Developer · Weekly · March 6, 2026

cloud fragility and surprise bills

AWS data centers in the UAE were hit by drone strikes and a fire, damaging two availability zones and impairing services like EC2, RDS, and DynamoDB, with AWS telling customers to migrate workloads to other zones or regions.[3]

The incident affected up to 25 managed services and kicked off a scramble for disaster recovery setups, including moves from the UAE to the Mumbai region.[4]

On the cost side, one team cut its AWS bill from $2,100 to $190/month by hunting down common leaks such as orphaned EBS volumes and idle NAT gateways.[5]

In contrast, a static site owner on S3 was hit with a ~$15,000 bill after a DDoS forced ~160TB of egress, highlighting how bandwidth can dwarf compute for public buckets.[6]

Outside AWS, Supabase being blocked in India and an $82,314 Google Gemini bill from 2,863 leaked API keys show that regional policy and API key hygiene can be as dangerous as outages.[7]

ai coding stack: claude, gpt‑5.4, copilot, cursor

Claude Code now authors about 4% of public GitHub commits, with projections it could exceed 20% by 2026, and usage is nearly on par with GitHub Copilot among engineers.[8]

Inside Anthropic, Claude Code is reportedly responsible for over 80% of deployed code, making it a primary implementation engine rather than a sidekick.[9]

Claude’s MCP server can cut code context consumption by ~98%, shifting more work into tool orchestration and less into raw tokens.[10] In parallel, GPT-5.4 landed in ChatGPT, the API, and Codex with 1M-token context and a /fast mode ~1.5× quicker, and is widely described as the strongest OpenAI model so far for reasoning and coding.[11]

Cursor has crossed $2B in annualized revenue and shown off multi-agent coordination that beats human-written solutions on formal math challenges, but users still report context-visibility problems on large codebases and are switching to Claude Code for better control, while enterprises continue to standardize around Copilot despite CLI-related malware concerns.[12]

local llms and hardware: ollama vs mlx, vllm, qwen, nvfp4

Ollama is getting hammered by experienced users for slow performance and garbage outputs under load, especially on larger models, making it feel more like a beginner-friendly wrapper than something you’d lean on for heavy workflows.[13]

On Apple Silicon, the MLX stack plus local models is hitting around 170 tokens per second on the Apple Neural Engine and showing big speedups like Qwen 3.5 DeltaNet cutting processing time from 21s to 7s, though some Qwen3.5 variants still drop to ~10 tok/s compared to older qwen3.[14]

Qwen3.5‑27B and the Qwen3.5‑35B‑A3B coding variants are tuned for 16GB NVIDIA GPUs, delivering ~57 tok/s and 55k context windows using quantization schemes like Q5_K_M and IQ2M plus q8 KV cache.[15]

Llama.cpp is adding true NVFP4 quantization support, with NVFP4 on Blackwell GPUs advertised as giving up to 2.5× lower latency and 16× more users per GPU at near‑FP8 accuracy and comparable quality to 8‑bit methods.[16]

Meanwhile, NPUs are inching toward relevance: Qwen3 9B runs at over 6 tok/s on a Samsung S25 Ultra’s Snapdragon 8 Elite, Apple’s ANE claims 38 INT8 TFLOPS, and AMD/Strix Halo NPUs hit ~19.5 tok/s at 20W, even as many developers question whether NPU software support is ready for serious workloads.[17]

ai agents, mcp, and openclaw’s security mess

OpenClaw, an open-source framework for personal AI agents, blew past React to become the most‑starred GitHub project with roughly 246,000 stars, yet its official container image ships with over 2,000 known vulnerabilities (10 critical), and scans show more than 220,000 OpenClaw instances exposed on the public internet without authentication.[18]

A reported “ClawJacked” attack lets malicious websites hijack OpenClaw sessions to steal data, and users say the system often needs heavy babysitting to get reliable automation.[19]

Broader scans across the agent ecosystem found over 220,000 AI agent instances lacking any auth and noted that 41% of official MCP servers have no authentication, granting any connecting agent full tool access.[20]

At the same time, MCP servers are where a lot of power is accruing: they can reduce Claude Code context by 98%, compress codebase knowledge into graphs with 120× token savings, and surface live uptime and incident data directly to agents.[21]

WebMCP and Chrome’s `navigator.modelContext` are entering preview to let sites expose structured tools and payments to agents, but commenters are already flagging the potential for new abuse classes if those interfaces are buggy or manipulative.[22]

self-hosting stack and data backbone: docker, homelabs, postgres/sqlite

Docker remains the default unit of deployment for self-hosting, with users praising the ability to isolate apps and DBs like Postgres and to rebuild stacks from a single compose file, while noting memory bloat in long‑running containers and a preference for compose+git as the source of truth over GUIs.[23]

Portainer is still widely used as the first app in Docker environments and can run on very low-spec hardware, but licensing limits, lack of safety nets for accidental edits, and no built‑in history have pushed power users toward tools like Dockhand, Komodo, and StackSnap for safer multi‑instance management.[24]

Proxmox clusters assembled from e‑waste PCs and mini PCs now commonly host Docker or Kubernetes along with AdGuard Home, Unbound, and Home Assistant, while WireGuard and OPNsense provide VPN access and network segmentation, yet people report that older hardware can cause subtle Kubernetes failures.[25]

On the data side, PostgreSQL is showing strong throughput and cost leverage, with one benchmark hitting 17,658 JSON inserts per second and another team cutting an AWS bill from $2,100 to $190 after tuning Postgres and cloud resources.[26]

SQLite is everywhere in the AI stack—powering FastAPI services with 10GB databases, Rust photo managers, agent platforms like KinBot, and telemetry or event processing pipelines up to 4.2M events per second—because of its simple deployment model and solid indexing.[27]