One prominent outlet tied the semiconductor selloff to a new DeepSeek model released the same day. At least five other independent outlets covering the identical price action did not mention DeepSeek at all.
The price action is not in dispute. On Thursday, September 10, Nvidia fell 2.37% to $218.36. Intel fell 5.57% to $100.32. Oracle fell 5.23% to $153.17. The 3x-leveraged Direxion Daily Semiconductor Bull ETF dropped 8.03% while its inverse rose 8.09%, confirming a genuine sector-wide move rather than a handful of idiosyncratic names.
The cause is very much in dispute, and the dispute is itself the more useful story.
What DeepSeek actually released
DeepSeek published DeepSeek-V4.1-Flash the same day, per its own official API changelog. The model carries 552 billion parameters, selectively activating 8 billion to 16 billion per token, with a 1 million token context window. API prices were cut by up to 32%. Third-party technical write-ups put its KV-cache compression at roughly 890 bytes per token.
The stated mechanism linking that to chip stocks is specific: a lower per-token memory footprint implies less high-bandwidth memory demand growth than current supplier capacity plans assume. That argument is coherent, and it applies most directly to memory manufacturers.
The competing explanations are specific too
Multiple independent outlets covering the identical session attributed the same moves to entirely different causes, none of them DeepSeek. Cited drivers included post-rally profit-taking, particularly for Intel after a run of roughly 188% year to date; the 10-year Treasury yield at a multi-year high near 4.95%; a crude-oil spike with Brent above $105; and, for Intel and AMD specifically, competitive pressure from Nvidia's newly launched Vera Arm-based server CPU, which is a story with no DeepSeek content whatsoever.
Oracle has the cleanest independent explanation of the four names. It reported fiscal first-quarter results the same evening, and two separate sources attributed its decline to pre-earnings positioning around rising AI capital expenditure, guided at $90 billion to $95 billion for fiscal 2027, and correspondingly negative projected free cash flow. Neither mentioned DeepSeek.
Most tellingly, coverage focused specifically on the Korean memory names at the center of the DeepSeek-demand argument attributed their declines to Wall Street's own macro-driven weakness rather than to the model release.
Why this matters beyond one session
DeepSeek-linked market narratives have a history of not surviving contact with time. The original January 2025 R1 release produced a far larger Nvidia decline and was widely reassessed a year later as having overstated its case. An April 2026 DeepSeek release caused Chinese chip stocks to rally, the opposite directional reaction.
Meanwhile, Nvidia chief executive Jensen Huang spent Wednesday at a technology conference reaffirming roughly 70% fiscal-year revenue growth guidance, defending the company's ecosystem investments with a "put in $1, get back $100" framing, and calling cybersecurity artificial intelligence's next blockbuster application. The stock fell anyway, both that session and the next.
That is the real question underneath the attribution dispute. A company reaffirming 70% growth guidance whose stock trades below its 52-week high of $236.54 is not being repriced on one model release. Rates, oil and positioning explain more of it than any product announcement does.
No named sell-side analyst has published a rating or estimate change citing DeepSeek-V4.1-Flash. Until one does, "DeepSeek rattled chipmakers" describes a coincidence in timing more defensibly than it describes a causal chain.