Now and then we come across small bits of information that are not quite worthy of writing up as a note, but interesting none the less.  Rather than disregard this information.  Rather than disregard it, we will put it under the subject title of “Little Bits” going forward.  Should there be feedback from readers about a particular Little Bit entry, we would be happy to dig deeper.
Shanghai Sansi Electronic Engineering (pvt), a Chinese LED lighting producer has officially announced its CoC (Chip on Ceramic)  its next generation ceramic streetlight series.  The lights have been certified in China to have a 100,000 hour lifespan (26 years at 10 – 11 hours/day or 11.4 years at 24/7).  Typically LEDs are mounted on a PCB which is attached to an aluminum heatsink with thermally conductive adhesive.  The Sansi concept is to mount the LED directly to a ceramic heatsink with no PCB, which accounts for their long life expectancy. And 67.4% energy savings over standard HID (High-intensity discharge) and HPS (High Pressure Sodium) lamps that are commonly used.  What makes this more interesting than the usual LED promo is that Sansi details the development process for this product (CoC).  Here’s what they noted, which is an indication that not all Chinese products are ‘garbage’ as has been recently stated by the current administration.  The lights are being used in a number of projects in China, particularly the Hong Kong-Zinhai Bridge that connects Hong Kong and Macao.  The bridge is 34.18 miles long with an undersea tunnel gap of 4.1 miles connected both ends to maintain open shipping lanes. 

• 106+                  Ceramic Structural Designs
• 496+                  Material Formula Refinements
• 2117+               Firing Process Trials
• 10,825              Optical Simulations
• 41,344              Performance Tests

As we noted last week LG Display (LPL) announced that they have reached the commercialization stage  of their blue OLED panels.  While these panels are not quite the final step in the blue saga, as they use a combination of blue fluorescent and blue phosphorescent material to achieve results, they are certainly a step toward the ultimate goal of a three color (RGB) phosphorescent stack (see our 5/1/25 note for more detail).  The development of this panel was conducted with Universal Display (OLED), who has been on the blue phosphorescent material development path for years and is the key supplier of red and green organometallic phosphorescent emitters to the entire OLED industry.

Why is blue so hard?
​

Commercialization of a blue phosphorescent emitter and host combination that does not rely on blue fluorescent support has proved to be a daunting task due to the high energy associated with blue photons (packets of electromagnetic energy).  These excited particles can break chemical bonds in their own molecules, degrading them, or can create new non-radiative molecules that reduce the efficiency of the blue emitter.  Additionally, the host material that the blue emitter sits in has to have a higher energy level than the blue emitter itself to keep energy from leaking back to the host as heat or non-radiative energy.  So finding a blue phosphorescent emitter that meets all specifications is only part of the process, as the host material  development can also be challenging.
So, we know the development of a blue phosphorescent emitter has been difficult to say the least, as some potential blue emitter materials have high efficiency and a deep blue color point but only last for a few minutes, while others have a longer lifetime, and proper color, but are too inefficient to be used commercially, and some have excellent efficiency and a long lifetime but can’t quite produce the deep blue that is needed.  While Universal Display has completed ‘commercial verification’ with LG Display, UDC continued to record blue emitter/host revenue in 1Q as ‘developmental’, which is required until the product using the material is commercially available.  As the timeline for LGD’s panel production is still unknown, the key to understanding whether the LGD panels are being used in a commercial device will be when UDC begins recording the blue material as ‘commercial’.
What about Samsung?
Obviously, there are other OLED panel manufacturers working to bring a full phosphorescent blue emitter to market, particularly Samsung Display (pvt), who is also working with UDC along with their own development team.  As the leader in small panel OLED displays, they have a very big stake in this process but tend to be a bit more ‘purist’ when it comes to OLED processes.  SDC did not believe that LG Display’s TV panel, which uses a single color OLED and a color filter to create red, green, and blue, was the right way to produce large panel (TV) OLED in 2013 and concentrated on smaller RGB OLED displays, eventually settling on a blue OLED with quantum dot s to create colors for their QD/OLED TVs.
As the LG Display panel uses both fluorescent and phosphorescent blue emitters, we suspect that the current blue phosphorescent host/emitter that LGD is using as part of its stack might not meet Samsung Display’s requirements yet.  Samsung would likely be most interested in using blue phosphorescent material in mobile devices (smartphones and tablets) where the higher efficiency of a phosphorescent blue emitter will be key to either a power consumption reduction or an improvement in brightness, but as mobile devices have individual sub-pixels for each color, we expect their requirements might be a bit more stringent.  That said, we do expect SDC will find a way to incorporate a blue phosphorescent system in some product this year.  It could be a similar fluorescent/phosphorescent blue emitter base for their QD/OLED TV/Monitor panels, or it could be a higher specification deep blue phosphorescent emitter for an RGB architecture for mobile devices, but we find it difficult to imagine that SDC will cede the first ‘blue year’ to LGD.
All of that said, changing from a blue fluorescent emitter to a phosphorescent emitter is much more complicated than just switching materials.  In a large panel (TV), the OLED materials are deposited across the entire panel and the driving circuitry is the same for every sub-pixel point, as each sub-pixel is the same (white) color until it reaches the color filter or quantum dot.  In current RGB (small panel) displays, the driver for the red and green sub-pixel can be the same but as the driving characteristics for the fluorescent sub-pixel (blue) as different, the circuitry for the blue driver is different, adding to complexity.  In an all phosphorescent RGB display, all three sub-pixel circuits can be the same (in theory), which means not only does the material stack change, but the driver circuitry also changes, adding another level of complexity to designing an all phosphorescent display.
Timeline?
Not only do all of these issues need to be worked out, but they also need to be tested both at the pilot level and in a mass production setting, and this can take time.  The issue then becomes where do they start?  Does the OLED producer have enough ‘spare’ capacity that they can convert a line to producing all phosphorescent RGB OLED displays, or are they capacity constrained enough that they cannot afford to dedicate a line to all phosphorescent OLED production?  As was the case when green phosphorescent emitter material became commercially available, adoption took time.  With the first commercial product using a phosphorescent green emitter was released in 2013, UDC’s green emitter sales increased but then stayed relatively flat for ~15 quarters, being adopted by one of two major customers. In 2017, sales increased as a second large customer adopted the material and continued to grow quickly through 2021.  While still growing to a lesser degree, as the industry has universally adopted green phosphorescent emitter  material, growth is more tied to capacity expansion and new product applications, although the adoption of multi-layer OLED displays could lead to incremental material sales.
Adoption?
So the question now becomes will the adoption take 15 months, as it did with green or will it be faster or slower?  These are essential questions for UDC’s longer-term prospects, as while OLED capacity growth continues, the addition of a third primary material revenue stream is a godsend for any material producer.  We expect the adoption of blue will be faster, but with some caveats.
Why faster?
Numbers – in 2013 there were two OLED producers, Samsung Display and LG Display.  Tianma (000050.CH) built their first OLED fab that year but did not ship commercial product and BOE (200725.CH), China’s largest OLED producer, did not build their first OLED fab until 2016, so the adoption of green phosphorescent emitter material was dependent on only two producing entities.  Now there are over a dozen producers, all of whom are looking to differentiate their OLED displays from others and blue is a perfect differentiator.
Experience –Samsung Display and LG Display had been involved with OLED display development for over 10 years when green phosphorescent emitter material was released commercially by UDC, yet much OLED production was still problematic, and yield was always an issue.  At that time making major changes to formulas, architecture, processes, and equipment meant a long learning curve before returning to decent production yields and carrying substantial losses that could erode potential funding and adoption.  The current experience level across the industry is considerably higher than 10 years ago and producers are more likely to see a change that could give them an edge over the competition as one they are willing to take after years of managing commercial production.
Quality – A true blue phosphorescent emitter will give display designers a greater ability to balance their systems.  As a more efficient material they can maintain brightness with less power and less power means longer battery life for mobile users and a longer lifetime for the material, putting a damper on the ever-present burn-in question.  They can maintain the current power level and produce a brighter display to compete with other display modalities that are encroaching on the OLED space, or they can use blue as a differentiator that will separate their display from those without blue phosphorescent emitters.
Advertising – The idea of the display industry is to sell displays, and in order to sell displays there have to be lots of products that use them.  As the display industry can find itself in a somewhat stagnant position, with few new enticements for consumers, any new technology affords the industry a shot at incremental unit sales. We expect the industry will be enamored with the promotion of blue when it starts and will start a new line of promotion for OLED devices to counteract Mini-LED, Quantum Dots, and eventually Micro-LED displays.  However unless there is a truly discernable difference between all phosphorescent displays and what we have now, price will remain the most important factor to consumers as the blue enthusiasm wears down.  UDC however will have a new revenue stream , one that can eventually be bigger than red or green.
Why bigger?
In order to produce white light in large OLED displays, one can combine a blue emitter and a yellow/green emitter and then send the light to a color filter to create red, green, and blue sub-pixels, essentially the way LG Display’s WOLED TV panels work.  Samsung Display’s QD/OLED panel is similar but based on a blue[1] OLED material that gets converted to red and green by quantum dots. Smaller devices use individual red, green and blue sub-pixels, directly creating all colors.  WOLED displays uses UDC’s yellow/green phosphorescent emitter with a blue fluorescent emitter.  If a phosphorescent blue emitter became available, UDC would have the potential to be able to put both materials in every WOLED TV.  In Samsung’s QD/OLED the blue material used is fluorescent, with UDC providing no substantial OLED emitter material.  If a phosphorescent blue emitter became available, UDC would have that potential new stream.  In RGB display (phones, tablets) the impact would not be as significant as UDC would only be adding a third phosphorescent emitter to the two they already supply, but the volumes are extremely high, so all in, UDC benefits unless someone comes up with a better phosphorescent blue.  That said, even in that scenario UDC still has device patents that cover the use of phosphorescent emitters in OLED devices, so they might lose the OLED material sale to someone else but should still be able to capture a device royalty stream as before.
Why Not?
Cost – Fluorescent emitter materials tend to be less expensive than phosphorescent ones.  In premium OLED displays, the additional cost can be absorbed, but as one migrates to lower price tiers, the cost will be more difficult to absorb, and adoption will be slower.  We expect however that many brands will bite the bullet and eat the additional material cost in order to compete, at least for some products.  The cost of converting formulas, structure, and process also must be considered, and some who have been producing OLED displays for years at a loss might hesitate, unless they can convince funding sources to foot the bill.
Complexity – While there are certainly issues that will make adding phosphorescent blue to OLED production more complex, at least at the onset, OLED producers are so used to phosphorescent materials that they will likely adapt to required changes more quickly
 So?
We note also that UDC has contracts with all major OLED producers.  Some are based on a flat fee license, and some are based on a per unit royalty, and some cover only current phosphorescent (red & green) emitters.  In some cases UDC will have to strike new deals for blue that follow current contract formats.  While developmental OLED materials are expensive their volumes are low, but when they become commercial, they tend to be priced according to volume, so large, early adopters could have an advantage over small lower volume producers, unless their current contracts cover ‘all phosphorescent materials’.  UDC will have to balance their production cost and volume tiers against their desire to encourage blue adoption, ideally setting smaller price/volume increments in the early years against the opposite in later years. 
All in, blue is good, especially for those who produce it, but regardless of the headlines that are calling for a new ‘blue’ era in the display world, we expect most investors will expect too much too soon.  Panel producers need to make money and if they are producing at profitable utilization levels, they are going to want to keep doing so as long as possible, putting aside any changes that might reduce volume or profitability.  Most will talk the ‘blue’ talk but the implementation might be a bit less than the rhetoric.  We believe the adoption of blue phosphorescent emitter material will certainly be a positive for the industry and for the consumer, but technology hype is just that whether it is AI hype, metaverse hype, or 5G hype.  How consumers see ‘blue’ will be the deciding factor as it always is.


[1] Actually a combination of fluorescent blue and phosphorescent green.

Warning…Thinking Caps on…
​

​In an OLED device, a voltage is applied to the cathode, creating electrons and the opposite electrode produces holes.  Think of them as the cute girl sitting at one end of the bar and the svelte guy sitting at the other end.  When they see each other, they are immediately attracted to each other (opposites attract) and both get up and push their way through the crowd (OLED stack).  They meet on the dance floor (emitting material) where the magic happens.  They hold each other in a warm embrace (forming an exciton, a combination of an electron and an electron hole that is in an excited state) and dance in the spotlight (produce light) until the music stops.  They gaze into each other’s eyes and quietly head back to their seats on either side of the bar.  OLED devices play out this scenario over and over as long as there is a voltage at the electrodes.

Simple enough, right?  Now let’s move this conceptual production to an industrial setting.  There are two companies in Pixeltown, both producing the same thing, excitons.  Fluorescent Inc. produces four excitons on each production run, One singlet exciton (red) and three triplet excitons (blue), but their process is old, and they are only able to sell the singlet excitons to customers, throwing away all of the triplets, leading to a 25% efficiency rating and a serious trash problem that the Pixeltown mayor is not happy about. Phosphorescent Inc. uses the same basic equipment and produces the same initial output of one singlet exciton and three triplet excitons.  However, the folks at Phosphorescent Inc hired some smart guys who came up with a way to get their triplet excitons to act like singlet excitons, which allows them to sell all three triplets and one singlet for each run, for a nearly 100% efficiency rating. 
Sooner or later the folks at Fluorescent Inc (Factory a) figured out that they are going to go out of business, having such a low efficiency rating, and the economic impact to Pixeltown would be catastrophic.   Management hired a hot-shot banker and put out  some feelers but there were no takers until the banker’s lowly assistant figured out that if you were to combine both fluorescent and phosphorescent materials together when making excitons, the result would be even better than the two individually. 
Here’s why.  If the materials are carefully matched, the ability of Phosphorescent Inc’s process to use both triplet and singlet excitons to produce light, allows some of the triplet excitons that Fluorescent Inc produces but throws away (heat rather than light), to become useful.  This means that the combined fluorescent and phosphorescent emitters could have an efficiency that is higher than 25% for the fluorescent excitons and remain at 100% for phosphorescent excitons, essentially improving the efficiency of the combination by about 15%.  Not all of the fluorescent triplet excitons can be converted and used by the phosphorescent emitter, but enough to make a difference.

Why is this important?
LG Display (LPL) made an announcement today that will undoubtedly shake up things in the OLED space, but the devil is in the details and it is essential to understand how OLEDs work in order to quantify the announcement.  In fact the structure that LG Display is speaking about is similar to the tandem system that the company uses for production of small OLED displays for ‘a large customer’.  Typically, in order to improve brightness, the dual stack approach is used, essentially squeezing two OLED stacks between electrodes instead of one.  This helps, but is an expensive solution as OLED materials, particularly phosphorescent emitters, are costly, especially if you are duplicating the entire (RGB) stack, and increases the number of steps involved in the deposition process, which has a tendancy to reduce yield.
We believe the LG Display approach is both similar in that it uses a multi-stack approach, but it is also a bit different.  We expect that the phosphorescent blue host and dopant combination that LGD is using  would not stand on its own commercially quite yet, as it could possibly fall short on a particular commercial specification, any of three major categories, lifetime, efficiency, or color point.  Developers must balance these three factors when trying to create a stable phosphorescent emitter and that has been a difficult task for all.  Materials that have the necessary color point (deep blue) might have a lifetime that is too short to use commercially or be lacking in efficiency (high power usage).  Other materials that have a more extended lifetime might not have the necessary color point.  You get the idea.  So while the concept of using a combination of blue phosphorescent and blue fluorescent emitters has promise, it is an interim solution until a truly stable blue emitter and host combination can be found. 
LG Display was careful to call this iteration ‘a step closer’ and not a final solution, but it will certainly get LG Display some acclaim and cachet from the announcement.  The response from Samsung Display (pvt) will be interesting to see as they have been working on the same blue phosphorescent emitter with Universal Display (OLED) for years and at one time, years agho, evaluated a combination blue Phosphorescent/Fluorescent combination.  We also expect a response from both the TADF community and those developing quantum dot EL displays. 
Here's the LG Display Press release: (our highlights in red)
LG Display, the world’s leading innovator of display technologies, announced today that it has become the world’s first company to successfully verify the commercialization-level performance of blue phosphorescent OLED panels on a mass production line. The achievement comes about eight months after the company partnered with UDC to develop blue phosphorescence and is considered a significant step closer to realizing a “dream OLED” display.
In the display industry, “dream OLED” refers to an OLED panel that achieves phosphorescence for all three primary colors of light (red, green, and blue). OLED panel light emission methods are broadly categorized into fluorescence and phosphorescence. Fluorescence is a simpler process in which materials emit light immediately upon receiving electrical energy, but its luminous efficiency is only 25%. In contrast, phosphorescence briefly stores received electrical energy before emitting light. Although it is technically more complex, this method offers luminous efficiency of 100% and uses a quarter as much power as fluorescence.
However, achieving blue phosphorescence has remained a major challenge even more than 20 years after the commercialization of red and green phosphorescence. This is due to blue, among the three primary colors, having the shortest wavelength and demanding the greatest energy.
LG Display has solved this issue by using a hybrid two-stack Tandem OLED structure, with blue fluorescence in the lower stack and blue phosphorescence in the upper stack. By combining the stability of fluorescence with the lower power consumption of phosphorescence, it consumes about 15% less power while maintaining a similar level of stability to existing OLED panels.
In particular, LG Display is the first to succeed in reaching the commercialization stage of blue phosphorescent OLED panels, where performance evaluation, optical characteristics, and processability on actual mass production lines should all be confirmed. The company has already completed commercialization verification with UDC.
LG Display has independently filed patents for its hybrid blue phosphorescent OLED technology in both South Korea and the United States.
The company will showcase a blue phosphorescent OLED panel featuring two-stack Tandem technology at SID Display Week 2025, the world’s largest display event, in San Jose, California from May 11th (local time).
At the show, LG Display will be unveiling a blue phosphorescent OLED panel featuring two-stack Tandem technology applied to a small and medium-sized panel that can be applied to IT devices such as smartphones and tablets. As more and more products require high definition and high efficiency such as AI PCs and AR/VR devices, the application of blue phosphorescence technology is expected to expand rapidly.
“The successful commercialization of blue phosphorescence technology, which has been called the final piece of the ‘dream OLED’ puzzle, will become an innovative milestone towards the next generation of OLED,” said Soo-young Yoon, CTO and Executive Vice President of LG Display. “We expect to secure a leading position in the future display market through blue phosphorescence technology.”
Based on LG Display’s IP here’s what we think the configurations might be… 

As the leader in both the smartphone and the TV set segment and one of only a few large memory producers, Samsung Electronics has considerable influence in the CE space.  They tend to be a leader in CE technology, particularly in the display space and produce both components and end end-user products.  As such they can be both an indicator and a lightning rod for change in the CE space, making them a valuable tool in understand the current an future status of consumer electronics.
Last night Samsung Electronics (005930.KS) reported 1st quarter 2025 results of 79.14 trillion won ($55.25b US) in sales, ↑4.4% q/q and ↑10.0% y/y and 1.3% above consensus with this quarter being the best in the company’s history.  Operating profit was 6.7 trillion won ($4.68b US), ↑3.1% q/q, ↑1.4% y/y, and ↑4.7% above consensus.  In order to better understand the table below, which breaks down sales and operating profit by Samsung division, we note:
DX (Device Experience) includes – TV sets & Monitors, Appliances, Smartphones and tablets, Network Equipment, and Health Products
DS (Device Solutions) – Memory, Processors & Sensors, Logic, foundry
SDC (Samsung Display) – small panel OLED and large panel QD/OLED displays
Harman – Automotive & Consumer audio

​Samsung also breaks out some components of its divisions.  The division formerly known as MX, which now is comprised of smartphone and tablet products (part of the DX division), the TV segment, also part of the DX division, and Memory, a part of the DS division.  While operating income is not given sales can be computed, as well as the share of sales of the company total,

The results were mixed, with some product categories outperforming and sone underperforming, essentially the way Samsung is supposed to operated in difficult times, although we expect many investors were hoping for a more positive quarter from the semiconductor segment, and a more optimistic feel for 2Q and 2H, which they did not get.  Samsung overall was careful to hedge any optimistic view of this year with the caveat of tariff uncertainty, and the division leaders were even more so, although the company was a bit more forthcoming about product rollouts and timelines than usual.
As a number of Samsung’s major products have been exempted from US reciprocal trade tariffs, they have only moderate exposure to direct import costs, but as a component supplier, they were cautious about raw material costs impacting component prices and how that would follow through the supply chain.  That said, without any hard US trade policy or realistic negotiations with major trading partners, and the prospects of another nuclear option in early July, they have little choice but to forge ahead as originally planned. 
As seems to be the case with many larger CE companies who have the option, they are considering shifting production from countries that have onerous tariff requirements to less onerous ones, but seem to be in no rush to make those changes.  We expect there will be lots of talk about how negotiations are progressing and how many deals have been agreed upon by July, but we also expect little confirmation, little detail, and even less about timelines for balancing trade.  While the full impact of tariffs has yet to be felt by consumers, the peripheral impact, such as a weak equity market, has already put consumers on edge, something mentioned by Samsung a number of times on the call, we believe a big part of Samsung’s cautious stand on 2Q and the rest of the year.
Getting all Samsung divisions to operate effectively and profitably is a complex task and one drenched in global macroeconomics, but a bit less ‘unknown’ might be helpful in getting the planets to align.  Below are our quick notes.  Comments in red are our own.

​General Comment Summary

• DX strength came from flagship smartphone strength (Galaxy S25 series) and high-end appliances
• DS weakness came from demand deferrals from High Bandwidth memory customers
• Capex – 12t won with 0.5t won for display (down substantially), as the Gen 8.6 IT OLED fab is completed and 10.9t won for the DS division overall.
• Repurchased 3t won of common and preferred as part of 10t won 2025 program, which continues in 2Q.

Memory - General

• Ai Server demand remained strong
• Some PC/Mobile demand improvement (Small, possibly China driven)
• SSD demand weak
• Some data center projects delayed

DRAM

• Bit growth higher than expected
• HBM export controls on AI chips and customers waiting for HBM3E release caused deferrals. 

NAND

• Bit growth down ~10% but above expectations as perception that market has bottomed sets in. More, less downside pressure than upside pressure.

Memory Outlook – 2Q

• AI Server demand to remain strong
• “Preemptive Purchasing” after tariff pause
• Memory for PC & Mobile inventory now normal because of China subsidy inclusion
• Overal,l expect some incremental demand but tariff remains a question

Memory Outlook – 2H

• Ai server demand to remain strong (Said a number of times)
• SSD to recover as deferred projects begin
• Incremental PC demand from Win10 end AI (Win10 replacement cycle theory seems dead – Ai better bet but still an unknown)
• Mobile demand improves due to AI

 
System LSI – General

• Weakness from delayed SoC adoption by major customer (Samsung)
• Strong demand for image sensors

LSI – 2Q Outlook

• Expect image sensor volume to decline
• SoC increase to offset sensor volume decline

LSI – 2H Outlook

• Limited mobile momentum
• SoC steady
• Will add product (Automotive sensors)

 
Foundry – General

• Seasonal weakness for mature nodes
• Inventory adjustments due to China trade tensions (meaning order reductions)
• New advanced node starts in 2H

Foundry - 2Q Outlook

• Subdued demand
• Ramping production for US automotive products (Tariff issues?)
• Tariffs could have big impact
• 2nm GAA production starts in 2Q, but small

Foundry - 2H Outlook

• Geopolitical Risk expected to increase
• Demand for PC & Mobile expected to weaken
• AI & HPC momentum still strong (advanced nodes)

 
Samsung Display – General

• Improvement in demand from major customers
• Favorable exchange rate
• But seasonally weak quarter
• Double digit monitor sales growth (QD/OLED share?)

SDC - 2Q Outlook

• Mobile – Conservative view due to tariff situation
• Will launch new ultra-high refresh rate monitors (gaming)

SDC - 2H Outlook

• Increasing uncertainty due to trade issues
• Competition increasing
• Weak consumer sentiment
• Mobile driver is AI
• QD/OLED will expand monitor line w. lower-priced models (Good news – when?)

Visual Display (TV) – General

• Demand was down q/q but up slightly y/y w. premium and ultra-large TVs driving growth.
• Raised prices and lowered material costs but…
• Overall TV demand remained weak, and the cost of competition was high

TV - 2Q Outlook

• TV demand flat y/y
• Expanding AI TV (Will consumers care?)

TV - 2H Outlook

• Demand for high-value products (Ultra-large and OLED) will remain

 
Q&A
Tariffs

• Semis, phones, tabs currently exempt
• Reviewing other products
• Potential to ‘manage’ global production

Stock Buyback

• Will cancel 2.5t won shares of 3t won current buyback

TV

• Intense entry-level competition
• Will add to 98”+ lineup

Memory

• NAND Bit growth up mid-teens in 2Q
• DRAM Bit growth up low 10%
• NAND for PC & Mobile price decreases to end – flat going forward

Foldables

• Differentiated AI for each foldable type

All in, we thought the quarter was just about as expected, although we were a bit surprised at the comments about data center deferrals, which was mentioned a number of times.  We were concerned that Samsung was as cautious about 2H, but we expect given the volatility of the situation, they have little choise.  The fact that they were able to find a way to the exemptions that will allow them to not have a disastrous year is a bit of an offset, but it is better to under-promise and overachieve than the opposite.

AI is unusual in that while we (humans) develop architectures and algorithms that make models work, we are not really sure how they do what they do.  But when we ask a model, in this case ChatGPT (OpenAI) to explain how they work, the model seems to be able to step back a bit in order to explain details.  This step back puts the LLM we are talking with in the unusual position of describing how it works as if it were not a model but an observer, although sometimes it seems odd when a model describes how it works by saying “models do this…” sort of ignoring the fact that it is  model, but we digress…
What we were trying to understand when we started our conversation with ChatGPT was how models represent information for each token as it learns.  We understand that the model (software called a tokenizer) breaks down text into tokens, typically a token for each word, although in many cases it can be a sub-word, such as a syllable or even a single character.  Each token gets assigned an ID number which goes into a master token ID list.
Example:
“The cat was running away from the dog.”

              
The list of unique tokens for a large model is fixed at ~100,000 tokens.  No matter how much data the model sees it only uses tokens from this list, breaking down unknown words into smaller known sub-word pieces, so the corpus of data the model sees could be 300 billion tokens.  The token ID list remains with the model after training, but the large list of tokens processed during training does not need to be stored, as the model learns from the tokens but does not need them later.
The part that is difficult to visualize comes as the tokens are first encountered by the model.  The model looks up the token in the token list and matches it to another list that contains that tokens vectors.  Think of vectors as a string of numbers (768 numbers for each token in a small model)

​ On the first run through, the dimensions for each token are set to random numbers, essentially ‘noise’, then the token sequence is passed to the first layer of the model.  These vectors are used to begin to ‘classify’ each token. If the model ‘sees’ that ‘cat’ and ‘dog’ appear in the same sentence often, it will adjust a particular dimension slightly for both the cat and dog token, and with each layer will further adjust that dimension, which we might call the “animal” dimension.  By the time the token has been cycled through all the layers, the ‘animal’ vector for both dog and cat will be close to each other, but not exactly the same.  That is how the model ‘knows’ that both dog and cat have the ‘animal’ relationship but are still different from each other.  If that vector was the same for both, the model would not know that while both are animals, they are different animals.
While this is a very simplistic look at how an LLM learns, one should understand that the model is always looking at the relationships between tokens, particularly in a sequence, and with over 700 vector dimensional ‘characteristics’ for each token, the model can develop lots of connections between tokens.  It is hard not to think of the dimensions as having specific ‘names’ as the semantic information that the dimensions contain is quite subtle, but it is all based on the relationships that the tokens have to each other, which is ‘shared’ in token vectors.
All in, this is just the tip of the iceberg in terms of understanding how models work and their positives and negatives, although even the best of LLNs still has difficulty explaining how things work internally when the questions are highly specific.  Sometimes we think its because it doesn’t really know how it works and other times it seems that it just doesn’t want to give that proprietary detail.  But we will continue to dig and pass on what we find out and how it affects AIs and their use in current society.  More to come…
 

The perfume and scents industry is not one that makes headlines often.  Perfume ads with celebrities tend to come and go but over the last 25 years there have been some scents for both men and women that have sustained themselves on the best-seller list and generated millions of dollars in revenue.  As an example in 2022 Dior (CDI.FR) Sauvage was selling at the rate of $4.6m/day for much of the year and last year the perfume market was estimated to be between $50.5b and $55.5b US, with an expected CAGR of between 4.7% and 5.9%[1].


[1] Sources: Estee Lauder, VMR.com, CB Insights

Scent developers start their process with an idea.  It could come from examining current popular scents to capitalize on a trend, or it could come from a creative point of view, maybe recalling  a travel destination or personal experience.  The process then moves to the selection stage where the perfumer, based on expertise, selects fragrances that they believe will represent their concept.  What follows is an extended trial and error process where the scents are blended to form a ‘top note’, the basis for the overall scent, ‘accords’ that push the scent in a particular direction (rose, marine, etc.), while making sure that the materials have a consumer-oriented longevity and projection (how far away the scent can be noticed), all of which are developed by trial and error.
There is software that can help perfumers, even AI based software like Philyra, developed by Symrise (SY1.XE) and IBM (IBM) and released in 2018.  The software contains a database of 3.5m legacy formulas and 2,000 raw materials, and, according to the company “…is able to guide perfumers towards exciting and surprising solutions, explore new combinations and materials without human bias, and help perfumers update and improve upon iconic fragrances.” In particular, Philyra helps perfumers to work toward using sustainable materials in their development.
While software platforms like this help the scent development process, it is a long and arduous process that takes many months or years until the right combination of scent and materials is reached.  Even with software providing assistance to perfumers and the expertise of a professional and experienced ‘nose’ (1st tier perfumers can make over $400K/year) commercial success is certainly not guaranteed and the cost of development, materials, and advertising can be quite financially burdensome, even for a large company.
But fear not perfumers, as a group of Japanese scientists have taken the idea of AI scent development further and created a Generative Diffusion Network for creating scents.  This new model uses mass spectrometry data from 166 essential oils to isolate 9 ‘odor descriptors’ that can be used to form scent combinations which are then tested for accuracy in a double-blind (human) process where participants had to match the AI aroma with the appropriate descriptors.
To illustrate: “As an illustration of the procedure, for the first sample of the sensory test two odor descriptors, Wood and Spicy, were selected. A random 201-dimensional vector of Gaussian noise was chosen as the seed for the OGDiffusion network. The network was then run in inference mode, generating a mass spectrum as the output. This mass spectrum was subsequently analyzed using non-negative matrix factorization to identify the essential oils required for the mixture. The analysis determined the following essential oils and proportions: Cypress (0.10), Angelica root (0.07), Cuminum cyminum (0.05), and Trachyspermum ammi (0.78). The specified amounts of each essential oil were pipetted into 5 mL vials and diluted with alcohol at a 2:1 ratio. The resulting mixtures were prepared for sensory evaluation in odor vials. Table S1 shows the essential oil recipes used in all sensory experiments.”
​

The idea behind this model is to eliminate problems that exist with current AI scent development systems.  Such systems are based on proprietary data and require expert human intervention, along with results being hard to reproduce.  While they are considered helpful to those in the scent profession, they are not automated and that is where this new model goes.  The system learns without needing prior chemical composition knowledge and is able to generate precise results that can be reproduced exactly, and mass spectrometry data can be easily represented as weighted sums, a function commonly used in LLMs.
So, will those wishing to become perfumers or scent specialists be out of a job?  In some ways the answer is yes, as there will be less need for the trial and error development system used today and that means less learning situations for those coming up in the industry, but again humans are essential, even in this automated scenario, as there must be someone who can test the combinations created by the AI, even if they were created without human assistance.  Without a ‘nose’ to smell the combinations there is no subjective point to attach to the scent.  So, in this case, such an AI system will reduce the amount of work associated with the development of scents but will still require a high-quality professional to make sure that the scent is a pleasant or exciting as expected.  The nose knows.

AGI or Artificial General Intelligence is a term that gets tossed about regularly, particularly by those in the AI business.  According to them, we are closing in on a point at which AI systems would have the capacity to reason, solve problems, understand complex ideas, and adapt to tasks that were not explicitly programmed.  This is both an exciting and scary prospect that is difficult to quantify, but we are sure that those in the Ai field will let us know as soon as it happens, or when they think it does.
There is only one problem, something called a world view.  A world view is a human being’s image of how the world works.  It’s not a single image but a collection of information that you starting amassing the second you were born.  From those first moments the human mind builds a world model by using sensory information to draw conclusions. A baby eventually understands that if you hold something and then let go, it falls to the ground.  The baby doesn’t know what the concept is called, but after a few (or many) things falling to the floor, the baby understands that if I let something go or push it off a table, it will fall to the floor.  Simply, our world model is experience based
Humans are also subjective.  Baby A will learn that if I let go of something it will drop to the floor and people will come over and make funny noises which are frightening, while Baby B learnsd that if I let go of this it will drop to the floor and people will come over and make funny noises which is funny.  Much of what human learn in implicit. It does not always require conscious effort.  It doesn’t’ take conscious effort to realize that your feet are going to blister if you walk barefoot on a hot surface.  Once it happens you don’t forget, but we also update our world model every second we are alive as our sensory input continues until death.
We are lucky to have the capacity to create a world model that helps us interact with our environment, as without it humans as a species would never have survived.  This is also true for animals who have to navigate through their environment by building a world model, albeit a much different one than we might have, although it is based on sensory input and a subjective interpretation of same, learned implicitly, and updated consistently.  While animal world models are different for each animal because of their sensory capabilities, they acquire the information and process it the same way we do.
AI systems don’t work the same way.  While many in the industry believe AI’s build their own internal world models, they are certainly unlike our own.  AI world models are quantitative not qualitative.  They are not based on sensory data but are based on numbers that have been labeled (mostly by humans) making the information explicit, and they are limited by the data they are trained on.  Of course, the typical response is, if you give them more data to learn on, they will get smarter, but we do not believe that is true because Ai systems do not have the ability to be subjective.  If two AIs are based on the exact same algorithms and taught with exactly the same data, they will arrive at the same answers, while humans will not.   AIs will certainly find patterns and relationships that we cannot, but unless they are told that a set of numbers represents an object falling to the ground, it is meaningless information.
Credit were due, AI systems are very good at finding relationships, essentially similarities that are extremely subtle.  In that way they can recognize that Dr. Seuss used certain words, certain rhyming patterns, certain letters, parts of speech, and other conventions that we don’t recognize.  In that way an AI can write ‘in the style of’ Dr, Seuss’, while we need some sort of sensory input to know that hearing “Sam I am” makes us think of Dr. Seuss.  But it doesn’t stop there.  The AI spits out an 8 line paragraph about a small environmentalist and moves on to the next task, while when we hear or see the word ‘Lorax’ we think of the happy times when that story was read to us as children or when we read that story to our own children.  That points to the difference in world models.:
In a world of bright hues, lived young Tilly True, who cared for the planet, the whole day through! With a Zatzit so zappy, and boots made of blue, she’d tell grumpy Grumbles, "There's much we can do!"
"Don't litter the Snumbles, or spoil the sweet air, Let's plant a big Truffula, with utmost of care!" Said Tilly so tiny, her voice like a chime, "For a healthy green planet, is truly sublime!"
We are not criticizing AIs here.  They are machines, essentially super calculators that have an almost infinite ability to follow instructions and find patterns but giving them more data doesn’t allow them to build a subjective world model.  While AIs can note that the color difference between two pixels in an image are different by 1 bit in a 16 bit number, our sensory (visual) input fits that color into our world view, and we say “Wow, those are beautiful flowers”. 
AGI, in our view, would require a huge amount of sensory input and the ability to place that input into a world view that is subjective, and at the moment, we don’t believe that is possible for any AI.  AIs can be better ‘pattern recognizers’ than humans and don’t get annoyed or tired, but they cannot ‘see’ or ‘hear’ or ‘touch’ anything and that is what keeps AGI from becoming a reality.  JOHO.
Side Note: Here is the image that we got when we asked Gemini, “How about you come up with an image that represents a human world view on one side and an AI world view on the other?”  That has to tell you something, right?

Samsung Electronics (005930.KS) and BOE (200725.CH) are rivals, not quite directly but Samsung Electronics’ affiliate Samsung Display (pvt) competes head-to-head with BOE in the small panel display market and to a lesser degree in the large panel TV space.  As we have noted, Samsung Display has been at loggerheads with BOE over IP issues and after a recent partial win concerning BOE’s misuse of Samsung trade secrets and IP, Samsung Display continues to fight BOE in the courts.  That said, Samsung Electronics also has issues with BOE.  As the largest TV set producer, Samsung Electronics, requires that those who use “We supply to Samsung Electronics” in their advertising, pay a royalty.  In 2022 BOE, who was the second largest supplier of TV panels to Samsung in 2021, refused to pay and Samsung has reduced BOE’s share as a TV panel supplier considerably since that time as they continued to battle over the royalty situation.
It seems that the President of Samsung Electronics TV division is scheduled to visit China in the middle of May, and BOE officials are expected to visit Samsung in Korea, with both expected to try to negotiate an agreement between the two on both panel prices and royalties.  The idea is that BOE can either lower panel prices to compensate Samsung or can leave panel prices the same and pay the royalty. 
While this seems reasonable, it might not be to BOE, who is also a major supplier to LG Display (LPL), Samsung’s local rival.  LGD has recently sold it’s last LCD TV panel plant (Guangzhou, China) to Chinastar (pvt), also a supplier to both Samsung and LG (066570.KS).  The large panel product that was being purchased from the LGD Guangzhou fab before the sale, would now become purchases from Chinastar.  Samsung has an internal requirement that no supplier of key materials can represent more than 30% share, and that means that it will have to maintain that 30% restriction, keeping it from purchasing panels from the Guangzhou fab now under the Chinastar banner.  While there are other large panel producers, such as AUO (2409.TT), Innolux (3481.TT), HKC (248.HK), and CHOT (pvt) that Samsung already buys panels from, Samsung tends to go with suppliers that have large capacity, leading to a secure supply, without violating the share limit..
At least to a degree, this puts BOE in the catbird seat or at least gives it some room to negotiate with Samsung, as Samsung Display is out of the large panel LCD business and supplies only QD/OLED TV panels to its parent which make up only a small portion of Samsung’s TV panel needs.  This leaves Samsung Electronics to outsource all of its LCD TV panel purchases.  As they cannot increase purchases from Chinastar without overstepping their limit, BOE is the obvious choice if they can come to an agreement over royalties.

In front of April 2, CE brands rushed to get containers into the US, only to find that ‘reciprocal’ tariffs proposed by the administration were to be postponed for 90 days.  One would expect brands to take a less aggressive approach to building US inventory levels given the lack of follow-through on these additional tariffs, yet it seems the postponement has done little to stem the inventory build in the CE retail space.  Given the exemptions the administration placed on a number of CE goods it can be seen that the administration is particularly aware of the impact that these potential tariffs would have on the US economy and the US consumer.  Yet brands don’t seem to be convinced that they will be at least partially and possibly fully exempt from new tariffs and continue to build inventory further as ‘protection’.
Sales for electronics & appliance retailers were down on a y/y basis in January and February and up in March (seasonally adjusted) and were above 5-year averages for February and March as shown below, so there was a bit of pre-buying by consumers.  That said, since the April 2 deadline and the subsequent postponement, inventory levels, at least in the PC space, have continued to increase. 
Typically 6 to 8 weeks of inventory are held across the CE retail supply chain, yet it seems that the average currently is about 10 weeks, with some mid and low price tier products in the 12 to 16 week range according to data from primary suppliers.  As we still have over 60 days before the next deadline, there seems to be little panic among brands, but more of a steady stream of additional product, with most traffic through standard shipping containers and little air transport.  But for containers to reach the US and pass through customs, they need to be on the water in early June and container prices to date are remaining steady, a good indication of the anxiety level of shippers at this point in time.  While they are not in a panic, they are not taking any chances.
We are a bit surprised that CE brands are continuing to build inventory after the postponement and exclusions, but it seems to indicate a distrust of the administration’s exemption fortitude and leans toward a worst case scenario in July.  We believe that the impact of reciprocal tariffs should they be implemented on CE products and therefore consumers, would be quite onerous and politically burdensome, making such an enactment very short-term (days/weeks) before exemptions come to the surface again.  While there are lots of countries that could see massive tariff increases, the primary US trading partners will likely be exempted again.  The harder question to answer would be how much the potentially high tariff status of countries like Vietnam, Indonesia, and the Philippines will affect prices if those tariffs are imposed, and whether India will be exempted as these are a key manufacturing centers for CE products.
In the interim, China remains the primary target for tariffs, and to offset the expectations for lower trade with the US, China will continue its efforts to increase local consumption.  We expect they will add additional consumer appliances to the subsidy programs that have been in place for months and will cooperate with provincial governments to add ‘smart’ home-oriented devices to the subsidy list, paid for with another round of bond sales of ~$41b US.  There has been some talk about the central government’s desire to consolidate the semiconductor equipment industry in China by merging China’s ~200 semi equipment companies into 20 primary suppliers.  While this will save subsidy capital and improve cost efficiency, it will take some time to implement and to see results, although it has been done in other industries in the past, typically through M&A financed with government support..

Buying a pair of glasses can be a traumatic experience.  For many, they are going to be wearing those glasses for most of the day, every day, for years, and a wrong choice can be devastating.  Now choosing glasses has become even more complicated as the glasses you choose that will help you read and keep you from stepping into traffic are not the only ones you need.  Not only do you need those prescription glasses (and maybe prescription sunglasses) but you also need a pair of AI glasses, now the hottest thing in China, where brands large and small are competing to grab consumer attention in this relatively new category.
In fact, similar glasses have been around for a while, but those were AR glasses that allowed you to overlay digital objects or text over what you see through ‘regular’ lenses.  Those are still a thing, typically dominated by Metsa’s (FB) Ray-Ban glasses, but as Ai becomes more embedded in our society, the drift is toward AI over AR, and in some cases both. 
So what are AI glasses?  Typically they look like slightly bulky sunglasses but have an integrated voice assistant that can understand what you want, similar to Siri, Alexa, or Google (GOOG) Assistant.  The voice assistant hears your commands through a number of microphones embedded in the frame and passes it to an LLM that parses speech the same way it parses text queries (actually not the same way, but similar).  The response can either be an answer or an action, typically responding through speakers also embedded in the frame or bone conduction modules that are touching your ear.  Most have some sort of image/video camera that can be activated to record an event or conversation, with some allowing direct livestreams to social media.
Of course, there are the applications that are usually on your phone, which is a necessary part of many AI/AR glasses, that communicate with the glasses, either by wire or wirelessly (typical) and allow the glasses to make calls, receive messages, and give you notifications, but when it really gets down to it, the applications available to each brand of glasses, whether AI or AR or both, are what makes them useful.
The most common application, aside from the basic messaging and notifications, is translation, which can be as complex as sentence by sentence instant translation that appears before your eyes (AR), or voice translation through the speakers.  This is not just for when you are traveling to another country, as anyone living in a metropolitan area is likely to face a few foreign speaking people each day.  They might not be talking to you (think nail salon, bodega, hospital, bus terminal) but it sure is nice to know what people around you are saying.  Existing aural applications like Spotify (SPOT), Apple (AAPL) Music, Amazon (AMZN) Music, or Deezer (DEEZR.FR) can be easily piped to your glasses, so no headphones or earbuds needed if you have glasses, but in the race to outdo other AI/AR glasses brands, there are lots of other applications that are finding their way into said glasses.
Object and scene recognition is one application that garners attention as it can be used for shopping (You see that person’s shoes? Find them for me”) or for navigation (“Tell me where I am -based on these buildings”), and while the navigation application seems to us to be the more important of the two, it is probably the other way around.  There are also health applications, with sensors that measure heart rate or oxygen levels and even some that are set up as hearing aids that use the embedded microphones and conduction systems to avoid having to stick obtrusive devices in your ears to hear.  There is even a set of glasses that can change their tint electronically and some that can read head or hand gestures, making it unnecessary to give a voice command unless a question needs to be answered..
As it is still very early in the ‘smart glasses’ game each new application or feature pushes that device forward into the public eye, only to be surpassed in days, weeks, or months by new features that catch the eye of consumers on another device.  Unlike smartphones however, which typically cost between $500 and $1000, smart glasses are less expensive and there are rumors that Chinese smartphone brand Xiaomi (1810.HK) is going to release their own branded smart glasses this year for just a bit over $200, making it difficult for smaller brands to compete.  While that might limit innovation a bit, it is certainly good for consumers who will benefit from low prices and feature competition similar to the smartphone space.
All in, we expect the smart (AI) glasses segment and the AR/XR segment to merge over the next two years and for new applications and features to drive expansion in the space.  But we also believe that in a relatively short period of time, most smart glasses sales will be based on large CE brands that exist today, with those brands focused on high unit volumes that will augment smartphone sales.  That said, it will be a delicate balance to keep smart glasses from eating into smartphone sales as some of that smartphone functionality shifts to the glasses.  We can also see a scenario where small inexpensive pocket computers, designed specifically for branded smart glasses, could replace smartphones altogether, but it is too early to make that call as consumers are just beginning to see the utility that smart glasses provide and designers are still trying to figure out the best ways to integrate AI functions.  It’s just the beginning of the cycle.