Samer Choucair: AI Performance Gap Narrows to 3% as China Challenges the Premium on US Models and Redraws the Investment Landscape
Samer Choucair, investment entrepreneur, said the artificial intelligence race has entered a new phase after a Chinese model moved to within approximately 3% of the leading US model in a recent LiveBench snapshot.
The development does not signal the end of American technological leadership. It does, however, challenge one of the sector’s most important valuation assumptions: that a substantial and persistent performance advantage will always justify a premium for US models.
Just 2.3 Points Separate the Leaders
DeepSeek V4.1 Flash Max Effort recorded a LiveBench score of approximately 81.1, compared with 83.4 for the leading US model, Claude Fable 5.1 Max Effort.
The difference was only 2.3 points, equivalent to approximately 2.8% of the leader’s score.
In agentic coding, however, the Chinese model reversed the ranking, scoring 77.3 points against 66.1 for its US competitor.
Choucair cautioned that these figures should not be interpreted as evidence of comprehensive parity. LiveBench measures a specific range of capabilities, while results vary across benchmarks, model versions, and testing methodologies.
Stanford estimated the gap between the leading American and Chinese models at approximately 2.7% in March 2026. By contrast, a US government assessment published in May estimated that DeepSeek V4 remained approximately eight months behind frontier models across a broader set of tasks.
The Real Contest Is the Cost of Intelligence
Choucair said the principal investment risk does not arise solely from model rankings. It lies in the economics of deploying those models at scale.
DeepSeek said V4.1 Flash uses an architecture designed to reduce cache memory requirements, lower operating costs, and offer off peak pricing at half the standard peak rate.
An analysis by J.P. Morgan Asset Management also found that Chinese models were processing a growing share of tokens on OpenRouter, although the United States retained the lead in overall capability.
Choucair warned against treating token share as equivalent to market share. OpenRouter is heavily used by developers and agent based applications and does not necessarily represent the large enterprise contracts that generate the sector’s most valuable and durable revenues.
From Model Performance to Profitability
According to Choucair, a narrowing performance gap combined with lower pricing could place downward pressure on the average revenue earned by US model laboratories.
It could also make returns on investments in data centres, semiconductors, and energy infrastructure increasingly dependent on actual utilization rather than projected demand alone.
Chinese competition remains constrained by other factors, including regulation and intellectual property disputes.
In September, US security agencies accused several Chinese companies, including DeepSeek, of using large scale model distillation to extract capabilities from American systems. China rejected the allegations.
A New Investment Opportunity for Saudi Arabia
Choucair linked these developments to Saudi Arabia, where HUMAIN has announced plans to allocate approximately 14 gigawatts of capacity to what it describes as AI factories, alongside agreements with international technology companies and a joint venture with AMD.
He argued that the most important indicator for investors will not be the volume of announced power capacity. It will be the utilization rate of that capacity and the price at which computing power is sold to customers.
Lower cost models could reduce the expense of operating government and private sector applications. At the same time, they could compress data centre returns if infrastructure capacity expands faster than commercially funded demand.
Three Scenarios for Investors
Choucair identified three possible paths for the market.
The first is that the performance gap remains close to 3% while model prices continue to decline.
The second is that US models reestablish a more substantial advantage in sophisticated and computationally demanding tasks.
The third is that regulatory restrictions intensify, fragmenting the global AI market along geographic and geopolitical lines.
Investors should therefore monitor benchmark performance after every major model release, usage prices per million tokens, token share relative to actual customer requests, new US regulatory measures, and the pace of contracting and capacity utilization across Saudi AI infrastructure projects.
Choucair concluded that the 3% gap does not mark the end of the race. It is, however, a powerful investment signal.
Competition is shifting away from the question of who possesses the most powerful model and toward a more commercially decisive question: who can deliver comparable intelligence at a lower cost and convert it into sustainable revenue?