In-depth analysis on AI infrastructure, inference economics, GPU pricing, agent tooling, and building production AI systems.
Amazon, Google, Microsoft, Meta, and Oracle have committed $660-690B in 2026 capex, roughly 75% targeting AI infrastructure. AWS broke two decades of declining cloud prices with a 15% GPU hike. Supply remains constrained through year-end as NVIDIA Rubin production ramps slowly and power grid bottlenecks add 24-72 months to new data center timelines.
NVIDIA's B200 cuts inference costs ~40% vs H100 for large models, but requires FP4 quantization and NVLink-72 fabric. Winners: large-scale deployers. Losers: anyone locked into H100 leases through 2027.
The three engines powering most production inference — benchmarked, compared, and mapped to the right workloads. Your choice of serving engine determines 30-60% of your inference cost.
On-demand H100 pricing dropped 15-20% across hyperscalers. Spot markets are volatile but offering 50-65% discounts. Reserved 1-year commitments remain the best value at 35-40% off on-demand. Tier-2 providers are undercutting on price but trailing on availability guarantees.
MCP (Model Context Protocol) has emerged as the de facto standard for connecting AI agents to external tools and data sources. With support from Anthropic, OpenAI, Google, and the open-source community, MCP servers now cover databases, APIs, dev tools, and enterprise systems. Builders should standardize on MCP now.
Stripe launched Agent Toolkit for programmatic payments. Coinbase and Circle are pushing stablecoin micropayment APIs. Meanwhile, API providers are building agent-native billing with per-action metering. The payment layer for the agent economy is emerging in parallel from fintech, crypto, and API-first companies.
Public benchmarks (MMLU, HumanEval, MATH) are increasingly gamed through training data contamination and cherry-picked configurations. Real-world performance on domain-specific tasks can differ by 15-30% from published scores. The only reliable approach: build custom eval pipelines on your production data.
Quarterly benchmarking of inference costs across major cloud providers and GPU generations. Covers pricing, throughput, and cost-per-token for leading open and closed models across 8 cloud providers.