AI Cost Watch - Week of 1 July 2026 (Issue 4)

AI Cost Watch is a recurring, dated note tracking the unit economics of AI infrastructure: whether the buildout's cost and demand assumptions are holding. Issue 4 (week of 1 July 2026) examines the cost side while the core signal (a down-revision in hyperscaler capital-expenditure guidance) stays quiet ahead of the late-July Q2 reports; the Big Four's 2026 capex guides still sum to about 725 billion USD, up about 77 percent on 2025. It measures headline efficiency claims at their true boundary and denominator: warm-water closed-loop cooling, which relocates the water footprint from the data centre to the power plant rather than removing it (USGS puts US thermoelectric withdrawals at 96 to 133 billion gallons a day in 2015, of which about 3 billion is consumed); an open-source vector-search memory gain of about 92 percent that is real at small scale but bounded past a million vectors; and the widening hundred-fold price spread across the July model wave, from Claude Fable 5 at 10 USD per million input tokens down to free open-weight models. It logs the week's clearest demand-side signal, Meta's move to resell excess AI compute alongside a raised 125 to 145 billion USD capex guide (Meta shares up about 9 percent; the neocloud providers CoreWeave and Nebius down about 12 percent each), the first hyperscaler admission of surplus capacity. It reads the custom-silicon layer where the cost floor actually moves (Anthropic's early-stage Samsung 2nm talks and OpenAI's Broadcom Jalapeno chip, both announced not delivered, against already-shipping hyperscaler ASICs); deflates the 850 billion USD announced data-centre lease headline against PJM's own vetting (about 60 GW submitted to about 38 GW accepted for 2030, energised nearer 20 to 22 percent); and shows that OpenAI's reported price cut is an added GPT-5.6 Luna tier at 1 USD / 6 USD rather than a reduction to the unchanged flagship or consumer price. It closes with the indicators that would revise the read at the Q2 reports. Independent analysis, not investment advice. Conflict of interest: produced with assistance from Claude Opus 4.8, an Anthropic model; Anthropic is among the companies this series tracks, and the note reports Anthropic's own custom-chip talks and model pricing, so any reading touching Anthropic is non-neutral and primary, vendor-independent coverage should be preferred.