"If all goes well, TSMC is aiming to start mass production of the glass core substrate in 4Q28–1Q29, to match the cadence of Nvidia's AI chip iterations."
👀
Breaking down TSMC's glass core substrate slide
On June 11, at JPCA Show 2026 in Japan, TSMC gave a roughly 40-slide presentation titled "Advanced Packaging Technology Essential to the Evolution of AI" (AIの進化に不可欠な先端パッケージング技術). One slide from the deck, titled
停止されているFebleについてかなり急展開ですが、韓国で開かれたAnthropicのイベントで関係者が「数日以内に(in the coming days)使えるようになると確信している」と発言していたようで、意外と早くFableが戻ってくる可能性。
トップのアモデイがG7でトランプや閣僚と直接話した結果かもしれない
A dinner at Versailles just became the setting for a major diplomatic breakthrough.
President Trump personally signed the Iran memorandum of understanding Wednesday while meeting with French President Emmanuel Macron, a White House official confirmed.
The agreement, which Iranian President Masoud Pezeshkian also signed, is now in effect to end the war and cancels the formal signing ceremony that had been scheduled in Geneva.
J.P.Morgan: Asia Passive Components, PCB/CCL, Substrate, and Testing
Passive Components:
> NVIDIA's upcoming Rubin platform will almost double the high-end MLCC consumption per compute tray compared to the Blackwell generation, alongside a higher blended ASP.
> Production resource consumption per highend MLCC unit is expected to double between 2025 and 2028 compared to lowend MLCCs.
> The supply/demand balance for low-end components is projected to plummet from a slight surplus in 2025 to a severe -15% to -20% shortage by 2028.
> High-end units only make up a small fraction of total volume (growing slightly from around 15% to nearly 20% by 2028). However, the "production resource consumption" chart reveals that by 2028, making those complex high-end units will eat up over 65% of all manufacturing resources.
> After bottoming out heavily around 2019 and staying sluggish for years, factory utilization for low-end parts for Yageo $2327.TW is coming back. It is climbing steadily back toward 90% through 2027 to meet the squeezed market demand.
> ASP for low-end MLCC is projected to jump nearly 45% by 2028 compared to the 2025 baseline.
> Yageo's MLCC revenue is projected to nearly double from 2025 levels to roughly NT$46 billion by 2028, pulling gross margins up toward a highly profitable 50%.
> Yageo holds massive, top-tier market share slices in both Tantalum and Resistors.
> Yageo's tantalum ASP will rise faster than resistors through 2028.
> For Yageo, by 2027, server/HPC applications are expected to make up roughly 45% of tantalum revenue and 35% of resistor revenue, dragging the total company-level exposure up significantly.
PCB/CCL:
> HPC CCL market is expanding at a massive 60% two-year CAGR, ballooning from $3.89 billion in 2025 to $9.98 billion by 2027.
> "Major AI server projects" are the undisputed locomotive here. They are growing at a 79% CAGR, jumping from $1.69 billion to $5.39 billion to command 54% of the entire market mix by 2027.
> Legacy materials (M7 and below) are seeing their market share collapse from 59% down to just 27%. In their place, M8 grade is becoming the dominant mainstream specification, growing at a 102% CAGR to reach $6.49 billion. Meanwhile, next-gen M9 material enters the frame, exploding 214% YoY in 2027 to capture an 8% mix as advanced platforms ramp up.
> CCL dollar value per chip steps up from ~$100 on Hopper to ~$200 on Blackwell, before surging to over $400 for Rubin. When moving to the Rubin Ultra generation with a full backplane configuration, the dollar content per chip skyrockets to nearly $800—an 8x increase from Hopper. Custom silicon is also driving this trend, with Google's $GOOGL TPU v7 and Amazon's $AMZN Trainium 2/3 hovering between $300 and $400 per chip.
> An H100/H200 rack requires roughly $2,000 worth of CCL. The NVL-style GB200/300 architecture bumps that to ~$10,000. Vera Rubin hits ~$30,000, and a Vera Rubin Ultra rack with a backplane configuration absolutely explodes past $200,000 per rack—a massive 100x increase in dollar content compared to a standard legacy Hopper rack.
> As networking speeds migrate from 100G up to 1.6T, the required PCB thickness and complexity scale dramatically, pushing average layer counts from ~20 layers all the way up to a 40-50 layer range for 1.6T architectures.
> As boards become thicker and more complex (scaling up to 40–50 layers), manufacturing fabrication yields are collapsing. Yields have plummeted from a healthy ~90% in 2022 to an estimated ~65% in 2026, and are projected to dive below 60% by 2028.
> NVIDIA's $NVDA Vera Rubin generation leverages M9 grade materials for its HGX structure UBB and LPU/LPX boards. The upcoming Vera Rubin Ultra generation pushes the limits even further for its midplane, LPU/LPX, and backplane architectures, introducing the potential to migrate entirely to M10 grade materials.
> Key premium suppliers for CCL—Elite Material Co. $2383.TW, Taiwan Union Technology Corp. $6274.TW, and Doosan $000150.KS.
> YoY demand growth is vastly outstripping supply growth. In 2026, demand growth peaked at over 60%, while supply growth crawled along at under 20%. This massive, widening gap confirms that supply could severely undershoot demand over the next two years.
> Intel server CPU shipments are projected to continue shrinking from 22.0M in 2020 down to 12.5M by 2027e. Meanwhile, AMD $AMD shipments are steadily climbing from 2.0M to 7.5M over the same period.
> Intel $INTC PCB ASP scales from $40 in 2020 to $322 by 2027e. AMD platforms command an even higher premium, with PCB ASP jumping from $42 to $400. This pushes the total Blended ASP up 11% to 22% YoY through 2027e.
> Server PCB TAM is set to more than double from $1.69B in 2024 to $3.46B by 2027e (growing 36% in 2026e and 29% in 2027e).
> Server CCL TAM (After Yield) mirrors this trajectory perfectly, compounding from $0.63B in 2024 to $1.28B by 2027e (assuming a steady 90% manufacturing yield baseline for these non-AI boards).
> For EMC, total infra revenue is projected to climb from ~$2.3B in 2025 to over $4.7B in 2026e, and ultimately spike close to $7.5B by 2027e. The largest individual expansion drivers are NVIDIA, Google, and generic switching/networking architecture.
> For standard M7 and below materials, EMC controls roughly 35% market share. That surges to ~73% for M8. By the time you reach ultra-low loss M9 materials, EMC commands a near-total monopoly with close to 100% market share.
> EMC's estimated 2026 market share across specific tech stacks. 38% for NVIDIA, but skyrockets across major Big Tech custom silicon and networking plays: ~66% for Google TPU, ~70% for AWS Trainium, and ~74% for high-speed Switch trays.
> While peers like TUC, Doosan, and Panasonic hold steady footprints, EMC's aggressive expansions mean it will single-handedly account for 50%+ of all global high-end CCL capacity starting in 2027e.
> Substrate CCL is rapidly emerging as a major secondary growth vector. It is projected to climb steadily toward the $3.0B to $4.0B range by 2027e.
Substrates:
> By 2028e, the demand curve forces the theoretical utilization rate up toward a staggering ~135%. Because physical factories cannot run above 100% capacity, this gap represents an unprecedented structural shortage where demand vastly outstrips the world's physical manufacturing footprint.
> In 2024, AI server & switch substrates made up just 26% of the global demand mix, while consumer applications dominated at 40%. By 2028e, that dynamic completely flips: AI server & switch will consume a staggering 75% of all global substrate demand, while consumer electronics wither away to just 11%.
> AI Server & Switch Hyper-Growth is growing at a massive 67% CAGR from 2025 to 2028. Volume demand jumps from 2,448 bn nm in 2024 to an incredible 21,522 bn nm by 2028e.
> Hopper sits at a baseline total size of 38,280 bn nm (12 layers). Blackwell steps up the complexity to 82,782 bn nm (14 layers). Rubin nearly doubles Blackwell at 144,918 bn nm (18 layers), while Rubin Ultra explodes to 322,040bn nm (20 layers). Feynman pushes boundaries even further, hitting 432,000bn nm with a staggering 24-layer substrate architecture.
> Unimicron $3037.TW and Ibiden $4062.T maintain their dominant positions at the top of the stack. Unimicron's capacity scales from 73 in 2024 to 117 by 2028e, while Ibiden expands from 65 to 107 over the same timeframe.
> Premium T-glass market is entirely cornered by the industry's two main titans. Unimicron captures the lion's share, securing 35-40% of total allocation, while Ibiden secures 30-35%. Everyone else is left to scramble over the remaining third. This secures Unimicron's raw material advantage during a time of extreme scarcity.
> J.P. Morgan's margin projections for Unimicron from 4Q24 through 4Q27e, driven by structural under-supply, Unimicron's core Substrate margins are projected to climb from ~20% in late 2024 to above 50% by 4Q27e. This carries the blended company margin up to ~40%.
> Chroma ATE $2360.TW overall semi/photon segment revenue jumps from around 10,000 NT$mn in 2025 to over 32,000 NT$mn by 2027E. The primary engines of this surge are SLT (System-Level Test) and Optical Comm testing, both expanding at a 100%+ CAGR. While optical comm YoY growth drops sharply after a 2026E peak, SLT YoY growth remains exceptionally strong, overtaking other legacy segments like Metrology and standard final test (FT).
Read more here: kc9rsmb8u8.feishu.cn/file/UIQybxDfx…
268 Followers 7K FollowingActivist short selling focused on frauds and promotes. Presume all tweets reflect short positions and biased opinion. Verify independently. Not inv advice.
1.4M Followers 2 FollowingWe're an AI safety and research company that builds reliable, interpretable, and steerable AI systems. Talk to our AI assistant @claudeai on https://t.co/FhDI3KQh0n.
2K Followers 270 FollowingJust a little goblin looking for the golden pot sharing information, discussing opinions, trading US and EU shares & options. No recommendations #DYOR
6K Followers 36 FollowingI research bottleneck stocks, hidden suppliers, and physical infrastructure constraints before they become consensus.
Semis · packaging · optics · power · NFA
84K Followers 206 FollowingActivist short selling focused on frauds and promotes. Presume all tweets reflect short positions and biased opinion. Verify independently. Not inv advice.
3K Followers 605 FollowingFormer energy balances geek at the @iea and energy analyst at the @ief_dialogue behind @JODI_data, now with @KPLER in quest of greater energy data transparency.
38K Followers 87 FollowingInvesting into the AI buildout ⚡️ 500% YTD
Head of AI | Product Manager 10+ years
DD and full thesis in https://t.co/4NlE3Jd7E7
NFA