret such events as 'cheap models hitting chip demand'. However, the report provides a more balanced judgment: powerful open models will indeed challenge the business models of closed-source models and semiconductor demand expectations, but they may also benefit AI adoption, application development, and enterprise customization.

The reason is that lower prices usually expand usage. Cheaper and more efficient models will enable more companies to integrate AI into customer service, office work, research and development, advertising, coding, data analysis, and internal processes. Companies that previously found API costs too high, data unable to leave, and models uncontrollable may start deploying AI due to open-weight models.

This aligns with the logic of the 'Jevons Paradox': as technological efficiency improves and unit costs decrease, overall consumption may actually increase. The increased efficiency of steam engines did not reduce coal demand but rather expanded the use of steam power. AI may experience a similar situation—while single inference becomes cheaper, the number of calls, application scenarios, and endpoints may increase significantly.

Therefore, computing demand may not peak; rather, the structure of demand may change. A few cutting-edge closed-source models will still require the strongest training and inference clusters; a large number of open models may be deployed on clouds, local servers, mobile phones, cars, and other edge devices. Chip demand may also shift from a singular pursuit of the strongest GPUs to a greater emphasis on inference efficiency, networking, electricity, storage, and edge computing power.

For application companies, open models may actually lower input costs. Enterprises can use cheaper models for vertical scenarios, software companies can develop more specific products on top of models, and cloud vendors can continue to profit by hosting open models, providing inference services, and enterprise support.

Closed and open models will coexist, but regulatory responsibilities remain unclear

The more likely scenario is not that open source defeats closed source, nor that closed source re-monopolizes, but rather that both coexist in the long term.

Closed-source models will continue to play the role of cutting-edge capabilities, industrial-grade quality, security assessments, and deep product integration. Open-weight models will force closed vendors to control prices, increase flexibility, and shift competition from 'whose model weights are stronger' to data, product design, memory capabilities, security, efficiency, and ecosystem services.

Model companies opening weights are not necessarily charitable acts. Developers can profit through hosting APIs, premium subscriptions, faster inference, enterprise support, fine-tuning, security services, and service level agreements; cloud service providers can sell the computing power needed to run open models; consumer platforms can embed model capabilities into recommendations, advertisements, and user experiences; application companies can build functionalities at lower costs.

What remains unresolved is the boundary of responsibility. Open-weight models give users more control but also place more responsibility on users for testing, security protection, network security, updates, and reliability. Once a model is downloaded, modified, and deployed locally, determining who bears the responsibility for misuse, failures, or security incidents becomes more complex than in a closed API model.

Regulation may also change the competitive landscape. Closed model vendors like Anthropic advocate for catastrophic risk testing, external evaluations, and ongoing disclosure requirements for cutting-edge AI developers; critics worry that excessive regulation may give well-funded large companies an advantage, thereby suppressing open ecosystems and small team innovations.

Thus, the 'second DeepSeek moment' brought by Kimi K3 is not a simple negative for AI. It is more like a stress test: the $700 billion AI expenditure, closed-source model pricing, U.S. technological advantages, and chip demand all need to answer the same question—where will the money in the AI industry come from when 'good enough' intelligence becomes cheaper?