DeepSeek V4 and the Crack It Reveals

A Widening Crack

In July 2025, DeepSeek V4 arrived on schedule.

On benchmarks, it once again delivered a remarkable scorecard — multiple metrics approaching or even surpassing the world's top closed-source models, while training costs maintained the kind of "Chinese efficiency" that keeps Silicon Valley uneasy. But if we fixate only on benchmark scores and parameter counts, we'll miss something far more significant: the true importance of DeepSeek V4 lies not in how powerful it is, but in how it has made an existing crack impossible to ignore.

What is that crack?

The power to define AI's value is shifting from those who build models to those who use them.

When Models Become Infrastructure

Looking back at the AI narrative over the past three years, the protagonist has always been the model itself. GPT-4's launch drew massive crowds, Claude 3.5 sparked debates about capability boundaries, and Gemini's multimodal abilities dazzled the world. The entire industry's attention, capital, and discourse have been concentrated on a single question: Who can build the most powerful model?

That question once mattered enormously. In an era of generational gaps in technical capability, the model was the moat, compute was the fortress, and parameter count was power. OpenAI could define an era with GPT-4 because no one could replicate it.

But the emergence of the DeepSeek series, alongside the rise of open-source forces like Qwen, Llama, and Mistral, is systematically dismantling that logic. When V3 approached GPT-4-level performance at a fraction of the cost, the industry shuddered. Now, as V4 further narrows the gap with the latest closed-source models, a reality is becoming increasingly clear: Top-tier model capabilities are being commoditized at an accelerating pace.

This doesn't mean models no longer matter. It means models are transitioning from "scarce assets" to "infrastructure." Just as electricity was once the exclusive advantage of a few factories before becoming a socket on the wall — the real value is no longer in generating power, but in what you do with it.

Two Sides of the Crack

On either side of this crack stand two fundamentally different worldviews.

On one side: "Model Supremacy." This camp believes that victory is determined by the absolute capability of the foundation model. Bigger parameters, more data, stronger reasoning — that's everything. Their business logic: build the most powerful model, charge through APIs, and maintain pricing power through technological superiority. OpenAI, Anthropic, and Google DeepMind have largely followed this path over the past few years.

On the other side: "Applications Define Value." This camp argues that as model capabilities converge, true differentiation will emerge at the application layer — whoever best understands user scenarios and can translate model capabilities into real productivity gains holds the pricing power. It's not about how high your model scores, but how much money you save or earn for your clients.

The release of DeepSeek V4 tips the scale further toward the latter.

The reason is simple: when an open-source model can deliver near-top-tier closed-source performance at extremely low cost, "model capability" itself ceases to be a scarce resource. Any developer, any company, can stand on the shoulders of DeepSeek V4 to build their own applications. Model providers' bargaining power is weakened, while application developers' freedom and influence expand dramatically.

The Migration of Value-Defining Power

This power shift didn't happen overnight, but it's accelerating.

We're already seeing the signals. In the enterprise market, more and more clients no longer ask, "Which model are you using?" but rather, "What problem can you solve for me?" The model has become a backend engine; front-end value is defined by scenario understanding, data assets, and workflow integration. An AI application deeply specialized in legal document review may carry far more commercial value than the underlying model itself, because it possesses deep understanding of legal scenarios and a continuously compounding professional data flywheel.

In developer communities, the winds are shifting too. The most active AI projects on GitHub are no longer model reproductions but model-based toolchains, Agent frameworks, RAG systems, and vertical-domain solutions. Developers are redirecting their energy from "understanding how models work" to "figuring out how to use models well."

Capital markets have also caught the scent of this shift. In the first half of 2025, application-layer companies captured a larger share of global AI funding than foundation model companies for the first time. Investors are realizing that the odds of winning the "cash-burning race" at the model layer are shrinking, while opportunities to find PMF (product-market fit) at the application layer are multiplying.

DeepSeek's Unique Role

In this power migration, DeepSeek plays a fascinating role.

It is both a standout among model builders and an active accelerator of model commoditization. Through open source and extreme cost optimization, DeepSeek is effectively doing one thing: continuously lowering the barrier to accessing model capabilities, thereby ceding the stage of value creation to the broader ecosystem.

This stands in stark contrast to OpenAI's approach. OpenAI seeks to build a closed moat at the model layer, maintaining high-margin API business through capability leadership. DeepSeek's strategy is more akin to "letting the water flow to the fields" — the model is the water, application scenarios are the fields, and when water is cheap and abundant enough, whoever owns the best fields reaps the greatest harvest.

This also explains why the discussion sparked by DeepSeek V4's release has long transcended the technology itself. It touches on the most fundamental question in the AI industry: Where does value truly reside?

The New Landscape Beyond the Crack

If this crack continues to widen — and it almost certainly will — we'll witness a profound restructuring of the AI industry.

First, the model layer will evolve toward a dual-track structure of "few major players + open-source ecosystem." A handful of institutions with top-tier compute and R&D capabilities will continue pushing frontier models, but their business models will increasingly rely on ecosystem rather than pure API revenue. Meanwhile, open-source models will become the default foundation for most applications.

Second, the application layer will experience a true explosion. When model costs are no longer a bottleneck, countless application scenarios previously deemed infeasible due to high costs will become viable. Education, healthcare, law, manufacturing — every vertical domain could give birth to its own "AI-native" companies.

Third, the core dimensions of competition will shift. From "whose model is stronger" to "who understands the scenario better," "whose data flywheel spins faster," and "who delivers the better user experience." For entrepreneurs, this is both a challenge and an opportunity — you don't need to train your own large model, but you need to understand your users better than anyone else.

Fourth, AI's geopolitical landscape may be rewritten. DeepSeek has proven one thing: the speed of catching up at the model layer can be remarkably fast. This means that AI competitiveness across nations and regions will increasingly depend on the richness and maturity of application ecosystems, rather than solely on foundation model capabilities.

Writing as the Crack Widens

Every shift in technological paradigm is accompanied by a migration of value-defining power. In the PC era, value shifted from hardware to software. In the mobile internet era, value shifted from operating systems to apps and services. In the AI era, we are witnessing value shift from models to applications.

DeepSeek V4 didn't cause this crack, but it delivered the blow that made it impossible to ignore.

For those still going all-in on the model layer, this is a moment for serious reflection. For entrepreneurs standing at the application layer, armed with scenarios and users, their era may be arriving faster than expected.

The crack has appeared. Water flows downhill, and value flows toward the user. This is not a prediction — it's what's happening right now.