DeepSeek-R1 Release Timeline: The Signals Before Nvidia's Selloff

Archive note: this retrospective was researched and first published in August 2026. The event time records DeepSeek's original R1 announcement, not MoveSurge's publication date.

Countdown to Nvidia's 17% drop: DeepSeek announced R1 on Monday, January 20, 2025—seven calendar days before Nvidia fell 16.9% on Monday, January 27.

DeepSeek-R1 did not appear from nowhere on the morning Nvidia sold off. The information arrived in separate channels over twelve days: the app launch, an official model announcement, a technical paper, a television interview, high-reach posts on X and a consumer app ranking. Each step made the story legible to a wider audience.

The verified timeline

What the first announcement actually established

The Monday, January 20 release—seven calendar days before Nvidia's drop—established that DeepSeek had produced a capable open-weight reasoning model and had exposed it through an inexpensive API. The Wednesday, January 22 paper, five days before the drop, supplied technical detail and benchmark results. Neither source said that the entire R1 research program cost the frequently repeated $6 million figure. That number came from the earlier DeepSeek-V3 paper and had a narrower accounting scope, examined in our cost-claim fact check.

Why the consumer ranking mattered

Code repositories and benchmark tables primarily reach engineers. A number-one App Store position reaches journalists, executives and retail investors at once. TechCrunch reported that DeepSeek was the top free U.S. app by Sunday afternoon—one calendar day before Nvidia's 16.9% Monday drop. The ranking did not prove long-term retention, revenue or a collapse in GPU demand. It did prove that a previously obscure lab could attract mass attention quickly.

Was Wall Street ignoring the story?

The evidence supports a narrower conclusion. Attention was fragmented before the weekend. Nvidia had already weakened on Friday, January 24, three calendar days before the 17% drop, and financial commentary was already asking whether cheaper models would change data-center spending. There is no public dataset showing that Wall Street desks collectively ignored R1 for a fixed period, or that trading algorithms could not process technical sources. The defensible claim is that the narrative became broader and more urgent as it moved from papers to television, X and the App Store.

That distinction matters because the Monday move was a reaction to a compressed story: frontier-level AI, lower reported cost, constrained chips and sudden consumer adoption. Several parts were real. Others had lost their accounting and technical qualifiers by the time they reached market headlines. The X amplification investigation traces that compression post by post.

The quantitative signal—and the result that weakens the easy story

OpenDigger's GitHub-event analysis reports that the R1 repository gained more than 2,000 stars on Monday, January 20—seven calendar days before Nvidia's drop—and then added roughly 2,000 to 4,000 stars per day through Sunday, January 26, one calendar day before the drop. Historical milestone sources place the repository near 10,000 stars by Wednesday, January 22, five calendar days before the selloff, and near 20,000 by Sunday, January 26. Those are ranges and thresholds, rather than invented exact daily counts.

A public series attributed to Appfigures estimates that DeepSeek app downloads rose from 78,674 on Monday, January 20—seven calendar days before the drop—to 677,870 on Sunday, January 26, one calendar day before it: an 8.6× increase and 2,039,761 estimated downloads across the seven days. This is directional evidence from one vendor lineage covering Apple and Google stores. It is not an Apple first-party App Store Connect export.

The app series also supplies an important disproof. Its largest estimated day-over-day acceleration was +173.75% on Monday, January 27—the selloff day, when broad news coverage and the market shock were already occurring. App downloads alone therefore cannot establish that consumers led newsrooms. The stronger precursor is the combination of GitHub growth, specialist discussion, Sunday's public App Store rank and cross-community X/Reddit diffusion.

What counts as a golden source in this study?

Mainstream media was early; broad saturation came later

The claim that mainstream media ignored DeepSeek until Monday is false. Nature published a substantive R1 feature on Thursday, January 23—four calendar days before Nvidia's drop. The Financial Times and The Wall Street Journal published substantive reporting by Saturday, January 25—two calendar days before the drop. The handoff's predeclared broad-saturation test asks for at least three selected independent Tier-1 originals across more than one outlet type on the same UTC date. Monday, January 27—the selloff day—is the first date that clears that threshold in the curated sample.

This changes the lead-lag thesis. Developer and social adoption preceded broad, multi-outlet saturation. They did not precede every mainstream article. Monday's shift was scale and translation: Bloomberg, Reuters, CNBC, AP, BBC and others turned model adoption, efficiency and App Store rank into a valuation question about chips and data-center capital spending.

Reddit shows the audience transition

On Monday, January 20—seven calendar days before the drop—a LocalLLaMA release thread drew more than 1,000 upvotes and more than 360 replies within five days in the contemporaneous secondary snapshot. Its discussion focused on licensing, quantization, local deployment, benchmark skepticism and the distinction between supervised fine-tuning and reinforcement learning.

On Saturday, January 25—two calendar days before the drop—a hosted-censorship demonstration had more than 9,800 upvotes and 700 comments in the contemporaneous snapshot. By Sunday, January 26—one calendar day before the drop—the subject had escaped specialist communities. A meme-explanation thread gained more than 7,000 upvotes within one day.

Reddit scores are mutable and fuzzed. These values document the breadth of the conversation rather than precise permanent totals.

The reusable narrative-detection lesson

Track separate layers instead of waiting for one headline count: primary technical releases, developer adoption velocity, consumer rank, social claim quality, cross-community diffusion, media breadth and market repricing. A signal becomes more useful when its evidence class is visible. An estimate should stay an estimate; an opinion should stay an opinion; a benchmark should retain its task scope. That discipline identifies a narrative's acceleration without turning attention into proof.