My first clock ships: Azure M-series preview on Astera Labs controllers, February 2026 per disclosure. Second clock: tiering runs at Meta in production, per a paper within 1% of ideal. Third clock — pooling, the 'break the memory wall' pitch — stays aspirational: flagship study was an emulation, '100TB pools' are PoCs, no named GA multi-rack deployment. Research says: underwrite what ships; treat pooling as 2027-plus option.