Who Pays for Attention?

Reading Time: 3 minutes

Attention Is Not All You Need: What Managers Should Take from the Transformer

Luv Bansal’s Medium explainer walks readers through the 2017 paper that introduced the transformer, the design behind most of today’s large language models. It is clear and well illustrated, and I would recommend it. But it answers an engineer’s question: how does the thing work? A manager has to answer different ones: what does it cost, where does it struggle, and what has to change in the organisation for people to use it well? On those, the article says almost nothing.


What the article gets right

It explains self-attention well. Every word is compared with every other word, so the model can use context from anywhere in a sentence, and the work can run in parallel instead of word by word as in older recurrent networks. That parallelism is the real business story. The original paper reports that its base model trained in about 12 hours on eight GPUs, at a fraction of the training cost of earlier systems.

I have one correction. The article says pre-trained Word2Vec embeddings are used. The original paper says it uses learned embeddings. It is a small error, but a useful reminder that explainers are secondary sources.


What it leaves out: the bill

The same all-to-all comparison that makes attention powerful also makes it expensive. The paper’s own comparison table puts the cost per layer at the square of the sequence length, and it notes that self-attention is cheaper than recurrent layers only when sequences are shorter than the model’s representation size. For long documents that cost dominates, which is why researchers keep proposing cheaper variants. One 2025 paper argues that alternatives such as state space models have stayed niche because of scaling and training problems, though its author is selling his own fix, so I read that claim with care.

The bill also shows up at system level. The IEA reports that data centre electricity demand grew 17% in 2025, against 3% for global electricity demand. It also notes that energy per AI query has fallen sharply, but heavier uses such as video generation, reasoning and agents keep growing.


My view: the bottleneck is organisational

So should managers learn the maths? I partly disagree with that framing. Knowing what a query or a key is will not tell you whether your team will use the tool. In an Aalto University study of a consulting firm with 600 employees, an AI tool was dropped after almost two years. Staff did not trust how it used their data, so they withheld information or fed it self-serving information, and its output got worse. A Gallup panel study of more than 30,000 US workers (Makridis, 2026) found that employees who say their organisation has a clear AI strategy were about 27 percentage points more likely to use AI frequently, and that frequent use without a clear strategy was linked to higher burnout. That is correlation, not proof, and the paper says so. But both studies point the same way: the constraint is trust and clarity, not architecture.


What I would do as a manager

  1. Ask for the cost drivers, not the architecture: context length, usage volume and the cloud or energy bill.
  2. Set the purpose and the rules first, and communicate them so people know what is allowed.
  3. Build trust before collecting data: tell employees what the tool sees and why.

Conclusion

Attention may be all a model needs, but it is not all a manager needs. A simple test: could I explain to my team what this tool costs, what it knows about them, and when they should overrule it? If not, the next step is a conversation, not a deeper technical explainer.

References

Bansal, L., “Transformer: Attention Is All You Need Easily Explained With Illustrations,” Medium: https://luv-bansal.medium.com/transformer-attention-is-all-you-need-easily-explained-with-illustrations-d38fdb06d7db Vaswani et al., “Attention Is All You Need” (2017): https://arxiv.org/abs/1706.03762 Filipek, “Sparse Query Attention (SQA),” arXiv (2025): https://arxiv.org/abs/2510.01817 IEA, “Key Questions on Energy and AI” (16 Apr 2026): https://www.iea.org/reports/key-questions-on-energy-and-ai Aalto University, “Why are 80 percent of companies failing to benefit from AI? It’s about the people not the tech” (23 Jan 2025), on the study by Vuori et al. in the Journal of Management Studies: https://www.aalto.fi/en/news/why-are-80-percent-of-companies-failing-to-benefit-from-ai-its-about-the-people-not-the-tech-says Makridis, “The Organizational Transmission of AI: The Role of Managers on AI Adoption and Impact,” CESifo Working Paper 12373 (Apr 2026): https://www.ifo.de/DocDL/cesifo1_wp12373.pdf

Research with Gemini Argon 4

Prompt used: Based on the article Luv Bansal, "Transformer: Attention Is All You Need Easily Explained With Illustrations" (Medium): https://luv-bansal.medium.com/transformer-attention-is-all-you-need-easily-explained-with-illustrations-d38fdb06d7db and my working thesis: popular transformer explainers teach how the architecture works (an engineer’s question) but leave out what a manager needs: cost, limits, and the organisational conditions that decide whether AI tools succeed conduct a complex research about similar topics and find 4 – 6 similar explainers or tutorials on transformers and attention.

Leave a Reply