Wall Street Tumbles on Kimi K3 Fears: Nvidia Down, Samsung and SK Hynix Face Margin Pressure
Moonshot AI’s Kimi K3 sent Nasdaq down 1.4%, Nvidia down 2.2%, and chip stocks into bear territory. Here’s what actually happened and why Samsung and SK Hynix are sweating.
When Moonshot AI unveiled Kimi K3 on July 17, 2026, investors didn’t need a history lesson. AI and semiconductor stocks dropped sharply on the news, as traders drew immediate comparisons to the DeepSeek moment of 2025, when a Chinese lab released a model that performed comparably to US rivals at a fraction of the cost. This time, the model in question is even bigger, even more open, and dropped during a week already loaded with bad news. The sell-off was brutal, global, and — depending on who you ask — entirely rational.
On the day of the release, the Nasdaq fell 1.40%, the S&P 500 fell 1.01%, and the Dow dropped 0.77%. Nvidia shed 2.2%, Applied Materials fell 5.6%, Intel dropped 2%, and SanDisk lost 4%. The Philadelphia Semiconductor Index fell into a bear market on July 17, down more than 20% from its June peak. Global semiconductor stocks shed $3.3 trillion in market value since June 22, nearing bear-market territory.
What K3 Actually Is
Kimi K3 carries 2.8 trillion parameters, making it the largest open-weight AI model in the world as of its launch date. Released ahead of the opening of the World Artificial Intelligence Conference in Shanghai, K3 features native multimodal capabilities that process images and text simultaneously, along with an ultra-large context window that handles up to 1 million tokens at once. The model ranks below Anthropic’s Claude Fable 5 and OpenAI’s GPT-5.6 Sol in overall rankings, yet beats Claude Opus 4.8 and GPT-5.5 across coding and general agent evaluations. On the Artificial Analysis Intelligence Index, an independent composite benchmark that aggregates model performance across reasoning, knowledge, mathematics, and coding, K3 scored 57, ranking above Claude Opus 4.8 and GPT-5.5, practically on par with Claude Fable 5 and OpenAI’s GPT-5.6 Sol.
K3 is a significant jump from its predecessor: the previous model in the series, K2, had 1 trillion parameters. Moonshot charges $3 per million input tokens, while Fable 5 costs $10. Full model weights are scheduled for public release on July 27, 2026. Once that happens, any lab with enough GPUs can download and self-host a near-frontier model instead of paying OpenAI or Anthropic. That prospect is exactly what spooked the market.
Two Fears, One Selloff
The market focused on two points. First is the possibility that as near-free open AI models rapidly advance, profitability could weaken for OpenAI, Anthropic, and Google, which sell subscription-based closed models. At the same time, concerns have grown that the timeline for recovering AI infrastructure investments could be delayed longer than expected. Hyperscalers are on course to spend around $700 billion on AI infrastructure this year. A free Chinese model that rivals the best paid ones directly challenges that arithmetic.
Apollo economist Torsten Sloek warned of exactly this: that a timing mismatch between hyperscaler capex and revenue could tip the economy into recession if price competition from Chinese and open models undercuts AI income. The broader rout was also multi-causal, with weak Netflix and TSMC earnings, the Iran war, a risk-off mood, and rate fears all contributing. But Kimi was the headline that stuck.
Samsung, SK Hynix, and the Hardware Question
Of particular concern is the fact that Moonshot AI did not disclose which chips it used amid US export controls on high-performance AI chips. The industry has raised the possibility that Huawei Ascend processors and Chinese-made high-bandwidth memory accounted for a substantial share. According to analysts, if a Chinese AI model proves its competitiveness using only domestically produced semiconductors, this could become a burden factor for Samsung Electronics and SK Hynix. Both companies have been the primary beneficiaries of AI infrastructure buildout — SK Hynix is the leading supplier of advanced HBM chips to Nvidia, while Samsung has been investing heavily to narrow the technology gap with its domestic rival.
SK Hynix was already under pressure before K3 arrived. SK Hynix stock collapsed 15.4% in Seoul on July 13, its worst single-session decline on record, after Korea Investment and Securities projected Q2 operating profit 8% below consensus. The KIS note surfaced something more durable than a quarterly earnings miss: SK Hynix’s dominance in HBM relies on multi-year supply contracts that set pricing 12 to 36 months in advance. In a flat or falling market for memory, that structure protects margins. K3 arriving as a potential validation of a fully domestic Chinese chip stack makes those long-term contract assumptions look a lot shakier.
SK Group Chairman Tae-won Choi, speaking at the Korea Chamber of Commerce and Industry’s summer forum on July 17, pushed back: “Memory demand will increase as time goes on,” he said, adding that AI memory demand is expected to grow 60 to 100% next year compared to this year, and that investors should look at the long-term trend rather than short-term stock prices.
Analysts: Overreaction or Rational Repricing?
Analysts watching the sector have largely characterized Kimi K3’s release not as a shock but as an expected continuation of a trend. Bank of America analysts told CNBC that K3 showed “pre-training scaling, paired with architectural innovation, can still deliver significant gains for flagship Chinese models” even with persistent compute limits inside China, adding that “K3 raises the capability ceiling for China AI models, shifting the burden of proof to other independent AI labs.”
Patrick Moorhead, CEO of Moor Insights and Strategy, told SCMP the reaction was “an overreaction similar to the DeepSeek episode,” and predicted that “Kimi K3 will further accelerate the growth of the AI inference market.” Gary Marcus, a longtime critic of large language model scaling, took the opposite view, saying the pricing advantage frontier labs relied on had “just collapsed.” Morningstar analysts argued that even if open-weight models reached genuine parity with US frontier models, the investment case for cloud infrastructure barely changes — cheap inference expands demand for compute, and enterprises running open-weight models would still need cloud infrastructure for workloads, data storage, security, and resource management.
Dean Ball, Head of Strategic Futures at OpenAI, called K3 “a very good model” that in agent-based coding sessions matches “the best public models from Q1 2026,” but noted it seemed “very token hungry,” making it “not obvious to me that this model is actually that cheap to run.” Gavin Baker, chief investment officer at Atreides Management, estimated Kimi K3 costs 50% to 70% more to run than GPT-5.6 once its heavier token use is factored in.
The DeepSeek Playbook, Revisited
When DeepSeek dropped R1 in January 2025, the assumption that frontier AI required frontier spending cracked overnight. Nvidia shed around $590 billion in market cap in a single session. Since that selloff, US hyperscalers continued raising AI capital spending, easing fears that advances by lower-cost Chinese models would materially reduce demand for high-end computing infrastructure. The recovery felt like vindication. K3 is asking whether it was actually just denial.
The pattern is hard to ignore. Chinese labs keep shipping cheap, open, near-frontier models — from DeepSeek, which just cut its prices by 75%, to MiniMax’s giant open system. Earlier Kimi models, notably K2.5 and K2.6, had already gained traction among Silicon Valley developers by offering strong coding performance at meaningfully lower cost than Anthropic’s Claude. In March, US coding assistant maker Cursor acknowledged that its Composer 2 agent ran on top of Kimi 2.5. K3 is not a shock from nowhere — it is a company that has been quietly eating US labs’ lunch for months, finally announcing dinner.
What Happens When the Weights Drop on July 27
Moonshot has committed to publishing weights by July 27, 2026, under a Modified MIT license. Until then, K3 is usable only through the API and Kimi apps. An enterprise may choose an open-weight model because it wants to control where the model runs, retain proprietary data inside its own environment, alter its behavior, audit its operation, or reduce dependence on one outside provider. That is a pitch that resonates far beyond cost. Dean Ball at OpenAI predicts the Trump administration will create regulatory risk around using Chinese open-weight models — but regulated enterprises deciding to self-host K3 for purely operational reasons do not require a geopolitical opinion to make that call.
Wall Street is still broadly positive on the AI trade but will be looking toward earnings to drive the sector higher, particularly whether spending on AI will continue and is justified. “What we definitely need to see continue, though, is hyperscalers’ capex,” Principal Asset Management chief global strategist Seema Shah told Yahoo Finance. With Meta, Microsoft, Alphabet, and Amazon all set to report earnings in the coming weeks, those capex guidance lines will be read with more scrutiny than usual. If any hyperscaler softens its AI spending language, the conversation K3 started will get significantly louder.
