The competitive landscape for AI large language models is undergoing a notable shift. With model supply on the rise and technologies like distillation lowering research hurdles, new entrants are now breaking into the high-performance segment with investment that is a fraction of what leading players spend. Jefferies argues this trend is pressuring API prices and challenging the logic that huge capital outlays are necessary to build a durable competitive moat.

According to a Jefferies research note released on September 16, four new models joined the global top 14 list tracked by Artificial Analysis (AA) in September, hailing from Singapore, South Korea, the UAE, and the US. These models come from research teams with generally modest funding, with some accumulating only around $20 million in total, yet they have already secured spots among the world's leading models.

Analysts led by Edison Lee suggest that the growing number of LLM participants, rising local support for homegrown models, and the broad adoption of distillation are further lowering barriers to entry. At the same time, narrowing performance gaps and intensifying API price competition are placing fresh pressure on the AI supply chain, creating a situation of high capital expenditure with uncertain returns.

New entrants are entering at low cost, with distillation lowering the bar. The report highlights that the four new models in AA's top 14 in September are Agnes 3.0 Flash from Singapore's Sapiens AI, Motif-3-Beta from South Korea's Motif Technologies, K2 Horizon from the UAE's MBZUAI, and Apodex 1.1, developed by a California-based company with a research team in Singapore. Sapiens AI has raised roughly $20 million, while Motif Technologies completed a Series B of about $16.6 million, a scale far below that of major AI players.

These examples show that the capital threshold for competing in large models is contracting. Rather than replicating the massive training runs of incumbents, some new players are moving quickly by leveraging open models, small teams, and more efficient training methods. Distillation is a key driver here. By using outputs from existing models to train new ones, developers can acquire advanced capabilities at a fraction of the cost. Jefferies believes this technique is hard to fully block, and as adoption spreads, the cost for latecomers to catch up with leading models could keep falling.

That said, low-cost entry does not erase the advantages of top-tier players. Training frontier models still requires enormous compute, data, and engineering effort; newcomers are mainly gaining edge in specific capabilities or niche applications. The real change is that the capital gap between entering the market and building the strongest model is widening.

API prices are coming under pressure as models shift toward price competition. The rise in model numbers is starting to feed through to API pricing. Jefferies data shows some new models are pricing their APIs well below flagship offerings, with Apodex 1.1 offering a blended API price of $0.30 per million tokens, among the lowest in AA's top 15. Meanwhile, leading vendors are not broadly cutting prices; instead, they are seeking higher pricing through capability upgrades in next-generation models. The report argues this pricing strategy not only reflects better performance but may also be a way to demonstrate commercialization strength to capital markets, using higher intelligence to justify price increases and improve expectations for margins and return on investment.

The tension is that price hikes and cuts are happening at the same time. Flagship models try to keep premiums with better performance, while newcomers grab developers and application demand with low costs. Jefferies notes the US market alone now has seven major LLM providers, and more participants mean API price competition could persist.

The competition is shifting from raw intelligence to efficiency. Another change is that the industry is moving beyond a singular focus on model smarts to an equal emphasis on inference efficiency. This month, Jefferies introduced the AutomationBench-AA benchmark to measure capability in agentic automation tasks and has recalibrated historical data accordingly. In this context, some model makers are cutting inference costs through architecture and memory optimization. On September 10, DeepSeek released the V4.1 Flash model, with a blended API price of $0.20 per million tokens, down 74% from its predecessor.

This model also uses a casual encoder-decoder design, an Engram conditional memory mechanism, and Compressed Sparse Attention 2 to slash hardware requirements: HBM needs are down by about a quarter and SSD needs by about an eighth. Jefferies sees this path as prioritizing compute and memory efficiency over just chasing raw intelligence. As a result, the competitive dimensions in AI models are expanding. With inference token demand growing quickly, model providers need to improve ability while also lowering the cost per generated token. For app developers, if cheaper models can handle enough tasks, price differences may become a more decisive factor than pure performance gaps.

Capital spending continues to climb, but returns need fresh validation. Jefferies' concern over the AI supply chain is not about disappearing demand, but whether capital investment growth can align with eventual returns. New players are entering global leaderboards with tens of millions of dollars, while incumbents keep expanding compute, data centers, and infrastructure. On the demand side, AI inference needs are still surging, and better models will unlock new use cases, so compute demand has not lost its growth base. The issue is that with abundant model supply and intense API price competition, how long will it take for new infrastructure investments to convert into sufficient revenue and profit?

Jefferies therefore suggests the AI industry may be moving toward a phase that stresses capital efficiency. For top firms, continuing to increase spending may still be key to staying ahead, but the broader sector needs to pay more attention to redundant builds, resource utilization, and the actual business value of new compute. Ultimately, Jefferies expects the industry to reduce overall capital outlays through consolidation, partnerships, and smarter resource allocation. AI infrastructure demand remains strong, yet demand growth does not automatically guarantee adequate returns on all investments, and this will be a key factor in how the market reassesses AI valuations in the next phase.