Google Offered $10M for a Dying Airline’s Data. How Can You Value Yours?

谷歌出价 1000 万美元收购一家垂死航空公司的“数据资产”，你该如何评估自己的数据价值？

Spirit Airlines collapsed. It started auctioning its assets. What would you normally expect to see in an airline’s bankruptcy auction? Aircraft, equipment, software, maybe even airport slots too. But what made the headlines was their data. Not any data- their operational data, such as emails and Teams messages. Google offered to pay $10M for it. Google wasn’t the only one bidding. Mercor, an AI-powered hiring platform, has offered $7.5M. This wasn’t their customer base or even invoice data. According to the reports, it comprises 100 million emails, 500 million Teams messages, and 30 million lines of code. That’s their day-to-day operations and decision-making conversations. Imagine a dying company’s conversations are worth $10M. Of course, the deal isn’t final yet. The airline staff has filed a petition against the deal over privacy concerns. But why did Google want that data? More importantly, how can you estimate the value of your data estate? What makes this estimation hard?

精神航空（Spirit Airlines）倒闭了，并开始拍卖其资产。在航空公司的破产拍卖中，你通常会看到什么？飞机、设备、软件，或许还有机场时刻。但真正登上头条的是他们的数据。并非普通数据，而是他们的运营数据，例如电子邮件和 Teams 消息。谷歌出价 1000 万美元收购这些数据。参与竞标的不仅是谷歌，AI 招聘平台 Mercor 也出价 750 万美元。这并非他们的客户群或发票数据。据报道，这些数据包含 1 亿封电子邮件、5 亿条 Teams 消息和 3000 万行代码。这些是他们日常运营和决策过程中的对话。试想一下，一家垂死公司的对话记录竟然价值 1000 万美元。当然，这笔交易尚未最终敲定。航空公司员工已因隐私问题对该交易提起诉讼。但谷歌为何想要这些数据？更重要的是，你该如何评估你所拥有的数据资产价值？又是什么让这种评估变得困难？

My company helps PE firms with commercial due diligence. Having spent 8 years here, I might be able to share some techniques to help you value your operational data. What makes Spirit Airlines’ data so attractive? Traditionally, structured data is perceived as high value. That’s because it’s easy to feed into analytical engines and train machine learning models. Converting unstructured data into a model-friendly format was cumbersome. Most companies didn’t think doing so had meaningful ROI. But that perspective started to change recently. GenAI could extract structure from this massive unstructured data.

我的公司专门协助私募股权（PE）公司进行商业尽职调查。在这里工作了 8 年后，我或许可以分享一些技巧，帮助你评估运营数据的价值。为什么精神航空的数据如此具有吸引力？传统上，结构化数据被认为具有高价值，因为它们易于输入分析引擎并用于训练机器学习模型。而将非结构化数据转换为模型友好的格式则非常繁琐，大多数公司认为这样做没有显著的投资回报率（ROI）。但最近，这种观点开始改变。生成式 AI（GenAI）能够从这些海量的非结构化数据中提取出结构。

What makes Spirit Airlines’ data more interesting is reinforcement learning. Reinforcement learning involves an agent interacting with an environment and learning from the feedback it receives. Over time, with more interaction and feedback, the agent optimizes for a challenging task. It has been very effective in teaching AI how to code. Because thousands of public GitHub repositories already provide a meaningful starting point for the agent. Agents can also run the code in a sandbox and get instant feedback by testing it. This instant feedback loop makes reinforcement learning possible. But coding is a special case. Creating a feedback environment is extremely easy there. But many other AI applications aren’t like that. And nothing is more lucrative than the airline industry.

精神航空的数据之所以更令人感兴趣，是因为强化学习。强化学习涉及一个智能体（Agent）与环境交互，并从接收到的反馈中学习。随着时间的推移，通过更多的交互和反馈，智能体能够针对具有挑战性的任务进行优化。这在教 AI 编程方面非常有效，因为成千上万的公共 GitHub 存储库已经为智能体提供了有意义的起点。智能体还可以在沙盒中运行代码，并通过测试获得即时反馈。这种即时反馈循环使得强化学习成为可能。但编程是一个特例，在编程中创建反馈环境极其容易。然而，许多其他 AI 应用并非如此。而没有什么比航空业更具利润空间了。

Engineers can create a simulated environment for the physical world to model aircraft operations. But decisions made during the airline’s day-to-day operations are invisible to the modeled environment. Simply put, engineers can simulate a storm and learn to navigate it with reinforcement learning. But whether it should fly through the storm or return to the previous airport needs more than survival probability. An airline operator must assess various factors and expert opinions before deciding. A physical simulation can’t help here. But what if engineers have access to decades of conversations that lead to such decisions? That’s what Spirit Airlines’ operational data means to Google. How did the airline survive the pandemic? What did it do during a sudden spike or drop in demand? How did it manage its supply chain issues? And many more questions that aren’t foreseeable or inferable by an outsider.

工程师可以为物理世界创建一个模拟环境来建模飞机运营。但在航空公司日常运营中所做的决策，对于模拟环境来说是不可见的。简单来说，工程师可以模拟一场风暴，并通过强化学习学习如何应对。但究竟是应该飞越风暴还是返回前一个机场，需要的不仅仅是生存概率。航空公司运营者在决策前必须评估各种因素和专家意见。物理模拟在这里无能为力。但如果工程师能够访问导致这些决策的数十年对话记录呢？这就是精神航空的运营数据对谷歌的意义所在。这家航空公司是如何在疫情中幸存的？在需求突然激增或骤降时他们做了什么？他们是如何管理供应链问题的？以及许多其他局外人无法预见或推断的问题。

How to value your operational data assets. Unlike physical assets, data, more specifically operational data, doesn’t have a direct value. But we can still use valuation principles to value data assets. But I want to make a distinction before moving on. The following section isn’t about valuing any data assets. I want to focus more on operational data, which has largely been overlooked because it didn’t have marketable value until recently. We primarily use three approaches to value any asset. The income approach: Assets generate income for the company. Wouldn’t it be wise to value assets based on their expected return over their lifetime? That’s the income approach. In general, we estimate the direct cash flow the asset generates, subtract the asset’s operating costs over an expected period, and discount those figures at the rate of return you could expect from the next-best alternative. This works perfectly for many asset classes, including physical, intangible, and even businesses. It works perfectly for your data too. But for operational data, this approach struggles with two aspects. Operational data doesn’t directly generate cash. The email sent a year ago and software logs are helpful, but they don’t make money on their own. The second issue is that there’s no next-best-alternative rate to discount it.

如何评估你的运营数据资产？与实物资产不同，数据（更具体地说是运营数据）没有直接价值。但我们仍然可以使用估值原则来评估数据资产。在继续之前，我想先做一个区分：以下部分并非关于评估所有数据资产，我更想专注于运营数据，因为直到最近，它才因缺乏市场价值而被长期忽视。我们主要使用三种方法来评估任何资产。收益法（Income approach）：资产为公司创造收益。根据资产在其生命周期内的预期回报来评估资产价值难道不明智吗？这就是收益法。通常，我们估算资产产生的直接现金流，减去预期期限内的运营成本，并以次优替代方案的预期回报率对这些数字进行折现。这对于许多资产类别（包括实物、无形资产甚至企业）都非常有效，对你的数据也同样适用。但对于运营数据，这种方法在两个方面存在困难：运营数据不会直接产生现金。一年前发送的电子邮件和软件日志很有用，但它们本身并不赚钱。第二个问题是，没有“次优替代方案”的利率来对其进行折现。

Yet, there are ways to think around them. Operational data doesn’t generate cash directly. But it saves. Think of emails; they don’t directly generate any cash. But the company wouldn’t have made revenue without them. Yet, emails that once helped generate cash aren’t as important as they once were. Also, not every email is worth the same. Client communication and important decisions carry more value than emails sent to inform staff about the year-end party. Yet these emails would have helped the company resolve issues faster and even prevented legal penalties. The challenge is how we could assign a monetary value to these. The answer isn’t straightforward. But the following steps are helpful:

然而，还是有办法绕过这些问题。运营数据虽然不直接产生现金，但它能节省成本。想想电子邮件，它们不直接产生现金，但没有它们，公司就无法获得收入。然而，曾经有助于产生收入的电子邮件，其重要性已不如从前。此外，并非每封邮件的价值都相同。客户沟通和重要决策的邮件比通知员工参加年终聚会的邮件价值更高。但即便是后者，也可能曾帮助公司更快地解决问题，甚至避免了法律处罚。挑战在于我们如何为这些赋予货币价值。答案并不简单，但以下步骤会有所帮助：

Step 1: Identify different data asset classes. By that I don’t mean the type of data. Even within email, email threads with customers, insurance providers, and staff comms must all be treated as separate classes. Likewise, treat a system’s performance monitoring differently from application logs. Step 2: Find future use cases for each data asset class. Each asset class has different future use cases. Think of performance monitoring logs. They help a team plan compute resources for the application. Application logs, on the other hand, will help explain customer queries and debug issues in the application logic. Step 3: Estimate the revenue it brings or the cost it saves.

第一步：识别不同的数据资产类别。我指的不是数据类型。即使在电子邮件中，与客户、保险提供商的往来邮件以及员工内部沟通也必须被视为不同的类别。同样，系统性能监控日志与应用程序日志也应区别对待。 第二步：为每种数据资产类别寻找未来的用例。每种资产类别都有不同的未来用途。以性能监控日志为例，它们可以帮助团队规划应用程序的计算资源。而应用程序日志则有助于解释客户查询并调试应用程序逻辑中的问题。 第三步：估算它带来的收入或节省的成本。