Skip to content

P4-C1 · DeepSeek Sell-off Complete Forensic

Core One-Liner

**NVDA -17% / -$589B in a single day — short-term panic or thesis truly broken? Using Part 3's five tools to dissect it all at once.**

Real Public Case — Using Part 3 Tools to Break Down 5 Real Events

P4-C1 (Part 4, Chapter 1). Entirely based on public data (NVDA / DeepSeek public releases + market reaction).

1. Event: 2025/01/27 DeepSeek V3 + R1 Release, NVDA -17% in a Single Day

Date Event
2024/12/26 DeepSeek V3 released (near GPT-4 capability, official training run cost $5.576M = 2.788M H800 hours × $2/hr per tech report; excludes prior research/ablations/data/failed experiments/hardware all-in cost) vs GPT-4 ~$100M+ (OpenAI did not disclose; external estimate only)
2025/01/20 DeepSeek R1 released (reasoning model, close to o1 capability)
2025/01/27 Monday Open **NVDA -17%**, SK Hynix -10%, Samsung -8%
2025/01/27 Intraday NVDA market cap evaporated -$589B (largest single-day loss in history)
End of Feb 2025 NVDA recovered 60-80%
2025/04/07 After "Liberation Day" tariffs, NVDA hit $86.62 stage low (second shock post-DeepSeek)
2025/06/25 NVDA's first new all-time high post-DeepSeek (~$154.31, Bloomberg data); kept making new highs through July

2. Using Part 3's Five Tools to Dissect — 4 Say Hold, 1 Says Re-evaluate

2.1 Hedge Fund Perspective (P3-C1)

Coatue / Druckenmiller-type Reaction (inferred from public interviews):

  • Step 4 monitor: earnings vs thesis. NVDA Q3 FY25 earnings (released 2024/11/20) showed $35B revenue, +94% YoY — still the most recent data visible to the market at the time of the DeepSeek event (1/27/2025). Q4 FY25 ($39B, +78% YoY) wasn't released until 2/26/2025, i.e. a month after the event — and that data further validated the thesis was intact. Supports not invalidated.
  • Step 5 exit triggers: not triggered (hyperscaler capex guide unchanged, OAI revenue still growing)
  • Action: hold, no trim

Note: cannot cite Druckenmiller as a "hold through" example — he had fully exited NVDA in 2024 Q3 and publicly admitted on Bloomberg in 2024/10 that it was a "big mistake". For PMs who actually held through 1/27, see the institutional 13F analysis below.

2.2 Buffett Perspective (P3-C2)

  • Step 1 circle of competence: AI economics still uncertain over 10 years. DeepSeek didn't change this.
  • Step 4 reasonable price: -17% brought fwd PE from 35x → 29x. Margin of safety improved.
  • Buffett framework: Don't enter (circle of competence unchanged). But Buffett doesn't own NVDA, so this is just observation.

2.3 Bottleneck-Finding Perspective (P3-C3)

DeepSeek is an algorithmic efficiency breakthrough — same capability with fewer GPUs. This is a Jevons Paradox candidate:

  • Step 1 physical bottlenecks unchanged (HBM / power / optical modules still there)
  • Step 2 supply: NVDA H100/H200 lead times haven't shortened
  • Step 3 demand: short-term panic but long-term compute demand rises more with efficiency (Jevons)

Jevons Paradox explanation: In 19th-century Britain, coal efficiency improvements led to increased total coal demand (because cheaper → more use cases). DeepSeek makes AI cheaper → more companies use it → total compute demand rises.

→ Bottleneck-finding perspective: thesis not broken, actually strengthened (inference demand will surge).

2.4 Multi-PM Perspective (P3-C4)

PM Perspective Post-DeepSeek View
Value (Buffett-style) Valuation + moat Improved (-17%) but circle unchanged, still don't enter
Growth (Coatue) TAM + positioning Strengthened (Jevons)
Macro (Druckenmiller) Capex cycle Unchanged (capex guide unmoved)

3 perspectives: 1 don't enter + 2 strengthened → mixed → net conclusion: hold or slightly add.

2.5 Anti-Thesis Perspective (P3-C5)

Strongest counter-argument (DeepSeek provides new ammunition):

"If training costs drop 90%, can OpenAI / Anthropic still raise at $300B valuation?" "If efficiency improves, can hyperscalers cut capex?"

Invalidation triggers (which did DeepSeek trigger?):

Trigger Triggered?
OpenAI valuation reset (primary $300B → <$150B) Not triggered (subsequent 2025 funding rounds still rising)
MSFT/GOOGL FY27 capex guide < 10% YoY growth Not triggered (post-Q1 2025 earnings, capex guide actually rose)
Microsoft AI business run-rate YoY < 50% (Copilot ARR not separately disclosed) Not triggered
Vertical AI major customer churn Not triggered

4 invalidation triggers all not triggered → anti-thesis didn't win, your long thesis holds.

3. Lesson — Short-term Price ≠ Long-term Thesis

Date NVDA Price DeepSeek Reaction Real Thesis Signal
1/24 (Pre-DeepSeek) $146 - Thesis OK
1/27 -17% $121 Panic Thesis unchanged (capex guide / revenue / earnings)
2/27 (1 month later) $135 Recovered 60% Market recognizes Jevons Paradox
2026/5/21 (Q1 FY27, 16 months after DeepSeek) $172 New highs Q1 FY27 earnings delivered on the thesis

Core lesson:

  • Retail approach: Sell on -17%, miss the +40% rebound in 3 months
  • Institutional approach: Use Part 3's five tools to confirm thesis intact, hold, some add (-17% provided an entry)
  • Key gap: Not analytical ability, but awareness of historical base rates like Jevons Paradox

4. vs. What Retail Actually Did at the Time

Public data (Robinhood / Fidelity retail flows):

  • 1/27 single day: **Retail net sold NVDA** (panic)
  • 1/28-30: Retail continued selling (-5% additional)
  • February: Retail net bought (but already from $146 → $135, missed entry)

Institutions (Whalewisdom Q1 13F):

Q1 2025 13F (institutional actions post-DeepSeek): - Druckenmiller: already at 0 (fully exited in 2024 Q3, not on the back of DeepSeek) - Coatue: continued trimming NVDA, rotating into other AI infra (ANET / AVGO) - Tiger Global: hold - Actual "buy the dip" players: not Soros / Druckenmiller; mainly long-only mutual funds + index funds picking up the float passively

→ Retail sold the bottom, institutions bought the bottom. This is Part 3 tools in action.

5. What Your Thesis Should Learn

Lesson How to Modify Your Thesis
Short-term panic ≠ thesis broken Add invalidation_triggers, don't be swayed by noise
Jevons Paradox is real AI efficiency breakthrough = total demand increases, not decreases
4 invalidation triggers not triggered = hold Write an explicit checklist (P3-C5)
Single-day -17% = adding opportunity Leave room for trim/add in starter → core (P3-C1 Step 3)
Earnings data > news sentiment On 1/27, the latest data was Q3 FY25 earnings (released 2024/11/20), all supports strong

6. Next → P4-C2

DeepSeek is an algorithmic wildcard. The next case is a supply chain wildcard — Samsung losing HBM market share.

→ P4-C2 · Samsung Loses HBM Market Share A real demonstration using P3-C3's 5-step bottleneck-finding approach.

7. Further Reading (this chapter — DeepSeek 2025/01/27 sell-off forensic)

All free sources, primary papers / primary announcements / primary earnings preferred

Primary papers (arxiv, all free):

Primary earnings / announcements:

Reuters / primary news (free):

SemiAnalysis (Dylan Patel, partial free):

Wikipedia (background, 3-5 min):

Podcasts (1-2 hr):

Pair with chapter self-check: After reading any 3-4 of the above, you should be able to answer the self-check more confidently — especially on "$5.6M vs $1.6B real capex" and the concrete arguments for why "4 tools say hold, 1 says re-evaluate".