Nvidia cancels 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 cards, no annual refresh—Jensen Huang will instead unveil the 1.6-nanometer “feynman” architecture at the March 2026 CTC conference, sources tell Coastal Code. The chips, taped out at TSMC’s bleeding-edge A16 node, push peak power beyond 2 kilowatts per dual-die package, forcing liquid cooling onto every buyer and liquid capital onto every data-center balance sheet.
The ritual is dead
For fifteen years gamers and OEMs could set their calendars to NVIDIA’s six-month cadence: new silicon in spring, “Super” or “Ti” in fall. That metronome stops now. Board partners learned in closed-door briefings last week that Blackwell Refresh is canceled; existing RTX 5090 inventories must cover the 2025 holiday window. The reason is brutally simple: AMD’s RDNA 4 and Intel’s second-gen Arc are landing at price points NVIDIA can’t match without gutting margins. Rather than chase a mid-cycle price war, Huang chose escalation.
Feynman—internally tagged AD1—already runs game builds at 3.4 GHz in the lab, 40 % faster than the 5090 while drawing 60 % more juice. TSMC’s A16 backside-power-delivery network eliminates front-side resistance, letting the shader array suck 1 000 W through a 0.7-millimeter interposer. Double the die, double the watts. The SPR rails keep voltage droop under 3 %, but they also keep HVAC contractors busy: a single 4U DGX-Feynman node needs 47 liters of coolant per minute.
Packaging without tsmc’s chokehold
NVIDIA couldn’t wait for TSMC’s CoWo-L capacity to triple in 2027. Engineers licensed Intel’s EMIB-T bridge tech, stacking HBM4 stacks side-by-side with the compute tile on a silicon interposer stitched with 25-micron micro-bumps. The hybrid package yields 96 %—high enough to keep costs below $18 000 per accelerator, a figure data-center buyers already whisper about as “bargain Ampere nostalgia.”
Liquid-cooled reference cards will ship with 360-millimeter rads pre-filled. AIB partners must certify their own loops for 2 000 W or face qualification failure. Early testers report 68 °C core temps at 2 100 W, but VRM hotspots still spike to 118 °C; capacitors rated 150 °C are popping on 5090 Founders boards already, a preview of the thermal gauntlet ahead.
Ai eats the power budget
Huang’s keynote will frame Feynman as an AI training monster first, gaming card second. Matrix units triple the FP8 throughput of Blackwell; sparse 4-bit inferencing hits 14 petaops. The rumored LPU (local-processing unit) tile is real: a 12-billion-transistor sidecar that crunches 1-trillion-parameter models without leaving the PCIe form factor. The catch—each LPU adds 350 W, pushing a full-fat workstation card to 2.35 kW. That is not a typo.
Wall Street analysts who flew to Santa Clara last month left with one takeaway: NVIDIA will sell every Feynman die it can tape out through 2028. Cloud giants pre-ordered 300 000 units on paper, locking in $5.4 billion in deposits. Gamers, meanwhile, will fight over whatever scraps slip into the GeForce channel—likely less than 15 % of total supply.
The last time NVIDIA skipped a mid-cycle refresh was never. The last time a gaming GPU demanded its own circuit breaker was also never. Feynman is both: a physics lesson in power and profit, delivered by a company that decided annual rituals are for companies with nothing better to ship.