TensorFlow and open source: what actually changed for companies
Why Google opening its ML framework shifts the conversation from 'we can't afford it' to 'we need data and an engineer'.
When Google released TensorFlow as open source in November 2015, most executives had the predictable reaction: that is for researchers, not for us. A few months on, I see the conversation shifting - but not always in the right direction.
Some companies read open source as a signal that machine learning is now cheap. That is a misreading.
What actually became free
TensorFlow and similar frameworks removed one barrier - expensive proprietary tooling. Before, a company wanting to build models either paid for costly licences or was locked into a single vendor. That barrier is gone.
But everything else stayed exactly where it was. Data still needs to be collected, cleaned, and stored. Engineers still cost money. Infrastructure for training models requires either powerful servers or cloud spend. And most importantly, the problem the model is meant to solve still needs to be clearly defined.
The tool became free. The work stayed expensive.
Why this still matters for non-research companies
Open source created something less obvious but more valuable: an ecosystem. Around TensorFlow a community formed quickly - tutorials, reusable components, domain-specific examples.
This means the entry threshold dropped not in licence cost, but in availability of knowledge. An engineer who had no idea where to start a year ago can now find a working example of a problem similar to theirs.
For a company, this means finding someone with practical experience is more realistic than before - not just a theorist from academia.
Where the trap appears
I often see this sequence: an executive hears about open tools, hires a data scientist or ML engineer, and expects results in a few months. The engineer sits down and finds no data, three disconnected systems, half the labels missing, and a business problem stated as "make it smarter somehow".
The tool is there. Everything else is not.
Open source does not change the order of work. It makes the technical layer cheaper, but it does not remove the need to define the problem first, then gather the data, then build the pipeline, and only then think about the model.
How to read this signal correctly
TensorFlow and open ML frameworks are a healthy signal that the technology is maturing. It is leaving the "Google-only" regime and becoming accessible to organisations without research budgets.
But mature tooling does not mean a ready task. A few questions worth asking before hiring the first ML engineer:
- Do we have a specific problem we want to solve - not a general idea, but a concrete input and a desired output?
- Is the data for that problem already in structured form, or does it need to be built first?
- Who in the company will use the result, and how does it fit into an existing process?
- Do we have engineering infrastructure - data pipelines, storage, governance - or will the ML engineer have to build all of that from scratch?
If the first three questions have answers, open tools genuinely open up possibilities. If not, it makes sense to work on the data layer before thinking about models.