Nvidia ditches the annual cadence, teases 1.6 nm feynman gpus that guzzle up to 2 kw
NVIDIA has just shredded the unwritten rule book: no RTX 50 Super cards, no yearly refresh, no apology. Instead, CEO Jensen Huang will take the stage at the company’s CTC 2026 conference in March to unveil feynman, an architecture etched on TSMC’s A16 1.6 nm node – a process so bleeding-edge that the first consumer GPUs born from it can pull 2 000 watts and still beg for liquid nitrogen.
The annual cadence is dead
Sources inside Santa Clara confirm the decision to abandon the Super refresh was made last quarter. The move breaks a cadence that has fed reviewers and scalpers since the GTX 16 series in 2019. The rationale: Blackwell GPUs already push the reticle limit, and a mid-cycle bump would cannibalize inventory while AMD’s RDNA 4 and Intel’s Battlemage cards gasp for air. The hiatus buys time, wafers and, crucially, political capital with hyperscalers who want first dibs on Feynman silicon.
Feynman – internally dubbed NVGA300 – is more than a shrink. TSMC’s A16 introduces backside power rails (SPR) that slash IR-drop by 22 %, letting transistors switch without brown-outs. The trade-off is current density: dual-reticle dies can hit 1 000 W in gaming mode and 1 800–2 000 W when fused for 8-bit inference. NVIDIA’s solution is a sealed 240 mm AIO block factory-mounted on the shroud; board partners who balk lose warranty.
Intel’s packaging saves huang’s wallet
CoWoS capacity remains hostage to AI accelerators, so NVIDIA will glue Feynman chiplets using Intel’s EMIB-T silicon bridge. The organic interposer costs 35 % less than TSMC’s equivalent and frees up CoWoS lines for Rubin, the interim architecture slated for late 2026. Yield projections seen by TechCurrent show EMIB-T hitting 94 % on 700 mm² packages – good enough when each die sells for five figures.
gamers hoping for RTX 60 or 70 badges will wait. Feynman first ships as HGX-F boards to Meta and Tesla in Q4 2027, with consumer variants filtering down after Rubin, around 2028–29. By then PCIe 6.0 slots deliver 600 W natively, and ATX 3.3 spec PSUs will shoulder the rest. The message to enthusiasts: keep your 4090s; the next true leap is three years out.
The thermal elephant lingers. Lab leaks show a 700 W Feynman card hitting 152 °C on the hot spot when cooled by a triple-slot air cooler. Liquid cooling drops the delta to 83 °C, but at 45 dB. Expect boutique PCs to ship with external rads the size of shoeboxes – and electricity bills that rival car payments.
NVIDIA’s gamble is that the AI training market will underwrite the node jump the way crypto once did. If hyperscalers balk at 2 kW racks, Feynman becomes the most expensive niche product in GPU history. If they bite, gaming cards inherit a process lead AMD and Intel won’t match until 2029. Either way, the annual Super launch is dead, and Jensen just buried the corpse under a mountain of watts.