$NVDA Jensen: "The $20 billion is for a standalone CPU. Remember....
...My sense is that we'll be supply constrained throughout the entire life of Vera Rubin."
$NVDA: "Vera CPU opens a brand-new $200 billion TAM for NVIDIA, a market we have never addressed before. Every major hyperscale and system maker is partnering with us to get it deployed.
We have visibility to nearly $20 billion in total CPU revenue this year, setting us up to become the world leading CPU supplier."
"GPT-5.5 was trained on the $ORCL Abilene supercomputer, part of our Stargate initiative and built on Oracle Cloud Infrastructure with $NVDA GB200 systems. A strong example of what deep partnership and ambitious execution can unlock."
Must watch.
$NVDA Bill Dally: "90% of power and data center is going to inference"
Many good stuff here. Watching it during my short lunch (gonna finish later)
Bible-level analysis from us on $NVDA
's Rubin NVL 72 and many other stuff, led by @thecopperwei and @dylan522p
Must read.
https://t.co/2zHD7AjFKx https://t.co/kKjenqKr4w
$NVDA mogs the shit out of tokens + TCO.
Done deal, I guess. Licensing approval for $NVDA H200 and $AMD MI325X. https://t.co/OwUM9nHagd
$NVDA CEO Jensen Huang on the future of AI.
- Frontier Models: Expects verticalization and domain specialization; major players with existing verticals ( $META and $GOOGL) will thrive, while open models will serve the rest.
- Next Breakthroughs: Focus on solving for infinite context memory and continuous learning for models
- Token Growth: Current 5x YoY growth was driven by reasoning; expects a future 50x inflection from agentic AI systems that use tools, planning, and simulation
$GOOGL TPU and $NVDA
2) Looking at it closely, $NVDA's contribution to SK Hynix’s revenue was approximately ₩4.7 trillion in Q1, ₩6.1 trillion in Q2, and ₩6.4 trillion in Q3.
The memory demand for AI so far show little sign of slow down (at least looking at it from $NVDA and SK Hynix).
1) SK Hynix disclosed that revenue from its largest customer—likely $NVDA —accounted for nearly 27% (₩17.3 trillion) of its total revenue (₩64.3 trillion) for the first three quarters of 2025.
Source: Company report https://t.co/UW9DrFbnA6
@mingchikuo:".....my latest supply chain checks indicate that the $NVDA VR200 NVL144 currently in development, testing, and assembly is using rails solely supplied by King Slide"
Key highlights from Wistron's Q3 call: "AI server demand remains very strong, with DOUBLE-DIGIT growth expected in Q4 2025 and throughout 2026. $NVDA
"Networking product shipments are projected to increase 10x next year (tho still a rather small part of its overall portfolio), and 2026 is expected to be a major inflection point for the networking business."
2) In Q3, we are seeing not only the ramp up of $NVDA AI servers, but also ASIC AI servers (roughly a 8:2 mix).
Rack shipment estimate for Q4 from sell side is going higher (should exceed some people's expectation back in Q3).
Hon Hai: "Cumulatively through the third quarter, AI server revenue has already reached NT$1 trillion ahead of schedule. With the gradual improvement of automation and testing systems, mass production of next-generation AI racks is expected to ramp up rapidly.
For the fourth quarter, AI rack shipments are projected to grow by a strong double-digit percentage quarter-on-quarter, aligning with our expectation of continued sequential growth in AI server revenue."
1) During the Q2 earnings call in August, Hon Hai guided for a 300% QoQ growth in AI rack shipments for Q3 (which was quite a impressive number to many back then). $NVDA
Today, the company confirmed it has achieved that target.
Hon Hai (Foxconn): "The AI industry continues to grow rapidly, driven by strong demand for AI computing power and the mass production of next-generation AI racks. This momentum propelled our overall AI rack shipments in the third quarter to achieve a 300% QoQ increase."
$NVDA CEO Jensen Huang revealed that the next-generation Rubin has already begun entering production, adding, “We’re now seeing Rubin on the production line.”
He also expressed appreciation for the significant capacity expansions by the three major memory suppliers.
When asked whether Samsung has begun supplying HBM for Blackwell, he simply responded that $NVDA has received latest HBM's samples from all memory manufacturers.
When asked about what $NVDA Jensen Huang discussed with him during his recent visit to Taiwan, Wei replied, “Jensen wants more chips.” As for how many chips Huang requested, Wei smiled and said, “That’s confidential.”
$TSM CEO C.C. Wei stated that the company has once again set new records for both revenue and profit this year, adding confidently that “not just this year, every year from now will set a new record.” $NVDA https://t.co/LLQsBpZAAY
2) Usually, I start with fundamentals. But I decided to just highlight the notable part of this call. So few things:
- Its October revenue is TWD 7B, ATH monthly revenue.
- Client SSD, Phison's controller shipment grow 72% QoQ and 380% YoY. Revenue grows 66% QoQ and 350% YoY. Facing a tight controller supply in October and see $TSM 6/7 nm supply to be higher in '26.
- Gaining share in both Client and mobile SSD. Now the hold biggest share in Android mobile storage. ("But now you see we gained a major share in the PC OEM, a major share in the mobile, and coming through in enterprise in AI..."
- OEM module grew by 2.5x YoY due to strong design winds and demand
- Mobile controllers up 65% YoY and 9% QoQ by unit, and revenue-wise is 120%.
- Acheived 60% global share in redriver supply (small TAM)
- Its Gen 5 retimer has entered $NVDA B300 supply chain and got "good amount of orders"
This is big if true. According to Taiwanese media reports, $NVDA additional AI chip orders using $TSM 3nm process are prompting TSMC to expand monthly wafer capacity from 100–110K to around 160K, representing a 45–50% increase. https://t.co/Srgoc5u5gC
$NVDA Jensen: “Why am I back? Well, business is very strong. So, I came back to encourage my TSMC friends (to work hard)… business is going very, very strong and I’m here to encourage TSMC. They’re doing a very good job for us and I’m here to thank them for all the hard work they’re doing.” $TSM
Hon Hai's (Foxconn) October 2025 results marked another record high, with NT$895.7 billion in revenue (+7.0% MoM, +11.3% YoY), driven by broad-based strength across computing, cloud/networking, and components. $NVDA
Computing Products benefited from new product launches and pre-holiday pull-in demand; Cloud and Networking gained from AI-related momentum; and Components and Other Products also rose sharply.
Smart Consumer Electronics remained flat MoM and slightly down YoY.
Overall, AI server demand and seasonal pull-ins drove strong top-line growth. This is expected to help Foxconn deliver a robust Q4 and potential record annual revenue in 2025.
KYEC is reportedly to be the only OSAT to secure advanced testing orders from $NVDA, $GOOGL, and $AMZN.
KYEC's AI-related business is expected to surpass 40% of total revenue once new capacity comes online in 2026. https://t.co/Pckrk1dnJU
" $ANET down $20 after earnings also doesn't help, among others, but ANET did lose some share back to $NVDA, where networking grew from $3B to 7.3B last 2 quarters after flatlining well over a year. Jayshree didn't really give the impression of denying that either.
ANET also isn't able to ship to demand due to ... memory constraints. And $AVGO switch chips. "Lead times on many of our components, including standard memory and chips and merchant silicon, and everything ... is ranging from 38 to 52 weeks. So, we are coping with that.""
$NVDA cooling supply chain partners:
"Morgan Stanley’s analysis estimates cooling costs for NVIDIA’s GB300 NVL72 racks at $49,860, rising 17% to $55,710 for Vera Rubin NVL144 due to denser GPU configurations and heightened thermal needs, with switch tray cooling costs surging 67%."
Interesting story by WSJ on Trump-Xi talks in Busan and $NVDA.
- Before meeting Xi, Trump weighed Nvidia CEO Jensen Huang’s push to allow Blackwell AI chip exports to China—a major potential policy reversal.
- Top officials, including Rubio and Lutnick, opposed the move, warning it would endanger U.S. national security while aiding China’s AI ambitions.
- United opposition from Trump's aide is reportedly making Trump decide not to discuss $NVDA chips during the meeting with Xi.
The US has allowed $MSFT to ship the latest $NVDA chips to the UAE for the first time.
$MSFT President Brad Smith: "$MSFT became “the first company to receive a licence under the Trump administration” to export Nvidia’s AI chips to the UAE.
$MSFT now plans to increase its UAE investment from $7.3B over the past three years, to more than $7.9B from 2026 to 2029, of which $5.5B will go on capital spending for AI and cloud infrastructure.
EMC Chairman Ding-Yu Dong: "EMC is currently operating at FULL CAPACITY AND ALL OUTPUT SOLD"
The company launched a three-year expansion plan last year, under which total capacity will increase by about 70% by 2027, with an additional 1 million sheets added in both 2025 and 2026. Output value will grow proportionally.
“This is how it will be for the next few years — full capacity, full sales,” he said, adding that the strong market demand continues and all of EMC’s production lines remain fully loaded.
In terms of product technology, EMC’s current flagship product line remains the M8 series, but Ding-Yu revealed that the next-generation M9 product has COMPLETED CUSTOMER QUALIFICATION and will begin ramping up production in the 2H26.
"EMC's M9 is the first to be qualified and will also be the first to enter mass production — just like our M8,” he emphasized, highlighting the company’s leadership in next-generation high-end materials."
Discussing AI development trends, Ding-Yu stated that data center and server demand will continue to surge, adding,
“It won’t be just $NVDA.” He expects that within three years, both laptops and database servers will move toward full white-label (ODM/OEM) models, with design and production increasingly dominated by ODM/OEM manufacturers.
He summarized EMC’s role in one sentence: “We are like cement — whether it’s a skyscraper or a thatched hut, everyone needs our raw materials.”
$NVDA VR200 NVL 144 midplane (M9) by EMC (and potentially Shenyi), Yes.
Doosan will supply M8 CCL for the compute board, while EMC/Shenyi (guess who holds the higher share) provide the CCL for the switch board (M8).
For CPX boards, it’s likely using M9 CCL from EMC/Shenyi (not sure if this is finalized yet).
The copper foil is expected to be HVLP4 grade.
The upstream material suppliers — glass cloth, resin, and copper foil — are probably a topic for another post.
6) The KV Cache offload is NOT a "NEW THING" (but attention on it has only increased over time this year), and in fact, many key players in the AI hardware supply chain like SK Hynix and $NVDA (check $NVDA Dynamo Distributed KV Cache Manager). https://t.co/4PJkm5y3yS
2) GF believes $TER is set to gain AI testing share, with $NVDA Rubin testers certified by 4Q25 (low-teens % share, shipments in 2Q26, Rubin Ultra mid-20s%).
It also claims that Mellanox and $AMZN shipments starting 4Q25, $AMD MI450 certification also in 4Q25, and early discussions ongoing with Meta.
1) Been following $TER for quite while. Share up 40+% since August 1st.
The note (#2) below from GF Securities caught my attention today (coverage initiated). Of course, one key development investors are paying attention to is $TER's potential entry in $NVDA Rubin supply chain (for chip testing).
So far, at least from public information, there is no straight forward evidence that $TER entered the supply chain, but rumors have been circulated for MONTHS.
But potential upside could be more than (see #2) just $NVDA (admittedly $NVDA still the biggest driver).
Earnings call on Wednesday hopefully offer more concrete details beyond previous GS CT conference, Citi TMT conference, and last earnings call.
Taiwan's AI server assembler Quanta has expanded production 8-TIMES already this year and it’s still NOT ENOUGH to meet insane AI demand. $NVDA
@elonmusk gave many interesting details on $TSM and Samsung, as well as $TSLA's AI chips during the earnings call today. Some takeaways (will add more details as I process the transcript).
1. $TSLA will work with both $TSM and Samsung on AI 5 chips (known).
2. In some metrics, AI 5 chips is 40 times better than AI 4....like, wow. Also love some details Elon gave (see #3).
3. Elon said its AI 5 dropped all the legacy components: no traditional GPU, no image signal processor. Those deletions let Tesla fit AI 5 on a half-reticle die with better signal margins between memory, AI accelerators, ARM CPUs, and PCIe blocks
4. In his view, he thinks technically the Samsung fab has slightly more advanced equipment than the $TSM fab.
"These (Tesla's AI 5) will both be made in the US, but in one TSMC in Arizona, Samsung in Texas"
5. Another interesting point is "oversupply." Elon believes even in there is oversupply of AI 5 chips, these chips can "always" be use in data centers, and combine them with $NVDA hardware.
Came across a note from KGI indicating that MediaTek’s ASIC revenue is expected to be lower in 2026 and 2027, which partly aligns with my view.
That said, I think 2027 might be somewhat optimistic (but depends on progress). Also, would prefer not underestimating MediaTek's ability to secure CoWoS capacity.
It’s indeed still unclear whether the previously guided $1B ASIC revenue for 2026 remains on track. We might hear more during the call at the end of this month.
"We believe $META AI ASIC project may be shifting to $MRVL, while $GOOGL v7e project has been progressing slowly.
The v7e has recently undergone another re-tapeout; although there has been considerable noise surrounding the project, we think the issues are still resolvable, albeit with slower progress.
In addition, since most of TSMC’s CoWoS capacity for 2027 has already been booked by $NVDA and $AVGO, even if the projects advance better than expected, second-tier ASIC vendors will likely struggle to secure sufficient capacity."
Inference explosion presented by $NVDA at OCP. https://t.co/GJovv8xy8w
It seems that some of the major opportunities in AI servers over the past 3 months are centered around substrates, CCLs, and PCBs. Drill is just another. $NVDA
AI server motherboard designs have increased from 20 layers to over 40 layers and now incorporate high-end CCL materials such as M8, significantly raising drilling complexity.
Industry sources note that previously, each drill bit could be used about 3,000 times, but its lifespan has now generally dropped to below 800 uses.
In the future, with the introduction of M9-grade midplane PCBs or ASIC motherboards, each drill bit’s usable count could shrink to as few as 100 times. In other words, producing a single AI motherboard will require several times more drill bits than before.
Demand for drill bits used in high-end AI circuit boards is growing far faster than the increase in PCB production value itself. Assuming PCB manufacturers’ annual revenue rises by 50%, the actual demand for drill bits may have already surged by 150%.
Sources indicate that China’s major PCB maker Shenghong Technology has begun large-scale drill bit procurement from Ding Tai, while Taiwan’s Topoint, which has a strong market penetration rate, is also observing clients aggressively adjusting their inventory strategies. If large manufacturers start stockpiling, supply-demand conditions could tighten further.
Drill bit suppliers have already shown early signs of price hikes. Industry sources note that adjustments are starting with lower-end “white bits,” while prices for high-end drill bits have yet to rise significantly. However, with new AI platforms ramping up in 2026 and material specifications becoming more advanced, broad-based price increases are highly likely. Moreover, since AI supply chains generally adopt only top-tier branded products, the market’s oligopolistic nature further strengthens suppliers’ pricing power.
Topoint has already launched another expansion phase, with its PCB drill bit monthly capacity expected to increase from 31 million to 35 million units in Q4 this year.
Looking ahead, drill bit manufacturers are positioned to benefit from both capacity control and enhanced pricing power.
According to an industry source, Foxconn has begun developing $NVDA Vera Rubin AI servers, with mass production and shipments planned for the 2H26.
Industry insiders revealed that Foxconn’s main shipments this year are focused on $NVDA GB200 AI server racks, which are scheduled for delivery completion by the end of this year. Meanwhile, shipments of GB300 racks began in September and are expected to continue through next year.
To prepare for the next wave of growth, Foxconn has already started developing NVIDIA’s next-generation Vera Rubin NVL144 MGX platform, with shipments projected to begin in the second half of next year.
KGI: According to our recent check, demand for $TSM CoWoS in 2026 has become even stronger compared with two weeks ago, mainly driven by $NVDA.
We estimate $TSM's CoWoS monthly capacity will expand from 70,000 wafers at the end of 2025 to 115,000 wafers by the end of 2026 (previously estimated at 110,000 wafers).
First time to $NVDA HQ, bruh this is beyond insane lol
Increasing power consumption across successive generations of $NVDA AI servers. https://t.co/BM4KSUEIbk
Jeffries says they found NO evidence that $TER enters $NVDA Rubin supply chain (referring chip testing).
Interesting.
$avantest https://t.co/e5i5kTUHWp
These incremental advancements are proving that Jensen's announcement of NVLink Fusion in May has been gaining momentum in the industry despite some initial doubt a few months ago. $NVDA
This guy's vision has been pretty spot on.
The bigger takeaways is $NVDA's NVLink Fusion ecosystem just add 3 new partners in less than 2 months—GUC, Intel, and Samsung.
It's Jensen's world, we are living in it.
Another W for $NVDA on networking (but honestly kinda expected). Ethernet? No problem.
“NVIDIA today announced that $Meta and $ORCL will boost their AI data center networks with NVIDIA Spectrum-X™ Ethernet networking switches.” https://t.co/AdDaFhETf1
AMD's deal now looks even more interesting comparing it with $NVDA and $AVGO.
$NVDA CPX and VR200 NVL144 are expected to feature a “cableless” design with midplane and CX9 board for both .
The "backplane" to be adopted in Kyber rack in 2H27.
Some incremental development in the $NVDA GPU and $GOOGL TPU, noted by GF Securities. VR 200 Switch tray upgrade should be well-known already.
1) PCB content in VR200 to increase to $1050/GPU vs $400/GPU for GB200
2) Shengyi Tech has gained market share from EMC and is now the largest supplier for $NVDA GB300 switch tray. Shengyi also enters $GOOGL CPU supply chain.
3) EMC is now the only supplier qualified for M9 for VR200, and gaining share from Panasonic for $GOOGL TPU v7
$NVDA: "$3-4 Trillion AI Infrastructure
Spend by 2030" https://t.co/asT9NUNldX
One interesting detail in the report is that the deal appears to include Samsung’s potential purchase of around 50k $NVDA GPUs (we'll see how this plays out).
It’s fascinating to see the concept of “compute financing” now extending into the memory space.
Fun time. https://t.co/CVlDJyBER8