US-China AI Round 2: Price, Not Performance, Takes Center Stage... What Lies Ahead for Semiconductors? [Techtalk]

Chinese Value-for-Money AI Reshapes Competitive Landscape
Price Wins Over Performance in the New Race
All AI Models Expected to Become Similar in Capability
"In the End, Raw Computing Power Will Decide the Winners"

Editor's NoteThis article aims to demystify the fundamental yet often unfamiliar technologies that are essential in our lives, from AI and semiconductors to telecommunications and biotechnology.

Chinese artificial intelligence (AI) models such as Deepseek and Kimi are shaking up the industry. By offering decent performance at prices far lower than cutting-edge models, the argument that there is no longer a need to insist on American AI is gaining credibility. Even the leading tech giants, once focused on performance, are now joining the price war. The new equation for AI competitiveness introduced by these Chinese models is expected to have a major impact not only on AI companies but also on the business outlook of semiconductor makers supplying the computing infrastructure.

Fail to Adapt, You Fall Behind... the 'Deepseek Kill Line'

A green square based on the price and performance of the Deepseek V4 Pro model analyzed by Artificial Analysis. Models included inside are likely to be outcompeted by open-source AI. X Capture

Recently, within the U.S. large language model (LLM) research community, the so-called 'Deepseek Kill Line' graph has drawn widespread attention. This graph classifies which models surpass or fall short of the latest Deepseek model based on per-token cost and benchmark performance. Any model inside the large box, defined by Deepseek's cost and performance thresholds, is essentially considered unnecessary for use.

The results were shocking. The latest models from so-called 'Frontier Lab' companies—including OpenAI, Anthropic, and Google DeepMind, which lead the AI industry—barely managed to clear the kill line, but the majority of other AI models were eclipsed by it. Internet users who saw the graphic posted reactions such as, "There are too many useless AIs now," "A bloodbath is coming to the industry," and "The chicken game is about to begin."

Frontier Labs Shift From Performance to Cost Competition

OpenAI has entered price competition with a strategy to differentiate GPT-5.6 according to performance levels, dividing it into Sol, Terra, and Luna. OpenAI

Until now, Frontier Labs have been obsessed with developing the largest, most powerful models, aiming for artificial general intelligence (AGI). Anthropic's high-risk, high-performance AI 'Mythos' boasted impeccable capabilities, but its enormous token consumption drew complaints from paid users.

However, as Chinese AI models—'not the most powerful, but just right'—infiltrated the low-cost market, even Frontier Labs appear to be rethinking their business strategies. OpenAI's recently released GPT-5.6 exemplifies this trend. The model is split into Sol, Terra, and Luna tiers based on performance: Sol and Terra handle complex work, while Luna, the lightest and cheapest variant, is dedicated to simple work, basic reasoning, and test outputs.

OpenAI made the bold decision to cut the API call costs for Luna by 80%, aligning the new prices closely with those of Chinese models.

If the Performance Gap Disappears..."The Return of Compute Power Competition"

In fact, Frontier Labs had already anticipated that the AI race would soon pivot from a focus on performance to cost. Instead of always pushing out their most powerful model, Google DeepMind is now prioritizing the release of lightweight models with maximized response speeds, such as Gemini Flash.

This shift is underpinned by the belief that, ultimately, AI will become commoditized.

OpenAI's Stargate Data Center. OpenAI

Joshua Akyiam, an AI scientist who until last month worked as a strategist for the future at OpenAI, said in an interview with the American tech podcast "MTS," "AI is gradually advancing toward self-updating, automated technology, and at a certain point, all models will likely reach a similar level." He added, "Even now, the gap between open-source AIs and Frontier Labs' models is only a matter of a few months of technology development."

He went on to predict, "In the end, everyone will be competing with very similar AIs." This means that individual model performance will be practically indistinguishable and treated like any other commodity. The decisive factor will be who can deliver a greater number of AI services to customers.

This shift in the competitive landscape will have significant ramifications for hardware companies supplying semiconductors to AI firms, such as Nvidia, Intel, and Samsung Electronics. Akyiam forecast, "In this situation, whoever possesses greater raw AI computational power will win." As AI development matures, the contest is expected to tilt back toward securing physical AI data centers.

This content was produced with the assistance of AI translation services.