(Looking at this ecosystem map, it’s clear that merging computing power and models to provide standard services involves far too many stakeholders)
Shanghai finally saw some sunlight in March. When the clouds parted, the temperature barely climbed to 16°C. Last night, as I stared at my screen reviewing a complex piece of AI-generated component code, I suddenly realized a rather spine-chilling detail.
Unprotected Business Logic
Last week, I visited a large enterprise and public institution for a technical exchange. Their IT infrastructure was highly advanced, with every developer on the team equipped with an AI coding assistant. Writing foundational architecture and business interfaces was as smooth as flowing water.
But then I took a look at their network request logs.
Most of their core database table structures, private protocol fields, and even some highly confidential government approval business logic were packaged directly as context and sent via API to overseas Claude and Gemini servers. These foreign models are undeniably brilliant, spitting out perfect business code in seconds. But frankly, this is like throwing your company’s financial ledgers onto the street and asking a stranger to do the math for you. No matter how accurate the calculation, your privacy is completely exposed.
National policies actually recognized the danger of such practices a long time ago. In recent years, regulations have been introduced to restrict the use of unvetted overseas LLMs in classified or critical infrastructure scenarios, and recent compliance reviews have become even stricter. But the reality is that many grassroots developers, driven by tight deadlines, are still secretly using overseas proxies. Once the code is written, as long as nobody mentions it during delivery, no one knows that these hundreds of lines of elegant logic were written by a foreign AI.
If you really dig into this, it’s quite terrifying.
It’s Not About Unstoppable Leaks, It’s About a Flimsy Foundation
Let’s think about this from another angle. Why does everyone risk policy violations just to use overseas models?
Because they work incredibly well. Developer tools are very honest: whoever helps you clock out on time is the one you use. It’s not that there haven’t been domestic alternatives before, but often, enterprises, for the sake of so-called “security and compliance,” forcefully deploy extremely clunky private setups on their intranets. I previously tested an early internal coding model launched by a top-tier domestic enterprise; it could handle simple loops, but the moment it encountered complex concurrency logic, it started speaking gibberish.
This leads to a contradiction: if strict checks on overseas access are enforced, crucial domestic government agencies and large enterprises must face the reality of a precipitous drop in AI programming efficiency.
Rely on every enterprise to build their own models? (Rubs temples) That’s clearly unrealistic. The training and inference costs of large models are terrifyingly high. Where would a regular company find the massive funds to buy expensive compute cards, or the army of algorithm engineers needed to specifically optimize code generation capabilities? The result is that everyone ostensibly complies on the surface while complaining bitterly behind closed doors.
Therefore, I believe what should genuinely be put on the agenda is a state-led, officially certified computing center and model service provider alliance.
Tearing Off the Packaging to Look at the Hard Metrics
Let’s do a horizontal comparison of the current solutions.
Currently, the market for enterprise-level AI programming implementations is roughly divided into a few categories. The first category is vendors directly selling API endpoints. The second category sells all-in-one hardware-software appliances, for example, those bundled with domestic hardware and certain open-source domestic models.
To put it bluntly, many of these all-in-one appliances are in an awkward position right now. A few days ago, I had coffee with a friend working in infrastructure at an industry-leading company. He complained to me that using domestic computing power to run domestic models for coding assistance currently involves an extremely painful underlying adaptation process. Forcing foundational logic originally written on mature foreign frameworks to migrate onto, say, heterogeneous architectures results in significant performance loss and frequent crashes. It might be fine for a small company just playing around, but for entities with extremely high stability requirements like financial systems or government networks, this kind of piecemeal private deployment is simply a disaster.
But here’s a question you must consider: what if an officially led alliance took care of all this dirty work?
For instance, led by the state, integrating several supercomputing centers, and bringing in the top domestic LLM vendors to unify interface standards. This would provide government and enterprise clients with a secure, isolated, and traceable “AI programming fundamental service.” In this alliance, the computing power would genuinely run in domestic data centers, the models would be strictly filed and verified by the Cyberspace Administration, and the data flow would pass through national-level encrypted gateways. The alliance would even have standardized cleaning protocols to prevent malicious open-source backdoors from creeping into generated code.
For crucial government agencies and institutions that need to write code, they wouldn’t need to set up a bunch of servers themselves, nor would they need to maintain a dedicated operations team just to monitor computing power. They could directly call the alliance’s service API on demand. It’s like how every household used to dig their own wells for water, but now they are directly connected to the tap water pipeline. Costs go down, and performance is guaranteed.
(With this densely packed industry chain structure, lacking official integration, it’s truly difficult for enterprises to create something usable just by relying on their own luck)
Opportunities Forced by Necessity
Sometimes I think, if this “officially certified computing alliance” could truly be rolled out on a massive scale, it might solve more than just surface-level information security issues.
It would actually transform into an immensely large and authentic pool of real-world scenario demands. Think about it: there are so many government and enterprise units, and state-owned IT outsourcing firms in China, generating a massive daily demand for code writing, refactoring historical “spaghetti code,” and code reviews. If all this traffic is forcefully directed to domestic model providers and computing platforms, wouldn’t this genuine, high-intensity business pressure directly force the capabilities of domestic AI to step up?
In the past, when writing code, if a domestic model got stuck or was hard to use, developers would just secretly switch back to a foreign API to get the job done. But if policies fundamentally cut off this escape route, and an official alliance provides the only legal and compliant channel, then domestic model vendors will have to bite the bullet and optimize their code generation quality. Underlying hardware vendors will also be forced to stubbornly grind through their headache-inducing heterogeneous compilers, squeezing out performance drop by drop.
Or maybe I’m overthinking it. Perhaps in the early stages, everyone will still try to break their code into even smaller fragments, or just run a pitifully small model locally to make do. But technological development often works this way: obstacles that seem insurmountable at first will gradually be leveled as long as they are fed with enough real-world scenarios. Developers who get used to the services of this alliance will ultimately polish these domestic tools until they become smoother and smoother to use. ¯\(ツ)/¯
This can be considered a forced mechanism accompanied by a few growing pains.
The faint sunlight outside the window has been covered by thick clouds again, making the sky as dark as dusk. Regardless, life goes on. I’m going to go fix that buggy piece of domestic component code from last night.
References:
- 2025 China Large Model Industry Development Research Report
- The Dilemma of AI Industry Compliance: Key Challenges from Multi-Country Policy Regulation to Practical Implementation
- China AI Computing Center Research Report 2024
- A Series of Landmark AI Achievements Released in Yizhuang, Beijing
—— Lyra Celest @ Turbulence τ.
