AI Compute Prices Surge Across the Board + Rise of Two Domestic Coding Tools: The Wave of Self-Reliance Arrives

The AI Industry Enters a Dual Movement of Cost and Autonomy

Recent developments in tech and finance reveal two clear main threads: first, the across-the-board rise in AI infrastructure costs, with prices climbing collectively from GPUs to cloud compute; and second, the accelerating "de-OpenAI-ization" of domestic AI coding tools, moving toward self-developed cores and open-source ecosystems. Though these two threads appear independent, they in fact point to the same proposition—the AI industry is transitioning from a period of technological euphoria into a new stage of cost rationality and self-controllable autonomy.

This article focuses on three of the most industry-significant events: the tightening of U.S. export controls on top-tier AI models, the same-day launch of two domestic AI coding tools, and the phenomenon of "compute inflation" running through both hardware and cloud services.

The U.S. Imposes the Strictest Export Controls in History on Top-Tier AI Models

The U.S. government has officially brought Anthropic's top-tier large model into its export control system, with unprecedented severity. What makes this move special is that it not only prohibits exporting the model outside U.S. borders, but also bars foreign nationals within the United States from using it—including AI researchers holding U.S. work visas.

A Signal Breaking Through the Boundaries of Traditional Export Controls

Historically, the U.S. export control system has primarily targeted hardware entities (such as chips and weapons equipment), with the Export Administration Regulations (EAR) and the Arms Export Control Act (AECA) as its core framework. The 2023 chip export restrictions mainly targeted GPU hardware such as NVIDIA's H100 and A100. This time, however, treating the large model itself as a controlled "technology item" is a disruptive extension of the traditional definition of "technology export"—the regulatory logic has upgraded from "controlling hardware" to "controlling capability," listing even "use by foreign nationals within U.S. borders" as a target of control, greatly increasing regulatory granularity.

By extending its regulatory scope from "cross-border flow" to "domestic use," this measure represents a major breakthrough of the traditional export control framework. The U.S. justification is that the relevant model poses "excessive risk," out of concern it could be used for military purposes or large-scale disinformation—so much so that it even claims the U.S. government itself should not use it.

This extreme level of control reflects the high alertness of the U.S. toward AI security risks, while also serving as a strategic manifestation of preserving technological hegemony. However, such measures may produce multiple counterproductive effects:

• Hindering global AI technology exchange and intensifying the division of technology camps;
• Damaging the international competitiveness of U.S. AI companies;
• Forcing other countries to accelerate investment in domestic AI R&D.

Far-Reaching Impact on Global AI Governance

This move is likely to prompt China, the EU, and others to accelerate independent AI R&D and standard-setting, driving global AI governance toward greater diversification. For enterprises, striking a balance between innovation and compliance is becoming an unavoidable strategic issue. Controls may slow the diffusion of technology, but they cannot stop the continued catch-up of global AI capabilities.

Domestic AI Coding Tools Enter the "Era of Two Giants"

On June 13, domestic AI coding tools reached a highlight moment with two giants launching on the same day, marking the milestone stage at which Chinese AI coding tools officially entered the era of "de-OpenAI-ization."

Zhipu AI: Switching to a Self-Developed Core, Surpassing GitHub Copilot

At the BAAI Conference, Zhipu AI released a new version of its AI coding tool. The biggest change is the switch from relying on external APIs (such as OpenAI) to a fully self-developed core, with deep adaptation to its self-developed GLM-5.2 large model. According to data released by Zhipu, the tool surpasses the current version of GitHub Copilot on mainstream coding benchmarks such as HumanEval and MBPP, supporting core features including whole-project code generation, intelligent debugging, code review, and automatic unit test generation.

Benchmark Background: HumanEval is a code generation benchmark set released by OpenAI in 2021, containing 164 hand-written Python programming problems. It measures a model's functional correctness using the pass@k metric (the probability that at least one of the k candidate answers generated by the model passes all test cases), and has become the most widely cited standard for evaluating coding ability in the industry. MBPP (Mostly Basic Python Problems), released by Google Research, contains 974 Python programming problems aimed at entry-level programmers, focusing on the application of basic algorithms and data structures. Combining the two benchmarks provides a relatively comprehensive evaluation of a model's overall coding performance, but there is also the risk of "benchmark-gaming" through over-optimization for the test sets—actual engineering-scenario performance must be judged in conjunction with real deployment data.

Moonshot AI: Open-Sourcing Kimi K2 Code, Focusing on Private Deployment

On the same day, Moonshot AI officially released and open-sourced Kimi K2 Code, a large model optimized specifically for coding scenarios, claiming to reach the level of the world's strongest open-source coding model in code generation and comprehension. The open-source strategy means enterprises and developers can deploy the model on their own compute, without relying on external APIs—which is especially critical for industries with extremely high requirements for data and code privacy, such as finance, healthcare, and defense.

The strategic value of open-source large models lies not merely in free access, but in building an ecosystem moat. Meta's LLaMA series rapidly gathered tens of thousands of developers through open-sourcing, forming a massive library of fine-tuned models and toolchains. Moonshot AI follows similar logic: once an open-source model is integrated into an enterprise's private deployment environment, the migration cost becomes extremely high, creating very strong user stickiness. For coding scenarios, open-sourcing also means the community can independently audit the model's security, bias, and backdoor risks—an irreplaceable trust endorsement for industries with stringent compliance requirements such as finance and defense.

The Three Driving Forces Behind De-OpenAI-ization

This trend is not merely a policy requirement, but a reasonable choice across three dimensions: cost, security, and customization.

Understanding the cost difference requires knowing the pricing unit of large language models, the "Token": it is not simply equivalent to one word, but the smallest linguistic unit in the model's vocabulary—in Chinese, one Token typically corresponds to 1–2 characters. For enterprise-grade coding tools with over a million daily calls, API fees can account for 30%–50% of operating costs, which is precisely the core economic motivation for domestic vendors shifting to self-development:

• Cost: OpenAI API call costs are roughly $15–30 per million tokens, while domestic models can be reduced to under $5 on self-developed compute;
• Security: Industries such as finance, healthcare, and defense require the entire code-generation workflow to be completed within their own or domestically-developed (Xinchuang) environments, and cannot upload source code to overseas APIs;
• Customization: Domestic AI has a naturally superior depth of understanding of Chinese programming contexts and local tech stacks (such as HarmonyOS, WeChat Mini Programs, and DingTalk plugins) compared to general-purpose English models.

Open-source means trustworthy, auditable, and privately deployable, which will significantly accelerate the penetration of domestic AI coding tools into security-sensitive industries. GitHub Copilot's market dominance is being shaken by the dual pincer of "open-source + self-developed."

The "Compute Inflation" Sweeping the Entire Industry

If the autonomy of coding tools is a proactive choice in industrial upgrading, then the compute price surge sweeping the entire industry is a real-world pressure that must be confronted.

GPU Inflation: Supply Contraction Rather Than Demand Explosion

NVIDIA's RTX Pro 6000 Blackwell series professional GPUs have seen across-the-board price increases, about 55% higher than the launch price. Notably, the core reason for the price hike is not a demand explosion, but supply contraction—NVIDIA has shifted the allocation focus of TSMC's CoWoS advanced packaging capacity from consumer-grade GPUs to AI training and inference GPUs (enterprise-grade products such as the H100 and B200).

CoWoS (Chip on Wafer on Substrate) is an advanced packaging technology developed by TSMC. Through a silicon interposer, it tightly integrates the GPU core with HBM high-bandwidth memory within the same package, greatly increasing inter-chip data transfer bandwidth—a key process for top-tier AI GPUs to achieve ultra-high compute power. This capacity is extremely scarce: only TSMC worldwide has the ability to mass-produce it, and its build cycle takes 18–24 months, making rapid capacity expansion impossible. This reveals the most fragile supply bottleneck in the current AI industry chain: the compute ceiling of the entire global AI infrastructure is, to a considerable extent, constrained by the single packaging process of one foundry.

Since AI GPU profit margins are as high as 70%–80%, far above the 50%–60% of consumer-grade GPUs, prioritizing limited capacity for high-margin products is a rational business decision. Analysts expect this strategy will not change over the next 6 to 12 months, meaning creative professionals and gamers will continue to face pressure.

Cloud Compute Prices Rise in Tandem, With Increases Up to 50%

The compute price surge is extending downstream along the industry chain. Kingsoft Cloud announced that due to the continued rise in hardware costs such as AI chips, electricity, and cooling, it will raise the prices of its AI compute services by 15%–50%, with high-end GPU compute seeing increases of up to 50%. Previously, domestic cloud providers such as Alibaba Cloud and Tencent Cloud had already successively raised AI compute prices, by increases ranging from 10% to 40%.

Rising compute costs have become a clear industry trend. This means the business models of AI applications will face new tests—against the backdrop of rising compute costs, inference efficiency optimization (through techniques such as KV Cache reuse, speculative sampling, and model quantization) is becoming a key variable determining an enterprise's cost competitiveness. Whoever can achieve higher inference efficiency and a better cost structure will gain the advantage in competition. This also reinforces, from another angle, the strategic appeal of the domestic model's "self-developed compute + private deployment" route.

Autonomy and Efficiency: The Dual Proposition of AI Industry Competition

From the tightening of AI model export controls, to the rise of two domestic coding tools, to the across-the-board compute price surge, these seemingly scattered events together sketch out the new landscape of the AI industry: technological autonomy is becoming an irreversible strategic choice, while cost efficiency is becoming the key variable determining who will have the last laugh.

For China's AI industry, control pressure may actually accelerate the process of local innovation, while the dual path of "open-source + self-developed" provides a realistic and feasible roadmap for breaking through international monopolies. Future competition will no longer be merely a contest of model capabilities, but a comprehensive battle over cost control, security compliance, and ecosystem building.

Key Takeaways