Are we dismissing AI spend before the 6x lands?
Despite current skepticism, massive AI compute capacity, projected to increase sixfold by 2026, is only now starting to yield results from prior investments.Summary
The article argues against dismissing current AI spending despite narratives of scaling walls and high capital expenditure, asserting that critics are judging models based on last-generation hardware.
Analysis of TSMC's CoWoS packaging allocations shows a projected surge from 117,000 wafers in 2023 to 1 million by 2026, translating to an estimated 6.16 exaFLOPs in 2023 to 122.6 exaFLOPs by 2026—a roughly 6x increase in global AI chip capacity.
However, significant lags exist: a minimum one-month delay from chip completion to online status (exacerbated by liquid cooling issues for new accelerators like the GB200), and a further six months from installation to training completion. Therefore, current advanced models like Opus 4.5 and Gemini 3, which show step-changes in capability (e.g., solving complex engineering tasks), were likely trained on the 2024 compute base (~36 exaFLOPs). The massive wave of compute ordered during the recent mania has not yet finished its training runs, suggesting the next generation of performance will be significantly more advanced, potentially leading to a 'zettascale future' by 2030.
(Source：Martin Alderson)