Your Advisor Left for a Frontier Lab. Who Reads Your Thesis Now?

The Problem

In 2026 alone, 22 professors left Stanford, Berkeley and Harvard for OpenAI, Anthropic, Meta and Google DeepMind. The split was roughly even: six to OpenAI, five to Anthropic, six to Meta, five to DeepMind, with five more announcing fall departures. At Harvard SEAS, 12 of 43 CS-affiliated tenure-track faculty currently hold or recently held industry positions.

The coverage stops at the departure. What it skips is who gets left standing. A departing professor typically has three to five graduate students three to six years into a thesis. The research question came from that professor’s interests. The experimental setup and data access live in that lab. The department assigns an interim advisor who is usually in a different subfield and cannot credibly evaluate the work.

In practice the gap gets patched socially. The departed professor keeps taking calls out of goodwill, or a former postdoc reads drafts unofficially. No contract, no compensation, no record. Students with strong networks survive, and when the relationship lapses the student either restarts the thesis or graduates one to two years late. Departments do not even track the loss.

Why Now

The migration stopped being individual and became structural. Training frontier models now requires GPU clusters costing hundreds of millions, so the compute gap between academia and industry is a canyon, and senior researcher compensation runs at multiples of a tenured salary. Neither gap closes soon.

Industry now has a reason to pay for the other side. The bottleneck in PhD-level hiring is not applicant volume, it is evaluation. A publication list does not show how someone actually runs research. Co-advising gives a lab months of direct exposure to a candidate’s reasoning, which is worth real money in recruiting terms. Lab researchers are already doing this informally. They are simply doing it for free.

Remote research collaboration is also settled practice. Code lives in GitHub, experiments run on shared clusters, meetings happen on video. The argument that supervision requires physical co-location carries much less weight than it did five years ago.

How to Build It

Start with the contract, not the matching. The real barrier for an industry researcher is not time, it is their employer’s moonlighting and IP policy. A standard co-advising agreement covering weekly hour caps, authorship rules, IP assignment and confidentiality scope removes most of the friction on its own. That template is the early product’s core asset.

Keep the MVP narrow. Pick one subfield, say ML systems, and recruit 30 researchers who moved to industry in the past two years plus 50 students with a supervision gap. Match by hand, no algorithm. Sessions run one hour biweekly in three-month blocks. The success metric is not satisfaction scores but papers submitted and committee milestones cleared in that window.

Do not bill the students. Graduate students in a supervision gap are the lowest-ability payer in the chain. Three plausible payers exist: departments that do not want to lose the student, industry labs buying a hiring pipeline, and funders who attach research continuity conditions to grants. Collect outcome data from day one so you can put a pilot result in front of all three.

flowchart LR
  A[Researcher now in industry] -->|Standard moonlighting contract| B[Co-advising match]
  C[Student without supervision] --> B
  B --> D[Biweekly sessions, 3 months]
  D --> E[Paper submitted, milestone cleared]
  E --> F[Department, lab or funder pays]

Success Criteria

Two assumptions carry the idea. First, will industry researchers make time when it is paid? The fact that some already do it unpaid is weak evidence, but at scale employer approval becomes the bottleneck. Negotiating the template with one large lab’s legal team settles this fast.

Second, will departments read this as an intrusion or a rescue? Losing a student is a metric hit, but handing supervision to an outsider is also a status question. Positioning strictly as secondary advising, never touching degree-granting authority, is the early survival condition.