martinalderson.com
|
ksl
|
|
Martin Alderson argues the AI compute crunch has arrived well ahead of schedule, with only about 1 to 2 percent of knowledge workers actively using AI coding and reasoning tools. Anthropic’s recent product degradations – reduced model effort levels, removed older models from Claude Code, disabled prompt suggestions – read less like product decisions and more like triage under GPU pressure. DDR5 memory prices nearly quadrupled between May and December 2025, and DRAM fabrication timelines cap total AI infrastructure at roughly 15 gigawatts until 2027. Alibaba Cloud’s leadership has publicly admitted they cannot deploy servers fast enough. The uncomfortable math is that current infrastructure cannot serve even modest adoption growth, and new fab capacity will not arrive before 2028.
