Nvidia shelves rtx 50 super, leaps straight to 1.6-nm ‘feynman’ gpus that guzzle 2 kw
NVIDIA just tore up its own playbook. No RTX 50 Super refresh, no annual cadence—straight to a 1.6-nm monster named Feynman that can pull 2 000 W and still smile. Jensen Huang will unwrap the blueprint at the company’s March 2026 CTC conference, and the first chips land inside RTX 60—maybe 70—series boards before the decade closes.
The math is brutal: TSMC’s brand-new A16 node crams wires so tight that parasitic capacitance drops 20 % versus today’s 2-nm pilot line. Translation: clocks sprint higher while voltage sags, exactly what AI training clusters crave. NVIDIA pairs the shrink with a dual-die package stitched together by Intel’s EMIB-T bridge, sidestepping TSMC’s CoWoS chokepoint that has starved RTX 5090 supply for months.
Why gamers should care about a physics lecture
Richard Feynman never coded a shader, but his namesake architecture redrafts the power curve. A single Feynman GPU can spit 100 TFLOPS of FP8—double a 5090—yet needs a 360 mm radiator and a 12 VHPWR cable thick enough to jump-start a truck. Liquid cooling stops being boutique; it becomes the PCIe bracket you screw in first.
Desktop terror aside, notebook vendors already whisper about “external Feynman docks.” Picture a slim magnesium chassis riding a fat pipe of PCIe 6.0 to a brick that houses the GPU and its own 1 200 W PSU. Legacy laptops suddenly outrun today’s flagships—if you don’t mind a desk that hums like a server closet.
Cost? Foundry wafers at 1.6 nm start at $30 000 a pop, up 70 % from 3-nm. Add Intel’s packaging tariff and board partners hint at $3 499 for the xx90 tier on day one. The consolation prize: multi-die yields let NVIDIA sell cut-down 60-class cards with half the cores enabled yet still trounce a 5090. Expect scalpers to pivot from sneakers to copper water blocks overnight.
AMD and Intel Arc sigh relief—until they see the slide titled LPU inference tile. A low-profile logic slice bolted beside the shader array chews sparse matrices for breakfast, giving NVIDIA a 5× perf-per-watt edge on transformer workloads. Rubin, the architecture scheduled for late 2025, now feels like a placeholder while Feynman preps for an 2028 market grab.
Power-grid engineers, start your spreadsheets. A 64-GPU Feynman pod peaks north of 120 kW, enough to brown-out a small office. Data centres are already lobbying utilities for 480 V three-phase drops, and European co-location parks quote €0.45 per kWh for AI racks—carbon tax included. The cloud bill you curse today will look quaint tomorrow.
So no, there won’t be a 5090 Ti to tide you over. NVIDIA bet the farm on physics so fresh the masks aren’t even dry. Gamers get frames, corporations get exaflops, and your wall socket gets a workout. March 2026 can’t arrive fast enough—just pray the grid holds its breath.