Global technology companies are moving away from 'tokenmaxxing'—the habit of using as much AI data as possible to show productivity. Instead, they are now focusing on strict cost-management strategies and ensuring that AI provides real value.

全球科技公司正逐漸擺脫「Token 極大化」——即透過盡可能使用大量 AI 數據來展現生產力的習慣。相反地，他們現在專注於嚴格的成本管理策略，並確保 AI 能提供真正的價值。

The industry has shifted from a period of cheap access to a system where companies pay for exactly what they use, which has caused a total review of AI spending. This change is driven by a global shortage of computing power. For example, Google has limited access to its Gemini models for high-volume clients like Meta. This problem is made worse by a lack of specialized memory and processing chips, which has increased the cost of renting hardware.

業界已從廉價獲取的時期，轉向公司需按實際用量付費的系統，這導致 AI 支出被全面審視。這一轉變是由全球運算能力短缺所驅動。例如，Google 限制了 Meta 等高用量客戶使用 Gemini 模型的權限。由於缺乏專業記憶體與處理晶片，增加了租用硬體的成本，使得問題更加嚴重。

To handle these financial pressures, companies are using different strategies. Coinbase, led by CEO Brian Armstrong, has started using cheaper AI models from China, automating how tasks are assigned based on difficulty, and improving data storage. Similarly, companies like Amazon and Uber have stopped using 'token leaderboards' because they realized that using more AI tokens does not actually mean an employee is more productive.

為了應對這些財務壓力，公司採取了不同的策略。由執行長 Brian Armstrong 領導的 Coinbase 已開始使用來自中國較便宜的 AI 模型，根據難度自動化分配任務，並改善數據儲存。同樣地，像 Amazon 和 Uber 這樣的公司已停止使用「Token 排行榜」，因為他們意識到使用更多 AI Token 並不代表員工的生產力更高。

Furthermore, many businesses are now switching to Small Language Models (SLMs) or hosting AI locally to avoid the high costs of larger models. In Australia, research from Elastic shows that about one-third of organizations have either gone over their AI budgets or stopped new projects because they could not prove the value. Consequently, there is a new focus on 'AI accountability,' where success is measured by real results, such as solved customer tickets, rather than how much data is consumed.

此外，許多企業現在轉向使用小型語言模型 (SLM) 或在本地端部署 AI，以避免大型模型的高昂成本。在澳洲，Elastic 的研究顯示，約三分之一的組織已超出其 AI 預算，或因無法證明價值而停止新項目。因此，目前出現了新的「AI 問責制」，衡量成功的標準在於實際結果（例如解決了多少客戶工單），而非消耗了多少數據。

The industry is now moving toward a model of strict financial control, ensuring that AI resources are only used when they provide clear operational value.

業界目前正走向一個嚴格財務控制的模式，確保只有在 AI 資源能提供明確營運價值時才會使用。