Coinbase bypasses Western high-powered computing, focuses on Chinese models GLM and Kimi
Coinbase, the blockchain exchange, announced it has shifted its core business operations to more cost-effective Chinese AI models like GLM5.2 and Kimi2.7. This move has significantly increased token usage while cutting the company's AI costs in half. Coinbase CEO Brian Armstrong revealed that although developers still retain the right to choose models, the newly established automatic routing system intelligently distributes requests based on task type, price, and caching potential, boosting the cache hit rate from 5% to an impressive 60%.
This strategic shift has quickly triggered a chain reaction across the Western tech industry. The CEO of startup Lindy recently switched to Deepseek v4, while cloud computing giant Snowflake is intensively testing Chinese models, positioning them as a cost-effective alternative to OpenAI and Anthropic. This move directly pressures Western top AI labs preparing for their IPOs, posing a severe test to valuation systems that rely on high unit prices to maintain growth expectations.

With high-performance reasoning models like GPT-5.x-Thinking and Opus4.5 hitting the market, enterprise token consumption is experiencing explosive growth. Under the "token maximization" trend tolerated by giants like Amazon and Meta—where employees are allowed to consume tokens without limits—Coinbase has introduced a more rational accountability mechanism, tying AI investment to actual business output.
Related article
AntGroup Unveils Ling-3.0-flash-VL, a 124B-Parameter Native Multimodal Large Model
The Ant Open Source team has officially launched Ling-3.0-flash-VL, the first native multimodal large model in the Bailing series. Built on the MoE architecture of Ling-3.0-flash, this model features 124B total parameters, with only 5.5B activated pe
OpenAI Delays IPO to 2027 Amid Valuation Surge to $1.2 Trillion
OpenAI, the pioneer in artificial intelligence, is currently negotiating private funding with investors to target a staggering $1.2 trillion valuation. This figure marks a dramatic leap from the $85.2 billion post-money valuation set in March, follow
Microsoft Classic Outlook to Integrate Copilot AI for Email Drafting on Windows 10 and 11
Windows Latest reports that Microsoft intends to bring the Copilot update to the classic Outlook for Windows 10 and 11 by late 2026, introducing an "email drafting" feature. This strategy signals Microsoft's push to extend AI tools from the new Outlo
Related Special Topic Recommendations
Comments (0)
0/500
Coinbase, the blockchain exchange, announced it has shifted its core business operations to more cost-effective Chinese AI models like GLM5.2 and Kimi2.7. This move has significantly increased token usage while cutting the company's AI costs in half. Coinbase CEO Brian Armstrong revealed that although developers still retain the right to choose models, the newly established automatic routing system intelligently distributes requests based on task type, price, and caching potential, boosting the cache hit rate from 5% to an impressive 60%.
This strategic shift has quickly triggered a chain reaction across the Western tech industry. The CEO of startup Lindy recently switched to Deepseek v4, while cloud computing giant Snowflake is intensively testing Chinese models, positioning them as a cost-effective alternative to OpenAI and Anthropic. This move directly pressures Western top AI labs preparing for their IPOs, posing a severe test to valuation systems that rely on high unit prices to maintain growth expectations.

With high-performance reasoning models like GPT-5.x-Thinking and Opus4.5 hitting the market, enterprise token consumption is experiencing explosive growth. Under the "token maximization" trend tolerated by giants like Amazon and Meta—where employees are allowed to consume tokens without limits—Coinbase has introduced a more rational accountability mechanism, tying AI investment to actual business output.
AntGroup Unveils Ling-3.0-flash-VL, a 124B-Parameter Native Multimodal Large Model
The Ant Open Source team has officially launched Ling-3.0-flash-VL, the first native multimodal large model in the Bailing series. Built on the MoE architecture of Ling-3.0-flash, this model features 124B total parameters, with only 5.5B activated pe
OpenAI Delays IPO to 2027 Amid Valuation Surge to $1.2 Trillion
OpenAI, the pioneer in artificial intelligence, is currently negotiating private funding with investors to target a staggering $1.2 trillion valuation. This figure marks a dramatic leap from the $85.2 billion post-money valuation set in March, follow
Microsoft Classic Outlook to Integrate Copilot AI for Email Drafting on Windows 10 and 11
Windows Latest reports that Microsoft intends to bring the Copilot update to the classic Outlook for Windows 10 and 11 by late 2026, introducing an "email drafting" feature. This strategy signals Microsoft's push to extend AI tools from the new Outlo





Home






