AI ROI Series | EP15 | IBM's AI Bill & What Does this Mean to Enterprise AI?
IBM just had its worst day since 1987. The stock fell more than 20%, about $55 billion in market value gone, on preliminary Q2 revenue of $17.2 billion against a consensus near $17.9 billion. Everyone read it as an IBM problem. It is not. It is the clearest sign yet of the shift reshaping every enterprise budget. Straight from CEO Arvind Krishna: in the final weeks of June, clients redirected capex toward servers, storage, and memory to lock in supply before prices climbed. Infrastructure fell 7%. Software deals slipped. Consulting was flat. The money went into AI hardware, and it came out of everything else. Accenture, Cognizant, ServiceNow, Adobe, and Workday all sold off in sympathy. AI spending is not cooling. It is cannibalizing. Every dollar going into compute, memory, and tokens is a dollar not going into software licenses and consulting. Which means every line item now has to justify itself against the AI line, including the AI line itself. Then I get into the numbers I promised on Friday, from Cursor's own usage data, and they surprised me: About 90% of coding-agent token usage is input, not output. You are paying the model to read your codebase, not write it. With caching, actual output is 0.6% of usage. Without smart caching, Cursor says the bill would be roughly 10x higher. Cost per request is misleading. Opus runs ~$1.57 an agent request against GPT-5.5 at ~$0.81, but on cost per accepted line of code they land in the same place. Route on cost per accepted output, not cost per request. In a single month, the share of developers letting agents commit code with no manual review went from about 10% to 40%. A cost story and a governance story at once. Plus what the AI natives are doing about it, why Perplexity is building a model-agnostic coding tool, and the uncomfortable question this puts to consulting: if agents do the work that used to be billable hours, what is an hour worth? Tuesday, we show the math. Directional as always. Check my work, and tell me if you see it differently. CHAPTERS 0:00 Why this video is late 0:45 IBM's worst day since 1987 1:40 Krishna's words: the capex reprioritization 2:40 Not an IBM problem: the sector sell-off 3:20 AI spending is cannibalizing, not cooling 3:55 Where the token money actually goes: input, not output 4:40 Caching is the whole ballgame 5:15 The correction: cost per accepted output, not per request 6:05 The governance bomb: 10% to 40% with no review 6:50 What the AI natives do: Perplexity's Teammate 7:35 The uncomfortable question for consulting 8:15 The move: proof is the only defensible position THE NUMBERS IBM: down more than 20%, ~$55B market value gone, worst day since 1987. Q2 revenue $17.2B vs ~$17.9B consensus. Infrastructure down 7% Cursor usage: ~90% input tokens; output 0.6% of usage with caching; ~10x cost without smart caching Opus ~$1.57 per agent request vs GPT-5.5 ~$0.81, similar on cost per accepted line Developers committing AI code with no manual review: ~10% to ~40% in a month MORE Read the full breakdown: [ARTICLE LINK] AI ROI Playbook and TokenWising Guide, or just say hello: https://olakai.ai/paul Connect with me on LinkedIn: [PAUL LINKEDIN] SOURCES IBM Q2 preliminary results and Arvind Krishna comments, July 2026. Coding-agent usage data: Cursor aggregated usage via The Pragmatic Engineer. Perplexity Teammate, reported. Figures are as reported and directional. #EnterpriseAI #AIROI #Tokenomics #IBM #AIspend #CodingAgents #Cursor #FinOps #AIgovernance #CFO
