From Side Project to Seven-Figure Token Burn
Peter Steinberger, creator of OpenClaw, offered a stark glimpse into modern AI token costs when he shared a CodexBar dashboard showing 30‑day AI API spending of USD 1,305,088.81 (approx. RM6,000,000+). That covered 603 billion tokens across 7.6 million requests, with one day alone (15 May) hitting USD 19,985.84 (approx. RM90,000+) for 19 billion tokens and 206,000 requests. The bulk of this spend reportedly went into building and testing OpenClaw and related AI tools, powered heavily by OpenAI’s gpt‑5.5‑2026‑04‑23. Commenters quickly framed his bill as equivalent to the annual compensation of multiple senior engineers redirected purely into token billing. Because Steinberger now works at OpenAI, he clarified that he is not personally paying this bill, underscoring how access to subsidised or free compute has become both a recruiting perk and a catalyst for token‑maximising experimentation.
Tokenmaxxing Culture and the Bubble Question
Steinberger’s spending sparked an online backlash that went beyond sticker shock. Critics argued that burning through over USD 1.3 million (approx. RM6,000,000+) in a month is only defensible if the resulting AI-generated outputs outperform what “USD 1MM worth of engineers” could deliver. Others noted this figure reflects subsidised AI pricing, implying the underlying compute cost is significantly higher than what today’s token billing reveals. Steinberger countered that switching off “fast mode” would make his usage roughly 70% cheaper, likening the spend to one employee rather than an entire team. The episode highlights a broader concern: AI token costs are accelerating faster than clearly measurable value in many projects. While some firms report dramatic productivity gains, the mismatch between token consumption and tangible ROI is fuelling debate over whether current AI usage patterns resemble a rational investment strategy or the early stages of an industry bubble.
Salesforce’s USD 300 Million Token Bet on Anthropic
If Steinberger’s dashboard shows the individual extreme of enterprise AI spending, Salesforce illustrates the institutional version. The company plans to spend USD 300 million (approx. RM1.38 billion) on Anthropic AI tokens in a single year, treating tokens as a core AI operational expense rather than an experimental line item. These tokens fuel Claude models embedded across engineering workflows, with AI now handling 30–50% of Salesforce’s overall workload. Management says AI tools have boosted engineering productivity by more than 30%, and the Agentforce AI business unit has grown to roughly USD 800 million (approx. RM3.7 billion) in annual recurring revenue. To manage costs, Salesforce is building an “intermediate layer” that routes routine tasks to smaller, cheaper models while reserving Anthropic’s frontier systems for complex jobs. This consumption‑based AI token strategy shows how large enterprises are institutionalising token billing and treating it like a utility bill for software development.
AI Tokens as a Substitute for Headcount
Both Steinberger’s personal token burn and Salesforce’s massive enterprise AI spending are reshaping workforce strategy. Observers of Steinberger’s usage noted that his monthly AI bill approximates what could fund a team of senior engineers, raising the question of whether tokens are effectively replacing human headcount. At Salesforce, the connection is more explicit: the company has announced an engineering hiring freeze, citing more than 30% productivity gains from AI tools such as Agentforce, Claude, OpenAI-based coders, and Cursor. Around 15,000 engineers are increasingly moving into supervisory roles, reviewing and integrating AI-produced code instead of writing it from scratch. AI token costs, in other words, are becoming a deliberate trade‑off against salary and headcount growth. For enterprises, the calculus is shifting from “Can we afford AI?” to “How much human work can expensive tokens credibly and consistently replace without compromising quality or resilience?”
Counting the Real Cost of Enterprise AI
High daily token bills, such as Steinberger’s nearly USD 20,000 (approx. RM92,000) spend in a single day, sharpen concerns about cost efficiency. For now, many of these expenses are shielded by subsidies and internal credits, masking the true price of large-scale AI compute. But as Salesforce’s USD 300 million (approx. RM1.38 billion) commitment shows, AI operational expenses are evolving into long-term, non‑negotiable budget items. The key question for enterprises is whether AI-generated outputs—code, documents, analytics—deliver consistent ROI compared with traditional staffing and infrastructure. If AI can compress months of specialised work into days, the economics may justify aggressive token usage. If not, organisations risk channelling vast sums into token billing without durable value. The emerging reality is that AI adoption is no longer limited by model capability alone; it is constrained by cash, governance, and the discipline to align token consumption with measurable business outcomes.