China’s AI Models Hit 4.69T Tokens—What It Means for OpenAI, Musk’s Space Compute

Key Data Points

China’s weekly token consumption: 4.69 trillion tokens (as of March 15), marking the second consecutive week the country surpassed the United States and now occupies the top three global slots.

JPMorgan forecast: AI inference token usage in China is projected to rise from roughly 10 quadrillion in 2025 to about 3.9 quadrillion in 2030—a five‑year increase of roughly 370×.

Why Token Volume Matters

Tokens are the smallest unit of computation for large language models; higher token counts indicate broader, more frequent real‑world applications and greater economic value. The shift from “parameter‑count races” to “token‑usage races” signals that the industry is moving toward sustainable, revenue‑generating deployments across finance, e‑commerce, gaming, and short‑video platforms.

OpenAI’s Talent Expansion

Insiders reveal that OpenAI intends to create 3,500 new positions, expanding its workforce from about 4,500 today to roughly 8,000 by the end of 2026. The commentary notes that talent depth will become a decisive factor in the next round of large‑model competition.

Musk’s "Terafab" Initiative

Elon Musk announced on X that SpaceX and Tesla will launch the "Terafab" project, targeting an annual production capacity of over 1 terawatt of compute (including logic chips, memory chips, and packaging). Eighty percent of this capacity is earmarked for space‑based applications, with the remaining twenty percent for terrestrial use, suggesting a future where the most powerful compute is deployed in near‑Earth orbit.

MiniMax M2.5 Performance

Five‑week streak as the global leader in model call volume.

Cost advantage: comparable overseas models cost many times more.

Technical edge: innovation plus low electricity cost (70‑80% of compute expense) creates a strong cost barrier.

Tencent’s Organizational Shift

On March 20, Tencent issued an internal notice cancelling its AI Lab while retaining an industry‑university‑research collaboration center. Key personnel, including AI Lab head Jiang Jie, were reassigned to the Hunyuan team under chief AI scientist Yao Shunyu.

Xiaomi’s Triple Model Launch

On March 19, Xiaomi unveiled three self‑developed models:

MiMo‑V2‑Pro : flagship base model for agent inference.

MiMo‑V2‑Omni : multimodal base model.

MiMo‑V2‑TTS : text‑to‑speech generation model.

Both the Pro and Omni versions, previously codenamed “Hunter Alpha” and “Healer Alpha” on OpenRouter, have already exceeded 1 trillion token calls.

Trend Insights

Token usage as a new benchmark : the industry is moving from competing on model size to competing on real‑world invocation frequency.

Cost‑performance as core competitiveness : dual drivers of technical innovation and energy efficiency shape market positioning.

Accelerated enterprise AI adoption : 2026 is viewed as a pivotal year for large‑scale corporate AI transformation.

Agent (AI‑assistant) surge : products like OpenClaw push large models from conversational tools toward productivity‑enhancing agents.

One‑Sentence Takeaway

China’s 4.69 trillion‑token milestone proves market value, OpenAI’s hiring surge shows confidence, and Musk’s Terafab paints a space‑centric compute future—AI competition is entering a phase of application‑driven growth and ecosystem building.