The entire company has been using Claude frantically for half a year, burning 300 million US dollars per year. Salesforce has finally started to do the math for its AI business: it has to figure out how to scrimp on tokens...

At the end of last month, Salesforce released its financial results for the second quarter of fiscal year 2027: In terms of revenue alone, it reached 11.35 billion US dollars, a year-on-year increase of 10.8%, which basically met Wall Street's expectations. Adjusted earnings per share were 5.90 US dollars, 80% higher than analysts' expectation of 3.27 US dollars. The full-year revenue guidance was also slightly raised from 46.05 billion US dollars to 46.25 billion US dollars.

After the earnings report was released, Salesforce share price rose by 11.8% immediately — no matter how you look at it, this is an impressive report card.

But the problem lies in the details: Salesforce's GAAP operating profit margin dropped from 22.8% in the same period last year to 20.5%, and the full-year profit margin guidance is only 20.1%.

This left investors confused: revenue is rising, but profits are being squeezed? Where did all the money go?

The answer to the question was revealed at a recent investor meeting. Mike Spencer, Vice CFO and Head of Finance at Salesforce, stated bluntly — all the money was spent on Anthropic's Claude model.

01

Six Months Ago, They "Unleashed" Claude

"About six months ago, we fully opened up the use of Claude across our R&D cycle," Mike Spencer said at the meeting. "This is part of the reason why we did not raise our full-year profit margin guidance — we need to cover the growing Token expenditure."

In other words, Salesforce's internal usage of Claude has grown so large that it has impacted the company's income statement.

But Mike Spencer emphasized that this is a deliberate "strategic investment". Salesforce's strategy is very simple: first roll out Claude comprehensively, see what breakthroughs the R&D team can make with it, and how much the product roadmap can be accelerated. "In the worst case, we can just pull back the investment," he said.

Facts have proved that this "worst case" did not happen — the efficiency gains brought by Claude are so significant that Salesforce is unwilling to scale back its usage.

But the problem has shifted to another one: how to make this huge expenditure deliver more value?

02

Faced with a $300 million AI bill, the company has entered a "cost-conscious" mode

Back in May this year, Salesforce CEO Marc Benioff revealed a staggering number on the All-In podcast: The company expects to spend about 300 million US dollars on Anthropic's Tokens in 2026, almost all of which will be used for programming scenarios.

What does 300 million US dollars mean? Judging from the growth rate that Anthropic's annualized revenue increased from about 9 billion US dollars at the end of 2025 to about 30 billion US dollars in March 2026, Salesforce is very likely one of Anthropic's largest enterprise customers.

Marc Benioff mentioned at the time that AI programming agents have brought "unprecedented" efficiency improvements. He even revealed that AI-driven productivity gains allowed Salesforce to reduce its customer service team from 9,000 people to 5,000 people. Now the same logic is playing out on the R&D side — faster product iteration, lower development costs, and output speeds that were impossible to achieve before.

But with a 300 million US dollar bill in front of them, no one can afford to take it lightly.

As a result, Salesforce's current strategy has changed. "We are entering what we call 'refinement mode'," Mike Spencer said. "The core is to select the right model for specific tasks — not every task requires the latest, most powerful model."

His logic is simple: software development or the construction of complex technology stacks may indeed require the most cutting-edge models; but for the "vast majority" of tasks, second-tier or third-tier models are more than sufficient.

This optimization is not limited to model selection, but also involves cross-vendor scheduling. Mike Spencer revealed that Salesforce internally uses models from OpenAI, Cursor and Claude at the same time, and is also testing X's Grok. "Different models have different cost structures, and we are exploring the optimal combination solution."

This idea of "selecting models for tasks" coincides with the "intermediate routing layer" that Marc Benioff advocated on the podcast a few months ago — assign complex reasoning tasks to cutting-edge models like Claude, and assign simple tasks to smaller, cheaper models.

03

AI Coding may be entering the "cost accounting era"

Looking back at Salesforce's experience over the past six months, it is very representative:

In the first stage, what enterprises are most concerned about is: can AI really improve efficiency? So they start using the most powerful models first, run Agents first, and let the R&D team fully embrace AI first.

In the second stage, after AI truly integrates into daily work, the question becomes: can the efficiency gains really cover the costs? So enterprises start counting Tokens, comparing models, optimizing routing, and even reconsidering open-source models.

Eventually, the third stage may be the "refined era" that Salesforce is now entering: AI is not better when it is more powerful, but to use a model that is strong enough and cheap enough for the right task.

Most importantly, Salesforce is not the only company doing this kind of cost calculation.

In June this year, Spencer Kimball, CEO of database company Cockroach Labs, told the media that they are using more open-source AI models internally to optimize expenditure, and are also exploring the use of persistent memory architecture to reduce Token waste — let AI Agents remember the context, so that there is no need to resend it in the prompt every time, which improves speed and saves costs at the same time.

Gartner also predicts that by 2028, the cost of AI programming tools will exceed the salary of ordinary developers. Its data shows that 23% of technology leaders are already paying 200-500 US dollars per developer per month for AI programming Token costs. Even, in extreme cases, a single developer's monthly Token bill can be as high as 20,000 US dollars.

Nitish Tyagi, Senior Principal Analyst at Gartner, said bluntly that AI programming bills have skyrocketed from 20-100 US dollars per month to 2000-5000 US dollars. However, "there is no direct relationship between the growth of Token consumption and productivity gains" — in simple terms, more usage does not mean more output.

For this phenomenon, his suggestion is basically consistent with what Salesforce is doing: adopt a model routing strategy, assign simple and high-frequency tasks to small models, and only use cutting-edge models for complex and high-value work.

So, what insights and suggestions do you have on the matter of "making every Token count"?

Reference link: https://www.theregister.com/ai-and-ml/2026/09/03/salesforce-blames-its-claude-addiction-for-denting-profit-margin-guidance/5294219

This article is from the WeChat official account "CSDN", author: Zheng Liyuan, published with authorization from 36Kr.