AI Companies Paying High Prices for Data: Google Spends Tens of Millions of Dollars on Bankrupt Airline's Records, Data Cost Accounting Still Opaque

Data left behind by a bankrupt airline drew a $10 million offer from Google.

The deal involves internal records accumulated over the years by U.S. budget carrier Spirit Airlines, including roughly 100 million emails, 500 million Microsoft Teams messages, and a large number of spreadsheets, calendars and other business materials. Google hopes to use this data to improve its AI models.

Why can the work records accumulated over many years by an airline that has ceased operations still sell for tens of millions of dollars?

At the same time, another AI data company, micro1, announced plans to invest $1 billion over the next 12 months to purchase or obtain licenses for operational data from real companies, in order to build reinforcement learning environments that let AI learn to complete complex business tasks.

Over the past few years, AI companies have spent enormous sums buying GPUs and building data centers, yet the data needed to train models has rarely had a clear price. Now, more and more price quotes are surfacing. A single highly difficult math evaluation question may be worth $5,000, a company's internal data can receive a $10 million offer, and a leading model company may purchase hundreds of millions of dollars' worth of services from a single data supplier in a year.

In the past, an important channel for large models to obtain data was the internet.

News reports, books, encyclopedias, forums, open-source code and specialized websites formed important corpora for early model training. An industry insider told Tencent Technology that in the early days of the industry, in order to obtain high-quality data, some people even bought old books in bulk and "tore them apart to scan them."

As the industry has developed, copyright issues have received increasing attention. Leading model companies obtain large amounts of content through the public internet, and also purchase licenses from publishers and professional databases, then clean, deduplicate and filter the data.

This market has always existed. OpenAI reached a multi-year content partnership with News Corp. According to a Wall Street Journal report at the time, the agreement could be worth more than $250 million over five years. OpenAI can use current and historical content from media outlets such as The Wall Street Journal and the New York Post, both to improve products and to display licensed content in ChatGPT.

However, as model capabilities improve, the targets of data procurement have begun to change.

At present, the external data sources of model companies can be roughly divided into three categories.

Companies such as Scale AI, Surge AI, Mercor and micro1 all belong to this market. They help model companies organize annotators and professionals, process raw data into materials usable for training, and produce the high-quality answers, evaluation questions and complex tasks that models need.

This business has already grown to a considerable scale.

In July 2026, financial documents obtained by The Information showed that Mercor achieved total revenue of $614 million in the first half of the year, of which about 91% came from foundation model companies, including OpenAI, Anthropic, Google DeepMind and others.

The revenue of these data companies can already reach the scale of hundreds of millions of dollars. They are also gradually extending, in line with changes in model training needs, from providing large-scale annotation labor to organizing professionals such as mathematicians, programmers and doctors to produce high-quality data.

The second category is people with specialized knowledge.

In February 2026, foreign media reported that some mathematicians were invited to provide highly difficult evaluation questions for AI, with a reward of about $5,000 per question.

These questions are not ordinary olympiad problems. They need to be enough to test the capabilities of frontier models, and they also need reliable answers and verification methods. For AI that is increasingly good at mathematical reasoning, posing a problem it cannot yet easily solve itself requires a high level of professional ability.

In addition, model companies also need engineers to design complex programming tasks, doctors to evaluate diagnostic reasoning, and legal experts to judge whether professional answers are reliable.

Compared with ordinary text annotation, this type of data is more expensive to produce and imposes stricter requirements on personnel.

The third category, and the one that has recently shown clear new momentum, is industry companies that hold real business data.

Spirit Airlines is a typical case. Before bankruptcy, this airline was not a company that provided data for AI. Its emails, meeting records, spreadsheets and business materials were simply records naturally formed through years of business operations.

But for model companies hoping to train AI to complete real work, these materials may offer another kind of value. For example, how does a company handle a supplier's delayed delivery? How does a customer complaint flow between different departments? How do employees adjust work plans based on new circumstances? These processes may be scattered across emails, contracts, meeting records and business systems.

The internet has a large number of articles introducing corporate management, supply chains and customer service, but it is very difficult to fully reconstruct the problems a company has actually dealt with.

micro1's latest procurement plan is also aimed at this type of data.

On October 9, micro1 announced the launch of its Company Data Partnerships program, purchasing or obtaining licenses for operational data from companies and converting it into reinforcement learning environments. Models can learn within them how to handle incomplete information, conflicting objectives, and various unexpected situations in real business.

This demand is clearly different from past large-scale web data procurement. Corporate work records therefore have new potential buyers.

But this type of data also faces obstacles. The fact that a company holds internal records does not mean it can freely sell information involving employees, customers, and partners; raw records are also not the same as training data that can be used directly. Buyers still need to complete authorization review, privacy processing, task extraction, and result validation.

How much budget model companies actually invest in data remains a highly opaque question to this day.

AI companies often disclose the compute and infrastructure investment used for training, but rarely publish a complete data ledger. Even for leading companies like OpenAI and Anthropic, the currently publicly available financial materials do not itemize expenditures on pre-training corpora, expert post-training, reinforcement learning tasks, and evaluation data.

This is first related to the composition of data costs.

A model company's data investment includes not only purchasing raw corpora and copyright licenses, but also data cleaning, deduplication, quality review, and privacy processing. Entering the post-training stage, it also needs to purchase demonstration answers, preference judgments, and complex tasks from experts; when training Agents, it may also need to develop executable environments and automated validation tools.

In addition, model companies are also producing data themselves. Internal researchers, annotation teams, synthetic data generation systems, and evaluation infrastructure may all generate related costs.

Looking at large model companies that have already gone public, in the first half of 2026, Zhipu's R&D expenses reached 2.131 billion yuan, up 33.6% year on year; MiniMax's R&D expenses were about $297 million, up 138.8% year on year. MiniMax explicitly stated that the growth mainly came from increased training-related cloud service costs brought by model iteration. However, it is very difficult to find a separately corresponding figure for data procurement in financial reports.

In the prospectus MiniMax released at the end of 2025, it disclosed a plan to invest about HK$64.9 million each year from 2026 to 2030 for data procurement, annotation, and compliance for text, vision, audio, and multimodal models, totaling about HK$325 million over five years.

Zhipu's prospectus shows that in the first half of 2025, R&D technical services and consulting expenses were about 52.21 million yuan, involving data cleaning, model evaluation, and other R&D outsourcing work. Because multiple projects were disclosed together, it is impossible to confirm how much of this was directly used for data-related services, nor to determine whether it included raw data procurement.

Model companies' compute and overall R&D spending have become increasingly transparent, but exactly how much they spend on purchasing raw corpora, expert-produced training data, reinforcement learning tasks, and evaluation materials still lacks a unified financial reporting standard. External procurement, internal data production, and related computing costs are also scattered across different line items, making it difficult to directly compare data investment across different model companies.

However, single external procurement deals by overseas model companies have already reached very high amounts. Google once planned to pay Scale AI about $200 million a year, and the hundreds of millions of dollars in revenue Mercor obtained in the first half of 2026 also shows that leading model companies are continuously purchasing external data production services.

According to industry insiders, the data investment cost for GPT-6 Astra is as high as several billion dollars, but this is only an industry rumor.

What is more worth studying is how model companies actually decide their data budgets.

In the traditional understanding, if a model needs more data, it should continue to expand procurement. But this approach may cause waste: a large amount of the same type of data has already been accumulated, and continuing to add more may not bring obvious improvement; what truly limits model capability may be another type of scarce data.

"Rethinking Data Mixture for Large Language Models," published in EACL Findings in March 2026, systematically reviewed methods for optimizing data mixing ratios in large model training.

It discusses "which data limited training resources should be spent on."

For example, the same generation of models needs to simultaneously improve mathematics, programming, language understanding, and tool use capabilities, but adding data in different domains has different effects on final performance. The research summarized how to adjust the proportions of various types of data under a fixed compute budget to improve model performance.

Put into actual commercial procurement, model companies also need to answer a third question: should this data be purchased, or produced in-house?

Suppose a model company wants to improve the success rate of complex programming tasks. If ordinary code can no longer significantly improve model performance, continuing to expand procurement may be meaningless; if the quality of synthetic tasks is high enough, the need for external procurement may also decline.

Conversely, when certain data can only be provided by professionals or real enterprises, and can indeed improve model performance, it may be worth purchasing even if the unit price is very high.

Therefore, the data budget is not just a fixed proportion of R&D costs. It is related to what capabilities the next-generation model is being prepared to improve, what data the existing model lacks, and how much additional benefit different data can bring.

On Mercor's recruitment page, the average contract rate for ordinary professionals is about $107 per hour, but some oncology experts quote $350 to $450, and machine learning experts $150 to $200. By comparison, basic tasks such as simple image classification and text annotation may cost as little as a few cents per item.

This price difference can show that, at present, model companies are willing to pay higher fees for professional judgment in specific domains.

As mentioned at the beginning of the article, the pay for each math problem is about $5,000. Pricing in the medical field is also telling. Mercor is currently publicly recruiting radiologists at contract rates of $200 to $400 per hour, with tasks including interpreting medical images, writing reference diagnostic reports, and evaluating model performance; some oncology specialist positions pay up to $450 per hour. These are expert fees, not the final unit price of data paid by model companies, but they already reflect the production cost of complex professional data.

Expert time is expensive and limited, so data companies need to extract more effective training signal from each hour of labor. Experts may be responsible for creating tasks, defining qualified delivery standards, or handling cases where the model and the automatic grader disagree. Data companies also have the task of measuring whether a dataset truly improves model performance.

Another growing demand is for reinforcement learning tasks and executable environments. Model companies need to let Agents complete multi-step operations in environments and determine whether execution succeeds through real results or verification programs.

This type of data costs not only from writing tasks, but also from environment development, rule setting, automatic verification, and repeated testing. The $1 billion enterprise data procurement plan announced by micro1 is intended to extract this kind of training material from real operational records.

A noteworthy price contrast has also emerged in the robotics field. On October 16, foreign media reported that data collection projects such as Human Archive are recruiting people in more than 30 countries to wear devices and record daily labor, paying $1 to $30 per hour.

The report also noted that demand for relevant real-world data has grown about tenfold within a year. Meanwhile, on October 15, foreign media reported that the domestic robotics company DexForce is using virtual simulation data to reduce training costs. The simultaneous development of the two routes means that demand for real operational data is expanding, but simulation technology may also gradually push down the collection cost of some data.

A more profound change is that models themselves are also becoming data producers. For math and programming tasks whose answers can be automatically verified, models can generate questions in batches, attempt to solve them, and then use code executors or other verification tools to filter the results. This will reduce the need for some manually created tasks while pushing procurement budgets toward materials that are harder to generate, lack reliable verifiers, or require human professional experience.

Of course, there is also a question: can the next generation of models generate at lower cost the data that today requires expensive experts to produce? If so, current high prices may not be sustainable in the long term.

What ultimately determines the price of data is whether it can provide a verifiable increment in capability beyond existing models and data sources, and how much it costs to produce data of the same quality.

Looking back today, "data" has already begun to develop into a mature industrial chain. As long as human activity exists, this industrial chain will be like an umbilical cord linking humanity and artificial intelligence, enduring and growing vigorously.

This article comes from the WeChat public account "Tencent Technology", author: Xiaojing, editor: Xu Qingyang, published by 36Kr with authorization.