As is our custom, we don’t rely on a single or even two estimates when it comes to smartphone shipments.  We average as many estimates as possible to arrive at a realistic estimate that smooths those that might be considered outliers.  Our summary totals for 1Q 2025 came to 296.9m units shipped, up 2.4% q/q but down 0.3% y/y.  The range topped out at 304.9m units and bottomed at 289m units.  Samsung Electronics (005930.KS) was the only brand that saw q/q improvement as the company pulled in shipments to avoid any tariff consequences, although 4 of the 5 top brands in terms of shipments saw positive y/y comparisons.  

​Given the uncertain tariff situation and that uncertainty leading to a weak consumer electronics purchasing cycle, full year estimates for smartphone shipments have been declining.  While most still see positive growth for the year, that growth rate is minimal to none, as opposed to estimates made late last year in the low to middle single digit range.  At this point in time, with expectations that there will be another extension on the tariff negotiations with China, we would be happy to see a flat to slightly up year, but our hope for anything positive is withering day by day.  China seems to finally be exerting a bit of leverage with their dominance in the rare earth market, which seems to be slowing down what we expect the Trump administration thought was going to be a quick settlement.  If the negotiations are extended more deeply into 3Q, we expect it will continue to dampen demand and leave 4Q as the ‘saving grace quarter.

AI is visible.  If it is not on the box of a CE product, it will most likely be mentioned in the promotional literature, the Set-up guide, or ‘Controls & Instructions”, and will certainly be mentioned (many times) in the advertisements promoting that particular CE device.  In fact it would likely be more difficult to find premium CE products that don’t mention AI, but aside from the chatbots and LLMs that are typically the face of AI, much of what we are told is AI is not seen but can still affect many of the device functions we use on a daily basis.  Smartphones are obvious AI candidates and the competition for ‘AIness’ is already underway, as it is in laptops, desktop computers, and even in tablets, but what about TV sets?  One doesn’t typically query your TV about “What’s happening at the Diddy trial?” and rarely does searches on your TV, other than when looking for a show or movie, but AI does play a part in the TV set space, even though you don’t see it at work.
It goes deeper.  Some TV set AI functions, while not controlled by the viewer, can be seem by the viewer.  Video processors in premium TVs are no longer just perform the same mathematical functions on each frame.  Now they adapt how they process images based on a neural model that was rained for that specific purpose while other Ai functions are totally invisible to the user, although they are able to influence users in other ways.  Non-AI image processors are traditionally rule-based (Think of a recipe for Lasagna).  They use fixed algorithms based on well-established formulas and are applied uniformly across images or sections of images.  They typically work toward improving picture sharpness, enhancing edge definition, and reducing noise by working on small squares (3x3, 5x5, etc.) of pixels.  AI image processors are constantly adapting to actual conditions (Think of a chef who constantly modifies a recipe to improve it).
As much of the content we watch is not high resolution, TV set processors spend considerable processing power trying to upscale video for current 4K displays, essentially trying to make a low resolution image look better on a high resolution display.  Typically this is done through one of three processes.
Bilinear interpolation – This function averages the 4 nearest pixel values to create a new pixel value to fill in extra pixels.  The result is a soft, blurrier image.
Bicubic interpolation – This function uses a larger block of pixels (16) and a more complex polynomial function to create new pixels.  The result is a sharper image but can create artifacts.
Nearest Neighbor – This function just copies the values of the nearest pixel.  It produces a blocky image but is the fastest of all.
There are also global functions, such as those that adjust the overall contrast of a scene based on the darkest and lightest parts, but older video processors lack the ability to understand the details, and processes like general noise reduction can also reduce detail.  Users typically have to adjust settings for different types of content (movies, sports, news, etc.) as the functions of typical TV set video processors are generic and unable to easily adapt to differing content or environmental conditions.
AI image processors are different, and while each brand has their own AI processor with its own characteristics, here is what sets them apart from the generic processors that are and have been used in TV sets.
Convolutional Neural Networks – These processors use deep learning networks, with CNNs trained on very large datasets of images and video that allows them to learn complex relationships.  Some of the training material consists of image or video pairs, with one being low resolution and the other high resolution.  When comparing the pair on a pixel by pixel, square by square and segment by segment basis, the CNN is able to develop the same, “This is the most likely next pixel” process that a trained LLM would use to figure the next token in a sentence.  By using this technique, it generates the ‘most likely’ pixel to fill in blanks when upscaling.  While these pixel choices are being made, the AI is also evaluating the scene on a frame by frame basis to better understand the image flow.  This can help the Ai to identify the content type (movie, sports, news, etc.) and make adjustments to match.  It can even go as far as recognizing objects in a frame in order to independently adjust their characteristics on the fly.
Depending on the AI, and its capabilities, the AI processor can also perform highly accurate color tone mapping, keeping object details accurate in both light and dark situations, essentially creating an HDR-like image from SDR content.  The AI can also be used for noise and artifact reduction, having been trained to recognize the difference between various types of noise (film grain as opposed to digital noise) and can even clean up motion blur that occurs with fast moving objects in lower resolution sports content.
Similarly AI processing can also be applied to TV sound where acoustic tuning, adaptive surround sound, and dialogue enhancement are all able to be accomplished using highly trained ML systems, and voice controlled remotes will improve greatly as AI LLMs better understand the context of commands.  But what about those mystery AI functions, the ones that work completely away from the user?  We know that many TVs track user viewing habits and that data is ripe for Ai’s to use for content recommendations.  Rather than the endless searching for attractive content that users wind up doing, Ai systems are easily tasked with real-time searches that are adapted to the user’s habits or past requests, but here’s one function that we can safely say would be a pleasure when implemented…live translation.  Instantaneous on-screen sub-title translations, and eventually (not that far off) direct translation to match the image without sub-titles, essentially lip-syncing the translation to the image.
There are a host of other TV set features, many of which are embedded in process applications that AI can improve, and we haven’t even touched on gaming, where there are many features that AI can improve by ‘learning’ from its users, but we would also be remiss if we did not mention the fact that along with the improvements in TV sets that AI is bringing comes a decrease in privacy and security.  ACR (Automatic Content Recognition) systems are always running when the set is on and identify all content, regardless of source, being displayed on the screen by capturing screenshots of content roughly two times each second.  Those screenshots are cross-referenced against a huge database of media content and advertising, providing a rich history of the viewer’s habits that is used  to enhance content recommendations but can also be sold to marketers, especially given how difficult TV set brands make it to disable the feature.  Passing this user information to an AI system, adapted to spot patterns and nuance, such systems would open the user to serious privacy concerns.
But when it comes down to the pluses and minuses, AI has the ability to improve TV set performance significantly and given the still early stage of AI and ML systems, brands have barely scratched the surface.  Brands are still experimenting with AI and applied AI processes to some of the more obvious functions we have mentioned.  As these processors improve, we expect there to be a shift from the more typical ‘Bigger is better” and “Higher resolution” competition to “…can your TV do this?”.  Differentiation is the name of the game in the CE space and feature cycles can dry up in an instant, so AI features, some of which the user may have to take the brand’s word for, open up a whole treasure trove of marketable features.  What brand would not look forward to that?

Sources:
https://www.samsung.com/us/televisions-home-theater/tvs/ai-tvs/
https://www.advanced-television.com/2025/03/26/the-future-of-television-how-technology-is-shaping-the-industry/
https://www.trendhunter.com/trends/copilot-ai-for-tvs
https://www.reliant.co.uk/blog/how-does-ai-improve-the-sound-of-a-tv/
https://www.lge360.com/iraq_en/lg-story/helpful-guide/ai-tv-vs-traditional-tv
https://insights.samsung.com/2025/06/03/how-does-ai-upscaling-work-on-direct-view-led-displays-2/
https://www.lg.com/us/experience/ai-tv-for-sound-and-vision
​

​While many consider the UN to be a powerless organization, it does have considerable cache and headline significance at the least, and a recent report by the UN Council of Presidents of the General Assembly on AI, more specifically AGI (Artificial General Intelligence) seems to be making a passionate call to UN members to initiate discussions and build a consensus opinion on the governance of AGI.  The report calls for dedicated UN General Assembly sessions  on AGI and the development of an ‘observatory’ to monitor AGI research, track development, and potentially issue warnings.  Along with these discussions would be the formation of a certification system to ensure responsible development of AGI and the creation of a dedicated International UN agency to oversee responsible AGI development and usage.
Unfortunately, while the concepts are all valid and meaningful, there is little actual power behind the UN, which finds geopolitical ambitions far outweighing more positive societal goals, and has little or no influence over countries known to be ‘bad actors’ relative to other potential world threats.  That said, here are the salient points from the report:

• AGI could emerge within the decade due to unprecedented financial investment and R&D.
• AGI holds immense potential (emphasis on public health) for scientific discovery, industry transformation,, general productivity, and sustainability goals.
• However, the risks are irreversible once AGI is achieved and integrating AGI with critical infrastructure could lead to a point where human oversight is no longer ‘reliable’ (implies “no longer acceptable”).
• AGI could be used by malicious actors to develop weapons of mass destruction (Chemical, biological, nuclear, etc.) and could be massed into swarms as a new WMD category.
• Critical infrastructure could be at risk if AGI is used in cyberattacks against national systems such as energy grids, communication systems, and financial markets causing widespread disruption.
• AGI development and control are currently concentrated among a small number of countries and companies and could exacerbate global wealth inequalities, lead to mass unemployment, the erosion of privacy, and the eventual heightening of geopolitical tensions.
• Lastly, AGI could create its own autonomous objective that is not aligned with human values, citing that the potential to overpower humans are not far-fetched hypotheses but risks that could take place also within the decade.

None of these concerns, positive or negative, are surprising or unusual, and have been bandied about  for years.  The report however seemed to indicate that the AGI timeline was much shorter than what was expected only a few years back, when AGI was predicted for 20250 or 2060.  Now a number of AI industry leaders expect AGI to be achieved within the next 5 years, with some as soon as 2 years, increasing the urgency of the report, especially when speaking about ‘irreversible consequences’.  That said, it is hard to imagine that any major nation is going to do much more than provide lip service to anything that might slow down that country’s push toward AI development and there is little that the UN could do to rein in bad actors, either members or non-members, making the entire report valuable form an intellectual standpoint but far less so from a practical one, regardless of how correct or incorrect it might be.
.

​According to President Trump’s post on Truth Social (see below) a deal with China concerning rare earths is done, although general news media tends to disagree, calling whatever has been agreed on a ‘framework’ rather than an agreement.  That said, it was enough to get China to resume rare earth exports after roughly two months of restricted (essentially halted) rare earth shipments, in return for a resumption of educational visas for Chinese students and a reduction of tariffs on Chinese goods from 145% to 55%, inclusive of a 10% baseline tariff, a 25% tariff remaining from Trump’s first term, and a 20% additional tariff related to Fentanyl trafficking.
OUR DEAL WITH CHINA IS DONE, SUBJECT TO FINAL APPROVAL WITH PRESIDENT XI AND ME. FULL MAGNETS, AND ANY NECESSARY RARE EARTHS, WILL BE SUPPLIED, UP FRONT, BY CHINA. LIKEWISE, WE WILL PROVIDE TO CHINA WHAT WAS AGREED TO, INCLUDING CHINESE STUDENTS USING OUR COLLEGES AND UNIVERSITIES (WHICH HAS ALWAYS BEEN GOOD WITH ME!). WE ARE GETTING A TOTAL OF 55% TARIFFS, CHINA IS GETTING 10%. RELATIONSHIP IS EXCELLENT! THANK YOU FOR YOUR ATTENTION TO THIS MATTER!
Jun 11, 2025, 8:04 AM
The WH was under pressure as Ford (F) was forced to halt production of Explorer SUVs at a Chicago assembly plant last month when a supplier was unable to obtain the rare earth magnets that are part of a number of key sub-systems and the US military began indicating concern over delays in certain programs that were highly reliant on rare earth materials. The Ford supplier has indicated that it has been granted a 6 month export license this month from the Chinese government and is resuming supply and we assume others will be doing the same, although the current agreement is not formal and will expire in August if terms are not agreed upon and those include general trade imbalances, intellectual property issues, and US business licensing issues on the mainland.
The US dependence on China for its rare earth supply presents an unusual challenge for the current (and future) administration, as the US imports 99% of its rare earth consumption and China controls ~70% of rare earth mining and ~90% of rare earth refining.  While the size of the US consumer market makes it extremely attractive to China and gives the US considerable trade leverage, the impact of even a short Chinese ban on rare earth exports shows China’s leverage over the US and how quickly we respond  We expect China will resume export licensing for rare earth suppliers this month and next, but will be careful not to let the US stockpile any significant amount of material in case the negotiations fail.

3D has come and gone a number of times, usually with less than promised results, but that does not stop CE companies from trying, and this time it is a combination of Hewlett-Packard (HPC) and Google (GOOG).  Google announced the idea in May of 2021 under the name Project Starline, with the goal or revolutionizing video conferencing, but had been working on the concept for a number of years before the announcement and demos.  Testing began in late 2022, with the idea that a more nuanced communication between two parties would improve their ability to communicate.  Now, in partnership with HP, the two have released the first commercial implementation of Google Starline, now known as Google Beam.
The system is based around a 65” TV display which allows for life-sized video, along with 4 speakers that produce spatial audio, meaning the position of the sounds is encoded  in the audio signal and can reproduce the precise 3 dimensional location of its origination.  This is accomplished with 12 microphones that enable the system to catch every word or sound and its precise location, all of which is part of HP’s Poly Studio A2 Audio Solution bridge for video conferencing.  The heart of the system is Google’s 3D imaging system and light field rendering, along with a volumetric video AI model, which needs some explanation.  A typical video image is 2 dimensional, with height and width but no depth information. 360° video does capture the depth information, but it is all oriented from a single  fixed point.  The volumetric model captures what is known as 6DoF (Degrees of Freedom) which not only captures the height, width, and depth as it moves through time, but also captures Pitch, Yaw, and Roll.  This allows the model to look at any ‘slice’ of video (time) from any perspective, no longer tied to a single point, which, in theory, means that you could look behind, around, or above or below the image without any additional hardware.  Of course you need a number of high-speed cameras and depth sensors to feed this information to the model, all of which are included in the system.  Then the Ai reconstructs and renders the data.
In order to play back this complex information the system creates a number of ‘views’ of each video slice, each with a slightly different perspective.  Each group of pixels in the display is covered with a microlens array that directs light in different directions.  The system can control which sub-pixel under each lens receives each piece of information and therefore controls the direction that information flows. So one eye will see a slightly different image than the other eye, creating the 3D effect as it does in nature.  It is a complex process, but with the Ai running the show the system should produce realistic live video that contains significant depth and looks as if the person were sitting across from the viewer.
Here's the problem.  While the idea of 3D communication might be attractive, it is expensive with a single system costing about $25,000.  The benefits are a more realistic conversation and the ability to notice considerable nuance that might not be visible in a 2D image.  But is this necessary?  Perhaps during tense geopolitical negotiations where every move of the hand or tap of the finger has meaning, but relatively little value in most video conversations, which commonly use relatively inexpensive or free technology.  Certainly not anywhere near the quality that the Google Beam system can generate, but att $25,000/unit (you need at least 2) it seems a bit out of the realm of most consumers and even a stretch for most businesses.  This leaves us to look at the system more as a ‘proof of concept’, showing it can be done.  As to refining it to more commercial pricing, it could be difficult given the number of microphones and cameras needed to create the 3D imaging along with the AI model and associated equipment.  It works, now what?  After 7 or 8 years of research, it seems a bit of a waste.
 

OpenAI (pvt) has announced a substantial price cut for its flagship reasoning o3 model, and when we say substantial, the cut was 80% for both input and output tokens.  This brings the model pricing down to a level similar to that of competitive models like Google’s Gemini 2.5 Pro, Anthropic’s (pvt) Opus 4, and DeepSeek’s (pvt) reasoning suite.  At the same time it points out the arcane science of model pricing, a sometimes opaque system that can involve a number of different parameters.  In the back of our mind however, every time we hear anything about model pricing a bell rings that asks “Will this be a major moneymaker for those that have spent millions or billions building infrastructure or training models?”, or will it become a ruthlessly competitive market where price competition will dominate and profits will be minimal?
We believe it is too early to answer the question but getting an understanding of how model pricing works and what drives pricing dynamics, at least at this point, is essential for having a basis for understanding how model pricing will play out over the next few years.  Here are the most common model pricing types:

• Token-based - The most often used pricing model for AI systems is token based.  Tokens represents the smallest unit of text that an AI model processes, which can be a word, a number, or even a punctuation mark.  For reference, 1,000 tokens equals roughly 750 words.  This pricing system is not only the most transparent for the user but ensures that the model is billing for actual usage, regardless of scale.
• Subscription based - Consumer-facing models tend to be subscription based, utilizing fixed rate pricing based on model access and type.  There are access levels that can allow unlimited access or unlimited usage and give the user a predictable cost within access parameters.
• User-based – More common in enterprise and team oriented situations, users are each charged a flat fee on a monthly basis which simplifies billing and managing multiple users.

However there are other factors that can influence model costs that are less obvious than billing schemes.  In particular, as mentioned above, input and output tokens are a key influence of model costs, with input tokens, essentially the text of the input prompt (query) being less expensive and output tokens (the answer) more expensive due to the processing resources that must be used to generate the response.  The context window is also a factor.  This metric is the maximum number of tokens that can be submitted to the model in one interaction.  This determines the maximum document(s) size that the system can handle, which determines the usefulness and the computing cost of the model.
In the timeline below we show the progression of  model context window size  from 1Q 2023 to 2Q 2025.  Early models had context windows that could handle less than 10,000 tokens or ~7,500 words.  Over time that has progressed to 1,000,000 tokens or about 6 ½  average size books or 2,500 pages and has become a key model selling point as it reduces the need to break up large  queries into smaller chunks that  have to be processed individually.   Large context windows do have a cost in terms of memory usage, computational overhead, and slower response time, and the model  has less ability to focus on specific details,  essentially wasting the model’s time sorting through a large token base, but large windows are ideal for  searches in very long documents where the entire document can fit into the context window in a single pass.

The model price comparison table shown below points out that that model pricing has already become a competitive market, but it remains a cost vs. performance evaluation for commercial users and enterprises.  There are ways to offset some of the costs (volume discounts, off-peak processing, etc.) but model selection (performance needs) and cost make up most of the decision process, however as the number of large models continues to increase there are additional lower cost alternatives that continue to bifurcate pricing tiers. 
The price comparison table shows that Google’s Gemini 2.5 Flash and DeepSeek’s V3 models are the most cost effective, while OpenAI’s GPT 4.5 (Preview) is the most costly.  We expect model brands will begin to stake out price/performance tiers more specifically over the next year, but with performance increasing rapidly against a backdrop of increasing training costs, models that can provide lower cost training without sacrificing quality will have an edge.  Deepseek is pursuing such a path but faces the fact that its servers are hosted in China and are therefore subject to Chinese PIPL and DSL rules, allowing the Chinese government access to said data under national security rules, ruling out US companies using these models for anything other than generic processing.

While the AI model business is new and in a high growth stage, it has the same characteristics as any other high volume, user-based, feature oriented systems and is already being molded by the competition between model brands, even though most of these brands are still in a non-profit mode (not all).  As the financial constraints of model training continue to expand, the need for profitability will become the primary driver and pricing will have to reflect that, but the need to drive user growth will make Ai pricing an even more competitive schema.  We expect China will be the low-price leader but constrained by politics, while the big players like Google, Meta (FB), and OpenAI will battle it out on a feature-by-feature basis in a “We might be expensive, but can your Ai do this?” world…

At a recent show a team from Postech (Pohang University of Science & Technology) in Korea demonstrated a system that produced sound as part of a 13” OLED display.  The team was very specific as to what this proof of concept system was and how it worked but for some reason a number of tech rags picked up the story and transformed it into one where the implication was that the OLED pixels were generating sound.  As far as we could see, this was not the case and the implication that ‘eventually every pixel in the display could act as a speaker’ was a figment of someone’s imagination and had no bearing on the concept shown.
The idea of embedding sound generation directly into a display has been around for years.  LG Display (LPL) has been demonstrating  its “Crystal Sound” system at shows for years, and Sony (SNE) has their ‘Acoustic Surface Audio’ system, both of which use electro-magnetic speakers attached to a frame that surrounds the OLED display.  These transducers (speaker) work, just the way your stereo speakers or even the speakers in your smartphone work, by sending an audio signal to a voice coil (an electromagnet), which interrupts a permanent magnetic field and moves a diaphragm.  That diaphragm creates sound waves, and you hear music or speech.  Unfortunately these ‘exciters’ are thick and not conducive to the ultra-thin displays in current use, but they also have another problem. 
External speakers, such as a soundbar or individual speakers placed in a room are isolated from each other, while the speakers in the systems mentioned above are attached to the same frame.  This causes them to interact with each other which creates peaks and valleys in the sound that are not part of the content or music.  As the human ear is very sensitive to loudness levels, this can make the sound less than perfect.  Some have tried using piezoelectric speakers, which convert an audio signal directly into diaphragm movement (no voice coil), which are thinner and less expensive, but they still interact with each other if mounted on the same display frame.
What the team from Postech did was to embellish the display frame by creating nine isolated sections in the frame with an exciter in each.  This reduced the interaction between each exciter and allowed each exciter location to operate independently of the other, creating a localized sound from each of the nine locations on the display.  This was certainly an accomplishment and indicated that it was possible to create a display that was able to localize sound across the display, without the distortion typically generated by a single frame.  Whether this is a everyday solution for display sound remains to be seen both from a practical and a cost standpoint, but it had little to do with the OLED pixels.  In fact the only reason an OLED display was being used as a test bed was because LCD and Mini-LED displays have too many layers that dampen the vibration of the frame.
So, taking the idea to its ultimate goal, attaching millions of piezoic speakers to a system that would ‘vibrate every pixel and turn each into a sound source’ was not really part of the demo, nor does it seem even remotely practical.  In order to create sound, the ‘millions of piezoic exciters’ would have to be attached to a frame, so we are talking about a frame that must be at least somewhat rigid yet able to adhere to millions of very tiny speakers and each of the millions of exciters would have to be isolated from each other, an almost impossible task.  Perhaps that can be accomplished someday using a lithographic process at a cost that is reasonable, but there is another problem.
In order for this system to originate localized sound at every pixel, the ultimate objective, the sound would have to be separated into a grid with a million localized horizontal and vertical points and each point in such a matrix would have to have a connection to one of the million exciters.  This is similar to how each pixel receives video information from a display system’s video processor, a complex task at best.  Doing this for audio would add considerably to the cost and complexity of any such system.  Again almost anything can be done with silicon, but at what cost and for what gain? 
The idea of localizing sound across a display is conceptually sound (sorry!)  but not nearly as necessary or as practical as it might seem, although object-oriented sound recording would work in favor of such a system.  If one were to use 1” square exciters for a 65” TV, it would require ~1,800 exciters, a frame that could isolate them all, the power necessary to drive 1,800 speakers, and a system to assign the audio matrix to each speaker.  That seems to be a bit of overkill.  More practical both from a structural and a practical one might be eighteen 10” exciters, each covering a 10” square section of the screen.  Do we really need to be that precise as to the location of a spoon falling to the floor?

Not surprisingly May was another uneventful month for Taiwanese panel producers as the ever volatile trade situation leaves both producers and consumers with little understanding of what to expect.   CE product status quo seems to be to keep inventory levels in the US relatively high, despite weak demand, anticipating a potential change in the current tariff situation.  As we get closer to July, when the most recent tariff postponement is  set to expire, we expect shipments to increase to get them over the US border by July 9, although some are beginning to doubt whether the ‘reciprocal’ tariffs will ever be implemented, at least for our more significant trading partners.
Panel sales in Taiwan, which are reported monthly, indicated that May m/m sales for the two large panel producers (AU Optronics (2409.TT) and Innolux (3481.TT)) were between -1.4% and +4.5%, relatively minor changes,  and with little or no change in May panel pricing, shipments also look to be flat, leading to another month of watching and waiting.   Given that inventory levels remain on the high side in the US, we expect brands will maintain current or slightly lower production levels until the 9th of July, before planning the seasonal holiday build.  Much will depend on the ongoing tariff situation, but we expect there will be less of a rush to build holiday inventories this year than last, and given the recent reductions in CE product forecasts for the year, we expect brands to remain cautious, regardless of the tariff situation .  

In response to US export restrictions on semiconductors and semiconductor production equipment, China has its own “Dual-Use” list of items that can be used in everyday items and in weaponry.  These items require a governmental review and the issuance of a license before they can be exported from the country, essentially allowing the Chinese government to use them to limit US advanced R&D and undermine the US military and semiconductor development efforts.   There are a substantial number of items in the Chinese list that are produced primarily in China, with only a small percentage of total consumption available from sources off the mainland, leaving little chance of the US stockpiling such materials or rapidly replacing newly restricted Chinese materials, and while the US government downplays the potential impact that these restrictions might have on various military projects and industries, even small shortages can slow or halt billion dollar programs and certainly contribute to rapidly rising costs.
Last week China abruptly restricted the export of Indium Phosphide substrates, likely in response to some  recent restrictive move by the Trump administration.  InP substrates are used in the production of laser diodes, photodetectors, and single-chip combination laser/amplifier/modulators, but most importantly InP is used to produce High Electron Mobility Transistors (HEMT) and a variety of RF and Microwave advanced communication components, along with high efficiency solar cells that are used for space communications.  The specific objective is not known but even with a number of suppliers of Indium itself, other than China, InP substrates will become harder to source.
Each addition to China’s list, especially in the rare earth category, puts more pressure on the administration to strike a deal with China, as alternate sources for a number of key materials are small but rarely does the WH indicate that China has any leverage in trade talks.  In fact, when it comes to rare earth elements, while the US Geological Survey does not break out individual rare earth  consumption and sourcing for the US, on an overall basis , the US imports ~80% of its rare earth consumption each year, with 56% of those elements coming from China.  Based on 2022 data, China is the leading supplier of 29 of the 50 most critical minerals and controls 85% of the processing for those minerals.

For example:

​Among the most critical is Yttrium, an element  that the US imports almost 100% of its needs from other countries, 93% of which comes from China.  The Department of Defense has identified Yttrium as essential for military applications and indicated that even a 90 day supply interruption could have a significant effect on defense contractors.  42% of yttrium consumption comes from its use in red display phosphors and LEDs, the radioactive Y90 isotope is used to treat cancer (64 day half-life), it is also used as a high-temperature metalogical additive for jet-engines, and as a source (Yttrium-Iron Garnet) in microwave radar systems, making it a very necessary material for both the military and commercial markets.
While Yttrium is an extreme case, there are many other strategic materials where production and refining are dominated by China and despite the 2025 Nation Defense Authorization Act allocation of $1.2 billion for stockpiling rare earths and $350 million for domestic development of rare earth materials, the US has a critical dependency on China for both rare earth minerals and a number of other important materials for both the military and commercial products.   Products like semiconductors, TVs, smartphones, and soybeans make the headlines when it comes to trade negotiations between China and the US, but underlying those popular products are many far more critical  materials that we believe are also a major part of the negotiations but receive far less publicity because of their sensitive nature.

China has a somewhat unique system for investors who wish to communicate with companies they are thinking of investing in or already own shares.  They have an online platform where investors can ask questions of company management and hopefully get answers .  Not all companies answer questions, and not all questions are answered, but it does give a unique window into what investors are thinking about specific companies.  We check those sites to see if there is any unusual change in sentiment or if the company has mentioned something significant (rarely) and we took a small poll at last night’s peak at the Q&A for BOE, China’s largest display producer.
We looked at the first 19 questions that were asked by investors that were answered by the company.  Out of those 19 questions, the score was as follows:
14         What are you doing about the poor return on shareholder value?/Market Value?
3           How will your all-in-one IoT products gain traction as there are many competitors already?
1           I question the authenticity of your financial data./Potential fraud.
1           Will you purchase production line equity as did Chinastar?
With 73.7% of the questions on the first two pages of the answered Q&A concerning the poor stock market results of the company shares, it seems quite obvious where the mindset of BOE investors lies currently, and we note that this is not as common as one might expect.  Looking back at other times, questions have been based on specific topics that are the focus of a BOE announcement or an article in the Chinese trade press.  While there is always some complaining about investor return, this time it was unusually large and rather vitriolic.   As noted, the company rarely makes specific comments as they might be construed as material information, but gives the usual party line, which they often repeat word for word in subsequent quesions. 
Here is an example (Fraud’ question):
Hello, Secretary Dong: I am an individual investor. Regarding your annual report released on April 15 and the first quarter performance forecast that met market expectations, the stock price has been falling. As an investor, I question the authenticity of your financial data. Is there any fraud? The state requires state-owned listed companies to do a good job in market value management. How do you do market value management? We adhere to value investment. How does your company reward investors?
Hello! The company strictly abides by relevant regulations and fulfills its information disclosure obligations in a timely manner, and there is no financial fraud. The company's achievements today are inseparable from the long-term trust and support of the majority of shareholders. In order to achieve continuous, stable and predictable shareholder returns, the company recently announced the "Shareholder Return Plan for the Next Three Years (2025-2027)", planning to distribute profits in cash each year not less than 35% of the net profit attributable to the parent company's owners in that year; the total amount of funds used to repurchase and cancel shares each year is not less than RMB 1.5 billion (share repurchases for other purposes such as equity incentives will be specially planned separately); consider mid-term profit distribution under certain conditions. In order to implement this return plan, the company plans to distribute cash dividends of 1.87 billion yuan in 2024, accounting for 35% of the net profit attributable to the parent company in the consolidated statement that year; at the same time, it plans to repurchase and cancel A shares of 1.5 billion to 2 billion yuan; in addition, it also plans to cancel treasury shares worth nearly 1 billion yuan, in a multi-level and combined way to effectively improve shareholder returns. According to statistics, from 2015 to 2024, the company has implemented cash dividends for 10 consecutive years, with a cumulative amount of nearly 22 billion yuan; from 2018 to 2024, the company's annual cash dividend ratio has remained above 30% of the net profit attributable to the parent company for 7 consecutive years, allowing investors to share the growth results with the company. From 2020 to 2024, the company's cumulative payment for repurchasing A shares exceeded 5.6 billion yuan, and the cumulative payment for repurchasing B shares was nearly HK$1 billion. With the company's further high-quality development, the company hopes to continue to share growth dividends with shareholders and achieve mutual benefit with shareholders. Thank you!

The Q&A forum is an interesting way to see what more aggressive investors are thinking, while also understanding that the forums can be used by short sellers, angry shareholders, and just about anyone, to plant negative stories and ideas, which then become obligatory for the company to respond to.  While like any forum, it’s IIWII (It is what it is), but it is an interesting idea for US companies, who will immediately say “it takes too much time to respond and most of the questions are ridiculous”, but it does give some insight into what rank and file investors are thinking.  Whether the company cares is another question.
Side note:  In asking Gemini to create the final image below, we used the following prompt: “Can you create an image of an angry shareholder holding a stack of papers while banging on a table at a corporate shareholders meeting?” Once the original image was created, we asked Gemini to change the investor to one who is Chinese as this references BOE.  It made the change, however not without a slight modification.  We corrected it for the final image, but it points to the fact that AI image generators do not have any understanding of the physical world, even if their training images depict it in a certain way.  They match what they have ‘seen’ in training data to what the prompt requested but sometimes forget that no matter how angry we might be at a company management, we don’t have 3 arms.