# HBM demand — X 热门讨论 (2026-09-16 11:44 UTC)

## @MartiniGuyYT (That Martini Guy ₿) · 09-16 10:34 · ♥32 ↻15 💬5 SK Hynix is reportedly in talks with Intel to manufacture memory chips in the US for the first time.

The deal could see SK Hynix use part of Intel’s planned Ohio facility to produce memory for the booming AI data centre market.

HBM demand is exploding as Nvidia and other AI chipmakers scale.

If this deal goes through, it could be a major shift for the US semiconductor industry. https://x.com/MartiniGuyYT/status/2100171362740682795

## @LaceyPresley (Lacey) · 09-15 23:12 · ♥30 ↻11 💬1 80% OF TERAFAB IS NOT FOR EARTH

That 80/20 split is the single most revealing metric of the entire Terafab initiative.

It reframes Terafab from a standard automotive or robotics chip factory into the primary manufacturing engine for off-world AI infrastructure.

Here is why that specific allocation ratio exists and what it unlocks:

Bypassing Earth's Physical "Energy Wall"

The Terrestrial Power Crunch: In 2025, U.S. compute demand reached 62 gigawatts, exceeding available data center power generation of 49 gigawatts. With U.S. grid capacity growing at less than 3% annually, attempting to power Terafab's target of 1 terawatt (1,000 gigawatts) of annual compute on Earth would overwhelm electrical grids and consume millions of gallons of municipal cooling water.

The Orbital Bypass: By allocating 80% of Terafab's capacity to space, SpaceX and xAI move the heavy electrical and thermal loads off-planet.

In Sun-synchronous orbits, solar panels receive more than 5 times the energy per unit area compared to terrestrial panels due to continuous, unfiltered sunlight.

Waste heat is rejected passively via deployable radiative cooling loops directly into the vacuum of space, incurring zero operational water or electricity costs.

D3 (Space) vs. AI5/AI6 (Earth) Architecture

The 20% Terrestrial Allocation: Dedicated to producing AI5 and AI6 edge-inference chips. These 2nm-class processors feature on-package High-Bandwidth Memory (HBM) stacking to power real-time vision loops for Tesla's Full Self-Driving (FSD v13+), Cybercab robotaxis, and Optimus humanoid robots.

The 80% Space Allocation: Dedicated to the D3 space-hardened chip family. Unlike terrestrial silicon, which succumbs to cosmic particle bombardment and thermal cycling in LEO, D3 chips incorporate physical radiation hardening, structural redundancy, and elevated operating temperature thresholds to prevent bit-flips without requiring heavy shielding.

The Starmind Backbone

The D3 chips produced at Terafab feed directly into SpaceX's Starmind constellation a network of up to one million solar-powered satellites operating at altitudes between 500 km and 2,000 km.

While Starlink functions as a high-speed data pipe, Starmind nodes act as active server racks in orbit.

They run AI model training and real-time inference directly in space and beam outputs back down to Earth via inter-satellite laser links, bypassing terrestrial fiber and ground data centers entirely.

By consolidating logic fabrication, memory production, lithography mask design, and 3D packaging under a single vertically integrated roof, Terafab creates a closed-loop system where engineers can fabricate, test, and iterate custom space silicon in days rather than months. https://x.com/LaceyPresley/status/2099999818521198789

## @Bearlovesbull (Bearlovesbull) · 09-16 03:11 · ♥39 ↻1 💬0 🚨 $MU SHOWS WHY AI MEMORY DEMAND IS BIGGER THAN HBM.

Today, Micron unveiled the world’s first 512GB DDR5 RDIMM, letting a 24 slot server reach up to 12TB of DRAM while using materially less power than equivalent lower density configurations.

AI servers are not just consuming more GPUs. They are also becoming far more memory dense and that means more semiconductor content per server.

$MU is positioning itself to capture more dollars from every next gen AI server through HBM, $DRAM and higher capacity memory modules.

Volume production is expected in 2H 2027. https://x.com/Bearlovesbull/status/2100059858812702739