全球半导体与半导体资本设备：你相信埃隆吗？

Global Semiconductors and Semicap: Do you believe in Elon? Over the weekend Elon Musk made his Terafab announcementas he begins anambitious project to scale humanity’s compute production to 1 terawatt per year (andseeks to push us closer to Kardashev-2 civilizational levels). While he intends to continuepurchasing chips from existing suppliers, he notes that 1 TW is ~50x the current globalcompute supply (~20 GW), and (correctly) observes that those suppliers are, shall we say,hesitant to attempt a capacity addition of this magnitude. Hence the requirement for Muskhimself to attempt the “Terafab.” Stacy A. Rasgon, Ph.D.+1 213 559 5917stacy.rasgon@bernsteinsg.com Mark Li+852 2123 2645mark.li@bernsteinsg.com David Dai, CFA+852 2918 5704david.dai@bernsteinsg.com Starting in Austin...The Terafab project will begin with an advanced fab in Austin,supposedly able to manufacture everything needed to supply advanced AI compute(compute engines and logic, memory, packaging, and mask production) enablingsupposedly fast iteration loops. He intends to focus on edge inference chips (for things likeTesla cars and Optimus robots) as well as space-optimized compute chips. Qingyuan Lin, Ph.D.+852 2123 2654qingyuan.lin@bernsteinsg.com What manufacturing capability would a Terafab actually require?It is instructive toask what magnitude of manufacturing capability would be needed to supply that muchannual compute. We have taken a (VERY rough) stab at it inExhibit 1. Using currentcompute paradigms (i.e.. NVDA racks) and just looking at the main semi requirements (HBM,GPU, CPU) we find that1 TW of annual compute would require somewhere between 7and 18 million 300mm wafer starts per month, dominated by HBM memory (Exhibit 2).For some context,this would be equivalent to 140-360 new 50K WSPM factories, or$5-$13 trillion dollars of capex spendat $35B per fab-equivalent (Exhibit 3). Arpad von Nemes+1 917 344 8461arpad.vonnemes@bernsteinsg.com Alrick Shaw+1 917 344 8454alrick.shaw@bernsteinsg.com Edward Hou, CFA+852 2123 2623edward.hou@bernsteinsg.com More capacity than the world has ever seen?Put another way, the required 1 TWcapacity wouldbe on the order of the entire current global installed semi capacitybase(around 16M 300mm eq WSPM) and would in factrequire many multiples ofcurrent installed capacity for “relevant” semis (memory plus leading edge logicwafers;~5M 300mm WSPM) (Exhibit 4). And note that in this analysis we are ignoringsemis outside of HBM, GPU, and CPU (i.e. networking, NVLink, optics, analog/power etc...) Yipin Cai, CFA+852 2123 2669yipin.cai@bernsteinsg.com Juho Hwang+852 2123 2632juho.hwang@bernsteinsg.com Harder than flying to Mars?Mr. Musk has admittedly done more than one thing thatnaysayers have called impossible at the outset so we wouldn’t necessarily count him out.That being said, a true Terafab feels like a stretch to us, especially if the current computeparadigm is followed. Perhaps we might see partnerships at some point with other currentmanufacturers if Elon decides he cannot make it happen on his own? Or perhaps Elon hassomething more off-the-wall in mind to improve things (we leave that as an exercise for thereader, as we don’t know what it would be...) Carmine Milano+44 20 7762 1857carmine.milano@bernsteinsg.com Francis Ma+852 2123 2626francis.ma@bernsteinsg.com What does this mean for the semi industry?For now, probably not all that much beyondthe hype. But if you believe in Elon for this, clearly you would want to buy semicap (and buyit and buy it and buy it). Perhaps Elon making his own chips might read negative for thecurrent incumbents though we think in a world where compute is this strong any playeris going to see far more upside than they could ever handle (this probably goes equallywell for the memory players). And including logic & memory, mask making, chip design,packaging, etc. together effectively makes it a “super IDM” (a model proven far less efficientthan foundry+fabless+dedicated memory IDM); we see little threat to foundry for now. Zheng Cui+852 2123 2694zheng.cui@bernsteinsg.com BERNSTEIN TICKER TABLE INVESTMENT IMPLICATIONS ADI (Market-Perform, $375.00):ADI is executing well with both cyclical and idiosyncratic drivers, though the shares remainexpensive. AMD (Market-Perform, $235.00):AI expectations remain high, but a new deal with OpenAI has the prospect to drive further(possibly substantial) growth. AVGO (Outperform, 525.00):A strong 2025 AI trajectory seems set to accelerate into 2026, bolstered by software, cashdeployment, and superb margins & FCF. INTC (Market-Perform, $36.00):Intel's problems have broken through to the forefront. NVDA (Outperform, $300.00):The datacenter opportunity is enormous, and still early, with material upside still possible. NXPI (Market-Perform, $240.00):The pace of recovery remains open for debate, with likely less torque in a cyclical rebound. QCOM (Outperform, $175.00):While memory headwinds are pressuring smartphone buil