A few days ago I had dinner with a friend who works in government-enterprise software. He said he's bid on seven or eight tenders this year, won one, and the quoted price was pressed down to one-third of what it used to be. It wasn't that the client was deliberately lowballing—someone really was bidding at that price, and the deliverable looked decent enough.
He asked me what was going on. I told him to look into how fast AI writes bid documents and builds software now.
This got me thinking for quite a while. The bidding system in the software industry might not hold out much longer.
How the Old System Worked
China's government procurement market is massive. In 2024, total government procurement nationwide exceeded 3.37 trillion yuan, with open bidding accounting for 76.63%. IT-related government procurement was roughly 143.3 billion yuan, down 4.5% year-over-year, but the number of projects actually grew by 21.5%—nearly 79,000 projects. In other words, more projects, but smaller individual contracts.
In the past, software bidding operated stably because of several barriers.
First, qualifications. This ensured bidding companies were legitimate, with corresponding capabilities and experience. But this barrier never blocked too many people—companies that wanted in always found a way. Restrict it too much and you lose the meaning of competition.
Then, information asymmetry. Many companies only learned about tenders after the bid was announced, leaving insufficient preparation time, or never knew about them at all. The window between announcement and deadline objectively kept some competitors out. Even if you knew, preparing a proper bid document required massive time and manpower investment; companies that weren't fully prepared suffered during technical evaluation.
Finally, commoditized comparison. This is the core design logic of bidding: treat software as a standardized commodity, write clear functional parameters, and score against them item by item. If you hadn't built similar systems before, you couldn't produce many features, couldn't describe the technology clearly, and certainly had nothing to show during demos.
These barriers combined to keep competition within a relatively controllable range. For a long time, the system worked well enough.
Barriers Loosening Simultaneously
AI is simultaneously loosening these barriers.
Qualifications need little explanation—they were never a real moat anyway.
What's changing fastest is information asymmetry. Quite a few companies are already using AI tools to scan bidding information across the entire web. Bailian Zhineng's "Zhiliao Bid Intelligence" covers over 100,000 bidding websites nationwide, accumulating more than 300 million bidding records, updating 20 million daily, with AI prediction accuracy for winning bids exceeding 70%. Qianlima Bidding Network integrated DeepSeek and other large models in 2025, updating 300,000 bidding information items daily, with AI monitoring the entire process from project approval to bid announcements, automatically screening high-potential projects.
You used to miss bids simply because you didn't know they existed; now AI is watching constantly. Don't have time to prepare the bid documents? AI can help with that too.
In bid document generation, the market grew 230% year-over-year in 2025. iFlytek's "Spark Bidding" claims to compress bid document preparation from 30 days to 3 days, improve document compliance by 90%, and increase win rates by 40%. Tai Toubiao claims to generate thousand-page bid documents in 30 minutes, extracting over 200 key elements in 3 minutes. Kuai Biaoshu AI directly promises rapid generation in 10 minutes.
I can't fully verify these vendors' numbers. But the direction is clear: the time and manpower needed to prepare a bid document is shrinking dramatically. The gap between well-prepared and poorly-prepared bids used to be huge; now that gap is narrowing fast.
The most critical barrier is the last one.
Bidding requires clear functional parameter descriptions, then comparing finished or near-finished products. Previously, if you hadn't built similar systems, you truly couldn't produce anything. But now, the clearer the functional description, the faster AI builds it. Throw the parameter requirements from the bidding document to a coding agent, spend a few thousand dollars plus a few days, and you can produce something demonstrable. Screenshots, feature points, even something running that evaluation experts can see.
Among Y Combinator's Winter 2025 batch, 25% of startups had 95% of their code generated by LLMs. YC CEO Garry Tan put it directly: you no longer need a 50-to-100-person engineering team. Cursor DAU exceeds one million, with annualized revenue breaking $2 billion by early 2026. 84% of developers are using or planning to use AI coding tools; 41% of code already involves AI generation.
Is that enough to handle bidding? Yes. The cost of subsequent refinement and delivery is also dropping significantly.
What Used to Be Worth $500k, Now $50k Gets You In
Previously, a software project quoting 500,000 yuan seemed normal—development and testing alone required considerable manpower and time. Now these costs are compressed to one-tenth or even less. Some dare to enter at 50,000 yuan, with AI-written bid documents, products built by coding agents, and only minimal actual costs borne by themselves.
Making matters worse, China's enterprise SaaS industry is already struggling. EY's 2024 report shows Chinese enterprise SaaS listed companies averaged negative net profit margins over the past four and a half years. Gross margins under 60%, sales expense ratio 30%, R&D expense ratio 20%. The industry as a whole is still losing money.
The software outsourcing field is worse. In 2025, the industry's real state was summarized as a "three-piece set": unfinished projects, wage arrears, and layoff waves. SaaS products targeting SMEs have become so homogenized they're competing on manpower. IBISWorld estimates 2025 industry profit margins at just 12.5%.
Against this backdrop, bidding quotes continue to be driven down. The entire system is starting to look somewhat absurd.
The Path LLMs Took Last Year
Actually, LLMs already walked this exact path last year.
In early 2025, DeepSeek R1 was released, matching or exceeding OpenAI o1 on key benchmarks, with API pricing at roughly 4% of OpenAI's. For the same inference task costing $100 on OpenAI o1, DeepSeek cost about $3.60. Training costs were reportedly around $6 million.
The Chinese market subsequently erupted into a price war. ByteDance's Doubao priced its Pro-32k model at 0.0008 yuan per thousand tokens, 99.3% below the industry average, with daily token usage breaking 500 billion by July 2024, up 22 times from May. Alibaba's Tongyi Qianwen slashed Qwen-Long's input price from 0.02 to 0.0005 yuan, a 97% cut. Baidu directly announced ERNIE Speed and ERNIE Lite would be free, with all Wenxin models becoming free from April 2025.
A RAND Corporation report found Chinese AI models cost roughly one-quarter to one-sixth of comparable US systems.
Top-tier models are all open source now; paying for a model itself no longer makes sense. The ending everyone saw later was that the model business became services around models. Helping enterprises use models well, doing post-training, industry adaptation—these have value. The model itself is no longer the transaction subject.
Software Is Taking the Same Path
Now looking back at software.
With production costs falling to sufficiently low levels, demand-side parties will eventually ask: why organize such complex procurement processes to buy this thing?
The logic is identical to models. When supply-side costs approach zero, organizing complex transactions specifically for it loses meaning.
How can the bidding system respond? One direction is to stop comparing functional parameters and instead compare experience and cases. How many clients have you served, how long has the system been running, do you have large-scale usage records? But this violates the original intent of bidding. The whole point of bidding was to compare things as standardized commodities to ensure fair competition. If it ultimately comes down to comparing experience and connections, how is that different from skipping the bid and directly appointing a vendor?
The problem is stuck here. Compare functionality, and AI can help you build it quickly—no differentiation. Compare experience and resources, and that's not what bidding is supposed to do.
Not Far Off
Many people think AI's impact on their industry is still years away. Bidding might not have that long.
In 2024, IT government procurement project numbers grew 21.5%, but total value dropped 4.5%. More projects, less money. Add AI's compression of software costs on top of this, and we hit the tipping point faster than most expect.
The last time I saw a similar situation was the LLM market in early 2025. From DeepSeek's open source to major vendors' comprehensive price wars, it took just a few months. Everyone was forced to transform, no longer selling the models themselves but selling services around models.
The software industry will likely follow the same path. The transaction value of software itself will continue to shrink. What can actually be sold for money are services surrounding software: helping clients clarify requirements, continuous operations and maintenance, data migration, process transformation.
Back to the friend's confusion at the beginning. He's thinking about how to adapt to the new price competition. But perhaps what he should be thinking about isn't how to win on price, but how much longer this game itself can last.