Don’t Be Fooled by “Proprietary Models”: Are You Spending Your Money Wisely, Bosses?
Dear bosses and CIOs, over the past six months, has someone always come to you with a PowerPoint, asking, “Should we develop our own large-scale model?”
The term sounds impressive, as if not having one would leave your company behind in the times. But to be honest, the money spent on these models feels somewhat uncertain. Today, let’s get down to the basics and calculate the costs in plain language.
To summarize the main point in one sentence:
So-called “proprietary models” are actually a mix of three types of “goods” that depreciate over time (self-trained models, fine-tuned models, and privately deployed models) and one type of “asset” that appreciates in value (your company’s unique context and business logic). Most companies mistake these goods for assets, only to find that their investment becomes less valuable over time. The real competitive advantage doesn’t lie in the model itself, but in the unwritten data and rules that make your business function effectively.
Let me break down this complex situation into five key aspects:
---
1. Unveiling the Four Maskes of “Proprietary Models”: What Are You Really Buying?
The so-called “proprietary models” on the market are like a basket that can hold all kinds of things. Let’s take a closer look; there are actually only four components, each with a completely different nature:
- The first component: Self-trained models (most expensive and hardest to develop):
- What this means: It’s like growing your own crops, grinding your own flour, and cooking your own food—training the model from scratch using the most basic code.
- Cost: The investment ranges from $100 million to $1 billion, plus a top-tier team of hundreds of people.
- Current status: Very few companies have actually achieved this; most can’t even reach the starting point.
- The second component: Fine-tuned models (the most “customized” option):
- What this means: You take an existing open-source model (like a general-purpose large model) and train it with your company’s data to make it understand your industry’s terminology.
- Cost: Just a single graphics card and a few days of work.
- Current status: This sounds appealing, as it can be used for customer service and contract review. However, it’s the most misleading option, as it creates the illusion of having a unique model.
- The third component: Privately deployed models (the safest option, but not necessarily the smartest):
- What this means: The model remains the property of the developer (e.g., Huawei or Baidu), but you host it in your data center. The data stays within your premises, so security is ensured.
- Cost: You need to buy the hardware and licenses.
- Current status: Banks and government agencies prefer this option. But be aware: you’re often getting the least intelligent model, just the one that doesn’t leave your premises.
- The fourth component: Context (the most valuable, yet invisible):
- What this means: Your company’s data, business rules, the expertise of experienced staff, and approval processes. These elements are not written in the contract and are crucial for the model’s effectiveness.
- Cost: They require long-term accumulation and organization.
- Current status: This is the true essence of what makes a model “proprietary.”
The key difference: The first three components are goods whose value is fixed at the point of purchase and will depreciate over time (as technology advances rapidly). The fourth component, however, is an asset that becomes more valuable with use; the more data you collect, the stronger your competitive advantage.
---
2. Why Do Suppliers Package “Goods” as “Assets” to Sell?
If only the fourth component (context) is an asset, then why do sellers constantly talk about “proprietary models”?
Because it’s a better sellable product at a higher price:
- Blurred boundaries: Suppliers rarely clarify which parts are purchased outright (assets) and which are rented (goods), or which need to be renewed next year. If the distinctions are clear, deals are harder to close.
- The truth about industry-specific models: Models like “HSBC Finance” or “iFlytek Medical” may seem customized, but the ownership still lies with the manufacturer; they are sold to multiple companies.
- Creating anxiety: The main goal is to make you feel that not investing in a model will put you behind.
To illustrate: A purchased car loses value quickly, but a family-owned house becomes more valuable over time. The current market is like selling you a “family-owned house” as a “new car” at a discounted price.
---
3. Analyzing the Mistakes Made with Bloomberg, Fine-Tuning, and Privately Deployed Models
Let’s look at some real examples to see where these approaches go wrong:
Example 1: Bloomberg’s “Self-Trained” Model
- Background: Bloomberg has the best financial data in the world. In 2023, they spent a lot of money (1.3 million GPU hours, over 500 high-end GPUs for 53 days) to train a 50-billion-parameter model.
- Result: Three years later, Bloomberg executives admitted the model was only for research and never used in any product.
- Lesson: Having excellent data doesn’t guarantee success against constantly evolving competitors. Self-training is a game for giants; ordinary companies should avoid it.
Example 2: The Failure Rate of Fine-Tuned Models
- Background: Fine-tuning seems cheap and straightforward (a single graphics card can complete the task in days).
- Reality: According to VentureBeat, 73% of fine-tuning projects fail or never get deployed.
- Problems:
1. Stuck in the development phase: 45% of companies get stuck here; the model works well in the lab but not in production.
2. Strict requirements: Success requires stable tasks, sufficient data, and timely evaluations—missing any of these makes the effort futile.
3 Static models: The model is most effective on the day of training; changes in business make it useless, and retraining is costly.
Example 3: The Misuse of Privately Deployed Models
- Background: 82% of companies prefer privately deployed models due to data security.
- Reality:
1. Low utilization: The average utilization rate of data centers is only 58.3%, and in some regions, it’s less than 30%. The machines are running, but the business benefits are minimal.
2 Pseudo-sovereignty: Countries like South Korea claim “sovereign AI,” but often rely on third-party technologies (e.g., Alibaba’s vision encoders). For example, NAVER used an Alibaba model but was later replaced by another provider.
- Truth: What you buy is often just the right to use the data center; the actual power and ecosystem still belong to the developer.
Conclusion: These approaches may look good on paper, but they often lead to wasted money and outdated models before they even see real use.
---
4. Why Do People Still Jump into These Mistakes? (Human Nature and Workplace Dynamics)
Despite the clear drawbacks, why do companies still invest? Because it’s not just a mathematical question—it’s also a political one:
- Unequal responsibility: If you make a mistake, it’s the company’s loss, and the boss might feel regret, but you’re unlikely to get fired. If you don’t invest, you’re held accountable for potential delays.
- CIO’s dilemma: Three-quarters of CIOs regret AI purchases, yet they signed the contracts. Nearly 30% are asked to explain AI results they don’t understand. Seventy percent worry about proving the model’s value, which affects their budget and position.
- The outcome: Investing is seen as meeting expectations; not investing means taking the blame. The products bought in a rush often end up being of little use.
---
5. A Way Out: Stop Buying Models and Focus on Building Assets
To break this cycle, you need a different approach. Stop asking “which model is best” and ask “where are my assets?”
What should you buy (in terms of goods)?
- Self-trained models: Don’t even consider them unless you have billions in budget and a large team.
- Fine-tuned models: Only proceed if the task is stable, you have sufficient data, and you can handle the evaluation process. Otherwise, it’s a waste of money.
- Privately deployed models: Use them if your data is sensitive (finance, government, healthcare). Ensure the data stays within your premises first, and then consider the model’s intelligence.
What should you build (in terms of assets)?
The real value lies in your company’s context. Here are two companies that got it right:
- SAP: They don’t sell models; they sell “Company Memory”—they collect customers’ decades of processes, rules, and approval records to guide AI decisions.
- Palantir: Their CEO says, “We’re not involved in the model itself; we sell the application layer.” They focus on the practical uses of the model.
- Harvey: They create legal AI by integrating legal language, cases, and processes into their models.
Action plan:
- For industries with strict data requirements (finance, government, healthcare): Start with privately deployed models to ensure compliance.
- For most companies: Buy the best models, as they will depreciate over time. Save the money and effort on building the necessary context: organize your data, clarify processes, and document the unwritten rules.
- For a few companies with rich data, stable scenarios, and skilled teams: You can try fine-tuning, but treat it as a project, not a strategic investment.
- Tap into expert knowledge: Why does one company get a better price for the same services? The unwritten expertise is the true asset. Extract and document it to feed into your models.
Final advice: Computing power can be rented, and models can be bought, but data and processes are the best investments. These efforts may not make headlines or require grand ceremonies, but they are the foundation for long-term success. Models may become obsolete, but your company’s unique context will not.
Next time someone talks about “proprietary models,” don’t rush to ask about the price. First, ask: “Which specific component are you referring to?”