Nat Friedman & Daniel Gross / NFDG + AI Grant
Nat Friedman (ex-GitHub CEO) and Daniel Gross (ex-SSI CEO, ex-Apple AI) — launched the NFDG fund in 2023 and run the AI Grant accelerator. Both were hired into Meta Superintelligence Labs in 2025 (Friedman co-leads MSL with Alexandr Wang; Gross joined the lab), with Meta partially buying out NFDG.
The marquee operator-investor duo of the model layer. Their ~$1.1B NFDG fund (reported ~4x in two years) holds the frontier-AI roster — SSI, Perplexity, Character.ai, CoreWeave — and their AI Grant program ($250k cash + $250k Azure credits) seeds the next cohort. In 2025 Meta hired both into Meta Superintelligence Labs and took a >$1B partial buyout of NFDG (no control, no portfolio info), making them the rare case where individual AI investors were effectively acqui-hired by a hyperscaler.
- Not open — Not an open-ended vehicle — analysis only, not an available allocation.
- Self-reported — Headline returns are manager- or firm-reported and unaudited.
Top holdings
Ilya Sutskever's lab; flagship frontier-model position (Gross was SSI CEO).
AI search; one of NFDG's headline winners.
AI-cloud/GPU infra — the compute-infra leg; now public (CRWV).
Consumer AI; backed before Google's licensing deal.
Design/AI-adjacent; long-held NFDG name.
Performance
July 2026: Fund-level not disclosed. Listed sleeve: CRWV -28.8% to $71.77 (Jul 31); SSI mark held ~$32B — No new fund-level figure was published, so the headline ~4x/~$1.1B are unmoved. Two real events did occur in July 2026: Nvidia committed $5B to SSI on Jul 27 (flat at ~$32B, plus a Vera Rubin compute partnership), and Meta launched Meta Compute on Jul 17, which drove CoreWeave -35% in a session. Perplexity marked up to ~$23B (Series E-6, Jan 2026) — the one supportive private move.
Recent moves
2025: Meta hired Friedman and Gross into Meta Superintelligence Labs and agreed a >$1B partial buyout of NFDG (no Meta control or portfolio visibility). AI Grant continues seeding startups ($250k + Azure credits). Reported NFDG stakes span SSI, Perplexity, Character.ai, CoreWeave.
Our take
Why it matters: NFDG is one of the highest-signal model-and-infra-layer books of the cycle (SSI, Perplexity, CoreWeave), and the founders now sit inside Meta's superintelligence effort — a unique investor-to-operator vantage. Edge: operator credibility (GitHub, SSI, Apple AI) translating into the best early AI deal flow. Caveats: the ~4x is press/self-reported and unaudited; nearly all positions are private (disclosed-not-weighted); the Meta hire creates a conflict/attention overhang — their day jobs are now at a portfolio-adjacent hyperscaler, and Meta explicitly took no portfolio information or control.
Read this as an operator-VC book, not a family office, and read the return honestly.
The headline ~4x-in-two-years is a paper TVPI mark, press/self-reported and unaudited — its credibility rests on private-round markups, with SSI at a reported ~$32B the load-bearing one. What makes NFDG unusual is that Meta's 2025 partial buyout actually CRYSTALLIZED a chunk of that paper into real DPI: LPs got liquidity on roughly half the book at ~4x book value plus undeployed capital back, while the rest stayed in to ride. Rare and genuinely good for LPs — most funds this vintage are still J-curve paper. But it also says the vehicle is done: no new deployments, principals gone, so there is no forward strategy to underwrite, only a legacy book to harvest. The tension the marketing won't volunteer: Gross co-founded SSI, the crown jewel, then left it for Meta's competing superintelligence effort — key-person risk realized on the single most important mark. Meta deliberately took no control and no portfolio info, which reads as a conflict firewall.
Thesis
NFDG is a ~$1.1B (2023 vintage) AI venture book run by two elite operators — Nat Friedman (ex-GitHub CEO) and Daniel Gross — whose edge was privileged deal access to the model/infra layer (SSI, Perplexity, CoreWeave, Character.ai). As of 2025 it is effectively a fund in RUN-OFF: Meta bought up to 49% of holdings (~$1B), undeployed capital was returned, new investing stopped, and both principals decamped to Meta Superintelligence Labs. Closed vehicle — not investable.
Concentrated, access-driven venture: deploy a large single fund into a short roster of frontier-AI names where operator reputation (GitHub, Apple AI, SSI, YC) buys allocation others can't get, paired with the AI Grant accelerator (~$250k uncapped SAFE + ~$350k Azure credits, ~$850k total incl. partner credits) as top-of-funnel deal flow. Not diversified VC — a high-conviction bet on the model-and-compute layer of one cycle, sized so a couple of winners drive the whole return.
Assessment
- Access edge was real: operator credibility bought early allocation in SSI, Perplexity and CoreWeave most funds couldn't get — the scarce input is a seat, not capital.
- LPs got actual liquidity: the Meta buyout turned paper marks into distributed cash on ~half the book at ~4x book — real DPI, not just an unrealized TVPI story.
- Barbell exposure to both the model layer (SSI, Perplexity, Character.ai) and the compute layer (CoreWeave) — captured the pick-and-shovel trade alongside the app bets.
- AI Grant gave a structured, low-cost top-of-funnel into the next cohort rather than relying only on inbound — a repeatable sourcing engine, not one-off angel checks.
- The ~4x is a paper TVPI on private rounds, unaudited and self-reported in spirit; the un-sold remainder is still marked to markups a down round could erase. Realized DPI is known only on the Meta-purchased slice.
- SSI carries the return and is now, from NFDG's vantage, founder-departed — Gross left it for a direct competitor. Concentration in one mark that just lost its most credible internal signal.
- Vehicle is in run-off: no new investments, both principals at Meta. There is no go-forward strategy to assess — only harvest of a legacy book, so past ~4x says little about any future NFDG.
- Marks depend on the AI private-valuation regime holding. A broad re-rating of frontier-lab valuations hits the residual book directly, and these are illiquid, hard-to-exit positions.
- Attention/conflict overhang: principals now work at a portfolio-adjacent hyperscaler competing with their own crown-jewel investment — the reason Meta explicitly took no info or control.
Record
Return3y ~4x, vintage ~2023–2025 — treat as a paper TVPI, not a track record. Sourcing: SaaStr/press and the principals' own framing; not audited, and private holdings are disclosed-not-weighted, so no position-level attribution is verifiable. The one hard data point is the Meta deal: it acquired up to 49% of holdings at roughly book (~$1B), reportedly delivering LPs ~4x on those positions plus ~$500M of undeployed capital — that slice is realized DPI. The rest is unrealized, marked to last-round prices, with SSI's reported ~$32B the dominant swing factor. Honest read: part of the ~4x is now cash for LPs (strong); part is still a mark that lives or dies with the AI private-market regime and a few names. Skill vs. cycle beta is impossible to separate from outside — 2023–25 lifted all frontier-AI paper. No YTD or 1y figure is disclosed.
Risks & fit
- SSI mark impairment — a down round or stall at Safe Superintelligence, now without Gross, would hit the single largest driver of the residual return.
- AI private-valuation regime reversal re-rating CoreWeave/Perplexity/frontier-lab marks across the illiquid residual book.
- Key-person gap is permanent, not a risk — both principals are at Meta; no one is actively managing new value creation.
- Exit/liquidity risk on the un-sold positions: private, concentrated, dependent on IPO or strategic windows that can close.
- Conflict overhang: principals inside a competing hyperscaler could constrain how the residual book is handled, even with Meta's info firewall.
The access-edge/skill framing is falsified if the residual, un-sold half fails to realize distributions near its carrying marks — SSI stalling or down-rounding without Gross, or an AI-valuation re-rating dropping eventual DPI well below the ~4x TVPI. It is further undercut if the ~4x proves cycle beta not skill: 2023–25 lifted essentially all frontier-AI paper, so index-like AI exposure matching the return would show the edge added little. If retained positions exit at or above marks, the skill claim holds. Vehicle is closed — no future NFDG record exists to test.
An analytical case study in AI operator-VC: how reputation converts to allocation, how a rare partial acqui-hire crystallizes paper marks into LP DPI mid-life, and why a ~4x TVPI is not a realized track record. Useful for studying return attribution (skill vs. cycle beta), key-person/concentration risk in one crown-jewel mark, and hyperscaler-vs-portfolio conflicts. Not investable — a closed, run-off private fund; this critiques strategy and record, not an allocation call.
Not disclosed; venture economics implied at ~25% carry (press estimated ~$62.5M each on crystallized gains) atop a typical management fee. As a closed private fund the terms are opaque and moot for outside readers — treat any figure as an assumption.