AI Infrastructure Costs Are Rising: What Businesses Should Change About Agent Deployment

What Happened

• Compute commitments grew. Anthropic disclosed $895.4 billion in purchase commitments and lease agreements, substantially all tied to compute and infrastructure. Its amended filing also reported nearly $17 billion in third-quarter revenue and a $23.3 billion net loss. Separately, Oracle plans to raise $20 billion through stock and bonds for data centers and chips. [17][26][24]
• Agent access faced tighter controls. Goldman Sachs is removing AI coding tools from traders and bankers while retaining access for engineers, citing productivity and risk management. Uber described an MCP gateway governing access across more than 800 servers and 5,000 agent tools. [2][28]
• Content rights became a more immediate product constraint. The New York Times sued Brave over alleged scraping and distribution of copyrighted content. Anthropic confirmed Claude Code limits intended to prevent recreations of specific existing games. The Brave claims are allegations, not findings of liability. [4][12]
• Infrastructure choices broadened. Google and Meta open-sourced TorchTPU to support PyTorch workloads on Google TPUs. OpenAI announced a program intended to reduce demand at future US data-center campuses when electricity grids are strained. [13][7]
• Data location moved up the agenda. India proposed a framework under which developers would train models in secure environments controlled by data providers, rather than moving data to developers. Implementation details remain unclear. [36]

Why It Matters to Businesses

The common issue is not whether an agent can complete a task in a demo. It is whether an organization can afford, authorize and audit that task at scale. Large infrastructure obligations and financing needs make capacity and unit economics material procurement questions; Goldman’s reversal shows that broad access to coding agents can be reconsidered after rollout. [17][24][2]

Rights and data controls also affect architecture. A system that retrieves content, invokes internal tools or trains on sensitive records needs enforceable permissions at each boundary—not just instructions in a prompt. The Brave lawsuit, Anthropic’s game-cloning restriction and India’s proposal each illustrate a different boundary that product teams must account for. [4][12][36]

Kimbodo Engineering Perspective

We would treat model choice, tool access and data access as separate decisions. A more capable model does not justify unrestricted access to production systems; conversely, strict controls that prevent legitimate work can erase an agent’s value. The practical trade-off is to grant narrow permissions for evaluated tasks, then expand them when measured outcomes support doing so.

Cost should be measured against work delivered and defects introduced, not tokens alone. Platform flexibility matters too: TorchTPU may reduce the effort of testing an alternative accelerator, but “minimal code changes” is not proof of equivalent performance or operating cost for a particular workload. [13]

How We Would Implement It

• Start with bounded workflows. Select tasks with an identifiable owner, approved data sources, a measurable output and a human escalation path. Establish baseline time, cost and error rates before enabling autonomous actions.
• Put a gateway between agents and tools. Register tools centrally; enforce per-user and per-agent authorization, scoped credentials, rate limits, sensitive-data redaction and response filtering. Uber’s gateway provides a concrete example of separating tool registration from execution. [28]
• Route models by task. Evaluate candidate models against quality, latency and total cost; reserve expensive models for work that demonstrates a benefit. Test accelerator alternatives on the actual workload rather than assuming framework compatibility guarantees savings. [13]
• Keep governed data in its approved environment. Where residency or confidentiality requires it, move training or inference into controlled environments with provider-defined usage rules and audit logs. India’s proposal is a useful design pattern, but not yet an implementation blueprint. [36]
• Instrument outcomes. Record tool calls, approvals, cost per completed task, rework, incidents and rights-related exceptions. Review those measures before widening access.

Risks, Costs and Security

Forecast agent costs under heavy use, not just average use; include tool infrastructure, human review and failed or repeated runs. For systems that act on external content, maintain provenance and licensing checks. For internal tools, prevent an agent from discovering a service it is not authorized to invoke, and retain logs sufficient to investigate actions without unnecessarily exposing sensitive data. [4][28]

Finally, avoid treating announced capacity, energy programs or proposed data frameworks as delivered capabilities. Anthropic’s obligations are commitments rather than an immediate cash bill; OpenAI’s grid program concerns future campuses; and India’s proposal still has implementation questions. Procurement and deployment decisions should rely on verified service terms, workload tests and enforceable controls. [17][7][36]

Sources

1. [2] Goldman Sachs pulls AI coding tools from traders and bankers
2. [4] New York Times and other outlets sue web browser Brave over alleged content scraping
3. [7] OpenAI launches the Grid Protectors Program, a US effort for its future data center campuses to reduce electricity demand when the grid is under strain (OpenAI)
4. [12] Anthropic confirms new Claude Code limits "to steer it away from recreating someone else's specific game", saying IP infringement violates its usage policy (Bruno Ferreira/Tom's Hardware)
5. [13] Google and Meta Open-source TorchTPU to Bring Native PyTorch Support to Google TPUs
6. [17] Amended S-1: Anthropic says it has purchase commitments and lease agreements of $895.4B, "substantially all of which is related to compute and infrastructure" (Ashley Capoot/CNBC)
7. [24] Filings: Oracle is selling $10B in common stock and $10B in bonds, as it rushes to get financing for its data center buildout and chip purchases; ORCL falls 5%+ (Kelly Cloonan/Wall Street Journal)
8. [26] S-1 update: Anthropic reports Q3 revenue up 12x YoY to nearly $17B, a $23.3B net loss, up from $12.1B in Q3 2025, and operating expenses up 292% YoY to $17.5B (Bailey Lipschultz/Bloomberg)
9. [28] Uber Builds MCP Gateway to Govern Access to 800+ Servers and 5,000+ AI Agent Tools
10. [36] India floats AI plan under which ‘The model travels to the data, rather than the data travelling to the model’