“Middle powers” 在 AI 关键指标上与美国和中国存在结构性差距。Artificial Analysis 的统计称，前 10 大 AI 模型全部来自总部在美国或中国的公司；在前 30 名中，South Korea 有“几个”模型，United Arab Emirates 也有“几个”，France 有 1 个；英国的 DeepMind 被称为“半个例外”，因为它属于 Google。文中也指出 Toronto 是重要研究中心，但并非前 30 大 LLM 提供商所在地。

差距不仅在模型榜单，还在算力集中度：美国与中国合计约占全球 AI 计算力的 87%，使“中等强国”即使拥有可用模型，也难以在前沿训练与迭代速度上追赶。Bloomberg GeoEconomics 分析师 Michael Deng 的判断是，前沿研发所需的资金与人才需求巨大，非美非中的竞争者很少能持续跟上；若 AI 模型成为未来经济的基础设施，则基础设施控制权将几乎完全落在美中手中，从而加深其他国家对两者之一的依赖，削弱 Carney 所倡导的战略灵活性。

即便存在“模型商品化”的观点，价值转向应用层，“中等强国”在应用生态上同样落后。BCG 在 2018 年的结论是：互联网公司前 20 名中有 18 家总部位于“美国西海岸”或“中国东海岸”，两条“gold coasts”长期主导互联网经济。由此推论，在这两大集群之外构建世界级 AI 应用将是上坡路；若要形成新的地缘政治集团，就必须在模型、算力与应用三条战线同时补课。

“Middle powers” are structurally behind the US and China on core AI metrics. Artificial Analysis is cited as finding that all top 10 AI models come from companies headquartered in the US or China; within the top 30, South Korea has a couple of models, the United Arab Emirates has a couple, and France has one, while the UK’s DeepMind is treated as an asterisk because it is part of Google. Toronto is noted as a significant AI research hub but not home to a top-30 LLM provider.

The gap extends beyond rankings into compute concentration: the US and China together account for about 87% of global AI computing power, leaving middle powers behind even if their models are “good enough.” Bloomberg GeoEconomics analyst Michael Deng argues that frontier development requires immense financial and talent inputs that few non-US, non-China firms can sustain; if models become foundational economic infrastructure, control will be almost entirely American and Chinese, deepening others’ dependence and reducing the strategic flexibility Carney urges.

Even under the view that models are commoditizing and value shifts to applications, middle powers still lag. BCG concluded in 2018 that 18 of the top 20 internet companies were headquartered on either the US West Coast or the east coast of China, two “gold coasts” that have dominated the internet economy. The implication is that building world-class AI applications outside these clusters will be an uphill battle, requiring coordinated effort across models, compute and apps to reduce reliance on the US or China.