Taboga's two-event comparison helps identify which AI exposures a trader should worry about after a Chinese model release. Its most useful table separates the losers into two groups. The returns support that distinction. They cannot establish the stronger claim that complementary assets protected incumbents, a limit the author acknowledges in several places.
The distinction starts with what a release tells the market. DeepSeek R1 in January 2025 suggested a change in the production function. Its API prices were $0.55 input and $2.19 output per million tokens, compared with $15.00 and $60.00 for OpenAI's o1-preview. The release also reported a base-model training cost of US$5.576 million. That self-reported figure excluded prior research and ablations, covered only the first stages of training, and was disputed by later analyses. The Stanford AI Index estimates compute cost for the GPT-4 training run at about US$79 million; those figures cover different stages. If frontier reasoning costs that little, projected chip demand looks too high. Compute suppliers should take the hit.
A release can instead announce another credible supplier at the prevailing price. Kimi K3 in July 2026 ranked third on the Artificial Analysis Intelligence Index with a score of 57. The index combines reasoning, coding and knowledge evaluations. Kimi's prices were $3.00 and $15.00. That kind of entry should hurt labs selling the same product. Taboga argues that open weights shift rents toward owners of the assets around the model: the orchestration harness, tools, retrieval, serving capacity, contracts, switching costs and proprietary data. As he puts it, "what users purchase is not a set of weights but a service, and the weights are only one of its inputs."
His evidence comes in four parts. The return table records raw close-to-close moves for 18 indices, stocks and futures, using one window per event drawn from market commentary. Alongside it are model cost and speed comparisons, a parallel with 1990s open-source software, and usage and revenue figures after the events. There is no abnormal-return model or significance testing. Taboga describes the result as "an interpretation of the observed pattern rather than a causal finding." He also explicitly calls the patterns "descriptive rather than causal evidence."
The case for a moat therefore rests on evidence beyond returns. From May to September 2025, OpenRouter data showed closed models accounting for about 80% of tokens at about six times the open-model price. Moonshot stopped accepting new Kimi subscriptions when demand outran serving capacity. Together with the 1990s open-source precedent, these observations give switching costs and serving capacity a credible role. Their durability is harder to defend: the paper's later figure shows top-four US closed models down to 23% of OpenRouter traffic by July 2026.
Similar index losses, different casualties
The index moves were almost identical. The S&P 500 lost 1.5% in both windows. The Nasdaq 100 declined 3.0% in January 2025 and 3.1% in July 2026; DS World Technology lost 4.2% and 4.4%. A book hedged against tech beta faced nearly the same index decline each time.
The losses beneath those indices were quite different. NVIDIA dropped 17.0% in the DeepSeek window, against 4.6% in the Kimi window. SOX lost 9.1% and then 5.8%. July's biggest casualties were Chinese pure-play labs: Zhipu AI declined 35.1% and MiniMax Group 17.1%. No sector index captures the Zhipu and MiniMax losses.
Did Kimi K3 deliver a cost shock?
No. This is the paper's most reusable distinction. Kimi charged half as much per output token as GPT-5.6 Sol ($15 against $30), yet running the full evaluation suite cost only 29.2% less ($2,437.41 against $3,442.81). Its verbosity explains the gap: Kimi produced about 130 million output tokens on the suite, compared with a median of about 63 million across models evaluated by Artificial Analysis. Running that suite on Kimi cost 3.8 times as much as on Grok 4.5.
Kimi was also the slowest model in the comparison. It generated 32.9 output tokens per second, against 68.7 for GPT-5.6 Sol and 176.8 for GLM-5.2. July therefore offered little news through the compute channel. For the next release, a screen should price a completed task rather than stop at the advertised token price.
Revenue mix could account for Alibaba and Tencent
The return-based case for protection rests heavily on the Chinese conglomerates. During the Kimi window, Alibaba lost 0.7% and Tencent 1.3%, while Xiaomi gained 3.9%. Zhipu fell 35.1%. Taboga observes that all three conglomerates escaped large losses even though their labs regularly release open-weight models (Qwen, Hy and MiMo). He attributes that protection to diversification and to the distribution, infrastructure, hardware and customer relationships through which they can monetise model capabilities.
He handles the US comparison more cautiously. Amazon, Alphabet and Microsoft own equity in private labs, and he describes their avoidance of comparably large declines as "informative, but not decisive." Those stakes sit inside much larger businesses. Cloud infrastructure, distribution, product portfolios and customer relationships may compensate for losses at the model layer.
Alibaba and Tencent deserve the same caveat. Diversification has two possible meanings here. As revenue dilution, it means a pure-play lab earns nearly all its revenue at the model layer, while a conglomerate earns only a small share there. Revenue weights alone can then explain a large decline in the former and a small decline in the latter, before a moat has any effect. As monetisation channels, diversification is itself the proposed moat. The returns fit both explanations, and the paper supplies no way to distinguish them.
The US platforms also challenge the account. Alphabet lost 6.5% and Meta 5.2%, larger declines than any of the Chinese conglomerates suffered, despite owning the bundles Taboga describes.
Two more gaps.
OpenAI and Anthropic pre-IPO futures fell 5.0% and 1.8%. These contracts trade on crypto exchanges. The author cautions that low liquidity may limit their information content and that their returns cannot be compared directly with equity returns. The table also has no DeepSeek-window observation for Zhipu or MiniMax. We therefore cannot judge how those names ordinarily move relative to the sector.
Two windows spanning one and two sessions
Taboga is explicit about other news. Both windows occurred during earnings season; July 2026 also brought semiconductor repricing and a re-escalation of the US-Iran war. He writes that these "may have amplified or offset the release-related effects; consequently, the observed declines cannot generally be interpreted even as upper bounds."
The windows have different lengths as well. From the close of 24 January to the close of 27 January 2025 covers a single trading session. From the close of 15 July to the close of 17 July 2026 covers two. None of the returns is beta-adjusted. SOX's 5.8% loss shares a window with broader semiconductor repricing. Before anyone sizes a position from Zhipu's 35.1% decline, that move needs a market model.
We could not rerun these results. Our price history ends in 2025. The paper's central comparison also requires Zhipu, MiniMax and the OpenAI and Anthropic pre-IPO futures, none of which we carry.
Compute demand kept growing as tokens got cheaper
The subsequent data hurt the January 2025 trade. Between January 2025 and July 2026, OpenRouter weekly usage climbed from 0.53 trillion to 58.00 trillion tokens. NVIDIA revenue rose from USD 44 billion in the quarter ending April 2025 to 82 billion a year later. Taboga calls this Jevons paradox. He offers no counterfactual, but the figures are enough to retire the simple proposition that cheaper models mean fewer chips.
The table also puts pressure on the switching-cost defence. Closed models from the four leading US labs (Anthropic, Google, OpenAI, xAI) accounted for 62% of OpenRouter traffic in January 2026 and 23% by July. Their absolute usage nevertheless grew from 4.61 to 13.19 trillion tokens a week over those six months. Taboga acknowledges that "user stickiness is not permanent." OpenRouter is a routing marketplace, so these shares describe the mix at one venue.
Carrying the distinction into the next release
The paper earns its place as an exposure map, with the limits that implies. When another Chinese model arrives, price the completed task first. R1 advertised output tokens at $2.19 against o1-preview's $60.00, though the paper gives no per-task cost for R1. Kimi's full-suite cost was 29.2% below GPT-5.6 Sol. A gap on the first scale directs attention to chip suppliers; a gap on the second directs it to model-layer pure-plays.
Then weight each company by its model-layer revenue share and remove sector beta before interpreting the return. I would change my view if conglomerate outperformance survived a control for that revenue share. Taboga's moat argument needs that test. This paper does not run it.