The AI price war is intensifying, and entry barriers are also decreasing: How long can high capital expenditures last?

The competition in large-scale AI models is undergoing new changes: model supply continues to increase, technologies such as distillation are lowering the R&D threshold, and new players are entering the high-performance model market with significantly less investment than leading companies. Jefferies believes this trend is depressing API prices and challenging the AI industry's previous logic of relying on massive capital expenditures to build competitive barriers.

According to a research report released by Jefferies on September 16, four new models were added to the global top 14 language model rankings tracked by Artificial Analysis (AA) in September, originating from Singapore, South Korea, the UAE, and the United States. The research institutions behind these models generally have limited funding, with some having raised only about $20 million in total, yet they have already entered the top global model rankings.

Jefferies analysts, including Edison Lee, believe that the increasing number of participants in the LLM market, growing regional support for local models, and the widespread application of distillation technology are further lowering barriers to entry. At the same time, narrowing performance gaps between models and intensified competition over API prices are placing new pressures on the AI supply chain, characterized by "high capital expenditures and low certainty of returns."

New players can enter the market at low cost; distillation technology lowers the barrier to entry.

The report shows that the four new models that entered the top 14 of the AA rankings in September are Agnes 3.0 Flash from Singapore's Sapiens AI, Motif-3-Beta from South Korea's Motif Technologies, K2 Horizon from MBZUAI in the UAE, and Apodex 1.1, which is headquartered in California, USA, with its R&D team in Singapore.

Sapiens AI has raised approximately $20 million in total funding, while Motif Technologies has completed a Series B funding round of approximately $16.6 million. Their overall investment scale is significantly smaller than that of leading AI companies.

These cases illustrate that the capital threshold for competing on large-scale models is decreasing. Some new players are not replicating the path of leading companies that train basic models on a large scale, but instead are rapidly iterating by leveraging open models, smaller teams, and more efficient training methods.

Distillation technology is a key driver of this trend. By training new models using the output of existing large models, developers can acquire some of the capabilities of advanced models at a lower cost. Jefferies believes that this type of technology is difficult to completely block, and as its application becomes more widespread, the cost for newcomers to catch up with leading models may continue to decrease.

However, low-cost entry does not mean that the advantages of leading manufacturers have disappeared. Training cutting-edge models still requires significant investment in computing power, data, and engineering. New players are gaining competitiveness primarily through specific capabilities and niche scenarios. The real change lies in the widening capital barrier between "entering the market" and "becoming the strongest."

API prices are under pressure, and models are beginning to compete on price.

The increase in the number of models has begun to impact API prices. Jefferies data shows that the API prices of some new models are already significantly lower than those of leading models. For example, the hybrid API price of Apodex 1.1 is $0.3 per million tokens, making it one of the lowest-priced models among the top 15 AA models.

Meanwhile, leading manufacturers have not universally joined the price reduction trend, but instead are seeking higher pricing through upgrades to the capabilities of their next-generation models. The report argues that this price increase not only reflects improved model capabilities but may also be a way to demonstrate commercial viability to the capital market—that is, supporting higher prices with higher levels of intelligence, thereby improving market expectations for profit margins and returns on investment.

The problem is that price increases and decreases are happening simultaneously. Leading models are trying to maintain their premium through performance upgrades, while new entrants are competing for developers and application demand by offering lower costs. Jefferies points out that there are already seven major model providers in the US market alone, and the increase in participants means that price competition in the API market is likely to continue .

Shifting from "competing on intelligence" to "competing on efficiency"

Another shift in model competition is that the industry is moving from simply pursuing model intelligence to simultaneously competing on inference efficiency. This month, Jefferies introduced the AutomationBench-AA benchmark to measure a model's capabilities in agent automation tasks and to recalibrate historical data accordingly.

Against this backdrop, some model vendors have begun to reduce inference costs through model architecture and memory optimization . DeepSeek released its V4.1 Flash model on September 10, with its hybrid API price dropping to $0.2 per million tokens, a 74% decrease from the previous version.

The model further reduces hardware requirements through techniques such as Casual Encoder-Decoder design, Engram conditional memory mechanism, and Compressed Sparse Attention 2: HBM requirements are reduced by about a quarter, and SSD requirements are reduced by about one-eighth. Jefferies believes that this approach emphasizes computational power and memory efficiency rather than simply pursuing model intelligence.

This also means that the competitive dimensions of AI models are increasing. With the rapid growth in demand for inference tokens, model vendors not only need to improve their capabilities but also reduce the cost of generating each token. For application developers, if lower-cost models can already accomplish enough tasks, price differences between models may be more important than simple performance differences.

Capital expenditures are still increasing, but the returns need to be re-evaluated.

Jefferies' concerns about the AI supply chain are not primarily about the disappearance of AI demand, but rather whether the growth rate of capital investment can match the final returns. On the one hand, new players are already able to enter the top global model rankings with investments in the tens of millions of dollars; on the other hand, leading companies are still continuously expanding their investments in computing power, data centers, and infrastructure.

From the demand side, the demand for AI inference is still growing rapidly, and the improvement of model capabilities will continue to create new application scenarios. Therefore, the demand for computing power itself has not lost its growth basis. The question is, as the supply of models becomes more abundant and API prices are suppressed by competition, how long will it take for new infrastructure investments to translate into sufficient revenue and profit?

Jefferies therefore believes that the AI industry may be gradually entering a phase that places greater emphasis on capital efficiency. For leading companies, continuing to expand capital expenditure may still be an important means of maintaining technological leadership, but the industry as a whole needs to pay more attention to redundant construction, resource utilization, and the actual commercial value corresponding to new computing power.

Ultimately, Jefferies predicts that the industry may reduce overall capital expenditures through consolidation, collaboration, and more efficient resource allocation. Demand for AI infrastructure remains strong, but "demand growth" does not automatically equate to "sufficient returns on all capital investment," which will be a crucial variable in the market's reassessment of AI valuations in the next phase.

Risk warning and disclaimerInvesting involves risk; please exercise caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Any investment decisions made based on this information are at your own risk.