Graphics & Text | Sister Tang

Before the markets opened this week, investment forums and communities were rife with pessimism, predicting an impending market downturn. Soaring oil prices have stoked inflation fears, while expectations of a Federal Reserve rate hike add further pressure. Rising interest rates will increase corporate borrowing costs and make it harder to justify high stock valuations.

Amid this macroeconomic squeeze, the AI sector itself is facing calls for a slowdown. On September 12 (U.S. time), Dario Amodei, CEO of Anthropic, published an article advocating for a coordinated deceleration in the development of the most advanced AI capabilities to allow more time for safety evaluations. Sam Altman of OpenAI and Elon Musk of xAI soon voiced their support for this initiative.

This has added another layer of concern for investors banking on the expansion of the AI industry: if cutting-edge R&D slows down, will the demand for data centers and chips diminish?

As anticipated, market sentiment turned sharply negative on Monday. The Philadelphia Semiconductor Index fell 5.9%, and Nvidia’s stock dropped 3.4%. On the same day, Nvidia’s Jensen Huang held a public discussion with Donald Trump at the All-In Summit in Los Angeles. While Trump acknowledged the need to approach AI risks cautiously, he emphasized that the U.S. should continue to lead in AI development.

Throughout the week, the Federal Reserve’s rate hike was confirmed, while oil prices retreated from their highs. By Friday’s close, the S&P 500 Index was slightly down for the week, but the Philadelphia Semiconductor Index, which had plummeted on Monday, had recovered its losses. Nvidia’s stock rose 1.82% for the week, SK Hynix gained 2.48%, and Lenovo Group surged nearly 20%.

Will cutting-edge AI R&D truly slow down just because a few leading U.S. companies have made non-binding statements? Even if it does, will the narrative of AI infrastructure driving stock gains become obsolete? And if the rapid growth of AI infrastructure ends, will new growth stories emerge to sustain capital market enthusiasm for AI?

To answer these questions, we must first understand how AI demand is evolving.

01 Aligning Market Adoption with Model Progress

From last year to this year, a notable shift in the large model industry is that inference is gradually overtaking training as the primary use case for computing power. Therefore, slowing down the training of cutting-edge models does not necessarily mean a rapid decline in demand for computing power.

Training cutting-edge models requires significant computing power to enable them to acquire new capabilities. Trained models that write code or organize files for users also continuously consume computing power—this is inference. Nvidia stated in its 2026 annual report that inference has surpassed training as the primary workload.

From a cost perspective, Zhipu AI also noted in its 2026 interim results announcement that in the era of inference, where models are called billions of times daily, inference costs now exceed training costs. Slowing down cutting-edge R&D may alter the scale and timing of the next round of training, but models already in use will continue to operate, consuming computing power.

With this demand foundation, a slowdown in cutting-edge development may not necessarily be detrimental. Beyond the need for safety evaluations, it could also create conditions for AI democratization:

Giving the entire industry time to absorb and integrate existing advancements.

E-commerce service provider Shopify revealed in April this year that it trained a dedicated model based on Alibaba’s Qwen3, enabling shop owners to create automated workflows for their stores using natural language. After offline testing met standards and the model went live, it was discovered that shop owners still needed to modify old workflows and set up emails—actual requests not fully covered in testing. The team supplemented training materials and adjusted tools based on real requests, closing the gap between testing and actual usage over two weeks.

Integrating models into business operations requires such an adjustment period, yet new options continue to emerge. From July 9 to September 10, in just two months, six models—GPT-5.6, Kimi K3, Qwen3.8, Claude Fable 5.1, GPT-6 Astra, and DeepSeek-V4.1-Flash—were released or opened up, with an average interval of about 13 days between releases. From September 1 to 10, Anthropic, OpenAI, and DeepSeek each launched new models in succession.

Just as enterprises finish testing what a model can do and prepare to integrate it into their business, new versions and models from other vendors arrive. Tasks that were previously difficult can now be handed over—or can they? If switching to a cheaper model, can the same level of effectiveness be maintained? The same applies to individuals: just as they figure out one way to use a model, they must reassess which tasks can be confidently handed over to AI—should they stick with familiar tools or spend time testing better options?

Model advancements are outpacing the market’s ability to absorb and utilize these capabilities.

Companies developing paid products for specific needs like tutoring or customer service may also face replacement of original functions by general-purpose models. Chegg, which charges subscription fees for services like exercise solutions and step-by-step explanations, has already admitted in its annual report that students increasingly view ChatGPT as a substitute for its services. Such application teams need time to redefine their products, distinguish which needs can already be met by general-purpose models and which are still worth investing in, and then adjust their services based on new model capabilities.

Even model companies themselves need such an absorption process. Within the same company, the R&D side pursues the next breakthrough, while the product side must find a market for existing capabilities, letting clients know what to use them for and how to use them cost-effectively.

If the leap in cutting-edge capabilities slows down a bit, users will have more time to become proficient with the tools, and product teams can improve services and accumulate clients based on usage feedback. More suitable products and willing clients mean existing capabilities have a better chance of becoming stable cash income; more daily usage will also drive inference demand, creating a virtuous cycle of application democratization, commercialization, and computing power demand.

However, just because the entire industry needs absorption time does not mean all companies are willing to slow down together.

02 Why a Unified Slowdown Is Difficult

Slowing down the development of cutting-edge models and giving the market time to absorb them also gives model companies an opportunity to sell their existing products better. However, even if a model company is willing to slow down R&D, it cannot guarantee that competitors will do the same; moreover, it takes time for user adoption to translate into real revenue growth. The potential benefits of slowing down must be weighed against the potential orders lost during the waiting period and the ongoing expenses that must still be covered.

For buyers of large models, another round of cutting-edge model updates may mean a new cycle of comparison and testing; for sellers, it could be an opportunity to secure business. Enterprises spending money to have AI help employees write code or answer customer questions need time to make good use of new tools; companies like OpenAI, Anthropic, and Zhipu AI, which develop large models and charge users, also hope clients will use their products effectively—but they must also guard against competitors launching stronger or cheaper models to attract clients away.

In his initiative, Amodei also proposed that coordinated slowdowns should allow companies time for safety research while preserving their commercial advantages and U.S. leadership. In his plan, how much U.S. companies can slow down depends on whether they can maintain their lead—meaning the closer Chinese large models get, the smaller the acceptable space for slowdown becomes.

Even without considering Chinese competition, more than just a few U.S. companies are involved in this race. Nvidia both sells chips and systems and trains its own large models. AI search company Perplexity has already used Nvidia’s Nemotron 3 Super, launched in March this year, for its search services, and Nvidia is also collaborating with other developers to train the next generation of open models.

Even if cutting-edge R&D truly slows down, cloud providers may not necessarily follow suit in slowing their expansion. Amazon both invests in Anthropic and provides the computing power for training and running its models through its cloud business AWS, then offers Claude services to enterprises; Microsoft is both a major shareholder of OpenAI and provides computing power through its Azure cloud platform.

In the second quarter, AWS’s operating profit grew from approximately $10.2 billion year-on-year to $16.6 billion, a year-on-year increase of about 64%; AI business revenue, annualized based on the current revenue rate, exceeded $25 billion, achieving triple-digit year-on-year growth. Amazon also explained that AI usage drives demand for traditional cloud services like storage and databases.

To maintain such growth, cloud providers must build data centers and purchase equipment in advance. Amazon explained in its 2026 shareholder letter that investments in data centers, power, and equipment typically precede customer billing by 6–24 months. A significant portion of AWS’s capital expenditures in 2026 is already supported by customer commitments, with much of the investment not expected to generate revenue until 2027–2028.

The willingness of a few model companies to slow down R&D is not enough to decelerate the entire industry. Cloud providers and chip manufacturers continue to support expansion through investments and procurements, and the success of these arrangements depends on whether AI sees more paid usage.

Nvidia disclosed in its latest quarterly report that some AI cloud providers procure its equipment, and in return, Nvidia commits to purchasing their cloud services. As of July 26, such commitments had reached $36 billion. These procurement commitments provide partial order support for cloud providers’ expansion; if other clients actually use the corresponding computing power, Nvidia’s procurement commitments decrease accordingly. Therefore, Nvidia hopes cloud providers will continue buying equipment and that more clients will use the computing power provided by that equipment.

Cloud providers have also resorted to debt financing for their expansion. Oracle disclosed in June this year, when introducing its AI cloud infrastructure investment plan, that it had raised $43 billion in debt financing in fiscal year 2026. For companies that have already borrowed for construction and made procurement commitments, expanding AI paid usage can both increase revenue and help recover investments and reduce payment pressures. Even if a few model companies slow down R&D, these cloud providers and equipment suppliers still have an incentive to drive existing models into more businesses.

03 Conclusion

Even if the rapid growth of AI infrastructure ends one day, the capital market may still find its next growth phase in the democratization of applications. The customer service business mentioned earlier is an entry point—existing models entering more daily work could both increase revenue for software companies and save costs for enterprises using the software.

However, the money saved by users will not automatically become profits for AI companies.

In the customer service business, if enterprises use AI to handle more customer inquiries, they may save on labor costs; software companies providing customer service software charge service fees but also incur costs for model calls, product maintenance, and customer service. If clients can easily switch to a similar product with comparable effectiveness at a lower cost, software companies will find it difficult to raise fees, and the savings may remain largely with the enterprises using the software.

For investors, what matters next is not only who can sell AI but also who can use AI to improve their own business. The next phase of growth may not be limited to companies selling AI.

Disclaimer: This article is for learning and communication purposes only and does not constitute investment advice.

Like, share, and repost! Your support is what keeps us updating!