Weekly Radar 005：AI 需求将 DRAM 产能推向服务器端

• Memory and storage supply remain tight well into 2027. DRAM stays Very Tight at 20 weeks versus an 8-week balanced level, while suppliers are shifting bit supply from PC DRAM toward server DRAM. CSPs are also moving to secure eSSD and HDD supply through 2027, with HDD lead times still at 50 weeks.

• Terafab puts foundry strategy back in focus ahead of TSMC earnings. Musk says Tesla and SpaceX plan to build and operate the fab themselves, with Intel still involved and any TSMC role potentially limited to a sublease. TSMC reports this week, putting capex, margins and the next phase of AI-driven capacity expansion back under the spotlight

• The OCP Global Summit opens this week, where the AI supply chain tackles its next set of bottlenecks. As racks move toward megawatt scale, watch for progress on 800 VDC power, liquid cooling, and optical networking, including co-packaged and near-packaged optics.

We track CPU, GPU, DRAM, NAND, HDD, ABF, and MLCC each week. Start with the status color, then compare the current lead time with the balanced-market benchmark and read the quick take.

MEMORY

What Happened: Samsung’s preliminary 3Q26 guidance points to a blockbuster quarter. Operating profit exploded 782.5% year over year to KRW 107.4 trillion, smashing through the KRW 100 trillion mark for the first time, while sales more than doubled to KRW 195 trillion, up 126.6%.

Why It Matters: The guidance adds to evidence that the memory upcycle is far from over, following Micron’s strong earnings in September. Conventional DRAM prices continue to climb as supply remains tight, and HBM4 is now a meaningful contributor, riding demand for NVIDIA’s Vera Rubin. However, the picture could be less rosy outside memory chips: soaring memory costs are squeezing Samsung’s smartphone and appliance business, and the market will watch closely when the full earnings release on October 29 breaks down results by division, with the DX unit’s losses expected to widen.

SEMICONDUCTOR

What Happened: Intel isn’t walking away from Musk’s Terafab. Responding to market chatter that TSMC could be tapped for the project, Intel CEO Lip-Bu Tan said the chipmaker will stay involved. Musk then moved to cap the TSMC speculation, saying the foundry giant could sublease part of Terafab “if they want, but nothing more than that.” At the October 10 X Takeover event, Musk called chipmaking “one of the hardest things to do in the world” and said mastering it could take his companies “quite a few years.” Even so, he set a goal of advancing around 10 chip designs a year, far above the industry’s roughly two years per major design.

Why It Matters: Terafab could remain a key external customer win for Intel’s foundry ambitions. Musk’s comments show SpaceX still plans to build and run the fab itself, leaving little room for TSMC beyond a possible sublease, especially as the foundry already has multiple fabs to build in Arizona. Analysts add that with the U.S. government now an Intel shareholder, it would be hard to see the chipmaker displaced to any significant extent.

MEMORY / SEMICONDUCTORS

What Happened: AMD CEO Lisa Su made a trip to South Korea to meet memory heavyweights Samsung and SK hynix, confirming that HBM4 shipments for the MI455 are getting underway. She also floated supply-chain partnerships spanning three to five years or more, a sign that the AI chip giant wants to lock in memory supply as shortages persist.

Why It Matters: Long-term agreements are fast becoming big tech’s playbook for securing scarce memory, and AMD seems to be following it. However, the reported deal could matter even more for Samsung: with AMD still heavily reliant on TSMC for its CPUs and GPUs, deeper memory ties could give Samsung leverage in foundry talks, as market speculation grows that AMD may adopt Samsung’s 2nm node.

Event: The 2026 OCP (Open Compute Project) Global Summit opens Oct 12 in San Jose and runs through Oct 15, expanded to four days under the theme “Scaling Innovation for the AI Era.”

Why It Matters: From hyperscalers like Meta and Google to chip heavyweights NVIDIA, AMD, Broadcom and Marvell and server rack makers, the AI supply chain gathers at OCP to tackle the bottlenecks of large-scale AI: power, cooling, networking and memory. A key focus is 800 VDC, a high-voltage power architecture NVIDIA, Google and Microsoft are working to establish as an open industry standard.

Key Watch Points: 800 VDC Power Architecture, Liquid Cooling, AI Networking & Optics:

As AI racks head toward megawatt scale, watch for progress on an open 800 VDC standard and vendor hardware, wider adoption of liquid cooling over air cooling, and optical interconnects, including co-packaged and near-packaged optics, to extend AI networks beyond copper.

Event: ASML earnings: the chip industry’s bellwether steps up

Why It Matters: ASML is the sole maker of EUV lithography machines, the tools Intel, TSMC and Samsung rely on for leading-edge chips. Its results offer a real-time read on the semiconductor cycle and how aggressively foundries and memory makers are pushing advanced nodes.

Key Watch Points: Guidance, 2027 Outlook, China:

After lifting its 2026 forecast twice, to €43–45 billion in sales and a 54–56% gross margin, will ASML raise it again? Investors will also parse its 2027 outlook for signs foundry and memory spending will last, and watch whether China, already down to 14% of 2Q26 sales, slips further amid U.S. restrictions.

Event: TSMC to report 3Q26 earnings, guided for record revenue of $44.6–45.8B while 2nm ramp and overseas expansion could weigh on margins

Why It Matters: With NVIDIA and AMD among its biggest customers, TSMC’s guidance is the clearest gauge of how long the AI chip boom can run. After adding $100B to its Arizona plan in July, investors will watch how tariffs and geopolitics, and rivals’ advances, shape its next moves.

Key Watch Points: Capex and Margins, Advanced Node/ CoWoS Roadmap, Possible Terafab Tie-up:

With another record quarter already guided, focus shifts to how higher capex (2026 budget of $60–64B, with 2027 plans in focus) and costly ramps will affect margins. Management has flagged 3–4 points of dilution from the 2nm ramp in 2H26, plus 2–3 points from overseas fabs at early stages. Rivalry with Intel continues on both foundry (A16/A14 vs. Intel’s 14A) and advanced packaging (CoWoS vs. EMIB), while a possible tie-up with Musk’s Terafab, which Intel has joined, adds another wrinkle.

No posts