Deep Research

NVIDIA (NVDA)

4 min read

Snapshot

  • Ticker: NVDA (Nasdaq)
  • Bucket: Compute & AI Megacaps
  • Q1 2026 position: 1,248,592 shares common, $217,754,445 — 4.35% of 13F, #6 position (calls fully unwound this quarter)
  • HQ: Santa Clara, California
  • What it does: Dominant designer of AI accelerators (GPUs), plus the networking, CPUs and software stack around them.

Business Overview

NVIDIA is the central clearinghouse of the AI buildout: its data center platform — Hopper, then Blackwell, now ramping toward Vera Rubin — captures the largest single share of global AI capex. Fiscal 2026 (ended January 2026) revenue was 68.1B (+73% y/y), with data center revenue of 500B of Blackwell-plus-Rubin revenue from the start of calendar 2025 through the end of calendar 2026.

Beyond the GPU itself, the company has widened its moat into Spectrum-X Ethernet and NVLink networking, Grace/Vera CPUs, and CUDA software. The February 17, 2026 multi-year Meta partnership is representative: millions of Blackwell and Rubin GPUs, the first large-scale standalone Grace CPU deployment, Vera CPU-only servers on the 2027 roadmap, and Spectrum-X integrated into Meta’s switching stack — against Meta’s guided 2026 capex of 5B stake in Intel and a $1B investment in Nokia (October 2025).

Financial Trajectory

Fiscal year (Jan-end)RevenueNotes
FY2024~$60.9BFirst full ChatGPT-era year
FY2025~$130.5BHopper peak, Blackwell launch
FY2026$215.9B+65% y/y; Q4 DC revenue $62.3B; networking +142%

Q4 FY2026 GAAP results and Q1 FY2027 guidance came in above expectations (“quashes AI bubble talk,” per Fortune’s coverage).

Why Atreides Owns It

Baker is publicly on the record that this is not a repeat of 2000, and NVDA is his test case. His Cisco argument: Cisco peaked at roughly 150-180x trailing earnings in March 2000; NVIDIA trades nearer 40x with far higher margins and — crucially — demand that is lit, not dark. The 2000 telecom bubble was defined by dark fiber, useless without the optics and routers on either side; today’s GPUs are deployed against measurable inference demand the moment they land. Both the earnings multiple and the utilization picture, in his view, refute the bubble analogy.

That said, Atreides has traded NVDA as much as owned it, and the options overlay is the tell. The fund initiated in Q1 2025 with 1.09M common shares plus calls on 2.3M underlying (~359M) plus calls on 3.5M underlying (1B of combined upside exposure, positioned ahead of the Meta Blackwell/Rubin partnership formally announced February 17, 2026, days after the Q4 13F window closed. In Q1 2026 the calls were unwound entirely and the common trimmed to 1.25M shares — consistent with monetizing the event and reverting to a core-but-not-maximal long. Within Baker’s “watts and wafers” framing, NVDA is the obvious beneficiary but not the mispriced one; the bigger weights sit in connectivity, optics and memory where multiples lag.

Position History

QuarterTypeShares/NotionalValue% of 13F
Q4 2024—not held——
Q1 2025Common1,085,628$117,660,3633.57%
Q1 2025Call2,300,000$249,274,0007.57%
Q2 2025Common689,813$108,983,5563.02%
Q3 2025Common574,723$107,231,8172.09%
Q4 2025Common1,925,231$359,055,5824.39%
Q4 2025Call3,500,000$652,750,0007.98%
Q1 2026Common1,248,592$217,754,4454.35%

Two distinct call campaigns bookend the history: entry calls in Q1 2025 (gone by Q2) and the big Q4 2025 overlay (gone by Q1 2026). The pattern is event-driven leverage layered on a core position, not buy-and-hold; common share count has swung from 575K to 1.93M and back to 1.25M across four quarters.

Risks

  • Customer concentration and custom silicon: the same hyperscalers driving the 100B multi-year AMD commitment.
  • AI capex digestion: any pause in hyperscaler/neocloud capex hits NVDA’s growth rate first and hardest; circular vendor-financing arrangements (investments in customers/suppliers) amplify cycle risk.
  • Supply chain dependence: TSMC CoWoS capacity and HBM supply are external bottlenecks; the Rubin ramp depends on both.
  • Export controls: China data center revenue remains structurally impaired and policy can tighten further.
  • For Atreides specifically: the alpha case is thinner than in their connectivity/memory names — NVDA is the consensus long, and Baker’s own framework argues the lagging multiples are elsewhere.

Sources

Backlinks

Graph View