martinalderson-com

As AI inference margins shrink, value flows to the hardware supply chain and consumers while model inference becomes a commodity. Frontier AI labs can escape the margin collapse by focusing on managed agents or maintaining a technological lead.

GLM 5.2 is reportedly the first open-weights model that can compete with proprietary frontier models (Opus, GPT) for agentic tasks, while costing roughly 15-20% of the price. The article argues this signals an imminent collapse of AI inference profit margins.

A technique called expert-aware quantisation profiles MoE models to find which experts matter for a specific task, then aggressively quantises less important ones. This achieves near-Q4 quality at near-Q2 size for local models.

Martin Alderson appeared on Radical's Built for Turbulence podcast to discuss how AI agents are reshaping software economics, the concept of the Figma Trap, and the growing recklessness of running human-written code without AI audit.

The article argues that the UK’s push for domestic AI datacentres, driven by concerns over sovereignty, latency, tax revenue, and control, is largely misguided. It contends that these justifications do not hold up under scrutiny, suggesting the perceived need for national data sovereignty may be overblown.

xAI is leasing vast GPU capacity to Anthropic and Google, blending elements of a data centre REIT with its AI lab mission. The strategy likely combines financial engineering ahead of the SpaceX IPO, a genuine compute shortage, and an underlying data-centre advantage.

KV cache compression techniques, including Multi-Query Attention (MQA), Grouped-Query Attention (GQA), Multi-head Latent Attention (MLA), and linear-attention hybrids, have evolved to reduce memory overhead in large language models. These developments have quietly enabled the long context windows required for modern agentic LLM applications by making key-value caching more efficient.

Google announced Gemini 3.5 Flash at I/O, a fast but costly model with mediocre coding performance. The article argues Gemini makes more sense as an internal Google tool, highlights Google's TPU advantage, and identifies coding agents as Google's actual weakness.

The article argues that while managed agents (cloud-hosted agent harnesses) offer significant power, committing to a specific frontier lab's platform carries risk. It advises against locking into such platforms and suggests alternative approaches instead.

The article warns that open-weight AI models, which currently keep major frontier labs competitive on pricing, are gradually disappearing. If this trend continues, the market could become dominated by a small number of oligopolists who can extract excessive consumer surplus without competitive pressure.

The article presents a fictional scenario set in 2026 to illustrate how AI could fundamentally alter cloud security dynamics, arguing that the purely technical explanation fails to resonate outside engineering audiences.

The author announces the launch of a new blog focused on AI-assisted development, MCP integrations, and building production software with modern tools like Claude Code and Cursor.

The article examines the current state of Model Context Protocol support across major AI platforms including OpenAI and Anthropic. It provides testing results and assesses the real-world implementation of cross-platform MCP capabilities in 2025.

The author tested Claude Code beyond software engineering by using it as a tax agent to analyze tax documents and navigate complex tax scenarios in real-time.

The article examines the actual costs of running AI inference at scale, suggesting through calculations that the economics may be more profitable for companies like OpenAI and Anthropic than commonly believed.

The article discusses using static analysis tools to provide AI coding assistants with comprehensive visibility into APIs and third-party libraries, moving beyond basic text search methods. This approach aims to give AI assistants complete understanding of code dependencies and interfaces.

AI coding agents currently feel slow and frustrating like dial-up internet, but ultra-fast inference could significantly improve developer productivity by making them more responsive and efficient.

Google's Gemini AI Studio API has experienced severe reliability issues for the past two weeks, with limited transparency about the problems on their status page.

The article explains how GitHub Actions runners can be slow and expensive, and how using bare metal servers can make CI/CD pipelines 2-10 times faster while costing 10 times less.

The MCP Dev Summit Europe discussed developments in agentic discovery, client compatibility challenges, and the emerging field of agentic experience design. The event provided insights into where the protocol is headed.

The author created a tracker to monitor the growth of MCP servers in the ecosystem. The data shows that the MCP server ecosystem is expanding at a faster rate than anticipated.

A non-technical CFO successfully built a production operations dashboard using Claude Code after previous attempts with low-code tools and agencies failed. This development signals a significant shift in who can create software.

The article analyzes whether AI datacenters are repeating the telecoms crash of the 2000s. It examines token demand growth, infrastructure utilization, and capacity constraints, concluding the economics differ from the earlier scenario.

Software engineers underestimate Excel's scale, but agents that work directly in spreadsheets could transform billions of dollars in business processes.

The author describes using Claude Code to manage production infrastructure through git-tracked markdown files. This approach provides a practical method for small teams to handle sysadmin tasks with version control and AI assistance.

The article describes how IPv6 can be used with Cloudflare to run multiple services on a single server without needing a reverse proxy. This approach leverages IPv6's abundant address space to assign unique addresses to different services.

Gemini 3 Pro's design capabilities and Opus 4.5's reduced babysitting needs represent a subtle but significant leap that traditional benchmarks completely miss.

Agentic coding tools are significantly reducing software development costs, with 2026 expected to surprise many people with these changes.

The article suggests that just as software transformed industries, AI agents are now poised to disrupt the traditional Software-as-a-Service model.

Claude Code demonstrates that minified JavaScript can be reverse-engineered quickly using ASTs and AI agents, showing that minification differs from true obfuscation. The process takes minutes rather than weeks to reconstruct readable code.