What actually happened - and what the headline leaves out
Moonshot AI presented Kimi K3 on July 16 and 17, 2026 at the World AI Conference in Shanghai: 2.8 trillion parameters, the largest open-weight language model in the world, ahead of Western frontier models on some early coding benchmarks. In parallel, semiconductor stocks fell worldwide - Taiwan around 6 percent, Japan around 4 percent, the Philadelphia Semiconductor Index more than 20 percent below its June high. The causal link between the two events was established in headlines within hours.
Trace the price movements back, however, and that chain has a problem. The selloff didn't begin on July 17 but around June 22-25, 2026 - a good three weeks before the model shipped. As early as July 7, nine days before Kimi K3, financial media reported that Micron, Samsung and SK Hynix had dragged memory stocks into a bear market. The drivers cited then were a price explosion in memory chips and general valuation concerns, not a Chinese language model that didn't yet exist.
A second point weighs even heavier: in the same week that semiconductor stocks supposedly collapsed on falling AI demand, two of the most important suppliers raised their guidance. TSMC lifted its capital spending forecast to $60-64bn, ASML its revenue outlook to EUR 43-45bn. Both are statements from companies sitting further up the supply chain than any share price move - and both point the opposite way. Several observers additionally noted that July 17 was overdetermined anyway: weak quarterly results, geopolitical tension and interest rate concerns landed on the same day.
None of this means Kimi K3 played no role. But the more defensible phrasing is: the model met an already-running selloff and amplified it rather than causing it. For judging the news, that is a substantial difference.
The market had long since grown used to Chinese frontier models
A second test of the story is even more revealing: if a strong Chinese model reliably triggers a semiconductor crash, that should show up in the preceding cases. It doesn't. Alibaba's Qwen3 and Qwen3-Max (2025) mainly pushed Alibaba's own share price up, with no demonstrable US chip selloff. Moonshot's own Kimi K2 (July 2025) and Kimi K2 Thinking (November 2025) - the latter with benchmark scores above Western frontier models in individual disciplines - had practically no market effect. Zhipu's GLM-5 and MiniMax M2.5 (February 2026) drove their own vendors' shares up rather than Western ones down.
The clearest case is DeepSeek V4 in February 2026: the successor to precisely the model that had triggered the original shock a year earlier was described in coverage as barely market-moving - the market had grown used to cheap Chinese models. Across a year and a half, then, a clear desensitization is visible. That Kimi K3 of all things triggered a sharp reaction again cannot be explained by model quality alone. The likelier reading is that an already fragile sentiment needed an occasion.
That yields a first practical insight beyond this single case: the severity of a market reaction to an AI news item is a poor indicator of that item's technical or economic significance. Mostly it measures how nervous the market already was when the news landed.
The same bet was already placed 18 months ago
The real advantage of this situation is that it isn't new territory. On January 27, 2025, Nvidia fell roughly 17 percent after the release of DeepSeek R1, losing about $589bn in market value in a single trading day - the largest one-day loss by a single company in stock market history. The thesis was identical to today's: efficient open models make part of the planned data center investment unnecessary. Unlike in July 2026, though, that thesis no longer has to be debated - it can be evaluated.
The result is unambiguous. Capital spending by the four large hyperscalers - Microsoft, Alphabet, Amazon and Meta - rose from a combined $226bn in 2024 to roughly $410bn in 2025. For 2026, announced budgets total approximately $725bn, an increase of about 77 percent. Individually: Microsoft's guidance for fiscal 2026 sits at around $190bn, Amazon named $200bn for 2026, Alphabet $180-190bn, Meta $125-145bn. Not one of these companies cut its budget citing DeepSeek - no such announcement could be found in the research.
Microsoft CEO Satya Nadella articulated the counter-thesis on the day of the drop itself: falling cost per use historically led not to less consumption but to more - the Jevons paradox. Whether one accepts that explanation is secondary. What matters is that the numbers of the following 18 months fit it, and not the market's scarcity thesis. Anyone betting in July 2026 on falling AI infrastructure demand is betting on a relationship that failed to materialize the last time it was tested.
The finding that hits the premise directly: prices aren't falling where you buy
Beneath the whole debate lies an assumption that is rarely examined: that AI keeps getting cheaper. In one specific reading, that's true. The venture firm a16z coined the term LLMflation for it, showing that the cost of a constantly held quality level falls by roughly a factor of ten per year - GPT-3-level quality cost about one thousandth of its original price within three years.
But no company buys the constant quality of two years ago. What gets bought is the current model in a performance tier - and there the picture looks different. A comparison of official list prices between early 2025 and mid-2026, per million input and output tokens:
- Budget tier, Anthropic: Claude Haiku 3.5 cost $0.80/$4; its successor generation Haiku 4.5 costs $1/$5 - up.
- Budget tier, Google: Gemini 1.5 Flash cost $0.075/$0.30; Gemini 2.5 Flash costs $0.30/$2.50 - up substantially.
- Frontier tier, OpenAI: GPT-4o sat at $2.50/$10 in early 2025; the current flagship is $5/$30 - up.
- Counter-example, Anthropic frontier tier: Claude Opus fell from $15/$75 to $5/$25 - down substantially.
- Actual price cuts on an existing model remained the exception: OpenAI cut the price of o3 by 80 percent in June 2025 without changing the model - a documented one-off, not a pattern.
The picture is mixed, and that is exactly the point: there is no reliable downward trend in the prices a company pays for current capability. On top of that comes an effect invisible in price tables: modern reasoning models consume considerably more tokens for the same task than earlier generations, because they produce intermediate steps. Falling prices per token and rising consumption per task can cancel each other out - no solid public figures exist on the net effect, which is why no quantification appears here.
And even if: the model is the smallest line item in your math
Suppose for a moment the market logic held and models did become ten times cheaper overnight. What would that mean for a concrete mid-market project? Considerably less than the excitement suggests - because the model price is rarely the dominant cost block there.
Boston Consulting Group summarizes the effort distribution of an AI transformation in a widely cited rule of thumb: roughly 10 percent goes to algorithms and models, 20 percent to technology and data, and 70 percent to people and processes. At the macroeconomic level the same order of magnitude shows up: in IDC's 2026 market forecasts, AI services account for roughly $589bn and AI software for roughly $453bn - the AI models category for about $26bn. That's roughly $40 of services and software for every dollar of model spend.
Research on failed projects points the same way. A 2025 MIT study covering more than 300 evaluated deployments found that only about 5 percent of pilot projects delivered measurable financial value - and the authors explicitly named the cause as missing process integration, not model quality. Gartner puts the share of generative AI projects abandoned after proof of concept at at least 30 percent; the list of reasons includes data quality, risk controls and unclear business value - the model price does not appear.
This renders the market news nearly meaningless for a concrete initiative. If 70 percent of the effort sits in processes and people and about 10 percent in the model, then even a spectacularly cheaper model shifts the total by a few percentage points. What genuinely moves the math appears in no market report: whether the use case is cleanly scoped, whether the data is accessible, whether someone in operations owns the result.
Where the rule inverts - and how to tell you're that case
This assessment has a clear exception, and it is well documented. In June 2026 the AI agent company Lindy migrated all of its traffic from a Western vendor to a Chinese model, cutting inference costs by roughly 90 percent - millions of dollars by its own account. The decisive detail from the founder: the company's inference spend had exceeded its personnel costs.
That's exactly where the line runs. The model price dominates total cost when inference is the product itself - for AI product vendors with very high token volume, where every end customer continuously generates model calls. A mid-sized company with an internal assistant, document processing or support automation is structurally the other case: one-time integration costs in the five-figure range, ongoing model costs often in the double to low triple digits per month.
Two details of that same case are rarely quoted alongside it. First, the founder described the migration as a hundred times more work than expected, with a months-long evaluation phase - so the switching costs landed precisely in the block that already dominates project cost. Second, he said he would switch back as soon as the previous vendor cut prices. Even in the textbook case for model-price dominance, the attachment to the cheap model is weak.
The overlooked finding: cheap models win volume, not budget
One data point from the research deserves more attention than it gets, because it shows how cheap models are actually deployed in practice. On Vercel's production gateway, open models accounted for roughly 29 percent of token volume in July 2026 - but under 4 percent of spend. On OpenRouter, Chinese models reached a record share of token volume among US firms in July 2026. At the same time, enterprise revenue per the Menlo Ventures survey stayed clearly with Western vendors: Anthropic 40 percent, OpenAI 27 percent, Google 21 percent of spend.
This gap between volume share and spend share is the genuinely useful information, and it contradicts the displacement story. Companies aren't replacing their expensive models with cheap ones. They're tiering: high-volume, low-stakes work - classifying, pre-sorting, summarizing, extracting - moves to cheap models, while the tasks carrying real business risk stay on the expensive ones. That easily explains why volume shares can flip without the spend distribution following.
For your own planning, that's the most usable takeaway from the entire Kimi K3 debate: the relevant question isn't whether a new cheap model replaces the incumbent, but which steps of a workflow need an expensive model at all. In practice that separation usually saves more than any vendor switch - and it holds regardless of which model makes headlines next week.
What this means for your decision
The five preceding sections point to a course of action noticeably different from the headline situation:
- Don't treat market reactions to AI news as expert judgment: the selloff began before the model, and suppliers like TSMC and ASML raised guidance in the same week. Price moves measure sentiment, not technical substance.
- Don't infer falling project costs from cheaper models: per BCG's rule of thumb, roughly 70 percent of effort sits with people and processes and only about 10 percent with the model. The lever is almost never where the news points.
- Test the pricing assumption against your own performance tier: AI is not getting cheaper across the board - list prices for Claude Haiku and Gemini Flash rose between generations. Calculate with the price of the model you would actually deploy.
- Ask about tiering before asking about switching vendors: which steps of your workflow genuinely need a frontier model, and which don't? That's the lever the volume-versus-spend gap exposes - and it works without a migration project.
- Check whether you're the exception at all: only if your inference costs approach the order of magnitude of your personnel costs does a model migration justify its own project. For an internal assistant, that is practically never the case.
None of this replaces a calculation for a specific initiative - that requires actual volume assumptions. What can be said clearly: anyone steering AI decisions by model headlines is reliably optimizing the smallest line item in their math. The question that decides economic viability in 2026 as in 2025 is less spectacular and much older: is the use case scoped tightly enough that the effort behind it pays off at all?