# AI capex — X 热门讨论 (2026-10-11 10:43 UTC)
## @USAnt_IDEA (USAnt 미국개미) · 10-10 23:03 · ♥35 ↻8 💬1 리사 수가 대만·한국 돌면서 칩 물량 더 달라고 싹싹 비는 중인데, TSMC 2나노 풀가동해도 감당 안 되는 역대급 쇼티지 터짐 ㅋㅋㅋ
10월에 대만 찍고 한국까지 날아간 리사 수 누님 보고 AMD $AMD 주주들 이제야 대가리 깨지고 현실 파악함.
리사 수가 "공급 대폭 늘리긴 했는데 아직도 턱없이 부족하다"고 대놓고 시인했거든. 2024~2025년부터 앞으로 몇 년간 칩 부족 심각할 거라고 떠들던 게 공포 마케팅이 아니라 진짜였던 거임.
지금 AMD 성장세 받쳐주려면 TSMC $TSM 설비투자(CapEx)를 50~70%는 더 꼬아박아야 하는 상황임.
과거 2023~2024년에 엔비디아 $NVDA 폭풍 성장할 때 TSMC가 케파 1순위로 몰아줬던 거 기억남?
지금 에이전틱 AI 메타 오면서 TSMC가 딱 그 소리 똑같이 하고 있음. 이번에 AMD가 물량 배정 증가율 1위 먹었는데도 리사 수는 대만 가서 "아직 배고프다, 물량 더 내놔라" 시전하고 온 거임 ㅋㅋㅋ
지금 파운드리 상황 보면 TSMC가 신주 팹 20이랑 가오슝 팹 22 포함해서 2나노 팹 5개를 동시에 올리는 중임. 초기 생산 케파가 3나노 때 같은 시점보다 대략 45%나 더 많은데, 고객사들이 2나노로 설계 넘긴 테이프아웃 건수는 3나노 동기 대비 4배나 폭증함. 공장 5개를 동시에 쥐어짜도 수요가 정신 나간 수준이라 감당이 안 됨.
웃긴 건 인텔 $INTC 이 CPU 쇼티지 온다고 제일 먼저 입 털었을 땐 월가 애널리스트 놈들 다 개소리 취급하고 무시했다는 거임 ㅋㅋㅋ 근데 이제 진짜 현실이 됐음.
에이전틱 AI 굴리려고 고성능 CPU 모셔가는 이 역대급 쇼티지 국면에서 앞으로 벌어질 일 정리해 줌:
• 가격 결정권(Pricing Power) 독점: 물량이 없으니 AMD랑 인텔 둘 다 부르는 게 값임. 빅테크 형들 줄 서서 사야 됨.
• 초장기 노예 계약 체결: 핵심 CPU 물량 먼저 뺏길까 봐 고객사들이 장기 계약으로 묶어둠.
• 마진율 수직 상승: 원래도 마진 깡패였던 EPYC(에픽) 서버 칩 덕분에 2026년 4분기부터 전체 믹스 마진 폭발함.
• 실적 퀀텀 점프: 잉여현금흐름(FCF), 영업 레버리지, EPS(주당순이익) 성장률이 전년 대비 160~300% 수준으로 떡상 예상됨.
• R&D 융단폭격으로 Arm $ARM 진영 압살: 번 돈으로 개발비 때려박아서 신제품 출시 주기를 9개월 이하로 단축시킴. 2028~2029년 에픽은 384코어 768스레드, 448코어 896스레드 괴물 스펙으로 나와서 기업이랑 소비자 에이전틱 AI 시장 싹 다 씹어먹을 예정임.
앞으로 3~5년 동안 성장 곡선 진짜 볼만할 듯. 쇼티지 난 반도체는 공급자가 깡패라 그냥 앉아서 돈 복사하는 그림임 ㅋㅋㅋ
Not financial advice https://t.co/6Mrx4GJirt https://x.com/USAnt_IDEA/status/2109057166959432007
## @FroehlichThors1 (Thorsten Froehlich) · 10-11 07:22 · ♥30 ↻5 💬3 Junk (CCC) ./. Champions League (AAA)
Spreads matter
The basement is starting to rott - or it gets crowded out by the AI capex boom, which one is it? https://x.com/FroehlichThors1/status/2109182910758592659
## @tslaming (Ming) · 10-11 08:55 · ♥30 ↻2 💬5 Every Robotaxi and Optimus bot needs memory to turn AI models into real-world action. As Tesla scales production, locking down that memory supply is becoming a critical pillar of its manufacturing strategy—and that’s precisely why Terafab matters.
According to Counterpoint Research, DRAM demand for humanoid robot "brains" is projected to surge roughly twentyfold between 2026 and 2030, reaching 21,500 TB (21.5 petabytes). As robots grow more capable and production ramps up, robotics will emerge as another major demand driver for memory makers.
Autonomous driving adds even more pressure. Micron notes that Level 4 and higher autonomous vehicles typically require over 200 GB of memory—more than ten times what today’s Level 2+ and Level 3 vehicles use. Meanwhile, Goldman Sachs estimates the global commercial robotaxi fleet could reach roughly 1 million vehicles by 2030 and 6 million by 2035.
At an estimated 200 GB per vehicle, those fleets would represent about 200 petabytes and 1.2 exabytes of cumulative onboard memory, respectively. While that reflects total installed memory across active fleets rather than annual chip demand, it underscores the staggering hardware footprint underpinning autonomous driving.
Leading DRAM suppliers already operate at massive scale. SemiAnalysis estimates that Samsung, SK hynix, and Micron will close out 2026 with monthly wafer capacities of roughly 720,000, 595,000, and 385,000, respectively. Combined, that comes out to about 1.7 million wafers per month—or roughly 20.4 million a year.
Yet those same fabs must also supply AI data centers, smartphones, PCs, and the broader automotive market. Because High-Bandwidth Memory (HBM) consumes significantly more wafer capacity per bit than standard DRAM, competition for cleanroom space and production lines is fiercer than ever. Even with planned capex expansions, Micron expects memory supply to tighten further through 2027 and 2028.
Tesla’s internal targets illustrate just how high the stakes are. Its Q2 2026 update pegs installed Cybercab capacity at over 125,000 vehicles annually. For Optimus, Tesla has outlined an initial Fremont line slated for 1 million units per year, with a second-generation line in Texas targeting a long-term run rate of 10 million.
Delivering on those ambitions will demand a silicon supply chain capable of feeding millions of autonomous systems every year.
Terafab is engineered for that scale.
With a stated target of 1 terawatt of compute output annually, the project is designed to power both terrestrial and orbital workloads—explicitly earmarking AI5 for FSD and Optimus, AI6 for next-gen Optimus, and D3 for space-based computing.
By co-locating logic, memory, and advanced packaging under a single roof, Terafab would hand Tesla end-to-end control over the hardware stack powering its autonomous future.
Rather than competing with Big Tech for external fab allocation and packaging lines, Tesla could synchronize processor yields, memory procurement, and packaging directly with vehicle and bot production schedules—all while shrinking design-to-silicon iteration cycles.
If Tesla pulls this off, the payoff is clear: more Robotaxis on the road, more Optimus units on factory floors, and a vertically integrated silicon engine built to sustain them both. > 引用 @tslaming: Terafab could bring Tesla’s experience building custom factory equipment into one of the world’s most demanding industries 🔥
During the X Takeover talk, @elonmusk described how he wants to get there: start with conventional equipment, learn the process, then build custom machines designed for much higher wafer throughput 🔥
Elon sees an opportunity to bring Tesla’s experience with custom manufacturing equipment into an entirely new industry:
“I have no idea how to make chips, but we’ll figure it out. Got to start somewhere. I had no idea how to make cars or rockets as well.
[…] The equipment’s all pretty much coming from the same vendors.
[…] Whereas if you tour a Tesla factory, we got tons of custom equipment.
We’re gonna start with conventional equipment and then make custom equipment that allows for much higher throughput on wafers.” https://x.com/tslaming/status/2109206118039986540