AI Just Got Cheaper. Your Data Did Not.

A Market Shock

On January 20, the Chinese AI lab DeepSeek released R1, a reasoning model whose published benchmark results were comparable to OpenAI's o1, with openly available weights under a permissive license. A few weeks earlier, the company had published a technical report on its V3 model stating that the final training run cost roughly $5.6 million in computing time, a figure that excluded prior research and experiments but was still far below what many had assumed such models required.

The reaction was swift. On January 27, Nvidia lost nearly $600 billion in market value, a record one-day loss for any US company, according to Reuters, as investors questioned how much expensive hardware future AI would really need. The same day, Reuters and Bloomberg reported that the DeepSeek app had reached the top of the free app chart in Apple's US App Store.

The debate about what DeepSeek really spent, and what it means for chipmakers, will continue.

The debate about what DeepSeek really spent, and what it means for chipmakers, will continue. For a manufacturer deciding how to use AI, the more useful question is what happens when capable models become cheap and widely available.

Models Are Becoming a Commodity

For the past two years, access to the most capable AI models has been treated as the scarce resource. Companies worried about which provider to choose, and providers competed on whose model was smartest.

DeepSeek's release, together with continued price reductions across the industry, suggests that capable models will be abundant and inexpensive. Openly available models can be run on infrastructure a company controls, which also addresses some concerns about sending sensitive data to outside services. When the model itself is cheap and interchangeable, it stops being a source of competitive advantage.

The Scarce Thing Is Your Context

What remains scarce is the information that makes AI useful for a specific business. A general model knows a great deal about manufacturing in general. It knows nothing about your parts, your customers, your machines, your suppliers, your quality history or the reasons behind your engineering decisions. That knowledge lives in your ERP, your quality and maintenance systems, your drawings, your work instructions and the heads of your experienced people.

AI applied to that knowledge can quote faster, find the root cause of a quality problem, answer a planner's question about why a part is late, or help a new technician troubleshoot a machine. AI without it can only give generic answers.

Most Manufacturers Are Not Ready to Use It

The uncomfortable truth is that many manufacturers' data is not in a condition that AI can use well. Item descriptions are inconsistent. Bills of material do not match production. Maintenance history is incomplete or written in free text nobody has standardized. Quality records sit in spreadsheets. Drawings are scanned images in folders. Critical knowledge has never been written down at all.

None of this is new, and the work to fix it is not glamorous. But it is now the main thing standing between a manufacturer and useful AI, because the models themselves are no longer the constraint.

What to Do Now

Start by identifying two or three business problems where better use of your own information would clearly pay off, such as quoting turnaround, first-time quality or maintenance response time. For each one, look at the data it depends on and assess its condition honestly. Clean and organize that data first, with owners and standards so it stays clean.

Make the data accessible in a controlled way. That means knowing where it lives, who is allowed to see it, and how an application could retrieve it securely. Then pilot an AI application on that narrow problem, using whichever model is suitable and affordable at the time, knowing that you can switch models later as the market continues to change.

The Bottom Line

DeepSeek's release showed that capable AI models can be built and run for far less than the industry assumed, and that they will be widely available. When the model is cheap and interchangeable, the advantage goes to the company with the best organized, most accessible knowledge of its own operations. For manufacturers, the most valuable AI investment this year may be the unglamorous work of getting their data ready.

At Cherry Street, this is the work we do. We help manufacturers choose the problems where AI will pay off, prepare the ERP and operational data those problems depend on, and build applications that can take advantage of cheaper models as they arrive.