Microsoft reports that the token-based cost of generative AI is now higher than the cost of paying human employees
Microsoft's fiscal year 2026 Q4 report has produced a critical data point for the industry: the token-based cost of generative AI applications is 18 percent higher than the hourly cost of a human employee producing equivalent work. Fortune magazine characterised the figure, presented in Microsoft's financial report under the heading 'AI Tax Margin', as the sector's first openly published comparison.
Microsoft CFO Amy Hood said in the earnings conference call: 'Generative AI infrastructure costs, GPU capital and energy consumption are higher than the AI service price passed to customers. This is a short-term imbalance; as AI infrastructure costs continue to fall, customer prices will be adjusted cautiously.' Microsoft spent $24.3 billion on AI infrastructure in fiscal year 2026 Q4 — a 78 percent increase compared with the same quarter of fiscal 2025.
The report is a concrete data point for academics and practitioners debating generative AI's economic logic. MIT Sloan School of Management Professor Erik Brynjolfsson wrote in a comment on Hacker News: 'This is the data point we have been waiting for in the AI sector. Until now we only heard it in Sam Altman's rhetoric at OpenAI; once Microsoft reports it in financial filings, the real-time economics of the sector become debatable.'
The figure in Microsoft's report implies a token-based cost of about $0.025 for GPT-4o-level output models — a daily cost of $250 for a mid-size enterprise customer consuming 10 million tokens a day. A human analyst capable of producing the same output, at an estimated $38 per hour as used by Microsoft's Foundation Engineering group, would cost $304 for an 8-hour working day. So AI cost is 18 percent below human cost; but Microsoft's report says that the real cost reaches $359 when infrastructure (capital amortisation, energy and cooling) is included — 18 percent above human cost.
The figure inverts the industry's standard line that 'AI is very cheap'. The headline services from Anthropic, OpenAI and Google are being sold below cost to customers. According to internal OpenAI documents shared with Bloomberg, the company is pricing access to the GPT-5 tier on average 38 percent below true cost. The same gap is 42 percent for Anthropic's Claude 4.7 and 35 percent for Google's Gemini 2.5 Pro.
Industry analyst Cassidy McGillicuddy of TBR Research commented: 'In this model, a hyperscaler providing token-allocated AI service to a customer is not sustainable. Either prices must rise through the rest of 2026 or mid-2027, or infrastructure costs must fall significantly. A third option is for the customer to absorb the loss margin — but that does not square with the sector's competitive dynamic.'
The asymmetry has three main drivers. First, the major energy cost of generative AI: a GPT-4o call consumes 30-40 times more energy than a Google search. According to Microsoft's December 2025 sustainability report, the company's energy consumption rose 75 percent from 2024; more than 80 percent of that growth came from AI workloads. Second, the cost of AI chips (Nvidia H100, B100 and the next-generation Rubin). Third, cooling systems for AI data centres — especially liquid cooling for the GPU racks — add a significant additional cost line.
What is the effect on Microsoft's customer pricing? The company announced that it will apply a 15-25 percent price increase to customers starting on 1 September 2026. This will affect Microsoft Copilot Enterprise (priced per user per month for M365 Copilot), Azure OpenAI Service and Foundry AI services. The company described the price increase as 'passing AI infrastructure cost on to customers'. Customers may consider switching to alternative AI providers in response.
Microsoft's asymmetry report also calls into question the AI labour-replacement thesis. McKinsey & Company's September 2025 report estimated that generative AI could replace 12-15 million salaried human jobs in the US by 2030 — about 8 percent of the total US labour force. Microsoft's new data point suggests this timeline could be delayed; the cost comparison between generative AI and human labour will not be equalised for some years longer due to price adjustments. Wall Street's reaction was mixed: Microsoft shares closed Friday down 2.8 percent, but FY2027 fiscal guidance has preserved investor confidence in AI's long-term role. Goldman Sachs analyst Kash Rangan said: 'It is known that AI costs will weigh on customer margins, but for Microsoft to report it plainly is a bold step — a potentially useful disclosure for sector confidence.'
The publication of the Fortune report coincided with a week in which Sam Altman was interviewed: on 21 May, OpenAI spokesperson Hannah Wong said: 'It is true that generative AI costs are higher than customer prices, but we believe this is a transitional situation. Once the Stargate data centres are ready by the end of 2026, costs will fall significantly.' Microsoft's Friday report indicates a more concrete arithmetic reality lies beneath OpenAI's rhetoric.
The next 12-18 months will be a real test for AI economics. After Microsoft's price increase decision, how rivals such as Anthropic, Google and OpenAI will respond is unclear. The whole sector may raise prices; or a capacity-intensive AI infrastructure race (Nvidia Rubin and successor chips, more efficient cooling systems) may bring costs down quickly. For customers, the practical question is whether the ROI on enterprise AI investments needs to be re-evaluated under the 2026-2027 pricing environment.
*This article is not investment advice. Make investment decisions based on your own research or by consulting an investment adviser.*