The Cost of Writing an ML Paper

Disclaimer.
The author is a PhD student in machine learning at a university in the United States.
As a result, the assumptions and numbers below may not generalize to other STEM fields or countries. This post focuses specifically on ML research in the US academic system.

Estimating the cost of writing a research paper is surprisingly tricky. There are many hidden assumptions, institutional details, and simplifications involved. In this post, I focus on machine learning papers submitted to top-tier conferences, using NeurIPS as the primary reference point.

NeurIPS is widely considered one of the highest-impact venues in ML, with extremely high submission volume and a low acceptance rate.

Context: NeurIPS scale

In 2025, NeurIPS received 21,575 submissions, with an acceptance rate of roughly 24.5%
(source).

Out of these, approximately 5,524 papers had at least one US-based affiliation
(source).

At this point, the accounting becomes messy: many papers involve cross-country affiliations (e.g., US + Europe, US + Asia). To keep the analysis tractable, I will restrict attention to papers affiliated exclusively with US institutions.

Initial intuition

I was honestly surprised by how expensive it is to produce an ML paper. Academic research is far from cheap.

To be fair, ML might still be less expensive than some experimental sciences (e.g., biology), but it is certainly not cheap, especially once tuition, labor, and compute are considered.

All estimates below assume:

• an international PhD student in ML
• enrolled at a US university
• producing research for a top-tier ML venue

Cost components

Writing a paper involves several major cost categories:

1. Labor (PhD student + advisor)

• A typical ML PhD student stipend: ~$35,000/year
• Roughly 2,000 working hours/year
• → ≈ $17.5/hour (student labor only)

This already ignores advisor cost, overhead, and unpaid overtime.

2. Compute

Compute costs vary wildly:

• theory vs. empirical work
• scale of experiments
• availability of institutional clusters

For simplicity, compute is acknowledged but not explicitly priced in the core estimate below.

3. Tuition

Tuition is often the largest hidden cost.

• Domestic or in-state PhD tuition: can be ~$7,000 per quarter
• International or out-of-state tuition: often much higher
• In many US universities, international tuition + fees ≈ $50,000/year

Some universities waive parts of this cost after advancement to candidacy, but this is not universal.

4. Travel

Conference travel (registration, flights, accommodation) adds additional cost, but is excluded from the main estimate to keep numbers conservative.

Putting the numbers together

Let’s consider a single PhD student working with a single advisor.

Annual cost per PhD student (international):

• Tuition: ~$50,000
• Stipend: ~$35,000
• Total: ~$85,000 per year

(This does not include advisor salary, benefits, or institutional overhead.)

Time allocation assumptions

• Officially: PhD students are limited to 20 hours/week
• In practice: often 40+ hours/week
• Direct advisor meetings on publication-related work:
  • ~2–3 meetings/week
  • ~1.5 hours/meeting

This excludes:

• off-the-clock writing
• revisions
• reviewer rebuttals
• rejected submissions

Output rate

A typical ML PhD student produces:

• ~0.5 to 1 paper per year

Final estimate: cost per paper

Given the above:

• Low estimate:
  ~0.5 papers/year → ≈ $170,000 per paper

• High estimate:
  ~1 paper/year → ≈ $85,000 per paper

➡️ Estimated cost per ML paper:
$85,000 – $170,000, excluding advisor salary, compute, and travel.

References

• UC Davis discussion on graduate student funding:
  https://biobeef.faculty.ucdavis.edu/2022/07/17/a-word-about-funding-graduate-students