​Back in May after trade discussions with China in Geneva, it was stated by both the President and a WH officials that a preliminary framework was reached with China concerning rare earths.  While the details were never revealed, the objective was to remove a number of export restrictions that the Chinese government had placed on the export of rare earth materials, particularly magnets containing rare earths.  Those trade restrictions were implemented by China in response to tariffs the administration had placed on Chinese goods entering the US.  The implication was that China would be resuming shipments of these materials to the US based on the US lowering tariffs on Chinese goods. 
It seems those negotiations did not produce the desired results as last night China reinstated a number of those trade restrictions, requiring a case-by-case review for and Chinese entity trying to export a number of rare earth materials, including the equipment needed to mine, process, or refine the materials themselves, and added a few new ones.  Specifically singled out were magnets:

• Neodymium/Iron/Boron magnets, which are the key component in hard drives.  They are also present in speakers, smartphones, and wind turbines.
• Samarium/Cobalt magnets which are used in high-temperature applications such as aerospace, military and medical
• Cerium, which is not a primary magnet material but is a primary component of Nickle Hydride batteries used in hybrids and portable tools.

The Chinese rules seem to indicate that all requests for export licenses for these materials will be denied if the destination has any overseas military implication or if the intended recipient is on China’s own control or watch lists.  That is an indication that these rules are in response to the US extension of its Entity List to include all subsidiaries of any company on the list.  Along with the rare earth restrictions, the Chinese edict includes restrictions on all logic chips 14nm or smaller and all memory chips with 256 layers or more.  Many of the restrictions go into effect immediately with the remainder starting 12/1/25.
While some of this could be Chinese and US posturing before the proposed meeting between President Trump and Xi Jinping, leaving some room for concessions, the necessity for rare earth materials and magnets is profound in the US with a number of major industries critically dependent on a constant supply .  As China is the primary source of US rare earth imports (2024 – 74% by value & 84% by volume) China has substantial leverage, and while it seemed, according to the WH, that a deal had been reached previously, it does not seem to have resonated with the Chinese government who has now more closely linked rare earth trade to China’s objectives in the semiconductor space.  The more the US limits China’s access to advanced semiconductors and tools, the tighter the limitations on rare earths become. 

It seems that its time for Jeff Bezos to retire.  He has been very successful and changed the way the world shops, but according to tech press pundits, it’s over.  It all because ChatGPT (pvt) announced “Instant Checkout” last week that some imply will put Amazon (AMZN) out of business.  Behind the hype is OpenAI’s ‘revolutionary’ application, which allows the user to ask the chat to recommend an item, and (drumroll) purchase it without leaving the chat!  No more does one have to open another window to buy the product, they just have to press the ’BUY” button in the chat and of course, pay for it, without all of that additional stress of opening a new window (“Having to do all that work makes me very sad and nervous!”)

Currently Amazon has 1.9m active sellers, 9.7m 3rd party sellers and, by some estimates almost1m 1st party sellers that operate under Amazon fulfillment.   ChatGPT has Etsy (ETSY), and while the make-up of the merchant base of Amazon and Etsy could not be more different, Etsy is thought to have over 5.6m active sellers.  If you want to buy a TV you can on Amazon but not on Etsy, unless it’s a home-made TV, but if you want to buy a sweater for a German Shepard with the word “Woof” on the side, you would go to Etsy, so the difference is night and day.
On Amazon, you have some idea of the criteria used for product placement and ranking, which are essentially recommendations; Sponsorship, reviews, and some mysterious Amazon algorithm.  ChatGPT is a black box, which makes it considerably more difficult to know if the recommendation is tailored to you or to another agenda, but be that as it may, in either case you have to make the final choice.  That said, you can buy as many items as you want on Amazon, while the ChatGPT service is single item only, at least for now.  ChatGPT is also working with Shopify (SHOP) to add their sellers to the service through Stripe (pvt), but as yet Shopify is in the ‘supported’ mode and not quite available to the public through the chat.
From the merchant’s standpoint the big selling point is ChatGPT’s ~800m MAUs, an attractive proposition, although more than likely all of those users have access to the vendor’s website directly, so it is hard to value that connection.  Vendors will also have to find their way in the ChatGPT recommendation ranking to become meaningful, and that is a difficult job, especially against a hoard of other vendors.  If a vendor knows it is rarely or never going to be the ChatGPT recommendation, its almost not worth the sign-up, or the ‘small fee’ that OpenAI will charge (unpublished amount) the vendor if the transaction is completed.
We don’t mean to be critical, but just because OpenAI developed the application doesn’t mean it will change the world.  Not everybody huddles into a ball and hides in a corner when they have to do something like opening a 2nd window on their phone and Amazon makes it very easy and convenient to make purchases, so we shouldn’t expect Amazon to shrivel up a die a quick death, especially if Amazon is able to develop its own version of the Instant Checkout app, or even decide to absorb the ‘small fee’ OpenAI charges and become an Instant Checkout vendor.  The OpenAI product is once again, a convenience (aka ‘productivity improvement’), not a game changer, and a good agentic system should be able to improve your shopping workflow even better.
Here are the instructions you give you Ai agent to do the same thing:
“I am interested in purchasing a 65” TV.  I am not sure whether to buy an LCD TV or an OLED TV and I don’t know what brand to choose.  Can you do the research, make 5 recommendations, listing the pros and cons of each.  When I make the choice, please go to the site and fill out the purchase form.  I will enter the credit card info and you can execute the purchase and send the receipt to my e-mail.  The price should be between $900 and $1600 and it should be deliverable within 5 days.”
 It will do a better job on the research….

Panel sales out of Taiwan for September were strong as expected, but that only tells a part of the story as it was well-known that on a general basis brands were filling inventory levels in anticipation of the upcoming holiday season.  That said, October demand for CE product in general has been weak and we already mentioned that a number of major panel producers have scheduled utilization cutbacks for this month in order to tighten supply/demand balance.  This lessens the value of the September and 3Q data as we expect a change for October.   Here’s the September and 3Q sales data for the three Taiwanese panel producers:

Based on 5-year averages, October is typically a down month (m/m) for the two large panel producers and an up month for Hannstar, as they are primarily a small panel producer and September is typically (5-year average) the month with the highest number of smartphone releases.  Here’s what October panel sales typically look like for the three Taiwanese panel producers, so we are not expecting October to be better than the averages based on the utilization rate cuts and weak demand.

Samsung (005930.KS), as the leader in both OLED display and semiconductor memory production, is the envy of competitors and the object of those wishing to take a shortcut to success by stealing sensitive information from the company.  Most recently three former Samsung employees were indicted in South Korea for leaking Samsung’s DRAM manufacturing process to Chinese semiconductor memory producer ChangXin Memory Technologies (pvt), China’s largest domestic DRAM producer.  We note that Changxin is a state controlled entity with the Hefei Industrial Investment Fund and the National Integrated Circuit Industry Investment Fund (aka ‘The Big Fund’)[1] as the largest shareholders.
The indictment indicates that the employees were recruited by CXMT to join a team that was verifying and testing DRAM manufacturing processes and were promised salaries that were up to 5 times higher than what they were earning, up to $2.14m/year according to the court documents.  The processes they were testing were allegedly stolen from Samsung by the Samsung employees and passed on to CXMT. 
This is not something new for Samsung, despite its numerous precautionary measures, which include HDAs and non-competes, no data transferred to flash drives, strict physical access control, the disabling of smartphones upon entry to all Samsung facilities, and most recently a limit of prompt size to 1024 characters for AI (after an accidental leak through ChatGPT (pvt).  Here are a few other Samsung breeches.

• 9/2024             Two Samsung execs accused of stealing 20nm DRAM technology and setting up a company in China with the data – Value $3.2b
• 7/2024             Samsung Display engineer convicted of stealing optical system plans for laser annealing and OCR tech to set up competing company – Value $24.5m
• 6/2023             Samsung exec and 6 employees accused of stealing factory blueprints and clean room designs. Attempted to duplicate the factory in Xian, China – Value $233m
• 2/2023             Seven employees of SEMES (pvt), a Samsung equipment affiliate, stole blueprints and component lists for a semiconductor cleaning tool.  Formed company and sold 14 units to Chinese customers – Value $60m
• 11/2021           Samsung researcher photographed 17 automated factory operation and digital display IP documents with a phone – Value $180m
• 11/2018           Executives of Toptec (108230.KS), a Samsung supplier, sold 3D lamination tech for flexible OLED displays to a Chinese company – Value $134m

What makes the most recent breech interesting is that CMXT announced in 2023 that it had become the first Chinese semiconductor company to produce 18nm DRAM, which was seen as a big step forward for the Chinese semiconductor industry.  The prosecutors in the current case, which originally took place in 2018, allege that CXMT’s technology announcement was not based on their own development but the DRAM process that was stolen from Samsung,  It’s hard to prove such an allegation without hard data from CXMT, which might be difficult to come by, but be on the lookout for a former employee of Samsung named Park, currently the subject of an Interpol Red Notice, who was suspected of copying hundreds of Samsung’s DRAM process steps for CXMT.


[1] ‘The Big Fund’  is a mandatory investor in all strategically important Chinese semiconductor projects.  It is financed and controlled by the central government.

Over the last two years we have droned on about both how stale many consumer electronics products have become and how brands look only to one-up rivals with a larger camera or higher resolution screen, rather than create something new that solves a problem..  AI has made things worse in that model builders have been primarily interested in how many parameters a model could have or the size of the prompt line, leading to more expensive models (training costs), more infrastructure to train those models, and more ‘conversational’ models. 
Model developers were successful in making models bigger and making users believe that the models were ‘thinking’ about how to go about answering questions, but at the same time they created sycophantic models (“Wow, what an intelligent question.  You got right to the heart of the matter.”) that still provide incorrect answers or even try to game the system into letting them off the hook.  Model builders did expand generative AI’s reach and  now we are on the verge of agents unburdening us from some mundane tasks (with reservations), but we have been waiting for the new killer AI application that can really make the average person look at AI as a real assistant, one whose hand you do not have to hold every time you ask a question.
We have two favorite AIs.  Gemini (GOOG), because it has access to Google’s search index and Perplexity (pvt) because it seems able to come up with more informative answers with less having to modify the question a number of times.  That said, neither are truly good at research, other than relatively simple Q&A.  Here’s why:
There is lots of data available, but much of it is unstructured, meaning it consists of many data types, images, text, binary, numerical, or video.  This makes it difficult for AIs to understand the data and its context unless it has been broken down  into digital format during training.  Unfortunately data on websites is not processed and typically exists as ‘pages’ that contain said unstructured information.  When asked to ‘read’ that information, the AI ‘reads’ the page but is unable to move to the next page.  Think of a list on an internet site.  You scan the list visually, but the page only goes halfway through the list, so you press ‘next’ and the remainder of the list appears (essentially a link to another page).  As the AI has no fingers, stylus, or mouse, it can only go so far and will likely report back that either it is unable to access the entire list, or it will only use the data which it is able to interact with (Page 1).
Of course, industrial or corporate AI can be programmed to access internal information, but outside (internet) data remains difficult for generative AI to navigate, so when we ask Gemini, “How many OLED monitors are expected to ship in 2026?”, it can only answer if the data is contained in text and cannot scan a database or data file for the actual information.  In cases like that, the user has to feed the data, sometimes page by page, to the AI in order for it to be processed, which is quite time consuming and prone to errors.
Yesterday Perplexity released ‘Comet’ a browser with an agentic AI built-in.  When you ask it a question like, “How many AR glasses were released in 2024?” and point it to a data source (it can find one if you don’t have one), it goes to the site, reads the first page, and essentially clicks the ‘next’ button for you until it cycles through all of the data, and then creates the answer.  This is very different from what other AI’s would do, which is search through documents (pages) that reference AR glasses that have been released this year, or have you copy each page from the database to the AI to get it to answer the question correctly.
While there is a significant improvement over other generative AIs, there are still some issues, the main one for us is the search engine that COMET uses.  This is a proprietary search index created by Perplexity, and while we find it the best of non-Google indexes, it seems a bit less robust at times.  In using COMET we have not found that the Perplexity index was unable to find specific information or sites, but if the COMET browser was also able to access Google’s index, it would assuage any concerns we might have as to internet reach.
All in, to us, this is a major step forward in the practical side of AI.  It’s great to be able to be conversational with an AI in order to express subtleties and nuance but giving the AI the ability to perform what is typically a human physical task to complete a project is a bonus.  Microsoft MSFT) Edge with Co-Pilot has some similar capabilities but nothing like COMET’s button pushing agent.  Perplexity has taken a step toward making AI more practical to the average person.  Still not in the ‘Save the World’ camp yet but at least making our world a bit easier to manage day-to-day.

AI hype continues to dominate the technology press and there is little to slow down this speeding AI locomotive as one AI company after another quotes massive gains in the number of active users they garner each month.  MAUs, or monthly active users, is a KPI (Key Performance Indicator) that is tossed around like it has monetary value, as if a MAU were a paying subscriber, but they are not.  What a MAU actually is as follows:
MAU

• A single person
• Counted once, no matter how many times they use the AI each month
• If they log-in or just visit the AI they can be counted, although this is a very weak metric
• If they send a prompt or query they are also counted, a much stronger indication of usage, as would be the case when using specialized Ai tools like image creators.

This is where things get subjective as there are no real standards or regulatory bodies that keep track of MAUs.  Many social media application use MAU metrics, as do AI developers, but whether a user that just opens the site counts as an ‘active’ user remains a question. Websites can also report MAUs, but in the SEO world impressions and page views are the real metrics.  A better metric for AI sites would be queries as, in theory, a developer could count anyone using their AI embedded in a 3rd party application as a MAU if they use the application at least once in a month.  It’s a bit of a stretch but technically they are an ‘active’ user.  You see how this can be a difficult metric to pinpoint.  So, while MAU statistics are an indicator of interest or curiosity in a product, it has little to do with revenue.
When it really gets down to it, given the valuations that the VC market has placed on AI companies, paying customers and how sticky they are is the real metric and that data is much harder to come by, particularly as individual consumers are able to pay monthly and cancel at any time.  Corporations are a different animal as they tend to look at AI from an organizational perspective.  Large companies look at AI as a tool, hoping to see a ROI across a large number of company AI users, so they represent a more stable income source and a stickier one, as long as they see ROI.  Corporations also pay for at least some AI infrastructure, particularly applications, while individuals tend to use AI on-line or through their phones, with little income generation outside of a monthly fee, if any.
So what really matters is the breakdown of income, which we believe makes MAU less relevant, but in the world we live in today, the average CEO has the ‘got to have it’ complex that says if you don’t have AI you have fallen behind your competition.  It is almost irrelevant whether you are gaining anything by having it, but if a competitor says they are making great strides because of AI, your board members will be asking why you have been so slow to join the real AI world. 
What makes this even more difficult is the subjective nature of AI evaluations.  There are measurement tools that calculate cycle time reductions for various worker functions, Average Customer Handle Time tools, Accuracy Analyzers and even Content Creation Volume measurement tools, all trying to pinpoint gains  generated by AI.  We certainly consider these tools valid as they have (hopefully) no bias toward particular results, but it is very hard to measure the overall effect AI has on productivity and even harder to measure how those changes affect ROI.  That is why we often  look to surveys whenever possible for a more personalized look at AI ROI.
There are hundreds of surveys, white papers, and expert reports on the benefits of AI, and a never-ending stream of technical journals that point to how AI is fundamentally changing us as human beings, but like product reviews on Amazon (AMZN), we gravitate toward the ‘one-star’ reviews to see what it is that might upset the apple cart in this brave new Ai world we live in.  We are not doing it to be glass half empty people, but to get a bit of a more realistic view of AI from those who either use it or pay for it.
A recent survey from Atlassian, an Ai company dedicated to removing corporate knowledge silos by creating AI-based enterprise search systems, revealed some interesting facts as to how AI is used at the enterprise level and its effectiveness and ability to generate a return on its investment. 
The bottom line is that AI does improve productivity, but by focusing on personal productivity over innovation the survey indicates that the Fortune 500 could lose $98 billion in lost returns on their AI investments.
Here's the data:
The survey classifies respondents into 5 categories:

Figure 1  shows that those with a positive view of AI’s capabilities saw their attitude improve this year (Type 2 – Type 4), while those with a more negative view (Type 0 – Type 1) saw that negative view decrease y/y, for an all-around improvement in AI perception.  However, while there was improvement y/y, what employees did with that extra time is telling, as shown in Figure 2 and even more so for executives in Figure 3.  Efficiency is certainly improved but executives see little other improvement for themselves and similarly for the organization.  In fact the detailed data shows that few executives see any transformational improvements due to AI.

• Only 3% of executives see an improvement in operational efficiency
• Only  2% of executives see an improvement in work quality
• Only 4% of executives see an improvement in innovation
• Only 4% of executives see an improvement in the ability to solve complex problems

When executives consider AI and how it affects internal teams rather than individual employees 37% said AI has wasted their team’s time and led them in the wrong direction, even though 76% said they saw an improvement in employee productivity.  Taking that a step further, the detailed data shows that organizational focus on personal productivity leads to an outcome that is 16% less likely to drive innovation, according to the respondents.    

​We were a bit surprised at the survey data’s negative take on AI across the organization, but we also understand that the objective of the study was to show how a singular focus on productivity improvement has limited results and bypasses how AI can help an organization become more innovative, something the company specializes in, so there is some inherent bias in the data.  However we were very surprised at how little executive thought of the employee productivity gains, and that they did little to improve corporate level innovation and ROI.
All in, the pressure managements feel to use AI seems to lead to metrics that support AI as a way to improve employee efficiency but does little toward the creative side of the business, which is essential in such competitive markets.   While the efficiency improvement is commendable, the lack of improvement at the corporate level changes the perceived ROI of large AI investments, at least currently.  Perhaps AI is missing the tools needed to afford the spectacular improvement across the corporation that model vendors are promising, but as yet improvements have been a little less than expected.

Smartphone brands are constantly trying to adapt to what is an extremely competitive environment.  Each year almost every brand has at least one major model release, and those with more extensive lines, can have two or three, with a number of brands also developing sub-brands  to maintain a certain brand image, while broadening their reach with consumers.  For instance Chinese smartphone producer Xiaomi (1810.HK) produces it own ‘Redmi’ smartphone line but also produces the ‘POCO’ brand of mid-range smartphones and the ‘Black Shark’ line of high-end gaming phones.
As consumers’ wants evolve, brands must adapt and offer a more or less diverse line of phones to broaden their customer base.  In some cases this means segmenting price points or feature sets, but shipments and share are the goalposts in the smartphone business and new models must perform or they drag down the brand and eat into profits.  The planning for new models is a complex issue as it rests on both the expertise of the brand’s marketing department and the company’s technical ability to satisfy evolving customer needs.  While brands can test new devices internally, using staff and focus groups, given the competitive nature of these products and the number of folks trying to get a peek before release, much depends on the expertise of the internal organization and even the most experienced product developers sometimes miscalculate, as it is very easy to convince oneself that such decisions are correct when your job depends on it.  The table below shows how Samsung’s Galaxy line models have evolved over the years.

Last year, years of progressively disappointing sales of the Galaxy S Plus series pushed Samsung to develop the Galaxy S25 Edge with the intention of substituting the Galaxy Edge for the Galaxy S Plus in 2026.  Unfortunately the sales of the Edge, which was a modified Galaxy S25+ that was 1.5mm. thinner, 27 grams lighter, had a bigger main camera, and a bigger battery for $100 more than the S25+.  Expectations were high pre-release, but early sales were disappointing, and Samsung seems to have set very low production volume for the remainder of the year for the S25 Edge, especially relative to the Galaxy S25+, which it was thought to be replacing next year, and the top Galaxy sellers, the S25 and the S25 Ultra.
According to industry sources here are the expected production volumes for the period between September and December:

• Galaxy 25 Edge                         300,000
• Galaxy S25+                               500,000
• Galaxy S25                                 2,900,000
• Galaxy S25 Ultra                       3,400,000

This substandard performance for the S25 Edge seems to have put the Galaxy S26+ program back into next year’s development lineup.  The question is whether Samsung has enough long-term confidence in whether the S26 Edge will sell more if the S26+ is not released.  Right now it seems they are hedging their bets by developing all four models, but it seems counterintuitive to release both the S26+ and the S26 Edge next year as they are quite close in price and function. 
Actual mass production for the displays begins later this month, although both the Galaxy S26+ and the Galaxy S26 Edge will use the same display, making early orders less dependent on a final decision.  That said, Samsung usually announces the upcoming Galaxy S series line in late January to early February, so a decision does need to be made soon.  The default would be to produce all four, but with the company trying to squeeze as much margin as possible out of the mobile division, three would likely be more profitable if they could boost sales for whichever model is chosen.  As both the S+ and the Edge sit between the Ultra and the standard Galaxy S, price will be more difficult to maneuver, which means features will likely be the selling point without overshadowing the Ultra.   No pressure…

Buying a TV is a chore.  TV brands list dozens of features many of which are esoteric and undecipherable for the average buyer and having sold TVs and stereos many years ago, one learns that there are many tricks retailers use to ‘enhance’ the look of in-store TV demos. Salespeople range from overly technical  (an 82 year-old grandmother, replacing a 24” CRT TV does not need to know about the quality of the upscaler) to overly folksy (My cousin bought this one and his wife loves it), with fiscal influences always in the back of the salesperson’s mind.  Even in stores where salespeople do not work on commission, memos about certain brands with ‘more attractive margins’ tend to influence recommendations and don’t always jive with what might be best for the customer.  We are not saying that salespersons are misleading, but bias is not easily controlled and can be a silent influencer.
Doing research on TVs for the average consumer usually consists of finding one or two sites that list ‘the best of…’ and jotting down a few names and model numbers, or (hopefully not) reading product reviews on Yelp (YELP) or Amazon (AMZN).  Most disconcerting to us are the TV ads that point to ‘Huge discounts’ or ‘big price drop’ that are meant to give the consumer the impression that they need to buy right away or they will lose those discounts.  The common practice is to always state the current ‘discounted’ price against the MSRP (Manufacturers Suggested Retail Price).  This tends to be the price point of the model’s initial release, which in most cases is the highest price the TV will ever see, so a ‘30% price drop’ rarely means a drop from wherever the set was priced most recently, but more likely a 30% drop from the MSRP.  Technically, despite how misleading it might be, this 30% discount could put the set at a higher price than its previous price point but it is still considered legitimate advertising (read the small print).
All of this makes one wonder  how much timing really matters when buying a TV, as it seems many influences can move prices up and down in an almost random pattern, yet nothing can be further from the truth.  TV pricing has a very specific yearly trajectory and pattern, which consumers should understand in order to make the ideal TV purchase.  Here are a few points to remember:

• Don’t buy on product release
• New TV models are typically released in 2Q, which means last year’s models tend to be the most attractively priced in 1Q
• Holidays are your friends when buying a TV
• Short-term discounts tend to be less important than the time of year
• Lowest price points for TVs (Premium) occur between day 300 (~October 28) and day 330 (November 26).
• TV set prices can go up quickly in December

We have been tracking all models of Samsung’s (005930.KS) Premium TV line (8K & 4K Mini-LED/QD, QD Only, and both QD/OLED and WOLED) since 2021 and the data proves this out.  So unless you are looking to buy a set for a specific event, it makes sense to have patience.  On a normalized basis, across Samsung’s entire TV line (Excluding OLED) from 2021 to 2024, having the patience to wait until late October/Early November would result in  a 26.8% average savings.  We note also that the 2025 year, as can be seen above, has been an anomaly, as even normalized prices have run above the average for the first half and below the average for the 3rd quarter.  Attribution for this unusual behavior is hard to pin down to a particular source, but we expect the bulk of the higher pricing is a result of both the monetary and psychological impact of US tariffs; less so from the tariffs themselves and more so on collective impact, with a dash of “everyone is expecting prices to go up, so let’s increase them.”  That said, we expect that prices across the Samsung line will decline in-line or greater than the norm for the remainder of the year.

​Note:  In order to compensate for new and deleted models and  feature changes we bring each year’s starting MSRP to a single point.  This normalization tracks throughout the year and allows for a better comparison as to how line price changes are occurring throughout the year and how they compare to previous years.  Again, this data is for Samsung’s Mini-LED/QD, QD only lines and does not include OLED.  OLED data is available on request.
While the data is helpful to consumers, it is also helpful to inventors who own public panel producer and display infrastructure investments, however we note that panel producer performance is greatly influenced by the build cycle rather than the buy cycle, with each product category having  specific periods during the year as to when new products enter production and inventories are being built for sale points during the remainder of the year.

The US National Defense Authorization Act (H.R. 3838 -119th Congress) is a massive appropriations bill that sets policies and authorizations for fiscal 2026 for the DOD, security programs for the Department of Energy and the Maritime Administration, and authorizes appropriations for the Defense Nuclear Facilities Safety Board and the Naval Petroleum Reserves.  Most importantly here, it modulates the US defense acquisition system to expedite the delivery of capabilities and materials to our Armed Forces.  The House of Representatives recently passed an amendment to this vast appropriations bill that could have relevance to the CE space.
While the final wording has yet to be published, what we have seen indicates that the bill could bar the Armed Forces from purchasing displays, particularly ‘advanced displays’, such as OLED or more specialized display modalities, if they are backed by the Chinese or Russian government.  Again the exact wording is important and yet to be revealed, but that could easily mean that any Chinese display producer that is receiving government subsidies could be considered to be ‘supported’ by the Chinese government and banned, and we believe all Chinese panel producers receive at least some government subsidies.
Using this amendment, the government could ban all display purchases from Chinese panel producers, giving an advantage to Samsung Display (pvt), LG Display (LPL), both based in South Korea, Sharp (6753.JP), based in Japan, and AU Optronics (2409.TT) and Innolux (3481.TT), both based in Taiwan, along with a number of small panel producers outside of China.  More specifically,  relating to OLED, the amendment would allow the US to ban both large and small OLED displays from all six Chinese OLED producers.  While the DOD categorizes display purchases by program, making it difficult to get a total display spending budget for a particular year, estimates for this year are substantially over $1b. Whether the US government will choose to use this amendment as a bargaining chip in its ongoing trade negotiations with China is yet unknown, but they have it in their pocket if the need it.

What do PCBs and glass have in common?  PCBs have glass fiber in them, but the commonality stops there, at least for now.  In this note we take a quick look at the use of glass as a core ingredient in future PCBs and as a key material for advanced semiconductor packing in the future.  The unique characteristics of glass allow it to solve a number of current and future issues with semiconductors and board level products that suddenly have become high priority issues as AI creates demand for servers and high-speed silicon devices at a rapid pace.  Before going further it is important to understand the basics about PCBs.
What is a PCB?

PCBs (Printed Circuit Boards) are typically made of FR-4[1], a widely used laminate consisting of a glass fiber reinforced epoxy, and copper foil, layered with an adhesive.  They can be single or multi-layer (up to over 30 layers).  FR-4 is inexpensive, has been standardized since the 1960’s and provides a rigid and stable framework for components, along with the ability to provide a structure for intricate traces (the equivalent of wires on a PCB) for component connections and power sources in most environments.  However as the number of layers and components increases so does crosstalk (the bleed of a signal from one part of the board to another) as does the amount of heat generated, and that heat reduces the life of board components.
While PCBs have been used for many years, the demands of modern servers, especially high powered GPUs used for AI training and inference have pushed the limits of this medium and research is being done on using glass as a substrate for a number of reasons.

• As board density increases, trace line widths decrease, making the lithography that produces these lines more difficult.  While PCB boards are rigid to a degree they have inherent flexibility and not necessarily perfectly flat and when it comes to lithography, even a small change in any direction, such as a slight bend in the substrate, can cause a defect that can lead to later issues or an outright board rejection.  Glass is over 1,000 times flatter than a typical FR-4 PCB.
• Thermal Expansion (CTE) occurs as the temperature rises in a material.  Glass has a CTE that is similar to that of silicon, which means it expands and contracts at roughly the same rate[2], while standard PCB CTE is 2 to 3 times higher, meaning it can expand and contract at a greater rate than the components.  This mismatched expansion can lead to solder joint fatigue or more serious warpage issues.
• In typical server operation, both the copper traces and the FR-4 absorb some of the signal.  As boards get larger that loss increases.  It also increases as the frequency of the signal increases, which becomes a significant factor with HBM (High Bandwidth Memory and High speed GPUs. Dielectric loss for glass is half to ¼ that of FR-4, and specialty glasses can reduce that loss by another order of magnitude.

 
Unfortunately, there are drawbacks with using glass as a PCB substrate or in advanced semiconductor packaging.  While FR-4 has been used for over 50 years and is well ingrained in the electronics production ecosystem, glass is a new material for these uses and has some unique manufacturing challenges.

• Glass is brittle and fragile and can be damaged during handling and processing, lowering yield and requiring more expensive handling equipment.
• Vias, copper clad tunnels that connect layers and vent heat, are more difficult to produce than with FR-4.  Lasers or ion etch is used to create vias in glass substrates, a slower and certainly more expensive process than the drilling used with FR-4 PCBs.
• The cost of substrate quality glass is considerably higher than FR-4 and the number of suppliers is extremely limited currently.
• The thermal characteristics of glass are such that it is a poor heat conductor, and as such, thermal management, already an issue with high-speed servers, becomes more so.

When Intel (INTC) first announced that it intended to develop and use glass substrates (9/2023) the timeline for commercialization was ‘in the second half of the decade’ or at least before 2030, and many expressed the opinion that it would take longer, if ever.  Since then the rapid expansion of AI and its concomitant need for greater server capacity and performance, have attracted a number of others to the glass substrate fold and the ecosystem needed to move this concept to commercialization continues to build.  We see four major categories for the glass substrate ecosystem:

• Specialty Glass Manufacturers
• Equipment Suppliers
• Fabricators
• Assembly & Test

 
Specialty Glass Manufacturers

This category is divided into two segments.  Those large commercial glass producers that are developing specialty glass for substrates and advanced semiconductor packaging, and those suppliers who are singly focused on this particular application.  While the most noise will come from the first category, it will take years for glass substrates to become meaningful in light of the other income streams these manufacturers have.  That is not to say it won’t become meaningful, only that it will take years to grow the application and the ecosystem.  Smaller suppliers, particularly Absolics, have been laying the groundwork for the industry, along with Intel.

Large Commercial Suppliers:
All of the companies listed under ‘large commercial suppliers’ are currently producing specialty glass products for advanced semiconductor packing R&D.  Most are at the qualification level as full scale packaging using glass substrates is just beginning to materialize.
Corning (GLW)
AGC (5201.JP)
Schott (pvt)
Hoya (7741.JP)
Nippon Electric Glass (5214.JP)

Small Focused Suppliers
These companies are more focused on the substrate glass market without the distraction of much larger other businesses and have a better chance to come up with specific solutions, as long as they are amply funded.
Absolics (011790.KS)
Plan Optik (P40.XE)
JNTC (204270.KS)

Equipment Suppliers
These companies are modifying existing tools and developing new ones that allow both glass producers and fabricators to utilize glass substrates.  As the technology is changing rapidly, there is no standardization on new tools at this time
LPKF Laser AG (LPK.XE)
Canon (7751.JP)
Yield Engineering Systems (pvt)
Applied Materials (AMAT)

Fabricators
These companies take th raw materials and create glass substrates to specification.  This is custom work and will likely remain so for the near future, which leaves open the expansion of the list to include a much larger base.
Samsung Electro-Mechanics (009150.KS)
Intel

Assembly & Test
Component placement (assembly) will be shared between fabricators and outsourced assemblers, while we expect test (other than at the line level) to be entirely outsourced to guarantee independence and impartiality.
ASE Group (ASX)
Amkor Technology (AMKR)
Powertech technology (6239.TT)
Toray Engineering (3402.JP)
​
A Deeper Look - Absolics
Absolics was created as an affiliate of SKC Ltd (011790.KS) in late 2021, with SKC being one of dozens of recognizable companies in Korea under the SK Group[3] umbrella.  Absolics was created to focus specifically on glass core substrates and to eventually become the world’s first large scale commercial manufacturer of glass core substrates.  By exploiting the characteristics of the company’s specialty glass formulations the company expects to solve many of the issues that limit chip-level progress.  The company has indicated that the use of its materials can provide up to a 50% power reduction and a 30% improvement in signal performance, which are two big issues for server farms.
The company began constructing on a $600m ~10,000 ft2 facility (Phase 1) at SKC’s campus in Covington, GA in November of 2022.  The first phase of the construction was a $240m small volume glass manufacturing facility, followed by a $360m high volume line.  The company received approval for up to $75m[4] in CHIPS Act funding for the phase 2 (120,000 ft2 total) facility in May of last year and up to $100m for R&D from the federal government’s National Advanced Packaging Manufacturing Program.  The government weighed the cost of building a glass substrate manufacturing facility from scratch and the administrative delays associated with such a project and made the investment with a highly experience manufacturer instead.  The company has targeted mass production by the end of this year with phase 1 production of 12,000 m2/year and 72,000 m2/year when phase 2 is completed,
The core of Absolics technology is its TGV (Through Glass Via) technology.  These ‘tunnels’ through the glass substrate provide inter-layer communication and further reduces signal loss by reducing the overall length of the signal path through layer stacking.  As chip production becomes more complex with multi-layer boards and interconnected chiplets the necessity to reduce signal interference and loss through vias and low loss materials becomes essential. This is particularly important in relation to power, which has to be delivered both horizontally across each layer and vertically to all layers and to passive components that can be embedded into the glass substrate itself. 


[1] FR stands for Fire Retardant while the 4 is a designation for the grade of the epoxy-glass laminate.

[2] Silicon ~2.6 ppm/°C; Glass – 3.2 to 7.5 (adjustable)

[3] Sk Telecom – the #1 telecom company in Korea; Sk Hynix the 3rd largest semiconductor company world-wide; SK Silton the only silicon wafer manufacturer in Korea to name a few

[4] NIST indicates that Absolics has been approved for $100m in CHIPS funding, which we believe is for R&D.

The research grant mentioned above is specifically targeted to TGV development in order to reduce the challenges associated with TGVs and to develop the specialized equipment to facilitate their production.  This is also an extension of the company’s collaboration with the DOD under their SHIP Program (State-of-the-Art Heterogeneous Integrated Packaging) which requires ultra-low loss package designs.  The company also benefits from a 12% investment SKC has made in Chipletz (pvt), a fabless substrate start-up in Texas.
The company is positioning itself as a neutral supplier to both major foundries and to design houses that do not want to purchase glass from competitors.  There are others who are developing similar technology and products with internally developed expertise, such as Samsung (005930.KS) and LG Innotek (011070.KS) so it will not be without risk for Absolics, but they have partnered with over 30 suppliers and external resources in order to compete more effectively, and continue to have the backing of SK Group, the 2nd largest manufacturing company in South Korea (Samsung is 1st).   There are still a number of major challenges that need to be reduced or eliminated, particularly low TGV yield and reliance on AI demand, but if glass substrates perform as well as expected in real world use, Absolics will directly benefit.
As Absolics is more focused on chip level and packaging glass substrates than board level products, we compare the parameters of using glass core substrates in advanced packaging against typical ABF and silicon interposers.

As with all new technologies, estimates vary considerably, with estimates for market size ranging from $4.3b to $7.8b in 2024 and $8.3b to $12.3b in 2032, with CAGR ranging from 3.6% to 7.8%.  That said we expect there is little commonality between these estimates as the term ‘Glass Substrates’ is far to general to mean only high purity glass substrates for PCB cores and advanced semiconductor packaging.  Therefore we reserve judgement on market size until we see some substantive numbers out of any of the glass suppliers listed above, although we don’t expect that soon as the technology and the products have to be rigorously tested before being adopted by any of the major foundries and put into production. When that happens others will follow and the market will grow quickly, especially if Ai demand continues to grow.
Glass substrates for PCBs and advanced semiconductor packaging solves a number of problems that put limitations on semiconductor growth.  We expect if the technology lives up to expectations it will become a necessary part of the advanced packaging space and find its way into the high-end of PCB board design.  It might be hard to tell who the winners will be this early in the game, but we will continue to update as the development progresses and mass production begins both at the glass level and at the application level.
​
Sources:
Absolicsinc.com
Eng.sk.som
PCBmake.com
Albanyceo.com
Kedglobal.com
Nist.gov
Yolegroup.com
Circuitinsight.com
Inbusinessphx.xom
Craft.com
Imasource.org
Electronicsforyou.biz
UnityPCB.com
Technews.tw