Why Are China's Open-Source AI Models Getting Stronger? A Deep Dive Into Three Comeback Strategies

Entering a Critical Juncture in the New Round of US-China AI Rivalry

As we enter a critical juncture in the new round of US-China AI rivalry, the United States is attempting to open up a lead through hardware generational gaps and ecosystem fragmentation, completing the closed loop of its own commercial path—while China launches wave after wave of cost-based counterattacks using open-source models and extreme cost-effectiveness. On the surface it looks like an ordinary competition, but beneath lies powerful undercurrents. According to in-depth analysis by relevant Bilibili creators, this is no longer a simple technology race, but a gamble over tech supremacy that neither side can afford to lose. Drawing on the original material, this article offers a forward-looking breakdown of the logic behind China's AI industry "taking a different path" to stage its comeback.

The Current State: No One Is Truly Ahead, but Everyone Is Fighting for Their Lives

If we still treat AGI as the sole marker of this AI revolution, then the current situation can be summed up in one sentence: No one is truly ahead, but everyone is fighting for their lives.

Within just 8 days in July, four top-tier models were released in rapid succession: Musk's Grok 4.5, OpenAI's GPT-5.6, Meta's Muse Spark, and Moonshot AI's Kimi K3. According to Artificial Analysis's scoring, the number of vendors crossing the 50-point threshold surged from two in June to six, while the gap between the top three was a mere 3 points. The shelf life of the word "leading" has shrunk from the "monthly" level after DeepSeek V3's breakthrough to today's "weekly race" level.

More important than the rankings is the fact that the rules of the game themselves are changing. According to data from OpenRouter, the world's largest model-invocation platform, the share of programming-type workloads has soared from 11% at the start of 2025 to over 50%, and a single agent request consumes roughly 15 times the tokens of an ordinary conversation.

The Deeper Logic of the Shift Toward Inference: The shift from the training side to the inference side is essentially a paradigm leap in the AI industry from "R&D-driven" to "deployment-driven." The training-side race centers on metrics like parameter count and FLOP consumption, while the inference-side race centers on tokens per second (TPS), energy per token (Joules/Token), and latency (TTFT, time to first token). The rise of agent workflows has dramatically amplified inference demand—a typical ReAct-framework agent, when completing a single task, requires multiple rounds of "think-act-observe" loops, each requiring an LLM call, and token consumption grows exponentially. The surge in programming workloads in OpenRouter's data is a direct reflection of large numbers of programming agents (such as Cursor, Devin, etc.) coming online in bulk. This trend means the strategic value of inference chips, KV cache optimization, and speculative decoding is rapidly surpassing that of training chips themselves.

This means the entire industry is shifting from the "training side" to the "inference side" at a speed beyond imagination, and the main battlefield of competition has moved from the lab to a comprehensive contest of combat strength across chips, electricity, capital, and application scenarios.

America's Open Hand of Advantages and China's Real-World Dilemma

The United States' current advantage is out in the open and stacked with buffs. In AI chips, the US controls over 70% of 2nm capacity and absolute control over HBM, backed by more than 20 years of the CUDA software ecosystem. According to CFR statistics (on a quality-adjusted basis), US advanced AI chip capacity is roughly 35 times that of China.

The Deeper Logic of the CUDA Ecosystem Moat: CUDA (Compute Unified Device Architecture), launched by NVIDIA in 2006, is currently the de facto programming standard in AI training and inference. Its moat comes not merely from hardware performance, but from the software ecosystem accumulated over 20 years: mainstream frameworks like PyTorch and TensorFlow are deeply tied to CUDA operator libraries (cuDNN, cuBLAS), and tens of thousands of open-source models and pretrained weights are optimized for the CUDA format by default. Developers migrating to other platforms (such as Ascend's CANN or Cambricon's MLU) need to rewrite or adapt large numbers of low-level operators, at extremely high migration cost. There are typically two paths to breaking through the CUDA moat: building a compatibility layer (such as ZLUDA, Triton, etc.), or leveraging the open-source community to drive native framework support—but both require trading time for space.

In terms of capital density, the five largest US tech giants have poured over $700 billion (about 5 trillion RMB) into AI over the past year—more than the GDP of the vast majority of countries. Wall Street statistics show that from 2025 to 2027, the tech sector will add roughly $1.5 trillion in new debt, almost all of it going all-in on AI. The result: in core capabilities such as parameter scale, multimodality, and long-context consistency, America's leading closed-source models still occupy the first tier.

China's difficulties, by contrast, are stuck at the root. Although DUV multi-patterning can squeeze out capacity, its cost is relatively high, and domestic HBM lags roughly two generations behind the world's most advanced technology, severely limiting memory bandwidth.

The Physical Bottleneck of HBM and Memory Bandwidth: HBM (High Bandwidth Memory) is one of the core physical bottlenecks in large-model inference. During inference, LLMs need to repeatedly load hundreds of GB of model weights into compute units (the "memory wall" problem), and memory bandwidth—rather than compute power—is often the true bottleneck determining inference speed. Take the H100 as an example: its HBM3 provides 3.35 TB/s of memory bandwidth, while ordinary GDDR6X offers only about 1 TB/s. HBM production is highly concentrated among three companies—SK Hynix, Samsung, and Micron—and is deeply coupled with wafer foundry, requiring advanced packaging technology (such as CoWoS) to tightly integrate HBM with the GPU die. China lagging roughly two generations in HBM means that even if it narrows the gap on compute cores against NVIDIA, the memory bandwidth bottleneck will systematically drag down large-model inference efficiency—this is one of the hardest chokepoints.

In single-card compute power, inter-chip interconnect efficiency, and software ecosystem maturity, China still needs time to catch up. This is not a question of "can it catch up" but a question of "having its hands and feet bound."

Yet the reality is exactly the opposite—each time Anthropic or OpenAI releases a flagship model, China's open-source products always manage to deliver answers approaching flagship level on certain performance points or in user experience. Why? The answer lies in three paths.

Path One: Open Source—Dismantling the Closed Loop of Closed-Source Pricing Power

To understand this open-source move, you first need to understand the business logic of America's closed-source approach: enormous capital expenditure trains the strongest model, then costs are recovered through monopoly pricing (selling APIs or subscriptions), funding the next round. The key precondition for this chain to hold is establishing an effective absolute monopoly on model capability—only a monopoly can secure pricing power. According to Bain's estimates, maintaining the current pace requires investing roughly $500 billion in capital per year, corresponding to a capital pool of about $2 trillion to make the game work.

China's open source strikes precisely at this vulnerability. The logic is simple: as long as the world always has a "second option," a monopoly cannot naturally close in a market-driven environment. Once the closed loop cannot be completed, the several hundred billion dollars of annual capex becomes a Sword of Damocles hanging overhead.

The data speaks for itself: DeepSeek V4 Flash's input price is just $0.14 per million tokens, and with caching, the actual cost can be pushed below 3 cents; GPT-5.5 is about $5, with the output-side price gap reaching over 100x at its peak. On OpenRouter, the share of tokens from US companies calling Chinese models has soared from an average of 4.5% in the first half of 2025 to a peak of 46% in April, and total invocations of Chinese models have surpassed those of US models. Palantir CEO Karp warned in an interview: if US vendors continue to respond by charging per token, there is a fundamental problem at the root.

The Price Anchoring Mechanism

Chinese open-source models need not become the sole dominant player; they only need to be "close to, or with hope of catching up" to exert enormous price pressure. Especially in commercial tasks such as customer service, coding, document translation, and agent workflows, once the "good enough" threshold is crossed, users become extremely sensitive to price, and the appeal of open source is amplified without limit.

The Sovereignty Boundary Mechanism

What open source opens is the weights, meaning users can fully deploy locally, with data never leaving the data center and not subject to US tech companies' censorship or value biases. This effectively hands the regulatory power over AI back to individual nations, enterprises, and individuals. For the vast majority of countries in the world, this is not only a security consideration but also a defense of the deployer's freedom of decision-making.

Path Two: Optical Interconnect—On the Same Starting Line for the Next Generation of Interconnect

After two years of intense industry competition, an awkward fact has emerged: single-card compute power is actually already strong enough, and data centers are being built at full speed, yet effective compute just can't scale up. The reason is—the more cards there are, the more critical inter-card communication becomes.

The Physical Essence of MoE Architecture and Inter-Card Communication: MoE (Mixture of Experts) architecture is the scaling paradigm commonly adopted by today's top large models. Its core idea is to split the model into dozens or even hundreds of "expert subnetworks," activating only a few of them per inference, thereby maintaining an ultra-large parameter count while greatly reducing the compute per inference. However, MoE's cost is communication overhead—when experts are distributed across different GPUs, every token must be routed over the network to the GPU hosting the corresponding expert, and the results aggregated back (i.e., the All-to-All communication primitive). As cluster scale expands from a single 8-card machine to thousand-card clusters spanning racks, this communication overhead grows roughly quadratically, causing a sharp drop in effective compute utilization (MFU, Model FLOP Utilization). This is exactly why improving inter-card interconnect bandwidth, in current large-scale MoE model inference scenarios, has strategic value that even exceeds improving single-card compute power.

MoE models' expert parallelism and the back-and-forth movement of KV cache all consume inter-card bandwidth; the more cards there are, the more effective compute is lost instead. Interconnect is always the shortest plank. By the 224G generation, the effective transmission distance of copper wire has shrunk to the one-meter level—once you leave the rack, you can only rely on optics. Every segment of copper wire an electrical signal travels adds a bit more loss, a bit more power consumption, and a bit more latency.

The Industrial Background of CPO and Silicon Photonics: CPO (Co-Packaged Optics) is the core technology path for the next generation of data center interconnect, referring to packaging the optical engine and the switch chip on the same substrate, replacing the long-distance copper connections between traditional pluggable optical modules and chips with extremely short copper wires. Traditional pluggable optical modules (such as QSFP-DD) require electrical signals to undergo multiple conversions along the "chip → copper wire → SerDes → optical module" chain, each conversion incurring loss and latency. By bringing the optoelectronic conversion point as close to the chip as possible, CPO can theoretically reduce power consumption by 30-50% and support higher transmission rates (currently mass-produced 800G is evolving toward 1.6T and 3.2T). Chinese optical module manufacturers (Innolight, HG Tech, Eoptolink, etc.) have deep expertise in this area, mainly because optical modules belong to the intersecting field of precision manufacturing and optoelectronic integration, with far less reliance on EUV lithography than logic chips. As a result, the impact of export controls is relatively limited, allowing Chinese companies to occupy a dominant share in global competition.

Therefore, silicon photonics, CPO, and even photonic computing are the consensus path for both the US and China, and Jensen Huang has also made silicon photonics a focus of the next-generation platform. And on this matter of "playing with light," China holds many strong cards: Chinese manufacturers hold over 70% of the global optical module market, roughly half of the global share of 800G silicon photonic modules, and 1.6T is already in mass production. The Ascend 950's 1024-card TB-level bandwidth and three-microsecond latency are themselves products of hybrid optoelectronic interconnect. The conclusion is clear: on single-card compute power, China may need to wait one or two generations to catch up, but on the optical interconnect battlefield, many of the rules, paradigms, and visions place everyone on the same starting line.

Path Three: On-Device Scenarios and Talent—What China Truly Holds in Its Grasp

If open source locks down the opponent's pricing power and optical interconnect stakes out a position in the next-generation architecture, then application scenarios, engineering talent, and on-device services are what China truly holds in its grasp right now.

China's market fundamentals are large enough. The user base for generative AI has already reached 515 million.

The Data Flywheel Effect and the Strategic Significance of the On-Device Market: The "Data Flywheel" is a classic positive-feedback growth model for internet products, given new meaning in the AI era: user interactions generate data → data is used for model fine-tuning and reinforcement learning (RLHF/RLAIF) → model capability improves → more users are attracted → more high-quality data is generated. This mechanism allows products with large user bases to build a hard-to-replicate "experience barrier" in model capability. The diverse, high-frequency, and localized interaction data generated by China's 515 million generative AI users is a structural advantage that US closed-source vendors struggle to make up through technical means. The penetration of products like Doubao and Kimi in vertical scenarios such as college admissions application, career planning, and local content creation means the data they accumulate is not only large in volume but also carries strong cultural and contextual specificity, forming a significant differentiation advantage over closed-source models trained on general English-language data in local application scenarios.

A vivid example: ordinary users are already using Doubao, asking Kimi, and even comparing arguments across different large models; when filling out college admissions applications, examinees ask AI directly, with feedback that it is "better than the vast majority of career planners on the market." The overall market cap of on-device AI applications has already reached 500 billion RMB.

This shows that China does not necessarily have to stubbornly chase leaderboards and benchmark scores in lockstep with American closed-source models. Instead, it should combine its own economic model, market characteristics, and user needs to make the business model work, letting domestic models get "stronger the more they are used" and letting commercialization "proceed ever more smoothly." Price increases by vendors like Kimi also reflect market recognition from another angle.

In addition, a series of physical-layer explorations are worth watching: China's first 8-inch diamond heat-sink wafer production line has come online, used for chip heat dissipation; carbon-based technology paths, compute-power coordination, and green-electricity support are all attempts to reshuffle the deck at the physical level, further driving down token prices.

Two Narratives, Converging on the Same Destination

Back to the original question: with the most expensive chips, sky-high investment, and the strongest models, why has the United States failed to crush China's AI industry—and even seen domestic open-source models grow ever more resilient?

The answer is now clear—powerful chips and higher scores are only part of the AI race, not the whole. Using open source against closed source so that the monopoly ecosystem cannot close its loop; using optical interconnect to break through the physical limits across racks; using innovations in power and cooling to make up for the single-card gap; using a massive market to drive one's own data flywheel—these are all proven effective tactics.

Of course, every variable could rewrite the ending: whether advanced process nodes can be broken through, whether HBM can achieve results, whether open source can be sustained, and whether optical interconnect can deliver on its promises. But what is certain is that the US and China on this track have never been engaged in a simple "catching up with each other," but are walking their own paths toward the popularization of intelligence in radically different narrative styles. That, in itself, is a story interesting enough.

Key Takeaways