Efficiency, Openness, and Multipolarity: The Next Phase of the Global AI Market

A Strategic Outlook for 2026–2028

By Yin Tong

Executive Summary

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.

I. The Shift in Competitive Coordinates: From the Scale Narrative to the Efficiency Narrative

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.

II. The Measured Migration of the Developer Ecosystem

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:

• October 2024: Chinese-developed models account for ~1.2% of platform token volume;
• March 2025 (after DeepSeek V3): above 10%;
• Q3 2025 (after Kimi K2, MiniMax and others): above 25%;
• February 2026: weekly token consumption of Chinese models surpasses that of U.S. models for the first time (~5.16 trillion vs. ~2.7 trillion);
• April 2026: ~45%; June 2026: ~61%; September 2026: four Chinese labs alone — DeepSeek, Z-AI, Tencent, and Xiaomi — carry ~52.4% of monthly tokens.

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.

Figure 2 — From scale to efficiency: unit cost becomes the core competitive variable

III. Differentiated Regional Adoption: One Product, Three Adoption Logics

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.

Figure 1 — The multipolar map: shifting share of global developer traffic

IV. De-Consolidation of the Infrastructure Layer: Chips, Runtimes, and Agents

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.

Figure 3 — Differentiated regional adoption: Europe, the Global South, and North America

V. Outlook 2026–2028: Base Case, Scenarios, and Indicators

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:

1. Regional token shares on neutral gateways such as OpenRouter (a leading indicator of developer preference);
2. Flagship API price curves (the most direct measure of efficiency competition);
3. Download composition of open-weight models on Hugging Face (demand for private deployment);
4. Enterprise adoption of multi-model routing architectures (the degree of vendor de-concentration);
5. Regulatory detail in major jurisdictions (EU AI Act enforcement, market-access policies in the U.S. and other economies).

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.

Figure 4 — Strategic monitoring dashboard: key indicators for 2026–2028

Conclusion

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.

Principal Sources

1. OpenRouter platform telemetry, with aggregation by Bloomberg / Exponential View (share milestones, Oct 2024–Sep 2026);
2. Friedrich Naumann Foundation / DGAP: China’s AI Sweeps Into Europe, Valentin Weber & Céline Nauer, October 2026 (EU 58%, Germany 61%, EU ~15% one year prior);
3. Microsoft AI Diffusion Report, January 2026 (DeepSeek shares of 11–14% across several African markets) and Q1 2026 edition (global usage 17.8%; North 27.5% vs. South 15.4%);
4. Sinolytics Radar (OpenRouter pricing: DeepSeek V4 Flash ~$0.28/M output tokens vs. Claude Opus 4.8 ~$25; DeepSeek V3 input ~$0.14 vs. Claude Sonnet 4.5 ~$3);
5. China Europe International Business School (CEIBS): How China overtook the US in AI token usage, April 2026;
6. DeepSeek Harness open-source project documentation (MIT license, released August 2026);
7. OpenAI corporate disclosures (ChatGPT weekly active users, early 2026).

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.