Thought Leadership
Salesforce was unable to raise its margin forecast, according to statements at an investor conference last week, because the company is spending substantial sums on using Anthropic’s Claude model.
According to its quarterly results, the operating margin under accounting rules stood at 20.5 percent in the second quarter (ending July), but the forecast for the full year is only 20.1 percent. Mike Spencer, Deputy CFO and Head of Finance at Salesforce, explained at the Deutsche Bank Technology Conference that spending on Anthropic, with whom Salesforce has now entered a new partnership, was sufficient to put investor expectations regarding margin trends into perspective accordingly. Spencer elaborated:
Part of the reason why the margin forecast for the year was not raised was that the company had to absorb a portion of the incurred token expenses itself. Back in May, Salesforce CEO Marc Benioff stated on the All-In podcast that the company had budgeted around 300 million US dollars in spending with Anthropic for the year 2026. The close cooperation between the two companies was recently demonstrated publicly: During the presentation of the quarterly results, Salesforce and Anthropic jointly presented Claudeforce, a co-developed interface that links Claude’s reasoning capabilities directly with Salesforce data, workflows, and permissions.
“Around six months ago, we turned Claude loose in our R&D cycle.”
Mike Spencer, Deputy CFO and Head of Finance at Salesforce
Salesforce: Targeted selection of cheaper models as a countermeasure
According to Spencer, Salesforce is now actively working to extract more value from its own AI spending, including by being more selective about which model is actually needed for which specific task. For the vast majority of tasks, an older, cheaper second- or third-generation model is reportedly sufficient, with the use of the latest model generation actually required only for certain more demanding tasks in software development. This optimization also extends across various providers; internally, alongside Claude, Salesforce also uses models from OpenAI as well as the coding tool Cursor.
Salesforce is not alone in this cost optimization: Other companies are also increasingly attempting to manage their AI spending more selectively to counteract so-called tokenmaxxing, an uncontrolled surge in AI-related operational costs. According to market research firm Gartner, the costs for AI-assisted coding assistants in some regions of the world could soon exceed the salary of a human developer.
(Editorial Team)
