2026-01-06
NVIDIA Vera Rubin
Six new chips as one AI supercomputer: Vera CPU, Rubin GPU, NVLink 6, ConnectX-9, BlueField-4, Spectrum-6.
Branches · HARDWARE
Next-generation silicon
AI hardware stopped being a graphics card. It is now a rack of several specialised chips working as one enormous machine.
The platform is designed as a whole: CPU, accelerator, switches and network chips built together. In parallel, every hyperscaler is building its own.
Status
TypeBranches
Marker
What exists today7
Researched
2026-02
Meta MTIA v3
Third-generation in-house silicon enters broad data-centre deployment.
2026-03
Rubin in production
By GTC 2026 the platform grew to seven chips in production; partner availability lands in H2.
In progress
2026
TPU v6, Trainium3, Maia 2
Google, Amazon and Microsoft scale their own accelerators — NVIDIA's share under real pressure for the first time.
препятствие
Memory is the shortage
The bottleneck moved from compute to data delivery: the chip idles waiting for memory. HBM supply is booked years ahead.
Planned
в планах
Optics between racks
Copper can no longer carry the required connectivity — interconnects move to light.
Distant horizons
впереди
A chip for one model
Silicon designed for a single architecture rather than a class of tasks: multiples of efficiency at the price of total inflexibility.
Branches off
To the line