A couple of days ago, I had dinner with a friend who works in government-enterprise software. He said this year he's bid on seven or eight tenders, won one, and the quoted price was squeezed to one-third of what it used to be. It wasn't the client deliberately driving down the price—there really were people bidding at that level, and their deliverables looked decent enough.
He asked me what was going on. I told him to look into how fast AI can now write bids and build software.
I thought about this for a long time. The bidding system in the software industry may not be able to sustain itself much longer.
Why Is the Software Bidding System Failing?
China's government procurement market is enormous. In 2024, total government procurement nationwide exceeded 3.37 trillion yuan, with open tenders accounting for 76.63%. IT-related government procurement was approximately 143.3 billion yuan. Although this dropped 4.5% year over year, the number of projects actually grew by 21.5%, reaching nearly 79,000. In other words, there are more projects, but each one is getting smaller.
The software bidding system used to run smoothly because of several barriers to entry.
First, qualifications. These were meant to ensure bidding companies were legitimate, with the right capabilities and experience. But this barrier never blocked too many people—companies that wanted in always found a way. Make it too strict, and you defeat the purpose of competition.
Then there was information asymmetry. Many companies only learned about a tender after it was published, leaving them with insufficient preparation time, or they never heard about it at all. The window between publication and deadline was limited, which objectively kept some competitors out. Even if they knew, producing a decent bid document required significant time and manpower. Companies that weren't fully prepared lost out during technical evaluation.
Finally, commoditized comparison. This is the core design logic of bidding: treat software as a standardized commodity, list the functional parameters clearly, and score item by item. If you hadn't built a similar system before, you couldn't produce many of the required functions, you couldn't write clear technical descriptions, and you had nothing to show during the demo.
Combined, these barriers kept competition within a relatively controllable range. For a long time, the system worked well enough.
The Barriers Are Loosening All at Once
AI is loosening all of these barriers simultaneously.
Qualifications need no explanation—they were never a real moat to begin with.
Information asymmetry is changing the fastest. Many companies are already using AI tools to scan tender information across the entire web. "Zhiliaobid" (知了标讯) from Bailing Intelligence covers more than 100,000 tender websites nationwide, with over 300 million tender records, updating 20 million entries daily, and its AI predicts winning bids with over 70% accuracy. Qianlima Tender Network integrated DeepSeek and other large models in 2025, updating 300,000 tender announcements daily, with AI monitoring the whole process from project approval to tender publication and automatically filtering high-potential projects.
In the past, you might never have known a tender existed. Now AI is watching around the clock. Don't have time to prepare the bid? AI can write that for you too.
In bid document generation, the market grew 230% year over year in 2025. iFlytek's "Spark Bidding" claims to compress bid preparation from 30 days to 3 days, improve document compliance by 90%, and increase win rates by 40%. TaiBiaoTou claims it can generate a thousand-page bid in 30 minutes and extract over 200 key elements in 3 minutes. KuaiBiaoShu AI outright 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 are shrinking dramatically. The gap between well-prepared and poorly prepared bidders used to be huge. That gap is now closing fast.
The most fatal change is to the last barrier.
Bidding requires clear functional specifications, followed by comparison of finished or near-finished products. In the past, if you hadn't built a similar system, you genuinely had nothing to show. But now, the clearer the functional description, the faster AI can build it. Throw the parameter requirements from the tender document into a coding agent, spend a few thousand yuan and a few days, and you can produce something demonstrable. Screenshots, feature points, even something that runs for the evaluation experts to see.
In Y Combinator's Winter 2025 batch, 25% of startups had codebases that were 95% AI-generated. YC CEO Garry Tan put it bluntly: you no longer need a 50- to 100-person engineering team. This is the change brought by vibe coding—describing requirements in natural language and letting AI turn them into runnable software. Cursor has over a million daily active users, and its annualized revenue broke $2 billion in early 2026. Eighty-four percent of developers are using or planning to use AI coding tools, and 41% of code already involves AI generation.
Is that enough to handle a tender? It is. The costs of subsequent refinement and delivery are also dropping sharply.
How Low Has AI Driven Software Costs?
It used to be normal for a software project to bid at 500,000 yuan, given the manpower and time needed just for development and testing. Now those costs are compressed to one-tenth or even less. Some people dare to enter at 50,000 yuan—the bid was written by AI, the product built by a coding agent, and their actual cost burden is minimal.
To make matters worse, China's enterprise SaaS industry was already struggling. An EY 2024 report showed that the average net profit margin of listed Chinese enterprise SaaS companies had been negative for the past four and a half years. Gross margins were under 60%, sales expenses 30%, and R&D expenses 20%. The industry as a whole is still losing money.
The software outsourcing sector is in even worse shape. The industry's real condition in 2025 has been summarized as a "three-piece set": unfinished projects, unpaid wages, and layoffs. SaaS products for SMEs have become so homogeneous that they compete on manpower alone. IBISWorld estimates the industry's profit margin in 2025 is only 12.5%.
Against this backdrop, bidding prices are being driven down even further. The whole system is starting to look a bit absurd.
The Path LLMs Took Last Year
In fact, large language models already went through the exact same playbook 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. The same inference task cost 3.60 on DeepSeek. Training costs were reportedly around $6 million.
A price war then broke out in the Chinese market. ByteDance's Doubao priced its Pro-32k model at 0.0008 yuan per thousand tokens, 99.3% below the industry average; by July 2024, daily token usage exceeded 500 billion, a 22-fold increase from May. Alibaba's Tongyi Qianwen slashed Qwen-Long input prices from 0.02 to 0.0005 yuan, a 97% cut. Baidu directly announced ERNIE Speed and ERNIE Lite were free, and starting April 2025, all Wenxin models became free.
A RAND Corporation report found that Chinese AI models cost roughly one-fourth to one-sixth of comparable U.S. systems.
Top-tier models were open-sourced, making it nonsensical to pay for a model itself. The eventual outcome everyone saw was that the model business turned into services around models. Helping enterprises deploy models effectively, doing post-training, doing industry adaptation—these things have value. The model itself is no longer the object of transaction.
Software Is Heading Down the Same Path
Now look back at software.
When production costs fall low enough, the buyer will eventually ask: why go through such a complex procurement process to purchase this thing?
The logic is exactly the same as with models. When supply-side costs approach zero, organizing complex transactions specifically for it loses all meaning.
How can the bidding system respond? One direction is to stop comparing functional parameters and instead compare experience and case studies. How many clients have you served? How long has the system been running? Do you have records of large-scale usage? But this goes against the original purpose of bidding. The whole point of bidding is to compare things as standardized commodities so that competition is fair. If it ultimately comes down to experience and connections, what's the difference from skipping the tender and directly appointing someone?
This is where the problem gets stuck. Compare functionality, and AI lets you whip it up quickly—no differentiation. Compare experience and resources, and you're doing something bidding was never designed to do.
Not Far Off
Many people think AI's impact on their industry is still a few years away. Bidding may not be able to wait that long.
In 2024, the number of IT government procurement projects grew 21.5%, but total value fell 4.5%. More projects, less money. Layer AI's compression of software costs on top of that, and the speed at which we hit the tipping point will be faster than most people expect.
The last time a similar situation was seen was the LLM market in early 2025. From DeepSeek's open-source release to the all-out price war among major vendors, it took only a few months. Everyone was forced to transform, no longer selling the model itself but selling services around the model.
The software industry will most likely follow the same path. The transactional value of software itself will keep shrinking. What can actually be sold for money are the services around software: helping clients clarify requirements, ongoing operations and maintenance, data migration, process transformation. The open-source community is undergoing the same shift—when production and distribution costs both approach zero, the original business model can no longer sustain itself.
Back to my friend's dilemma at the beginning. He's wondering 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 be played.