Deep Research

CoreWeave (CRWV)

4 min read

Snapshot

  • Ticker: CRWV (Nasdaq); IPO March 2025
  • Bucket: Compute & AI Megacaps
  • Q1 2026 position: 810,535 shares, $62.8M, 1.26% of 13F
  • HQ: Livingston, New Jersey
  • What it does: The largest pure-play “neocloud” — GPU cloud infrastructure purpose-built for AI training and inference, contracted to hyperscalers and frontier labs.

Business Overview

CoreWeave pivoted from Ethereum mining to GPU cloud and became, by its own description, the fastest cloud in history to 5.13B versus 2.08B, up 112% year over year, with adjusted EBITDA of 740M (−36% margin) as depreciation and interest ramp ahead of revenue. FY2026 guidance is $12–13B.

The model is contract-backed leverage: a 24.9B of total debt, with 22B), Meta, Anthropic, and others. The attempted Core Scientific acquisition (≈1.3GW of power) was voted down by Core Scientific shareholders in October 2025, a setback for the power-ownership strategy. In June 2026 CoreWeave announced it would be first to deploy Nvidia’s Vera Rubin NVL72 — its speed-to-new-silicon edge remains the core differentiation.

Why Atreides Owns It

CoreWeave is the most direct public pure-play on Baker’s “watts and wafers” constraint thesis: its entire business is converting scarce power and scarce accelerators into contracted compute revenue. Atreides was part of the crossover IPO flywheel here — it held CoreWeave as an IPO-cohort position, appearing in the 13F with 1.88M shares (40 offering price.

What followed says as much about Atreides’ risk management as its conviction: the fund sold ~92% into the post-IPO melt-up in Q2 2025 (CRWV roughly quadrupled from the IPO price by June 2025), cut to a 63K-share stub by Q3, then rebuilt to 606K and 810K shares across Q4 2025–Q1 2026 at much lower prices. Baker has repeatedly framed neoclouds as levered, cyclical expressions of AI demand rather than structural winners — which is exactly how this position has been traded: rented aggressively around valuation, never held religiously. The fund’s simultaneous small Nebius position suggests a basket approach to the neocloud category rather than a single-name bet.

Financial Trajectory

PeriodRevenueYoYNotes
FY2024$1.92B~+730%Pre-IPO
FY2025$5.13B+167%Backlog $66.8B at year-end; Microsoft ~67% of revenue
Q1 2026$2.08B+112%Adj. EBITDA 740M; backlog 24.9B

Position History

QuarterTypeShares/NotionalValue% of 13F
Q4 2024—not held——
Q1 2025Common1,875,000$69,525,0002.11%
Q2 2025Common154,563$25,203,0430.70%
Q3 2025Common62,904$8,608,4120.17%
Q4 2025Common606,110$43,403,5370.53%
Q1 2026Common810,535$62,792,1461.26%

A violently traded IPO-cohort position: 1.88M shares at the Q1 2025 IPO, sold down ~92% into the Q2 2025 spike, reduced to a near-token 63K shares by Q3, then rebuilt 13x over the next two quarters as the stock de-rated. The round trip captured the IPO pop and re-entered cheaper — a textbook example of the fund treating neoclouds as trades around a structural theme rather than buy-and-hold compounders.

Risks

  • Leverage against depreciating assets: ~$25B of debt funds GPUs whose economic life may be shorter than contract tenors assume; refinancing risk if AI capex sentiment turns.
  • Customer concentration: Microsoft was ~67% of FY2025 revenue; OpenAI/Meta diversification is mostly backlog, not yet revenue.
  • Counterparty quality of backlog: the $99.4B figure includes commitments from cash-burning AI labs whose own funding is circular with the GPU trade.
  • Margin compression: GAAP losses widened in Q1 2026 even as revenue doubled; pricing on renewals versus hyperscaler in-house silicon (Trainium, TPU) is an open question — and Atreides itself is loudest on Trainium’s rise.
  • Power-acquisition setbacks: the failed Core Scientific deal shows securing watts at scale is contested and expensive.

Sources

Backlinks

Graph View