Scale-up and scale-out fabrics, optics, and the standards fights that decide which vendors get paid. Resolution runs through issuer filings and public consortium releases.
Sharpest call. 75% that By 2031-12-31 at least two of Amazon, Microsoft, Meta, and Oracle publicly disclose a production AI training or inference deployment whose scale-up domain of 64 or more accelerators is connected by an open-standard fabric (UALink or an Ethernet scale-up specification published by UEC or the Open Compute Project) rather than NVLink or a proprietary in-house interconnect. The card.
No UALink Consortium member announces general availability or volume production of a switch ASIC implementing the native UALink 200G specification (not an Ethernet-tunnelled or bridged variant) on or before 2027-12-31.
By 2028-12-31 at least three of NVIDIA, Broadcom, Cisco, Marvell, and Arista have announced general availability or volume production of a switch with co-packaged optics (CPO) on their own newsrooms, and at least one names a hyperscaler production deployment.
By 2031-12-31 at least two of Amazon, Microsoft, Meta, and Oracle publicly disclose a production AI training or inference deployment whose scale-up domain of 64 or more accelerators is connected by an open-standard fabric (UALink or an Ethernet scale-up specification published by UEC or the Open Compute Project) rather than NVLink or a proprietary in-house interconnect.