Atreides Management (Gavin Baker)
Gavin Baker — Managing Partner & CIO of Atreides Management, LP (founded 2019). Ex-Fidelity, where he ran the OTC Portfolio for ~13 years. Among the most-followed AI/semi voices on FinTwit (@GavinSBaker, ~400k+ X followers); widely circulated for technically deep podcast appearances (e.g. Invest Like the Best) on chip architectures, memory, and the AI compute stack.
Concentrated long-biased growth across the AI/semiconductor value chain — compute (NVDA), connectivity/interconnect (Astera Labs, Credo, Ciena, Coherent, Lumentum), memory (Micron), plus AI-software and cybersecurity. Known for fundamental, architecture-level conviction (why memory configs and optical interconnect matter for AI) and for willingness to express it with leverage (NVDA call options).
- Not open — Not an open-ended vehicle — analysis only, not an available allocation.
- Partial 13F — The disclosed book is longs-only and ~45 days lagged — a slice, not the strategy.
Top holdings
Largest 13F position; in early 2026 Atreides added leveraged NVDA call options on top of the equity — flagged in press as 'one of his most aggressive bets yet' on the AI-compute leader.
Signature high-conviction connectivity name — held since the Series C private round. Added ~1.75M shares in Q1 2026 (to ~3.37M, ~$369M), a top buy of the quarter; AI back-end interconnect / memory-pooling play.
Optical networking beneficiary of datacenter interconnect / coherent optics demand from the AI buildout.
AI-memory (HBM/DRAM) conviction long; added in Q1 2026 alongside other 'smart money' loading memory as Micron revenue grew ~57% YoY on AI demand.
Photonics / optical-components supplier to the datacenter interconnect chain.
new buy — Top-five Q1 2026 buy; active electrical cable / SerDes connectivity for AI clusters — same interconnect thesis as ALAB/CIEN.
Lasers / optical transceivers for datacenter optics.
Recent moves
Q1 2026 13F: exited Microsoft entirely (sold all ~181k MSFT shares); fully exited HubSpot and Intuit, cut most of Snowflake and trimmed GitLab. Added/initiated heavily in the interconnect + memory complex — Astera Labs, Credo, Ciena, Coherent, Lumentum, Micron — plus new positions in Zoom, Akamai and Palo Alto Networks. Layered leveraged NVDA call options on top of the equity stake in early 2026.
Loading disclosed positions…
Our take
The most institutionally credible name in this 'viral KOL' set — a real ~$7–8B book with a quarterly-audited 13F trail, run by an investor with a ~two-decade Fidelity pedigree and genuine architecture-level fluency on AI silicon. The signature edge is the interconnect/connectivity layer (ALAB, CRDO, CIEN, COHR, LITE) plus memory (MU), not just owning NVDA. Caveat: 13F shows only US long equity at a lagged quarter-end — it omits shorts, the options/leverage he's known to use, non-US names, and the private book, so the public picture understates both the risk and the strategy. Fund-level returns are not auditable from outside.
The most useful thing here is how unreliable the public picture is.
For the same Q1 2026 filing (period 2026-03-31, filed 2026-05-18), trackers disagree materially: HedgeFollow ~$4.1B / 51 holdings, Insider Edges $5.00B / 51 holdings, an Insider Monkey April-2026 snapshot ~$8.18B / 20 positions. Insider Monkey, citing the firm's Form ADV, reports ~$6.49B discretionary RAUM as of 2025-12-31 — secondhand, not the ADV itself, and none of these is the fund. They differ on options treatment, aggregation and snapshot date, so any 'Baker portfolio weight' quoted to a decimal is a vendor artifact. Substantively, the interconnect thesis is real and was early. But it is one trade: ALAB, CRDO, CIEN and MU do not diversify each other; they are correlated expressions of hyperscaler optical and memory capex. The Q1 2026 full exit from Microsoft (181,468 shares) alongside new Zoom, Akamai, Credo and Astera Labs positions reads as rotating out of SaaS-at-risk-from-AI into AI infrastructure — coherent, and fully undiversified.
- 2026-03SEC data puts Q1 2026 at $5.00B across 54 holdings; the $8.18B / 56-holding figure is the Q4 2025 filing rather than a vendor disagreement about Q1. Book value is quarter-stamped, not disputed.
Thesis
Atreides is a long-biased concentrated tech fund whose distinctive claim is not owning NVIDIA but owning the plumbing around it — interconnect (Astera Labs, Credo, Ciena) and memory (Micron). The bet is that AI capex bottlenecks migrate from raw compute to moving and storing data, so second-derivative suppliers re-rate harder than the chip vendor. A genuine, testable mechanism — and also a single macro factor expressed across a dozen tickers that mostly rhyme.
Concentrated long-biased public equity plus a private book, run by a fundamental technology PM (at Fidelity 1999–2017, PM of the Fidelity OTC Portfolio 2009–2017; firm founded 2019). Holdings cluster along one value chain — compute, interconnect, memory, AI software. Expression includes derivatives: the Q4 2025 13F (period 2025-12-31) disclosed call options on an underlying 3.5M NVDA shares layered on the equity stake, so realised risk exceeds what share counts imply.
Assessment
- Mechanistic rather than narrative: names why memory configuration and optical interconnect gate AI throughput, which is falsifiable against shipment and capex data.
- Positioned one layer off the crowded trade — interconnect and memory suppliers, not just the obvious GPU name.
- 18 years at Fidelity (1999–2017), eight of them running the Fidelity OTC Portfolio (2009–2017) — domain fluency that predates the current AI cycle.
- Exited Microsoft outright in Q1 2026 while adding AI-infrastructure names — thesis discipline over incumbency.
- Correlation masquerading as diversification: ALAB, CRDO, CIEN and MU are one factor — hyperscaler capex — held several ways.
- The 13F is longs-only, US-only and ~45 days lagged; shorts, non-US names and the private book are invisible, so the disclosed book is not the strategy.
- Public 'portfolio value' for the same filing ranges ~$4.1B to ~$8.18B across trackers; weights quoted anywhere are vendor-dependent, not fund facts.
- The Q4 2025 call overlay on 3.5M underlying NVDA shares creates drawdown convexity that share-count-based reconstructions structurally understate.
- Single key person: Atreides is the CIO's view at scale, with no publicly evidenced succession plan or independent risk seat.
- 2026-07July 2026 inverted this: NVIDIA +0.3% and Broadcom +3.1% held while the layer-off names led the rout — SOX -20.6%, Micron -28.7%, Astera Labs -35.6%. One layer off the GPU was the risk that month, not the edge.
- 2026-07Microsoft rose 24.6% in July 2026 after holding capex flat — among the month's best large-cap outcomes — while the AI-infrastructure names bought in its place fell hardest. Defensible as thesis, expensive so far.
- 2026-03The largest disclosed Q1 2026 line is a put on QQQ — 1.4M shares underlying, $808M, 16.15% of the book, and the top holding in Q4 2025 too. That index hedge is visible in the 13F, so the long bias is overstated.
Record
Fund-level net-of-fee returns are not disclosed and cannot be verified from outside — Atreides is a private partnership. Circulating '1-year return' figures are third-party reconstructions of the long-only 13F book, and they disagree: Insider Edges shows +59.99% as of the Q1 2026 report date, earlier trackers cited ~+47%, and HedgeFollow separately shows Q1 2026 at -3.53%. These models assume static quarter-end snapshots, ignore intra-quarter trading, exclude shorts and the private book, and mis-handle options — they measure the vendor, not the manager. Attribution matters more than the number: across 2024–2026 the AI-infrastructure complex rose sharply on its own, so a long-biased book concentrated there prints large gains from theme beta plus option leverage. Isolating skill would need the shorts, hedges and private marks — none public.
Risks & fit
- A hyperscaler capex digestion pause hits compute, interconnect and memory simultaneously; the book shows no offsetting exposure in the 13F.
- Memory is cyclical and has historically round-tripped; MU-type positions can give back multi-year gains in two quarters.
- The disclosed call overlay amplifies drawdown and can decay to zero even if the directional view is eventually right.
- Custom-silicon or co-packaged-optics roadmap shifts could disintermediate specific interconnect suppliers while the broad thesis stays intact.
- Capacity: a book this size in mid-caps like ALAB and CRDO faces real liquidity constraints on exit.
The thesis breaks if AI datacenter spend keeps growing but value stops accruing to interconnect and memory suppliers — hyperscalers vertically integrating optics and custom NICs, or co-packaged optics collapsing the merchant-silicon layer, so ALAB/CRDO/CIEN revenue decelerates while NVDA and hyperscaler capex keep rising. That divergence shows the bottleneck migrated somewhere the book does not own. Second falsifier: if positions were mostly bought after the re-rating, the 'early on interconnect' claim reduces to trend-following. Both checkable against supplier revenue and 13F entry quarters.
Readers studying how a concentrated single-theme technology book is constructed, and how little a 13F actually reveals about one. Useful as a case study in reading the AI value chain layer by layer rather than owning the headline name, and as a worked example of why third-party 'hedge fund returns' derived from 13Fs should be treated as vendor models, not results. Atreides is a private fund; nothing here describes an available or advisable allocation.
Fee terms are not publicly disclosed; the firm's own site publishes no AUM, strategy detail or fee schedule. Funds of this type typically carry management-plus-incentive structures, but no Atreides-specific figure is verifiable from public sources.