Discussing DeepSeek V4 and Meituan LongCat with America's AI Big Three

An Unexpected Cross-Pacific Dialogue

When we presented the latest technical breakthroughs of DeepSeek V4 and Meituan LongCat to individuals associated with America's AI "Big Three" — OpenAI, Google DeepMind, and Anthropic — the feedback far exceeded expectations. Not only did they show keen interest, but the exchanges also revealed an emerging industry consensus: The dominant theme of AI competition is undergoing a fundamental shift.

The first half was about competing on compute resources; the second half is about competing on physical foundations. What once sounded like an insider quip is fast becoming the industry's lived reality.

DeepSeek V4: The "Efficiency Beast" That Made Silicon Valley Take Notice — Again

Since the viral success of R1, DeepSeek has become one of the Chinese AI forces that Silicon Valley watches most closely. The arrival of V4 has once again reset external perceptions of Chinese teams' engineering capabilities.

DeepSeek V4 is understood to incorporate multiple deep optimizations at the architecture level, continuing the company's consistent philosophy of "doing more with less." A researcher with ties to Google DeepMind admitted: "DeepSeek's engineering efficiency is impressive. Under constrained compute, they've pushed architectural innovation to the extreme — a path diametrically opposed to Silicon Valley's 'pile on resources first, optimize later' approach."

Even more noteworthy is DeepSeek V4's exploration of new training paradigms. Analysts suggest that V4 may have further deepened the application of MoE (Mixture of Experts) architecture, striking a new balance between inference cost and model capability. This "efficiency-first" technical philosophy is forcing American counterparts to re-examine their own technical roadmaps.

Meituan LongCat: China's Dark Horse in the Long-Context Arena

If DeepSeek V4 was a "pleasant surprise within expectations," then Meituan LongCat was the true "unexpected discovery" of these conversations.

Meituan, a company best known for local life services, has produced the LongCat model through its AI lab, demonstrating formidable competitiveness in long-context processing. According to publicly available information, LongCat excels in ultra-long text comprehension and multi-turn dialogue memory retention — directly addressing one of the most critical pain points of current large models: the effective utilization of context windows.

A technical expert with an Anthropic background remarked: "Long context isn't just about stretching the window longer. The key is whether a model can maintain precise information retrieval and logical coherence across ultra-long sequences. LongCat's technical approach deserves serious study."

Meituan's pursuit of long context has its own unique business logic — local life service scenarios inherently involve the cross-processing of massive amounts of structured and unstructured information, from merchant data and user reviews to delivery dispatching. These real-world scenarios provide LongCat with training ground that Silicon Valley labs would find hard to replicate.

First Half: Compete on Compute. Second Half: Compete on Foundations

Throughout conversations with U.S. AI practitioners, an increasingly clear verdict emerged: AI competition is shifting from a "compute arms race" to a deeper contest over "physical foundations."

The concept of "physical foundations" encompasses multiple dimensions:

• Chips and hardware ecosystems: Who can command full-stack capabilities from chip design to cluster scheduling
• Data flywheels: Who possesses richer, higher-quality real-world scenario data
• Engineering systems: Who can achieve peak efficiency in training and inference at the systems level
• Energy and infrastructure: "Hard constraints" such as data center power supply and cooling solutions

Over the past two years, the dominant narrative was "whoever has more GPUs wins." But DeepSeek has proven through tangible results that compute is not the only variable, while Meituan LongCat demonstrates that scenario data and engineering optimization can also build formidable moats.

A former OpenAI engineer reflected: "In 2023, we thought China was 18 months behind in AI. In 2024, we felt the gap was narrowing. By 2025, looking at work like DeepSeek V4 and LongCat, we're starting to realize this isn't a single-dimensional catch-up game — it's the parallel evolution of multiple technical paths."

Reassessing the Competitive Landscape

From a global perspective, the AI competitive landscape is becoming more diverse and complex:

First, the efficiency-driven approach is gaining legitimacy. Across successive generations of products, DeepSeek has proven that world-class results can still be achieved under compute constraints. The success of this approach is attracting more teams to join the "efficient training" camp.

Second, application-scenario-driven technical innovation has been underestimated. Behind Meituan LongCat are the real needs of hundreds of millions of users. This R&D model of "coming from real scenarios and returning to real scenarios" may hold more long-term value than pure benchmark competitions.

Third, the Big Three's anxiety is real. During the exchanges, it was palpable that leading U.S. AI companies are paying significantly more attention to their Chinese counterparts. They are no longer just making polite inquiries — they are seriously reading papers, reproducing experiments, and analyzing architectures.

Outlook: New Variables in the Age of Physical Foundations

Looking back from mid-2025, the competitive logic of the AI industry has undergone profound changes. Compute still matters, but it is no longer the decisive factor. The real differentiators are shifting to deeper layers of the physical world — chip energy efficiency, data quality, engineering systems, and energy supply. These elements, "not glamorous enough" yet utterly critical, will determine who goes further in the next phase.

The emergence of DeepSeek V4 and Meituan LongCat is not only another demonstration of China's AI prowess but also a powerful challenge to the global AI competition paradigm: When compute is no longer the sole bottleneck, where does the real moat lie?

The answer may well be hidden in that simple verdict — in the second half, the contest is over physical foundations.