What the main fellowships ask for
| Fellowship | Research document | Limit | Judged on |
|---|---|---|---|
| NSF GRFP | Graduate Research Plan Statement, plus a separate Personal, Relevant Background and Future Goals Statement | 2 pages for the research plan, 3 for the personal statement, including references and figures [source] | Intellectual Merit and Broader Impacts, under separate headings [source] |
| Google PhD Fellowship (next cycle) | Research or dissertation proposal | Up to 3 pages, excluding references [source] | Innovation, relevance to Google's research areas, robustness and potential impact [source] |
| NVIDIA Graduate Fellowship (next cycle) | Research proposal | Up to 2 pages plus a bibliography, which does not count [source] | Student quality, research quality and relevance to NVIDIA [source] |
| Hertz Fellowship | Written application, followed by two rounds of interviews [source] | See the application | Hertz describes a rigorous selection process [source] |
The limits are strict, and NSF counts references and figures inside its two pages [source]. Write to the limit of the fellowship you care about most, then adapt.
NSF GRFP: two statements, two criteria
The NSF Graduate Research Fellowship gives three years of support over a five-year fellowship period. For each of those years NSF pays a $37,000 stipend, plus a $16,000 cost-of-education allowance to the university [source].
The application has two statements. The Personal, Relevant Background and Future Goals Statement can be up to three pages, and the Graduate Research Plan Statement up to two. Both limits include all references, citations, charts, figures, images and lists of publications [source].
The rule that catches people out is the headings. NSF requires you to address "Intellectual Merit" and "Broader Impacts" under separate headings in both statements, with no other terms and without combining them [source]. NSF defines Intellectual Merit as the potential to advance knowledge, and Broader Impacts as the potential to benefit society and contribute to specific, desired societal outcomes [source]. Reviewers are told that both criteria are needed and that neither is enough by itself [source].
Reviewers look at the whole application, including your educational and research record, leadership, outreach, service and future plans [source]. So the research plan does not have to carry everything. It needs to show you can design a credible project, and the personal statement shows the record behind it.
Google and NVIDIA: relevance matters
The industry fellowships want strong research that fits their own interests.
For the Google PhD Fellowship, most regions require your university to nominate you, and the application is submitted by an official university representative [source]. Every nominee sends a CV and a research or dissertation proposal of at most three pages, excluding references [source]. Google says proposals are judged on innovative concepts relevant to its research areas, on robustness and on potential impact. It also asks you to describe where the work is going as well as what you are doing now [source]. It publishes research areas with specific topics and says topics outside them will still be considered [source]. On funding, our listing quotes Google: Funding varies by region, up to US $85K a year towards tuition, fees and living costs in some regions.
The NVIDIA Graduate Fellowship offers Up to $60,000 per award for stipend, tuition and mandatory fees, with a required in-person summer internship. Its research proposal can be up to two pages, plus a bibliography that does not count towards the limit. Your name goes at the top, and text and images must fit within the two pages [source]. You also need two or three recommendation letters, including one from your thesis advisor [source]. For the 2027–28 fellowship, applications and letters were due on 30 October 2026, so the next chance is the following cycle [source].
For both, name the link to the company's work explicitly. Do not leave the reviewer to work it out.
A structure that fits in two pages
- The problem, in two or three sentences. What does not work today, and why does it matter? Be concrete: "language-model agents fail at multi-step tool use when..." is better than "AI is transforming the world".
- Your question or hypothesis. One sentence a reviewer could repeat to a colleague.
- Why you. A short line linking the question to something you have already done, such as a paper, a project, a replication or a result.
- Approach, in aims. Two or three aims, each with a method and a way to tell whether it worked. Name the datasets, models or benchmarks.
- Risks and fallbacks. One line per aim on what you will do if the main approach fails. Reviewers notice when this is missing.
- Impact. For NSF, this is the Broader Impacts heading. For industry fellowships, say what changes for users or for the field if you succeed.
- References. Keep them short. NSF counts them in the page limit and asks for abbreviated titles with no URLs or DOIs [source].
A two-page plan usually splits into about a third of a page on the problem and your question, a page on the aims, and the rest on impact and references. If one part grows, cut background first and aims last.
Write for a reviewer who is a scientist but not in your sub-field. NSF panels and company committees both read across many topics, so define your terms, keep acronyms to a minimum and put the key idea in the first paragraph. A figure can help, but at NSF it counts towards the two pages, so use one only if it saves words.
The research plan is a plan, not a promise. Reviewers know that PhD projects change. They are judging whether you can turn an interesting question into a project that could work.
Common mistakes
- Too broad. "Making AI safe" is a field, not a project. Narrow it until you could start next month.
- All background, no plan. If the literature review takes up a page, there is no room for the aims.
- No evaluation. In machine learning especially, say what you will measure and what you will compare against.
- Ignoring the format rules. Missing NSF's required headings or going over a page limit can count against you or get the application returned. Check every rule on the official page.
- Different stories. Your statement, CV and letters should describe the same person. Share your draft with your recommenders.
When to start
The big US fellowships close in October and November, for funding the following academic year. The Hertz Fellowship's application deadline for this cycle was 30 October 2026 [source], and NVIDIA's was the same day [source]. The Google PhD Fellowship's 2026 window ran from 5 March to 30 April 2026 [source]. Start your draft at least two months before a deadline. Your recommenders need time too.
The fellowships we track that are waiting for their next round, such as Google PhD Fellowship (next cycle) and NVIDIA Graduate Fellowship (next cycle), have a "Tell me when it reopens" button on their listing. For UK PhDs, see funded AI PhDs at UKRI CDTs. For the wider picture, see AI funding for students. For after the PhD, see postdoctoral AI fellowships compared.
