January 2018 was the peak of the last memory cycle.  Then began a decline that continued through the year, a result of a significant amount of capacity oversupply.  The decline continued in 2019, with a slight recovery during 2020 due to COVID work-from-home demand and in 2021 (through the 1st half) that demand source continued.  In the back half of 2021 however COVID related demand began to wane, and DRAM prices began to slide again, and 2022 saw DRAM prices almost in freefall.  Toward the end of 2022 the key DRAM suppliers began to make investment and production cuts and in 2023 they began to make substantial cuts to reduce inventory that had been building for many quarters, leading to a gradual but steady recovery through mid-2024.

The AI wave began to give DRAM manufacturers relief last year which has continued into 2025 as we have noted previously, but with a vengeance, and by mid-year weekly double digit price increases were becoming common.  Over the last month, while prices have not declined they are no longer rising, having flattened for the last 6 weeks.  That said, projections for 3Q are for another 10% to 15% price increase for ‘generic’ DRAM and a 15% to 20% increase when including HBM.  Expectations for NAND in 3Q are more moderate but still project a 5% to 10% increase for the quarter.
Much rests on HBM as it is the current growth engine for the DRAM market with sales expected to almost double this year, and while AI continues to drive the HBM market, estimates, even those more than 5 years out continue to increase.  The enthusiasm behind AI growth, which should drive HBM products like DDR5 are also dragging along increased pricing for mainstream and legacy products like DDR4, which do not have the demand pull from AI.  In fact the price increases in DDR4 and similar memory products is also due to constrained supply as memory producers shift more production capacity to high-end HBM products which generate sales that give the share gains that memory manufacturers are striving for, despite the mantra that the industry is now focused on profitability over share.
The question then becomes whether the capacity and production management that DRAM suppliers are currently working under is enough to satisfy their desire for profits and share or will the desire for bragging rights drive the same cyclical expansion that devastated pricing during the last cycle?  One would like to think that the memory of severe price reductions would be fresh enough in their memories (sorry) that they might avoid the capacity expansion that caused the problem, but business is business and shareholders are shareholders, and the demands of same tend to cause memory loss (sorry again). 
As it turns out all three major memory producers have announced plans for capacity expansion.  Of course it is justified by the fact that most of that capacity is oriented toward HBM where both short-term and long-term estimates are rapidly increasing.  However, as in every cycle there are always estimates that justify an almost infinite growth path, as is the case for HBM and AI generally.  Even the perturbation of DeepSeek’s (pvt)  low-cost training model made little difference to memory pricing on both short and long term memory capacity plans, so we are left to watch the tea leaves and wonder whether the current memory price plateau has meaning or whether it is just another step in the stairway to the stars.  We know that recent capacity increase plans will take a number of years to implement, which those who see the glass half full look at as time needed to build even more HBM demand, while we expect the scenario that the memory business will be facing by the time that new capacity is in full production will be far different than it is today.

Using LLMs is what is known as passive reception, essentially the opposite of active inquiry and critical thinking.  You ask a question, and you get an answer.  Any critical thinking that needs to be done is done after the answer is provided.  The Socratic method is an active one.  Rather than the ‘learner’ asking a question and receiving an answer,  the ‘learner’ asks for information and the response comes back in the form of a question.  This creates an active dialogue between the teacher and learner that allows for the dynamic interplay of ideas rather than the passive uni-directional flow of information.
The folks at OpenAI (pvt) like Socrates and have created a Socratic Method model to help students, currently college students, learn more effectively.  Here’s how it works:
You start with a topic.
“Teach me about multi-dimensional models”
The model responds by formatting your question into a series of questions that it poses to you. With a small introduction describing how multi-dimensional models are able to discover subtle differences and similarities that are not easily apparent in low-dimensional space, adding the idea that dimensions are like ‘features’ in humans.
Then you get a question like:
“What might you expect ‘features’ to be if we are thinking of people?”
You answer:
“Height, weight, eye color, hair color”
The AI comes back with how good an answer that was and describes a bit more about low-dimension (1 - 3) features, higher dimension (5 – 100’s) features, and how the more dimensions an object has the more similarities (and differences) between them can be discerned.  Then it asks the next question:
“What might be the downside to high-dimensional models?”
You respond:
“More dimensions means more comparisons for each item, which means longer compute time for comparisons and more expense”
Again, the model responds by telling you how insightful your answer is, adds a bit of color to the answer and follows with another question.  This repeats five or six times as you get a bit deeper into the concept until the AI reaches some sort of learning module break point and the session ends with a relatively simple summary.  While we have tried this system, we have yet to go past two consecutive sessions to see how deep the AI will go into the topic without a specific prompt from the user, but on the surface it seems to be a rather ‘lite’ version of learning, like having someone quiz you on Cliff Notes.
Our main objection, although we like any form of learning, is that the model is obsequious and sycophantic, at least with us, which we expect Socrates might not have been.  Too many “Excellent answer, you are right on target” and too few “Close, but you are missing an important point”.  Perhaps we are at the age where we don’t need constant reassurance that we can absorb complex ideas and concepts and would rather be challenged, which might be different than a college student who feels pressure to excel and justify a six figure tuition or a grant, but in any case learning is learning.  With a few refinements the system could improve the interaction between LLM and user and stimulate better understanding of topic related learning.  That said, because it is currently free, and extremely easy to use, we expect it will have an impact on educational AI systems that are currently sold to large institutions and organizations.  We expect there will be a higher level version at some point that will carry a price tag, but maybe OpenAI will make it open source for the good of all students across the globe…
​

Normally our focus with Corning (GLW) is on two of the company’s 6 broad product categories, display and specialty materials, the segments that supply substrate and cover glass to many CE products.  However, the world is now focused on AI and anything that smacks of AI gets all the attention.  Such is the case at Corning, and rightly so as the company’s optical products are in the sweet spot of the race to build out AI training and inference capacity.  At the same time carriers, who have been working down inventory for many quarters (more than everyone, including Corning thought) have finally returned to ordering in line with actual expansion, releasing the pressure that has contained carrier optical sales.  Other than display, which saw sales down 0.8% q/q/ and down 11.4% y/y all Corning divisions beat consensus sales  estimates.
Here are the sales and net income metrics for 2Q by division:

​We will have more on Corning as we get more complete data.

AI image detection is a difficult task and one where there is a clear leader, and it is not human.  Microsoft’s (MSFT) AI Lab for Good did a study as part of an AI image recognition development project that was both telling and not surprising.  What sets this research apart is that it relies on a large sample of human evaluations through an online game. To date, the game has logged more than 82,500 distinct sessions. For the  analysis below however the data comes from those that played between August 1st and August 8th 2024.  This correlates to12,500 participants around the world, all of whom completed the full game, which represent 17,340 unique games, producing roughly 287,000 image evaluations.
The images displayed in the game come from the collection of 350 copy-right free ‘real’ images and the generation of around 700 diffusion-based images using DallE-3, Stable diffusion-3, Stable diffusion XL, Stable diffusion XL inpaintings[1], Amazon Titan v1 and Midjourney v6. Also added were GAN-based fake faces. Each user is shown one image at a time, for 15 rounds. The 15 images in a game are randomly selected from all images, both AI and real. The statistics are derived only from the results of completed games - in which all 15 images were evaluated by a user.
The game was designed to raise awareness about AI-generated images and how challenging it can be to spot them, and Microsoft could have cherry-picked AI images that can fool most  people most of the time, but the goal was to provide a realistic panorama of the AI images that people are likely to be exposed to, an “average” output of generative AI.


[1] Inpainting are images that are added to both real and AI images as items that were not in the original image

Certain types of images are more easily detected by us humans, such as
:
·       People                                                          65%
·       Vehicles                                                       63%
·       Objects                                                        62%
·       Urban                                                           61%
·       Nature                                                          59%
 
But when it comes down to really being able to discern which are real and which are not, we humans are not so good, while AI detectors, such as the one Microsoft is developing as part of this study, are.  Here’s the data:

We hate to admit it but we took the test and ended with a 53% score, quite a bit below average, although the system was kind enough to tell us that our score was still better than 18% of the participants.  (It did little to make us feel much better).  The scary part is that as Generative AI images improve, it will get harder for humans to tell them apart from real images and the necessity for watermarking or other embedded forms of identification becomes greater.  The problem however is not with most rank and file content consumers, it is with those interested in disrupting news and media cycles or influencing consumer sentiment, those who are not going to watermark their devious works.  It’s a hard problem to solve and one that will continue to get more difficult, but there is no free lunch.  Everything in technology has a price.
Here are some of the images used in the survey.  You can convince yourself that you are better than average by looking at them, but it means nothing unless you play the game:
https://arxiv.org/html/2507.18640v1
​

Yesterday we noted points concerning the recently released “Winning the Race – America’s AI Action Plan”.  While the document has its positives and negatives, along with the accompanying three EOs that mandate certain changes, today we compare it to the also recently released Chinese Global AI Governance Action Plan.  Before we go further, we point out the fact that both are political documents, written to further the standing of both entities in the global eye.  They are filled with all sorts of carefully written platitudes that are positioned to aggrandize a particula position.  That said, there are some very obvious differences between the ‘attitude’ in the documents, and that is what we are looking at here.  Much of the rest of the world will likely be more interested in what they can get out of the two largest technology suppliers by playing them off each other.
Core Philosophy

• China – Sees AI generally as a new part of human development and an international ‘public good’ that should benefit humanity.  The emphasis is on safety, reliability, and controllability, and a desire to align with the United Nations Pact for the Future & Global Digital Compact, with a goal of establishing a framework for global AI development.
• US – Framed as a “race to achieve global dominance in AI”, with an emphasis on national security, economic competitiveness, and American innovation.  The direct aim is for ”unquestioned and unchallenged global technological dominance” to usher in a new golden age of human flourishing, economic competitiveness, and national security, with the goal of winning the AI race and securing American global leadership in AI, primarily through accelerating private-sector-led innovation and building robust domestic infrastructure.

Regulation

• China – Looking for an innovation friendly policy balanced within a framework of safety and risk control, balancing progress with social ethics.  Promotes ‘bold experimentation’
• US – Prioritizing removing government red tape and onerous regulation to the extent of limiting federal aid to those states with ‘burdensome AI regulation’.

Open Source

• China – Focus on exploring a global platform for data sharing and creating high-quality datasets, while safeguarding personal privacy and supporting the building of ‘cross-border open source communities’.
• US – Focus on building world-class scientific datasets and providing open source models for start-ups and academic research, along with providing large scale computer power to same, all while creating a ‘healthy financial market for compute’.

Infrastructure

• China – Speed up the construction of global clean power networks and data centers to support access to AI infrastructure for less developed nations
• US – Building a ‘vast AI infrastructure particularly streamlining permitting for data centers while restoring the American semiconductor effort.

Cooperation/Diplomacy

• China – Focus on ‘global solidarity’ and ‘open cooperation’ for AI capacity building, using the UN as the main way to create an inclusive digital global governance system..
• US – Simply “export American AI to Allies & Partners”

First, while it might seem that we are biased against the US AI doctrine, we are not.  The wording in most of these encapsulations are taken directly from the US plan.  Second, in no way do we believe that the Chinese AI Action plan is honest, unbiased, and apolitical.  We expect that since the Trump administration has made no bones about the idea that the US must rule the AI world, pretty much at any cost, it was an easy read for the Chinese government to take the opposite path.  This dovish and broad-minded approach sets China apart from the US on a global basis, making sure to make obvious the distance between the philosophies, sort of a kinder, gentler China, as opposed to a no-holds-barred US approach.
​
The Chinese attitude is as phony as a $3 bill borne out by years of government rules and oppression  so we grudgingly give credit to the current administration for being nothing if not honest about its AI intensions.  But when it comes down to how secondary and tertiary countries will view the intentions of both, we expect China will have the psychological advantage of being slightly more believable than the somewhat volatile current US administration.  While China is certainly in the AI race to win it, they have a level of trust (maybe trust is the wrong word) that Trump does not, despite the fact that both are pretty much untrustworthy.  China can point to the fact that Alibaba’s (BABA) Qwen family  has the world’s largest open source community, beating out Meta’s (FB) Llama community, and has surpassed the US in open source models, despite the initial US lead in 2023.   When it comes to AI and global plans from leading developer countries, its the lesser of two evils.

The Trump administration has released its “Winning the AI Race: America's AI Action Plan”, a design for winning the AI race and becoming the dominant global AI power.  As we noted last week, rather than focusing on developing a rational strategy that includes insuring the health and safety of the US population and the planet, the focus of this strategy, not surprisingly, is to win, win, win and not worry about the consequences (until after we win). 
There are over 90 ‘points’ in the document, although many of these are recommendations and goals rather than legally binding actions.  At the same time, three EOs that are legally binding were also signed, one of which we have already reviewed (Promoting the Export of the American AI Technology Stack).  The other two are, “Accelerating Federal Permitting of Data Center Infrastructure” and “Preventing Woke AI in the Federal Government”.  The former removes any hope of climate change restrictions when it comes to data centers, and the latter is another anti-DEI tirade specific to AI and data centers.  Intermixed are plans for government financing for large data centers, although much is boiler plate and a few small tidbits for the former tree huggers that now want to avoid a global power shortage, climate change, or the enslavement of the human race by AI overlords!.
Rather than walk through all 90+ recommendations in the plan, we looked at it according to 14 government agencies that were singled out in the plan.  However, unless the agency was specifically mentioned as a focus in the three additional EOs, the plan does not mandate any of these actions, it only recommends them.  So for those actions in the plan that are not part of the EOs life can go on as it has thus far without assigned change.  In other words, as it was before the action plan.  That said, looking at each agency against the EOs, anything underlined and in italics becomes mandated, anything not is a suggestion.
When all is said and done, the ‘Action Plan’ is nothing more than a reiteration of the three executive orders that preceded it, as it carries no weight other than what is mandated in the three EOs (underlined here).  Simply put they are:

1. Prevent woke AI in the Federal government
2. Accelerating Federal permitting of data center infrastructure
3. Promoting the export of the American AI Technology stack.

Anything else is padding to support those three ideals or give false credibility to the plan.  Obviously there is no concern for climate or security, in fact just the opposite, given that either might slow down the ‘AI race’ to beat China. 
The only reference to anything outside of direct promotion of the EO mandates and the AI race was the mention of open source as a way to promote the development of Ai applications, an important point that is a bit out of character for the current administration’s policy of capital formation at all costs.  Open source models allow developers, especially smaller developers, to utilize models that would typically be available only to large institutions and enterprise customers, and as we have said in the past, the race is not to build the best model, it is to build the most practical applications. 
The average smartphone user likes the idea that AI is helping to take better pictures but could do without it if necessary.  However if someone came up with an AI based application that could sort through your e-mails, put them in priority order, write responses to those that needed them (with your approval), make notations to your calendar, read all of your texts and note important events and personal nuances, and make sure you were aware that the #2 subway was stalled at 14th street (police investigation) and an alternative might be necessary, and Microsoft (MSFT) did not own it (free is good), then there would be real value in AI to the average consumer.  Other than that its all promotion.
Here are the ‘Plan’ details:.
I. Office of Management and Budget (OMB)

• Deregulatory Actions:
  • Work with all Federal agencies to identify, revise, or repeal regulations, rules, memoranda, administrative orders, guidance documents, policy statements, and interagency agreements that hinder AI innovation and adoption.
  • Work with Federal agencies that have AI-related discretionary funding programs to ensure, consistent with applicable law, that they consider a state's AI regulatory climate when making funding decisions and limit funding if the state's AI regulatory regimes may hinder the effectiveness of the funds.
  • Issue guidance to agencies regarding "unbiased AI principles" for federal procurement of LLMs.

II. Department of Commerce (DOC)

• National Institute of Standards and Technology (NIST):
  • Revise the NIST AI Risk Management Framework to eliminate references to misinformation, Diversity, Equity, and Inclusion (DEI), and climate change.
  • Lead NAIRR (National AI Research Resource) pilot efforts to give startups and researchers better access to compute, data, and software (in coordination with OSTP and NSF).
  • Convene public and private sector stakeholders to accelerate the development and adoption of national standards for AI systems (e.g., in healthcare, energy, agriculture).
  • Publish guidelines and resources for federal agencies to conduct their own evaluations of AI systems.
  • Develop a formal guideline and companion voluntary forensic benchmark for deepfake evaluation (building on Guardians of Forensic Evidence program).
  • Convene the NIST AI Consortium to develop scalable, interoperable measurement techniques and metrics.
  • Support theoretical, computational, and experimental research to foster new AI breakthroughs.
• National Telecommunications and Information Administration (NTIA):
  • Convene stakeholders to drive adoption of open-source and open-weight models by small and medium-sized businesses.
• Economic Development Administration (EDA) / Other DOC bureaus:
  • Launch an initiative to provide financial support (loans, grants, tax incentives, off-take agreements) for qualifying data center, semiconductor manufacturing, and energy infrastructure projects.
  • Continue CHIPS program funding and simplify administrative processes.
  • Partner with industry to deliver secure, full-stack AI export packages to allied nations.
  • Establish an American AI Exports Program.
  • Implement new export controls on semiconductor manufacturing subsystems and expand US oversight on foreign-made items that rely on US technology.

III. Department of Labor (DOL)

• Workforce Development:
  • Prioritize AI skill development as a core objective of education and workforce funding streams.
  • Establish an "AI Workforce Research Hub" to evaluate AI's impact on the labor market.
  • Create a national initiative identifying high-priority occupations critical to AI infrastructure (e.g., electricians, HVAC technicians) and partner with state and local governments to support industry-driven training programs.
  • Expand Registered Apprenticeships for AI-critical occupations and update career and technical education programs.
  • Create early career exposure programs and pre-apprenticeships for middle and high school students.

IV. Department of Defense (DOD)

• AI Adoption and Research:
  • Transform senior military colleges into specialized hubs of AI R&D.
  • Integrate AI-specific curriculum across various academic majors to foster military AI expertise.
  • Refine its Responsible AI and Generative AI Frameworks, Roadmaps, and Toolkits.
  • Prioritize and invest in the construction of secure data centers for national security applications.
  • Prioritize agreements with cloud service providers and computing infrastructure operators for priority access to computing resources in a national emergency.
• Defense Advanced Research Projects Agency (DARPA):
  • Launch a technology development program to improve AI interpretability and control.

V. Department of Energy (DOE)

• Energy Infrastructure & Scientific Computing:
  • Expand the number of agencies participating in the Department of Energy's PermitAI project to expedite permitting for data centers.
  • Support automated, cloud-enabled labs across various scientific fields in collaboration with National Laboratories.
  • Promote dispatchable, next-generation energy sources (enhanced geothermal, nuclear fission, fusion).

VI. Department of Homeland Security (DHS) / Cybersecurity and Infrastructure Security Agency (CISA)

• Cybersecurity & Incident Response:
  • Establish an AI-ISAC (Information Sharing and Analysis Center) to circulate AI security threat intelligence.
  • Develop guidance for the private sector regarding remediating and responding to AI-specific vulnerabilities and threats.
  • Promote the development and incorporation of AI Incident Response actions into existing incident response doctrine and best practices.
  • (With NIST) Include AI in their standards and guidance related to incident response.

VII. Office of Science and Technology Policy (OSTP)

• Policy Coordination & Deregulation:
  • Lead a Request for Information from businesses and the public about current Federal regulations that hinder AI innovation and adoption.
  • Coordinate with NIST and NSF on NAIRR pilot efforts.
  • Participate in CAISI intelligence collection (with NSC, IC, DOE).

VIII. Department of State

• International Diplomacy & Exports:
  • Partner with industry through the Commerce and State Departments to deliver secure, full-stack AI export packages to allied nations.
  • Actively counter China's growing presence in international AI governance bodies.
  • Strengthen global alliances to maintain trust and promote responsible AI practices.

IX. Department of the Treasury

• Tax Incentives:
  • Issue guidance clarifying that many AI literacy and skill programs may qualify as eligible education assistance under the Internal Revenue Code, allowing employers to offer tax-free reimbursement.

X. Department of Justice (DOJ)

• Legal System & Synthetic Media:
  • Consider developing a formal guideline and companion voluntary forensic benchmark for deepfake evaluation to equip the legal system.
  • Combat industrial espionage and theft of AI technology.

XI. National Science Foundation (NSF)

• Research & NAIRR:
  • Participate in NAIRR pilot efforts (with NIST, OSTP).
  • Support theoretical, computational, and experimental research.

XII. White House (Cross-Agency Coordination/Directives)

• Chief Artificial Intelligence Officer Council (CAIOC): Formalize as the primary venue for interagency coordination.
• Cross-Agency Directives:
  • Direct federal agencies to identify and eliminate rules that may slow the deployment of AI technologies.
  • Direct federal agencies to assess and expand their use of AI to improve efficiency and effectiveness.
  • Direct agencies to build a pipeline of AI-skilled federal employees.
  • Direct federal agencies to identify and make available appropriate federal lands for data center construction.

XIII. Center for AI Safety and Innovation (CAISI) (within DOC/NIST)

• AI Safety Evaluation:
  • Prioritize, collect, and distribute intelligence on foreign frontier AI projects with national security implications (in collaboration with DOE, NSC, IC, and OSTP).
  • Evaluate frontier models from China for alignment with Chinese Communist Party (CCP) censorship and talking points.
  • Hold biannual CAISI-led meetings for federal agencies and researchers to share AI evaluation best practices.

XIV. Other Specific Mentions:

• Federal Communications Commission (FCC): (Potentially) Review state AI regulations that interfere with its mandate.
• Federal Trade Commission (FTC): Review and reassess ongoing AI-related investigations and final orders that burden AI innovation.
• Securities and Exchange Commission (SEC) and Food and Drug Administration (FDA): Encouraged to use regulatory sandboxes.
• National Science and Technology Council's Machine Learning and AI Subcommittee: Directed to develop minimum data quality standards for scientific disciplines.

Struggling panel producer Japan Display (6740.JP) has decided to pull in the closing of its Mobara fab that was scheduled for March 2026.  As we have noted previously, the Mobara fab is being sold to Japan Display’s largest shareholder Ichigo Asset (2337.JP) as part of a refinancing deal (https://scmr-llc.com/blog/another-try).  The original plan was to transfer some of the fab’s equipment to the company’s remaining fab in Ishikawa and to sell some to China’s HKC (248.HK).  It seems those plans have changed and JDI will no longer be transferring the Mobara equipment to Ishikawa, a result of its decision to exit the smart watch panel production business, including its supply to Apple (AAPL).  The company will build inventory before the closure, which we assume will be at the end of this year, in order to fulfill production obligations.  Japan Display is still looking for partners to re-enter the LCD panel business as a fabless supplier and to promote its eLEAP photolithographic OLED process.
More on Japan Display
Taking the eLEAP? - https://scmr-llc.com/blog/taking-the-eleap
Another Try - https://scmr-llc.com/blog/another-try
Not According to Plan - https://scmr-llc.com/blog/not-according-to-plan
Sold - https://scmr-llc.com/blog/sold5732273
​

Samsung Electronics (005930.KS) announced that it had received a semiconductor contract with an undisclosed “large global company” worth 22.764 trillion won or ~$16.544 billion US.  The value of the contract is equal to 7.6%  of Samsung’s sales in 2024  and will run from 7/24/25 to 12/31/33.   While details were not given in the Samsung release, speculation is that Tesla (TSLA) is the customer and Elon Musk confirmed that Samsung is currently producing the A14 chip (7nm) and will be producing the A16 (2nm), although TSMC (TSM) will be producing the A15 (3nm), which has just completed the design phase in Taiwan and Arizona and is slated for production late next year.
The A16 is expected to be produced at Samsung’s Taylor, Texas fab, which is scheduled to begin production in 2H next year.  The Tesla A16 evaluation is loosely scheduled for 2H 2027.  Interestingly, Musk indicated that one factor in the contract decision was the fact that Samsung agreed to allow Tesla to assist in “maximizing manufacturing efficiency” at the fab.  Musk indicated, “I will walk the line personally to accelerate the pace of progress as the fab is conveniently located not far from my house.” (~40mi from what we can see).   While the contact amount was specified in the formal release, Musk indicated that it was a ‘minimum’ and gives Samsung’s new fab a substantial revenue base.

Yesterday the President signed another Executive Order titled “PROMOTING THE EXPORT OF THE AMERICAN AI TECHNOLOGY STACK”, different from the “PREVENTING WOKE AI IN THE FEDERAL GOVERNMENT” that the media seems to have focused on.  The “woke” order is socio-political while the “promotion” order is all business, which is where we focus, letting the media try to sort out the implications of the “woke” order.
The “promotion” order begins the process of setting up a framework where the federal government will take part in promoting and exporting US AI technology, standards, and governance through the creation of an American AI Export Program, administered and financed through the Department of Commerce.  To establish the program, industry-led groups will submit proposals that detail AI ‘packages’ that can be marketed to other countries.  The packages should include:

• AI-Optimized Hardware
• Data Systems
• AI Models
• Security Protocols
• Specific AI Applications

The package proposals will be evaluated by a ‘number of government agencies’ and ‘priority’ packages will be selected and supported through various federal financing tools, such as:

• Direct loans
• Loan Guarantees
• Equity Investments
• Co-financing
• Political Risk Insurance
• Credit

The overall program will be implemented by the Secretary of Commerce and will encompass the creation of EDAG (Economic Diplomacy Action Group) that will be chaired by the Secretary of State and be tasked with developing a unified strategy, aligning federal resources, coordinating international partnerships, and addressing market access issues.
Unfortunately the EO does not contain any details such as export targets, financing budgets, and any real specifics about how the program will operate, leaving those details up to the offices of the Secretary of Commerce and the Secretary of State.  No mention of where actual funding will come from, although we believe at least the operational side will come from the Commerce budget (unconfirmed). 
On an overall basis, the idea of promoting US AI technology is good, but we wonder why the federal government needs to be involved given the vast sums available to the AI community from the investment community (public & private).  We also wonder how and by who the proposed ‘packages’ will be evaluated.  Additionally the lack of clarity as to who or what companies or groups will be putting together the proposed packages and the details on how final decisions will be made leaves us with the feeling that the project is open to side deals, PAC contributions, and ‘retirement funding’ for those in charge. 
No doubt there will be something in this program for Musk-owned xAI (pvt) and the others who have made the required ring-kiss trip to Mar-a-Lago,  along with those who have already paid homage, as we expect there will be little public oversite on this program.  We are a bit negative here as the program seems to have little focus on public transparency and AI safeguards, and more on generating financial gain for those who are able to garner favor with the powers that be, but maybe its just years of watching technology programs slip into the Black Hole of federal government, never to be heard from again unless Ronan Farrow finds a string to pull and there is something on the other end that has a fishy smaell.

Pew Research is a well-known data provider that has been delivering data through its surveys and panels for 35 years.  As a highly reputable data source we pay close attention to their work when it relates to CE and related tech topics.  In March of this year they did a survey of the browsing habits of 900 respondents to gain insight into their interest and engagement that consumers have with AI and how AI influences their internet browsing.  As we have a keen interest in understanding the impact of AI on consumers and the CE space, when Pew speaks we take interest. 
The survey tracked (with their permission) all of the browsing that participants performed during the month, which came to ~2.5m website visits and ~1.1m unique URLs, for a daily average of 89.6 site visits/day and 53.8 unique URLs/day per person over the period.
At its simplest, during March, 58% of participants viewed a search page where an AI generated summary appeared and 65% of participants visited a search page that referenced AI somewhere on the page, while 13% visited the site of a Generative AI tool (such as https://bard.google.com/ or https://chatgpt.com/) , while only 10% searched for a site about AI related content itself.  This shows that while AI is certainly a media topic, only a relative small percentage of respondents had an interest in getting information about the topic or using Gen AI tools directly.  While this was surprising to us as regular AI tool users, we understand how much hype the media can create about a topic and how little the average global citizen cares.
Survey Basics

• Survey Participants:                900
• Survey Period:                           3/1/25 – 3/31/25
• Webpage Visits:                        ~2.5m
• Unique URLs:                            ~1.1m

In the results, it was found that:

• 93% of participants visited sites where AI was mentioned, yet, only
• 49% of participants visited sites where  AI was the main focus

This was refined further as:

• 52% of participants visited news websites, but only
• 8% of those participants visited news sites where AI was the main focus.
• 54% of participants visited shopping websites but only
• 2% of participants visited shopping websites where AI was the main focus.

For reference, we show the percentage of times that various AI topics, companies, and generic terms were found in search results, to point out more detail as to what the public is interested in regarding AI:

However, the most interesting data in the survey was not about how interested participants were in AI and related topics but how the AI summaries affected their browsing, a topic we have been focused on, as it has the potential to disrupt the highly concentrated search market.  The survey data separated results between participant search pages that had an AI summary and those that did not.  [Note that these are all Google (GOOG) searches.] 
At the onset, the most telling point was that Google search users are more likely to end their browsing session after visiting a search page with an Ai summary than a page without a summary.  According to the survey, this happened on 26% of the pages with an AI summary, compared to 10% of the pages without an AI summary, essentially traditional search pages.  This is not what those who advertise on popular websites want to hear but reflects an actual lesser need for detailed information than content producers might imagine.  In a typical web search, without an AI summary, the user will click on at least one link to clarify the search results, but as the summary contains bits of information and data from a number of sites, much of what the user was looking for is made available without a next-level visit.  While the summary is the essence of the search, it obviates the need for more information in many cases.
The most frequently cited sources in both typical Google searches and Google AI summaries are Wikipedia (non-profit), You Tube (Google), and Reddit (RDDT).  They accounted for 15% of the sources listed in AI summaries and 17% of sources in typical searches.  The data shows that when examining the first three links of Google summaries and the first page of traditional Google searches, Wikipedia, xxx.gov, and reddit came up more often in summaries than they did in traditional searches.  You Tube was more common in traditional searches than in summaries.

Overall 18% of Google searches produced an AI summary and 88% of those summaries cited three or more sources (Only 1% cited a single source).  In passing, the shortest summary produced was only seven words long, while the longest was 369 words, and while that sounds like a less than useful metric, it turns out that longer searches (words) are more likely to produce summaries.  Only 8% of one or two word searches generated a summary, while searches with 10 or more words generated a summary 53% of the time.  More detail revealed that searches that contain a noun and a verb generate an AI summary 36% of the time while those searches that start with a question word generate a summary 60% of the time.

The data is very helpful in gaining a more realistic understanding of the consumer’s view of AI and a more nuanced look at how the generation of AI summaries is affecting the monetization of search results. As the premier search engine developer, Google has the most to lose, although in the same regard it has the most ‘staying power’, especially in comparison to smaller search engines.  But while Google might not feel that what is a small loss of click through’s is meaningful publicly, advertisers are likely more sensitive, since their advertising rates are based on metrics like CTR and conversions, in reality Google is definitely concerned.
Google, at least on a public basis, has indicated a focus on ‘higher quality clicks’, and higher user satisfaction, but on a long-term basis they have to be concerned over any potential for slower growth.  Google has one point in its favor and that is that the algorithm that creates the AI search summaries is Google’s own, allowing them to make any modifications they want and that leaves the summaries open to advertising (likely a premium category) and highly visual summaries for commercial queries.  It is a hard to balance between improving user click quality and not destroying per click income but also a situation that was inevitable.  If Google did not create the summary algorithm and incorporate it in its search results a competitor would have.  Now they have the most control over how to find a solution 
That said, what the survey shows is that consumers are not nearly as enamored with AI as the tech press would like us to believe, essentially willing to use anything that causes them to have to do less work.  If they can get their search answer from a summary and don’t have to wait the 2.5 to 3.2 seconds it takes for the average page to load, they will, especially after  few years of becoming used to summaries.  If Google can find a comfortable way to monetize summaries they should be able to compensate or even exceed any standard advertising loss, but in the interim it could be a bumpy ride.

​