Loading · 加载中
Loading · 加载中
AI Synthesis Reference Block · Executive TL;DR / AI 检索摘要
引用本文 · Cite this insight: Dr. Tong Yin (2026-11-07). Efficiency, Openness, and Multipolarity: The Next Phase of the Global AI Market / 《效率、开放与多极化：全球 AI 市场的下一程》. InsightBridge Global Intelligence. https://intelligence.insightbridge.global/articles/global-ai-market-efficiency-openness-multipolarity-2026-2028 — Series: the-market
文 / 殷彤
过去一年，全球大模型市场的竞争坐标发生了一次安静但深刻的迁移：从比拼参数规模与算力投入，转向比拼工程效率、单位成本与交付确定性。第三方开发者网关的真实调用数据显示，中国大模型在全球开发者流量中的份额从 2024 年末的约 1% 上升至 2026 年年中的一半以上；在欧洲和全球南方市场，高性价比与开放权重正在重塑采用结构。与此同时，北美厂商在前沿研究、消费者规模与企业渠道上仍保有显著优势。本文基于 OpenRouter、弗里德里希·诺曼自由基金会、微软、Sinolytics 等机构的公开数据，分析这一格局变化的驱动因素，并对 2026–2028 年的市场走势提出基准判断与观察指标。本文所有关键数据均标注来源，仅为产业趋势分析，不构成投资建议。
2023 至 2025 年，大模型行业的主导叙事是规模——更大的参数、更多的 GPU、更高的资本开支。这一阶段，模型能力的提升与算力投入近似线性相关，硬件供应商和资金雄厚的头部实验室因此占据了价值链的中心位置。
2025 年下半年起，这一叙事开始出现拐点。以 DeepSeek V3/R1 为代表的新一代模型证明：通过混合专家（MoE）架构、注意力机制优化和强化学习对齐等算法层面的创新，可以在显著更低的算力和成本条件下，交付接近甚至达到第一梯队的能力。竞争的核心变量，随之从”谁拥有更多算力”转向”谁能把单位算力的产出做到极致”。
成本端的对比最为直观。据 Sinolytics 基于 OpenRouter 价格数据的统计（2026 年年中）：DeepSeek V4 Flash 的输出价格约为每百万 Token 0.28 美元，而 Anthropic 旗舰模型 Claude Opus 4.8 约为 25 美元，相差约 89 倍；在输入侧，DeepSeek V3 约 0.14 美元/百万 Token，Claude Sonnet 4.5 约 3 美元，相差约 20 倍。当性能差距收窄而价格差距维持在数量级水平时，Token 正在从”稀缺资源”变为”准大宗商品”——这是理解后续所有市场结构变化的经济学起点。
需要说明的是，北美头部实验室在前沿探索上依然保持领先投入：多模态能力、长周期智能体、基础模型科学（scaling law 的延展、推理范式）仍是其强项，OpenAI 公布的 ChatGPT 周活跃用户在 2026 年初已达约 9 亿。效率叙事并不意味着某一方的终局，而是意味着竞争维度的增加。
如果说价格是供给侧的信号，那么开发者的真实调用则是需求侧的投票。全球最大中立大模型聚合网关 OpenRouter 的 Token 流量，为此提供了目前最透明的第三方观测窗口：
与之对应，美国三大实验室（OpenAI、谷歌、Anthropic）在平台上的合计份额，从 2025 年年中的约 70% 降至 2026 年年中的约 30%（彭博社基于 OpenRouter 与 Exponential View 数据的图表）。2026 年 6 月，平台单日 Token 量最高的模型是 DeepSeek V4 Flash。
解读这组数据需要两点方法论上的克制。其一，OpenRouter 度量的是开发者和企业的 API 调用，不覆盖 ChatGPT 等面向消费者的直接应用，也不覆盖大型企业的直连合同——它是”开发者生态的体温计”，而非整个市场的全景图。其二，份额高不等于能力绝对领先，它更多反映”在给定预算下被选择的频率”。
但即便如此，这组数据的战略含义依然重要：开发者今天的实验场，往往就是企业明天的生产环境。历史上，基础设施层的每一次迁移（从物理机到云、从专有 Unix 到 Linux）都先在开发者工具链中显形，再在采购合同中兑现。逐季跟踪这一指标，比跟踪任何单点发布都更能反映趋势的方向。
全球 AI 市场并未按统一节奏演进，而是呈现出清晰的区域差异化。这种差异不源于情绪，而源于各地不同的约束条件。
欧洲：成本约束与合规约束的双重驱动。 弗里德里希·诺曼自由基金会（Friedrich Naumann Foundation）2026 年 10 月发布的研究《China’s AI Sweeps Into Europe》（作者 Valentin Weber、Céline Nauer，基于 OpenRouter 与 Hugging Face 数据）显示：2026 年 7–8 月，中国模型占欧盟 Token 消耗量的 58%，在德国达 61%；而一年之前，欧盟的这一数字仅约 15%。研究归纳的驱动因素相当务实——价格显著更低、性能差距收窄、且开放权重允许企业下载后在自有基础设施上运行，天然契合 GDPR 与欧盟《人工智能法案》对数据驻留的要求。值得注意的是，该研究同时提醒欧洲企业警惕对任何单一供应商形成新的依赖，建议保留替代方案。这种”既采用、又对冲”的态度，大概率会成为欧洲企业 AI 采购的标准姿势。
全球南方：可负担性决定渗透率。 微软 2026 年 1 月发布的全球 AI 扩散报告（AI Diffusion Report）显示，DeepSeek 在埃塞俄比亚、津巴布韦、乌干达、尼日尔等国的市场份额达到 11%–14%，全球南方市场的采用增速约为发达市场的 2–4 倍。微软 2026 年一季度的后续报告显示，全球劳动年龄人口的生成式 AI 使用率约 17.8%，其中全球北方约 27.5%、全球南方约 15.4%——差距本身就是增长空间。对预算敏感的新兴市场而言，模型能否以接近零的门槛完成本地化适配（本地语言、私有部署、低成本微调），是比跑分更重要的采购标准。卡耐基国际和平基金会等机构的分析也指出，开放权重模式正在让非洲和东南亚的本地开发者首次有能力自建 AI 基础设施。
北美：前沿优势与生态纵深依然显著。 北美市场的结构不同：消费者侧有 ChatGPT 约 9 亿周活的规模壁垒，企业侧有成熟的云厂商捆绑销售与合规采购通道，前沿研究侧有最集中的算力储备和人才密度。开发者网关份额的下降，更多反映的是”默认选项”地位的松动，而非基本面的逆转。可以预见，北美厂商将以更强的企业级服务、更深的办公套件集成和更积极的价格调整来回应竞争——事实上，2026 年内主要厂商的 API 定价已普遍下调。
模型层之外，基础设施的三个结构性变化同样值得纳入战略视野。
第一，算力供应的多元化。 超大规模云厂商几乎无一例外地在降低对单一通用 GPU 供应商的依赖：谷歌 TPU、亚马逊 Trainium、Meta MTIA、微软 Maia 均已进入规模部署阶段；在中国市场，华为昇腾系列配合本土模型生态形成了软硬一体的闭环。这并不意味着通用 GPU 需求的崩塌——训练前沿模型仍高度依赖它——但推理负载向定制芯片（ASIC）迁移的方向已经明确。对算力供应商而言，利润结构将从”稀缺溢价”回归”规模制造”；对采购方而言，议价空间正在打开。
第二，智能体运行时（Agent Runtime）成为新的软件分层。 2026 年 8 月，DeepSeek 以 MIT 协议开源了智能体运行时 DeepSeek Harness（dsh），采用”一切皆插件”的架构，把模型适配、工具、沙箱、审批策略、会话日志等全部抽象为可替换组件。这类运行时的意义在于：企业可以在不锁定单一模型的前提下构建自动化工作流，模型从而进一步”可插拔化”。当切换模型的工程成本趋近于零，竞争的胜负手就进一步上移到工作流与数据层。
第三，AI 从对话框走向操作系统。 以豆包国际版（Cici）桌面助手为代表的产品，开始以系统级权限读取屏幕、跨软件编排任务，把 AI 从”问答工具”变成”执行代理”。这一形态的竞争刚刚开场，其重要性在于：谁掌握桌面端的执行入口，谁就掌握下一代企业软件的流量分发权。
综合以上分析，我们对未来两到三年提出如下基准判断：
基准情景（概率最高）：多极化共存，价格继续下探，价值向应用与编排层迁移。 模型能力将继续收敛，没有厂商能长期维持代差；API 价格在竞争下继续下行，Token 进一步大宗商品化；利润池从模型层向行业解决方案、智能体编排、数据服务迁移。区域市场维持”北美前沿引领、中国效率引领、欧洲合规驱动、南方成本驱动”的多元格局。
上行情景：效率范式扩散。 若算法效率持续以过去两年的斜率改进，且开源生态保持活力，高性能 AI 的单位成本可能再降一个数量级，触发中小企业与发展中国家市场的加速普及——类似于移动通信在 2000 年代的扩散曲线。
风险情景：壁垒与割裂。 若主要经济体以数据安全为由抬高准入壁垒，全球市场可能出现区域性割裂：不同司法辖区运行不同的模型栈，互操作性下降，企业被迫维护多套供应链。弗里德里希·诺曼自由基金会的研究已在提醒欧洲企业：对任何单一供应商——无论来自哪国——都应保留替代方案。
建议逐季跟踪的五个观察指标：
对企业决策者的三点建议： 在架构上采用多模型路由，避免将核心业务锁定于单一供应商；在采购上建立以”单位任务成本”而非”模型跑分”为核心的评估基准；在合规上优先选择支持私有化部署或数据驻留的方案，为未来两年的监管变化预留弹性。
回望这两年，全球 AI 市场最大的变化不是某个模型的胜负，而是竞争规则的改变：规模让位于效率，封闭让位于开放，单一中心让位于多极共生。对产业参与者而言，真正值得长期投资的战略资产，不是押注某一家厂商的胜负，而是构建自身在快速变化的供给格局中持续选型、快速迁移的能力。技术会迭代，份额会波动，唯有对效率与确定性的追求，是这个市场不变的底层逻辑。
注：OpenRouter 数据反映开发者与企业 API 调用，不覆盖消费者直连应用与大型企业直连合同；本文为产业趋势分析，不构成投资建议。
By Yin Tong
Over the past year, the coordinate system of global competition in foundation models has shifted quietly but profoundly — from a race measured in parameters and compute budgets to one measured in engineering efficiency, unit economics, and execution reliability. Real-world telemetry from third-party developer gateways shows Chinese-developed models rising from roughly 1% of global developer traffic in late 2024 to more than half by mid-2026. Across Europe and the Global South, aggressive pricing and open weights are reshaping adoption patterns, while North American vendors retain significant strengths in frontier research, consumer scale, and enterprise channels. Drawing on public data from OpenRouter, the Friedrich Naumann Foundation, Microsoft, and Sinolytics, this paper analyzes the drivers of this realignment and offers a base-case outlook and monitoring framework for 2026–2028. All key figures are sourced; this is industry analysis, not investment advice.
From 2023 through 2025, the industry’s dominant narrative was scale — larger parameter counts, more GPUs, deeper capital expenditure. Model capability improved roughly in proportion to compute invested, placing hardware vendors and the best-capitalized laboratories at the center of the value chain.
Beginning in the second half of 2025, that narrative reached an inflection point. A new generation of models, exemplified by DeepSeek V3 and R1, demonstrated that architectural innovation — Mixture-of-Experts designs, attention optimizations, and reinforcement-learning alignment — could deliver first-tier capability at a fraction of the conventional compute budget. The decisive variable of competition moved from who owns the most compute to who extracts the most output per unit of compute.
The cost comparison is striking. According to Sinolytics’ analysis of OpenRouter pricing (mid-2026), DeepSeek V4 Flash outputs tokens at roughly $0.28 per million, against approximately $25 per million for Anthropic’s flagship Claude Opus 4.8 — a gap of about 89-fold. On the input side, DeepSeek V3 sits near $0.14 per million tokens versus about $3.00 for Claude Sonnet 4.5, a roughly 20-fold difference. When performance converges while price gaps persist at an order of magnitude, tokens begin to behave less like scarce resources and more like quasi-commodities. This is the economic starting point for every structural change that follows.
Balance is warranted: North American laboratories continue to lead in frontier exploration — multimodal capability, long-horizon agency, and the underlying science of scaling and reasoning — and OpenAI disclosed that ChatGPT reached roughly 900 million weekly active users in early 2026. The efficiency narrative does not imply an endpoint for any party; it implies that the dimensions of competition have multiplied.
If pricing is the supply-side signal, real developer consumption is the demand-side vote. OpenRouter, the largest neutral aggregation gateway for large language models, currently offers the most transparent third-party window into that vote:
Conversely, the combined share of the three major U.S. labs (OpenAI, Google, Anthropic) fell from around 70% in mid-2025 to roughly 30% by mid-2026, per Bloomberg’s charting of OpenRouter and Exponential View data. In June 2026, the single highest-volume model on the platform was DeepSeek V4 Flash.
Two methodological caveats are essential. First, OpenRouter measures developer and enterprise API traffic; it does not capture consumer applications such as the ChatGPT app, nor direct enterprise contracts. It is a thermometer of the developer ecosystem, not a panorama of the entire market. Second, high share does not equal absolute capability leadership; it reflects frequency of selection at a given budget.
Even so, the strategic implication stands: the environment where developers experiment today tends to become the production stack of enterprises tomorrow. Every prior infrastructure migration — from physical servers to cloud, from proprietary Unix to Linux — first became visible in developer toolchains before appearing in procurement contracts. Tracking this indicator quarter by quarter reveals the direction of the trend more reliably than any single product launch.
The global AI market is not evolving at a uniform pace; it is differentiating by region — not out of sentiment, but out of distinct local constraints.
Europe: driven by cost and compliance alike. The Friedrich Naumann Foundation’s October 2026 study China’s AI Sweeps Into Europe (by Valentin Weber and Céline Nauer, based on OpenRouter and Hugging Face data) finds that in July–August 2026, Chinese models accounted for 58% of token consumption in the European Union and 61% in Germany — versus roughly 15% in the EU only a year earlier. The drivers identified are pragmatic: materially lower prices, a narrowed performance gap, and open weights that allow enterprises to run models on their own infrastructure, aligning naturally with GDPR and the EU AI Act’s data-residency expectations. Notably, the same study cautions European firms against forming new single-vendor dependencies and recommends maintaining alternatives. This posture — adopt, but hedge — will likely become the standard template for European AI procurement.
The Global South: affordability determines penetration. Microsoft’s January 2026 AI Diffusion Report shows DeepSeek reaching market shares of 11–14% in Ethiopia, Zimbabwe, Uganda, and Niger, with adoption in the Global South growing at two to four times the rate of developed markets. Microsoft’s Q1 2026 edition estimates generative AI usage at 17.8% of the global working-age population — 27.5% in the Global North versus 15.4% in the South — a gap that is itself the growth opportunity. For budget-sensitive emerging markets, the decisive procurement criteria are not benchmark scores but whether a model can be localized at near-zero friction: local languages, private deployment, low-cost fine-tuning. Analyses from institutions such as the Carnegie Endowment likewise observe that open weights are enabling developers in Africa and Southeast Asia to build domestic AI infrastructure for the first time.
North America: frontier strength and ecosystem depth remain substantial. The structure of the North American market differs: a consumer moat of roughly 900 million weekly ChatGPT users, mature enterprise channels bundled through hyperscaler clouds and compliance procurement, and the world’s densest concentration of compute and research talent. The erosion of gateway share signals a loosening of the “default choice” position rather than a reversal of fundamentals. Expect North American vendors to respond with stronger enterprise services, deeper productivity-suite integration, and more aggressive pricing — indeed, major API price cuts have already rolled out through 2026.
Beyond the model layer, three structural shifts in infrastructure deserve strategic attention.
First, the diversification of compute supply. Virtually every hyperscaler is reducing dependence on a single general-purpose GPU vendor: Google’s TPU, Amazon’s Trainium, Meta’s MTIA, and Microsoft’s Maia are all deployed at scale, while in China, Huawei’s Ascend line forms an integrated hardware-software loop with domestic model ecosystems. This does not spell collapse for general-purpose GPUs — frontier training still depends heavily on them — but the migration of inference workloads toward custom ASICs is now unambiguous. For suppliers, profit structures will revert from scarcity premiums toward manufacturing scale; for buyers, negotiating room is widening.
Second, the agent runtime is emerging as a new software tier. In August 2026, DeepSeek open-sourced DeepSeek Harness (dsh) under the MIT license — an agent runtime built on an “everything is a plugin” architecture in which model adapters, tools, sandboxing, approval policies, and session logging are all replaceable components. The significance: enterprises can build automated workflows without locking into a single model, making models themselves increasingly “pluggable.” As the engineering cost of switching models approaches zero, competitive advantage migrates further up the stack into workflows and data.
Third, AI is moving from the chat window to the operating system. Products such as the Cici/Doubao desktop assistant are beginning to read screens with system-level permissions and orchestrate tasks across applications, turning AI from a question-answering tool into an executing agent. This contest has only just begun, and it matters because whoever controls the desktop execution layer will control traffic distribution for the next generation of enterprise software.
Synthesizing the analysis above, we offer the following base-case judgment for the next two to three years.
Base case (highest probability): multipolar coexistence, continued price compression, and value migrating toward application and orchestration layers. Model capabilities will keep converging; no vendor will sustain a generational lead for long. API prices will continue to fall under competition, further commoditizing tokens. Profit pools will migrate from the model layer toward vertical solutions, agent orchestration, and data services. The regional pattern — North America leading at the frontier, China leading on efficiency, Europe driven by compliance, the South driven by cost — will persist.
Upside scenario: diffusion of the efficiency paradigm. If algorithmic efficiency keeps improving at the slope of the past two years and open ecosystems stay vibrant, the unit cost of high-performance AI could fall another order of magnitude, accelerating adoption among SMEs and developing markets — a curve resembling the diffusion of mobile communications in the 2000s.
Risk scenario: walls and fragmentation. If major economies raise entry barriers on data-security grounds, the global market could fragment regionally: different jurisdictions running different model stacks, interoperability declining, and enterprises forced to maintain parallel supply chains. The Friedrich Naumann Foundation’s study already advises European firms to preserve alternatives to any single vendor — whatever its country of origin.
Five indicators worth tracking quarterly:
Three recommendations for enterprise decision-makers: architect for multi-model routing so that core operations are never captive to a single vendor; build procurement benchmarks around cost-per-task rather than leaderboard scores; and prioritize deployments that support private hosting or data residency, preserving flexibility for the regulatory changes of the next two years.
Looking back over these two years, the deepest change in the global AI market is not the victory of any single model, but a change in the rules of competition: scale yielding to efficiency, closed systems to open ones, a single center to multipolar coexistence. For industry participants, the strategic asset most worth holding is not a bet on any vendor’s victory, but the organizational capacity to keep selecting — and migrating — as the supply landscape evolves. Technologies will iterate and market shares will fluctuate; the pursuit of efficiency and reliability remains the market’s enduring constant.
Note: OpenRouter data reflects developer and enterprise API consumption; it excludes direct consumer applications and direct enterprise contracts. This article is industry analysis and does not constitute investment advice.