AI costs more than it produces: the problem nobody wanted to admit

Today’s most important AI news

Sunday arrived loaded with uncomfortable news for the industry: Microsoft publicly admitted that the cost of its AI agents can exceed that of human employees, while cancelling Claude Code licenses from its internal stack. AI profitability stopped being a rhetorical question and became the week’s central topic. In addition, Nvidia presented a technical approach that promises to generate text at radically different speeds. It’s a good day to do the math.

Microsoft and the problem the industry preferred to ignore: AI is expensive

According to a Fortune article citing internal Microsoft data, the cost of operating AI agents at scale can exceed what it would cost to pay human employees for the same tasks. The problem isn’t any particular model but token consumption at scale: when agents chain calls, consult long context and run tasks in loops, inference spending adds up in ways few product teams anticipated when designing their flows.

This comes just as Microsoft cancels Claude Code licenses — Anthropic’s development-assistance tool — for its internal teams, as reported by The Verge. The decision has multiple possible readings: it could be a move to consolidate the use of Copilot and GitHub within its own ecosystem, or it could be a sign that the ROI wasn’t justifying the spend. With 453 comments on Hacker News, the news struck a real nerve in the developer community.

The site isaiprofitable.com, which aggregates public data on the margins and costs of the main AI providers, became a reference over the weekend with more than 250 points on Hacker News and 196 comments. The general conclusion emerging from that debate is that profitability depends enormously on the use case: AI for high-frequency, low-context tasks can be cheap, but complex agents are still a luxury that’s hard to justify on spreadsheets.

The practical implication is direct: if your team is evaluating scaling agents in production, you need a cost model based on real token consumption — not demos — before committing to an architecture. The model’s list price is just the tip of the iceberg.

Specialization vs. scale: how to choose the model you actually need

A post by Dharma-AI on Hugging Face makes an argument that should be part of every AI procurement conversation: specialized models often beat large general-purpose models on specific tasks, and that advantage rarely makes it into standard comparisons. Most purchasing decisions are made by looking at general benchmarks — MMLU, HumanEval, and the like — that don’t capture real performance in the domain where the model will work.

The argument carries particular weight in the context of the costs Microsoft just acknowledged. A smaller specialized model can not only be more accurate in its domain, but significantly cheaper to operate. The combination of lower cost per token and a higher task success rate can completely change the ROI equation compared to a general-purpose frontier model.

For teams in Latin America working in specific verticals — legal, health, finance, retail — this approach is especially relevant: large models trained mainly in English on generalized global data may perform worse than more modest alternatives tuned to the domain. The complementary resource here is models.dev, an open-source database that aggregates technical specs, prices and capabilities of models in one place, ideal for making informed comparisons before making an architecture decision.

Diffusion for text: Nvidia bets on radically faster generation

Nvidia published a technical post on Hugging Face about Nemotron-Labs’ diffusion language models, with the stated goal of approaching the theoretical maximum speed of text generation. Traditional language models generate text autoregressively: one token at a time, left to right. Diffusion models for text work differently: they start from noise and refine the entire output in parallel over multiple steps, which opens the possibility of generating complete responses with considerably lower latency.

This technical direction is relevant because inference speed is one of the most real bottlenecks in production applications, especially for conversational use cases or mass content generation. If diffusion models for text manage to maintain quality comparable to autoregressive ones at higher speed, they could significantly reduce inference costs — directly attacking the problem Microsoft revealed this week.

It’s still early-stage research and the quality trade-offs are being evaluated by the community, but it’s a clear signal of where Nvidia’s infrastructure teams are looking: not just more capable models, but models that are economically viable at scale.

In a nutshell

The pattern emerging today is a single question that can no longer be postponed: how much does this really cost? For two years, the industry operated under the assumption that costs would fall fast enough for profitability to sort itself out. Microsoft just showed that this assumption doesn’t survive contact with scale. What’s coming is a cooler consolidation phase: fewer experiments, more rigor in cost-benefit metrics, and a bigger premium for teams that know how to design efficient architectures. AI isn’t going away, but the era when showing a demo was enough to justify the budget is over.

---

Sources used: https://fortune.com/2026/05/22/microsoft-ai-cost-problem-tokens-agents/, https://www.theverge.com/tech/930447/microsoft-claude-code-discontinued-notepad, https://isaiprofitable.com/, https://huggingface.co/blog/Dharma-AI/specialization-beats-scale, https://github.com/anomalyco/models.dev, https://huggingface.co/blog/nvidia/nemotron-labs-diffusion