The world was less than amazed when the first text-to-image model was released in 2015.  It (alignDRAW) was developed by a team at the University of Toronto, and while it gained recognition in technical circles, the general public was largely unaware it existed.  That changed in January of 2021 when OpenAI (pvt) introduced DALL-E1, the first TTI model to catch the eye of the general public.  DALL-E was able to understand prompt context and create images that were both creative and unusual, beginning the public’s odyssey into the world of AI image creation.  Since then there have been a multitude of such models, some of which are in the public domain, although most are primarily available to subscription users or for a per image fee.
As we have noted previously, the world of text-to-image generators has expanded rapidly and has moved from a niche application to social media star (https://scmr-llc.com/blog/tiny-tostones) status, with new TTI models are being released almost monthly.  We count over 20 public models this year to date, with Microsoft (MSFT) the most recent, releasing Mai Image 1, the company’s first fully in-house developed image model and the company’s 3rd overall.   Other large companies have jumped on the bandwagon quickly, leaving smaller model developers and labs by the wayside unless they are able to find a niche where they excel over more general TTI models.
We have put together a list of many of the current public text-to-image models (not all) that are available to consumers, most for a fee (see the “open Source” column for those that are free).  There are some that are intentionally missing, such as OpenAI’s Sora and Sora 2, as Sora is only available to ChatGPT Plus & Pro subscribers in the US and Canada, and Sora 2 is an invite only model.  We note also that all models listed have a release date in 2025, other than Dalle-E3 as the TTI space is technically in toddler status relative to other CE products.  We also show a ‘vote’ column, a capture from an image generator rating site that allows users to vote for the generator that they like best.  Given the short lives of some of these models it is a bit difficult to reconcile popularity and time from release, but there are some that are obviously more popular than others, and not all are from major AI companies. 

Users of these models can vote for their favorite, and while this is a biased and unscientific survey, it is more of an indication of how useful these models are than many of the benchmarks that are commonly used for image generators.  While Google is the obvious dominant force, much smaller Black Forest Labs (the team that developed Stable Diffusion) beat out all others, including OpenAI in this very practical user survey.

Again, while we see these rankings of equal value to benchmarks, when it comes to what users expect when using text-to-image models, there are very specific features/areas that are key to a TTI model’s success as shown below:

• Adherence – A measure of how accurately the image reflects the details given in the prompts, essentially how well the model ‘listens’ to the prompts.
• Realism – The overall visual believability and appeal that the image has, whether it was specified in the prompt or not.
• Typography – The ability of the model to generate clear, legible, and correctly spelled text within an image.
• Control – The tools given to the user to make adjustments, fine tune, or edit an image
• Ease of Use – How easy it is to operate the model and make changes
• Cost – While model builders talk about open source, they rarely offer same.  We note there are only a few models that are not fee based.  Stable Diffusion has been open source since its initial release in 2022.

Image generator model pricing is far from standardized and there are as many pricing configurations as there are image models.  Other than open source models, there are those that require only a monthly subscription, usually to a premium tier, while others charge on a per image basis, although most of the per image pricing is based on the size of the image (pixel count) or the resolution.  Some charge based on the number of tokens in the prompt, some charge for in-painting, essentially editing the image itself, and there is always the ability for some models to be self-hosted by an enterprise, which requires a full enterprise license.
There is little said about the profitability of text-to-image models and there is a train of thought that says they are a necessary part of a full model stack, even if they generate a loss, but along with generative AI they also require a massive amount of compute power during training.  The good news for developers is TTI models typically have fewer parameters than generative models because they are so specialized, and Image models don’t process every pixel of an image by working in ‘latent space’, a compressed representation of the image, reducing the amount of computing power necessary to train the model.  Generative AI models need to process every token sequentially to understand its relationships, pushing the amount of processing power and costs to near theoretical limits.   Image models are therefore less expensive to train than generative models, however at the inference level, when the model is interacting with users, the compute power necessary to understand the prompt and then create an image is greater than that of generative  AI, where the result is just a relatively simple text response.  Net, the TTI model is less expensive to train but more expensive to run than a generative AI model.
In the relatively short time we have been using TTI models we have already seen a big improvement in performance, particularly recently with the release of Google’s Nano-banana model, which allows users to make modifications on an image without changing the basic structure of the image, something many models struggle to do.  Nothing is more frustrating to a user than working on a complex prompt, getting the correct image, only to find a spelling error or physical glitch.  When trying to correct the issue, the AI changes the image and rarely can find its way back to the correct one.  Nano-banana has been able to maintain a stable image while making most corrections.  This might seem a basic function for an image model, but until recently it was a dream rather than a reality.
As is obvious, the TTI space is evolving very rapidly and will continue to advance, perhaps faster than generative AI, and while those that decry Ai image generation as a job killer or an invasion of personal privacy are missing the concept that almost anyone, even those whose drawing ability is so poor that a straight line is a challenge, can now let their creative juices flow through a text-to-image model.  The better you can describe it, the better the image will be…
Here's an example:
Can you create an image of a gray striped tabby cat? 
​

Can you create an image of a gray stripped tabby cat with blue eyes and silver ear tufts, sitting on a wooden crate in front of a red barn?
​

Can you add two gray and white kittens sitting next to the cat, both with blue eyes and silver ear tufts, but make one with a bit of orange in the fur color. Can you make it a bit sunnier and make the barn a more intense red? 
​

Can you keep everything the same and put a bush next to the crate?
​

Can you keep everything the same but put a washed out "Borden's Evaporated Milk" label on the front of the crate that the cats are sitting on?
​

​In 2024 the US District Court for the District of Columbia ruled that Google (GOOG)  had violated Section 2 of the Sherman Act (prohibits any single entity from monopolizing, attempting to monopolize, or conspiring to monopolize any part of trade or commerce in the United States, making such actions a felony punishable by fines and imprisonment), based on a decade of distribution agreements with browser developers, original equipment manufacturers, and wireless carriers to be the out-of-the box, default general search engine.  According to the court, Google paid billions to lock up these channels which prevented rivals from competing and discouraged entry into the search market.  Google used these agreements to build scale and generate what the court (United States v. Google LLC – 2023) called ‘monopoly profits’, essentially freezing the search market.
That said, since then, the courts acknowledge that Generative AI is changing the search market, providing information through chatbots, rather than search, which has put Gen AI in a prominent place at the ‘remedies’ hearing that determines what Google will face as a result of its monopolistic behavior.  The plaintiffs want to make sure that Google’s dominance in search does not carry over into Gen AI and have proposed a number of remedies, while the court is interested in finding an equitable solution.  Google has proposed the idea of an injunction that prohibits it from entering into exclusive distribution agreements, including its GenAI products, the Gemini app and Google Assistant.  Google says it has already stripped from its existing contracts the provisions that the court found anti-competitive and says Gen AI is already highly competitive, and the company should not be restricted in areas where it has not been proven to be anti-competitive.
The plaintiffs want more, stating that they are trying to revive the search market and to protect the Gen AI market from a similar fate, however the court has decided to walk a more moderate path.  Here are the details:
Positives

• Google will not be required to divest Chrome; nor will the court include a contingent divestiture of the Android operating system in the final judgment.  This is a big win as Google, as a company, remains intact, despite the requests from the DOJ and a recent ‘helping out’ offer from PerplexityAI (pvt).
• Google will not be barred from making payments or offering other consideration to distribution partners for preloading or placement of Google Search, Chrome, or its GenAI products as the court found that cutting off payments from Google will impose substantial, in some cases, crippling downstream harm to distribution partners, related markets, and consumers.  Another big win as ending the practice of payments for placement would materially change the mechanics of the search business.
• Google will not have to present users with choice screens on its products or encourage its Android distribution partners to do the same.  This is a win for Google as both the plaintiffs and the DOJ asked for ‘choice screens’ to be part of the ‘remediation’.   When a device (phone, PC, etc.) or application is set-up for the first time the user is given a choice of what browser and search engine they want.   Most users pick the default, which is why plaintiffs wanted to force a choice screen.  Now if Google is already the default, no other choices need to me shown.
• Google will not be required to share granular, query-level data with advertisers or provide them with more access to such data. Nor will it have to restore an “exact match” keyword bidding option.  Query-level data are the search terms that a user typed in before clicking on an ad and Google does not make that detail available in order to keep advertisers using Google’s ad ecosystem.  The keywords in that data give Google more control over ad placement and give advertisers less transparency, forcing them to bid on a wider variety of search terms.  A loss here would have been a serious blow to Google’s ad bidding system, making the Google system more transparent and therefore less competitive for advertisers.
• Google will not have to underwrite a nationwide public education campaign. The court noted that such a remedy does not fit Google’s anti-trust violations and its terms are too indefinite.
• Google will not have to modify its policies to offer website publishers more choice in how Google uses their content.  Another win for Google, with the court deciding not to compel Google to give publishers alternatives as to how their content is used.  Publishers have been concerned that AI search summaries reduce the click-through rate to their content, so there was contention over whether Google should be forced to allow content providers a way to opt out of Ai summaries while still being included in Google’s search index.  Again the court felt this did not relate to Google’s anti-competitive practices.
• Google will not be subject to an investment reporting requirement. This too bears no relationship to Google’s anticompetitive conduct.  Another win for Google as the DOJ had asked the court to force Google to report any transactions or investments, particularly in AI, citing the fact that Google’s resources were so large that it could acquire smaller competitors or proprietary technology before its rivals.  The court did not agree.
• Google will not be subject to anti-retaliation, anti-circumvention, or self-preferencing provisions.  The DOJ requested that Google should be prevented from retaliating against partners that choose to work with Google competitors,  Anti-circumvention, also a DOJ request, would have restricted Google from creating a new deal form, sort of a  ‘quasi-exclusive’ deal that would not violate the restrictions but still allow Google to maintain an exclusivity with its partners.  Self-preferencing refers to situations where a vertically integrated company, operating at a number of different levels of the supply chain, uses its influence in one area to force preferences in another.  The example might be Google’s search algorithm, which puts Google products at the top of the search list, over those of competitors, regardless of the actual rankings.  The court did not consider any of these practices as necessarily prohibitive to the competition under the ruling.

 
Negatives

• Google is barred from entering or maintaining any exclusive contracts relating to the distribution of Google Search, Chrome, Google Assistant, and the Gemini app or enter or maintain any agreement that conditions the licensing of the Play Store or any other Google application on the distribution, preloading, or placement of Google Search, Chrome, Google Assistant, or the Gemini app.  Probably the best Google could have hoped for in light of the anti-competitive ruling.  It eliminates all exclusive placement contracts.
• Google will have to make available to Qualified Competitors certain search index and user-interaction data, though not ads data.  The court narrowed the datasets Google will be required to share to tailor the remedy to its specific anticompetitive conduct, but this means that competitors will now have access to Google’s search index and, in theory, no longer have to scrape the internet to create their own indexes.  Creating an index from scratch is an expensive process and as Google’s index is, by far, the largest (2020 data says ~400 billion webpages, while Microsoft’s (MSFT) Bing is ~14 billion pages) this is a major plus for the competition and a blow to Google’s ability to compete.
• Google has been barred from receiving revenue share payments for the placement of a Google application (e.g., Search, Chrome, Google Assistant, or the Gemini app) on the placement of another such application, or conditions the receipt of revenue share payments on maintaining Google Search, Chrome, Google Assistant, or the Gemini app on any device, browser, or search access point for more than one year, or prohibits any partner from simultaneously distributing any other GSE, browser, or GenAI product.  This allows a multi-year non-exclusive agreement where Google pays for placement but does not allow Google to require multi-year guaranteed payments.  This keeps Google from establishing multi-year non-exclusive contracts that act as exclusive contracts because the partner cannot afford to lose the revenue by switching  to another competitor after a year.
• Google must offer Qualified Competitors search and search text ads syndication services to enable those firms to deliver high-quality search results and ads to compete with Google while they develop their own search technologies and capacity.  This has practical implications… New competitors, who have no index or monetization can send queries to Google search.  The search results from Google appear as if they were the competitor’s results, along with Google ads.  Should a user click on an ad, the new competitor shares in the revenue stream with Google.  Sort of ‘renting’ Google, without the Google logo attached.
• Google will be compelled to publicly disclose material changes it makes to its ad auctions to promote greater transparency in search text ads pricing and to prevent Google from increasing prices by secretly fine-tuning its ad auctions.  This means Google must disclose changes to its ad pricing, rather than secretly increasing or artificially inflating ad prices.

The term of the judgement is six years.  Google asked for three.  Plaintiffs asked for ten years.  While the results of the actual ruling itself were certainly negative for Google, we see the remedies as a win for the company when compared to the potential for some of the more draconian measures that the court could have implemented.  In a sense AI was Google’s savior in that the courts saw AI as a bit of a more level playing field and allowed Google to keep both its Chrome search engine and its Android OS.  Exclusivity was a key point in the original trial so prohibiting exclusive contracts was assumed, although the ding on the data side, where Google must share its search engine and some data with competitors will sting after many billions were spent on developing Google’s massive search index.
The parties involved, the DOJ and a number of State Attorney Generals, have the ability to appeal the court’s rulings 60 days after the final judgement on the remedies is filed.  The appeal would go to the Court of Appeals for the DC Circuit, but the plaintiffs cannot appeal because they are unhappy with the outcome, as the argument must be that the trial judge made an ‘error of law’.  For example they could argue that the judge abused his discretion by creating remedies that were not sufficient to remove Google’s anti-competitive behavior, or they could argue that by allowing Google to keep its auction practices the remedies did not strip Google of the benefits it gained from its illegal activities.  Should an appeal be filed (likely), if either side is then unhappy with the Appellate Court ruling they can petition the US Supreme Court, although it has no obligation to hear the case.  All in, while Google came out ahead in this round, losing the battle (initial decision) but so fr, winning the war (remidies).  

Estimates for new markets in the CE space tend to be optimistic.  Those that publish estimates hope that an optimistic view will help them gain clients who are associated with the success of the category, and, of course, estimates can always be changed once a client is a subscriber, especially as clients want to know the detailed reason for the lower estimate.  We don’t believe intentionally over optimistic estimates are the modus operandi for most, but as we have been involved with the CE space and the estimates that come with it for over 30 years, we have a sense for who is showboating and who is serious about their estimates.  Of course, accuracy is the ultimate goal when it comes to estimates and forecasts, but anyone in the CE space, especially in reference to new product categories, understands that estimates are only good on the day they are made, and this year’s tariff issues and how they have colored estimates is just one example of estimate volatility.
Foldables displays are a relatively new category with a commercial debut in late 2018 with the Royole (bankrupt) FlexPai, followed by Samsung’s (005930.KS) Galaxy Fold 10 months later.  Since then there have been over 85 foldable smartphone models released, with over 13 brands releasing at least one foldable smartphone.  Given the fact that only a few models were released in the early years of foldables, unit volume expectations were muted, especially given the crease issues that plagued early models.  However by 2022 longer-term estimates began to surface, which we examine to give some clarity to the most recent foldable estimates and forecasts.  

This is not an exercise to point out how far off estimates were, quite the opposite as we were surprised at how accurate the estimates were up to 2024, but more to point to out year forecasts, using the 2022 forecast results as a reality check.  The data from 2022 is from our composite system that collects estimates from multiple sources and aggregates and averages the results for each period.  In this way we reduce the effect of outliers that might skew the thinking of those examining only one or two sources.  The data below compares estimates and forecasts from June of 2022 to current estimates and forecasts for foldable devices.  We note that the current estimates and forecasts can include foldable devices other than smartphones, while we expect the earlier forecasts did not, although most current estimates do not break out device types.  We also note that Apple (AAPL) is expected to release a foldable iPhone in 2026 or 2027, which also skews the forecasts depending on the absolute release data, most certainly the price, and the overall success of Apple’s foldable marketing program.

As can be seen from the above table, forecasts for 2022 and 2023 were quite accurate, while the outyears in the 2022 data were not, as the growth of foldable smartphones has slowed, with most calling for a flat or down year this year, which would put foldable shipments this year at ~36% of the 2022 forecast.  As noted, growth in 2026 and 2027 is predicated, in part, on Apple’s entry into the market (assumedly in 2027 based on estimates).  Under the assumption that 2027 is Apple’s entry year, if we average 2026 and 2028, the normalized volume in 2027 (ex-Apple) would be ~44.6m units, with the actual estimate ~22.5 million above that mean. 
We have to assume much of that outperformance comes from Apple.  This overage, if it were all Apple, would represent ~9.6% of Apple’s yearly iPhone shipments.  In 2024, Samsung’s foldable smartphone shipments represented between 2% and 3% of its smartphone shipments and given that the product is not new for Samsung, this seems a reasonable base.  Excluding non-smartphone foldables (laptops) we could see a 3% to 4% foldable share of Samsung’s smartphone volume this year, although our gut says closer to 2% to 3%, which makes the 9.5% assumption for Apple stand out as a bit excessive.
We do expect Apple’s foldable initial year to be a big one, as the Apple fan base has been clamoring for such a product for years.   Samsung’s has a much broader customer base, inclusive of low and middle price tier customers that would not have an interest in a foldable.  By comparison Apple’s base is less eclectic and more willing to pay a premium, but we see the incremental 22.5m (9.5%) Apple units as high, with a more reasonable range between 13m and 14m units, putting the 2027 estimate at 58.1m, strangely the same as the 2029 composite forecast.  That said, much will depend on the timing of Apple’s release and the uptake of other foldable devices, however, while foldable displays overall have improved dramatically over the last few years, they are still quite expensive, which keeps the customer base narrow and limited to top tier consumers. 
Unfortunately for foldable smartphones, much of the attention given to such devices in the early years has been usurped by AI, which has a number of use cases and potential applications.  Foldables do give the user additional workable real estate, but don’t have a particular application under which they provide a proprietary solution, other than perhaps content creation, and thus far such an application has not appeared.  This leads us to see foldables as a product still in development, not from a hardware perspective but from an application perspective.  Perhaps gamers will adopt foldables at some point in the future and catapult the category to stardom, but for now we are still looking for an application to justify such a purchase.
. For those that need the screen real estate, the comparison below, with 4 non-foldable upper tier smartphones with similar features, the Galaxy Z Fold (and similar foldables) are the least expensive in terms of price per square inch of incremental screen space, but the initial price is almost 47% higher than the Galaxy S25 Ultra, the next most expensive comparable non-foldable.  This means foldables, at least at this juncture, need a justifiable use case to make fiscal sense. In the early days of foldables, we expect many purchases were made to show off the fact that the purchaser was a technology leader, but as foldables become more mainstream, the early adopters play out and the hard realities of the CE space and the smartphone market chime in, and with that we have to look at current forecasts with 5+ years duration with a bit of skepticism.  They are certainly possible if all of the right assumptions actually take place at the prescribed time, but they rarely do in the CE space.

Perplexity.AI (pvt), whose valuation currently sits around $18 billion, has made an unsolicited bid for Google’s (GOOG) Chrome browser of $34.5 billion (all cash), under the concept that Google might be forced to divest Chrome under ongoing anti-trust proceedings with the government.  The company claims to have significant financial backing for the bid from a number of investment funds to put the process in motion should it be accepted, but also considers the bid a long shot as Google has not indicated that it is willing to sell Chrome despite a 2024 ruling in the US District Court of the District of Columbia that declared that Google (Alphabet) has maintained an illegal monopoly over online search, keeping others from developing competitive products..
Perplexity has indicated (no formal release yet) that it will maintain the Chrome engine (Chromium) as open-source and has pledged $3b to maintaining and develop the Chromium engine over the next two years also pledging not to change user search engine defaults should the deal go through.  In reality, if Google showed any interest in an offer, Perplexity would face real competition, not at the browser level, but at the bid level, as OpenAI (pvt) would likely counter with a higher bid considering its valuation is currently somewhere between $300 billion (last round) and $500b (potential employee share secondary) should the Perplexity bid spark any interest from Google. 
We expect the bid is more to prove the point that Open AI is not the only rodeo in town and others can call up the necessary capital to become big fish in the AI ecosystem, and to make sure the Judge in the court case knows that there will be bids should it become necessary to pull apart Google search.  It also gets the Perplexity name into the headlines as sort of an $18 billion dollar “Little Engine that Could”.

Over the last few months we have witnessed some radical price decreases in AI model pricing.  OpenAI’s (pvt) GPT 4 Turbo saw a massive price cut when GPT-4 Turbo was replaced by GPT 4.o in May of last year and again after the release of GPT 4o mini.  Anthropic (pvt) also made similar reductions, with Claude 3.5 Sonnet dropping by 25 in January of this year, Claude 3 Haiku dropped in November of last year, and Claude 3.5 Haiku dropped by 20% in December of 2024.  Google’s (GOOG) Gemini 1.5 Flash saw a massive reduction in price  last August and the newer Gemini Flash 8B model was half of the price of Gemini 1.5 Flash before the end of last year

​These price reductions tend to be in response to the real price leaders, Chinese AI models, where price reductions are almost continuous, the most recent, this week, with Baidu (BIDU) moving a number of its Ernie 4.5 models into the open source category, essentially making them free for users.  This follows the Ernie 4.5 and X1 models initially being priced at 50% of the price of the DeepSeek R1 model (already a steeply discounted model) in March of this year and then followed by an 80% price reduction for the Ernie 4.5 and 50% for the X1 turbo models (April), both of which are now free.
Given the rapid changes seen in model pricing we dig deeper into the AI landscape in order to understand how model pricing will match model costs, both initial and ongoing, and what other costs are associated with model infrastructure and TCO, especially in light of the performance metrics that are used to justify model pricing by developers.  As organizations adopt AI, they cannot afford to allow AI to become a ‘Black Box’ without real cost justification and those that incorporate AI into consumer products need to understand its real cost and benefits.  It is a complex and rapidly changing picture where the usual technological advantage over Chinese companies does not exist, which makes for a more level playing field when it comes to model competition, but the commercial aspects of AI are so new that there is no clear company, technology, or regional winner, making it a truly open playing field.
In order to understand how the AI pricing model is evolving, we started by looking at common AI pricing models, with the predominant pricing model for generative AI being token-based pricing. In this pricing format costs are calculated on a unit of text (token) basis, roughly corresponding to the number of characters. For both input (query or data) and output (response) the costs are figured in million token units and allow AI application developers and users to pay only for what they actually use.  However, when it comes to AI services, such as image generation, per-unit or per-request pricing is most common.  The well-known DALL-E image generation model charges $0.02 for a standard resolution image, while others offer a small number of free images and then a per image fee.  Speech-to-text services (transcription) typically charge a per minute rate,  although volume can bring down that cost.
Subscription based pricing is more common for team oriented development and can be a management favorite as it allows for consistent budgeting, while the newest pricing model, pay-by-task, where a monthly fee is charged for projects like fixing a bug or adding a feature, something with measurable results, as a form of incentive oriented pricing.
Regardless of the pricing mechanism used, models have to justify their pricing, and we see a large gap between performance metrics and real-world performance, especially as LLMs are quite different from machine learning applications where answers are quite clear.  LLM responses can generate a number of ‘correct or acceptable’ answers to a common query, making evaluation a more subjective process.  Some core performance metrics, primarily those obtained through human ‘raters’ reflect the values and preferences of those raters whether they are aware of those biases or not .  They act as judges, deciding whether answers are factually correct and how well they relate to the real world.  They have to evaluate coherence and fluency, whether the output is  safe, and how ‘creative’ the model is when answering, all of which are qualitative and subjective evaluations.
There are also quantitative metrics used to develop pricing models, based on algorithms that compare output to a reference text or some other form of reference.  These can be complex formula that measure a wide variety of  model capabilities. Here are samples:

• Perplexity – How well an LLM predicts text when presented with new data.  Higher scores indicate less ‘surprise’ when a model encounters new data.
• Embedding-Based Semantic Similarity (BERTS) – This metric evaluates the semantic similarity between generated text and reference text.  Thought to capture the deeper meaning of text rather than just relying on string matching.
• Generative Quality Metrics (MAUVE) – These metrics evaluate how closely the model output resembles human language distribution.

Then there are the operational metrics like latency, throughput, resource utilization, token usage, error rates, etc., all of which play into the establishment of a per million token rate that the model builders establish when the model is released, but pricing models certainly don’t end there.  The metrics get even more explicit when measuring specific applications, particularly image generation applications, where a slew of more specific metrics like Frechet Inception Distance[1], CLIP Score[2], and Compositional Quality[3] are used to evaluate models and set base pricing.
While we can go deeper into pricing models for individual AI applications or the rationale behind the various pricing models themselves, the real truth to pricing models is simple, “What will the market pay?”  If the metrics were identical for every model, pricing would be simple, but developers are quick to point out that ‘their’ model is a bit better than another and should carry a higher price, usually backed up with carefully chosen or even specifically designed metrics that reflect their optimism.  However given the subjective nature of AI metrics, trying to evaluate pricing from such a perspective is like trying to juggle bowling pins while balancing on a ball that is sitting on a seesaw.  But for once the pricing paradigm is moving in the right direction for the consumer/business and the influence of competition seems to be having a greater impact on AI pricing than any other factor in the AI space.
From the perspective of the developer, price reductions are a gift to consumers brought about by technological improvements and efficiency, while behind the florid prose of those generous developers, they still have one eye on what the competition is doing.  If it seems necessary to discount in order to attract or maintain customers, model developers look as if they are willing to adjust their pricing, especially at a time when model company valuation has little to do with profitability and more to do with momentum.  From the consumer side, businesses are still trying to figure out where AI sits between being a cost center and a profit center, but for now it is almost a prerequisite that almost any business has some sort of AI development project underway, and with prices coming down precipitously, the pressure is off, at least for a while.
However that price deflation also comes at a cost, as while commodity AI prices are declining, specialized AI services are getting more expensive, creating a more complex picture for AI buyers.  Is it cheaper to build than buy?  Is it cheaper to Host or outsource? Do I want to build an AI development team or buy packaged products?   Google recently included Gemini AI into Google Workspace, which led to a price increase for the entire suite, while Gemini Flash 8B, released only 2 months after Gemini Flash 1.5 saw an input price drop from $0.075 to $0.0375 and an output price drop from $0.30 to $0.15.  This means that static pricing plans are likely a thing of the past and will change with every new model, perhaps leading to the new position of ‘VP of AI Procurement’.  But in all seriousness, Ai has moved from being an ‘on the box’ logo to being a necessary component of operational R&D, and a complex path for businesses who have had little experience with both the technology and AI cost structure, and little experience with building AI applications.


[1] Standard metric for assessing image quality that measures statistical distribution of generated images against a set of real images.

[2] This measures the semantic similarity between an image and its corresponding text prompt.

[3] Measures how well the generated image accurately reflects details, objects, spatial relationships, and atributes that were defined in the text prompt.

​Consumers are in a more advantageous position, with LLM-like applications being offered for free or at minimal cost and a highly competitive industry with key players looking to gain share over near-term profitability.  Consumers can take advantage of said competition with embedded AI applications or LLMs at no cost, while businesses who have serious AI aspirations are caught in the middle of a key determination as to build or buy.  One would assume that with deflationary AI model pricing the lean would be to the build side, but with TCO and ROI still a bit on the fuzzy side, a clear path for businesses is less apparent.  The good news is that at least one side of the equation is decreasing.  As to the implications for model builder companies, that’s for another day.

AI chatbots are not new, despite the recent flood of conversational LLMs.  ELIZA, developed by Joseph Weizenbaum at MIT AI lab in 1964 is considered the ‘first’ AI chatbot as it used pattern matching as the basis for its responses, but over the years much of what were known as chatbots were based on pre-written scripts and rule based systems such as the “Ask XYZ” systems that have been available on many websites for years.  The current crop of AI LLM based chatbots are far more sophisticated and based on large, high-performance models.  More recently chatbots features have been extended to included internet search capabilities and early predictions were that search enabled chatbots would spell the end to standard search systems.
While there certainly has been impact from Chatbot search, the numbers reflect only a small incursion on regular internet search visits.  A recent survey[1] indicated that over the period between April 2023 and March 2025, there were 1.863 trillion search engine queries, down 0.51% y/y.  The obvious dominant search engine was Google (GOOG) Search with an 87.6% share, and a1.41% reduction in search volume.  Bing (MSFT) was a big gainer in search volume, up 27.77% during the period, but even as the number two search engine, Bing only has a 3.2% share of the search market.


[1] Onelittleweb.com
 

The chatbot search top 10 is similar in that ChatGPT (OpenAI) is the dominant player, however when one looks at the number of searches generated by chatbots against generic search engines, the search engine traffic is over 33 times greater, so as of today, chatbot generated searches are only 2.96% of what is generated by generic search engines.  However chatbots offering search capabilities are new, as can be seen by the ‘Search Added’ and the ‘Current Period Change’ columns, where newer chatbot show extremely large y/y increases but small shares.  That said, most chatbots that have been around for more than a year have more reasonable search growth rates.

​Based on the current data it would seem that chatbot searches have had little effect on overall generic searches, but it is far to early in the cycle to make a long-term judgement.  In fact we expect at least Gemini, Google’s LLM/chatbot, chatbot related searches will be incremental to generic search traffic.  Likely, although to a lesser degree, Co-Pilot will have the same effect on Bing.  On a longer-term basis, as AI chatbots become more embedded in operating systems, it would seem logical that unless you were looking for an answer to a question that you specifically request includes an internet search, the AI would try to answer the question using its most recent training or update data.  If, and that is a big if, the data corpus of the AI is broad enough, the answer might not require the chatbot to conduct an internet search, and in that way could weigh on internet search growth. 
Much in that scenario depends on how ‘transparent’ the Chatbot is.  If it is always available on a home page and looks and acts like a search bar, users will gravitate toward the chatbot/AI.  If it has to be chosen, is slow to answer, or comes up with skimpy answers, users will remain generic search fans.  But there are other factors that come into play. 
It is understandable from the standpoint of the AI owner that they keep the compute time as low as possible, so the default would likely be a query first run without an internet search.  However as time goes by, we expect most chatbots will default to  include internet search results in query answers, but it is not quite that easy because those searches are not always free.
Currently, some (Perplexity (pvt), Co-Pilot, You.com (pvt), Komo (pvt), Andi (pvt), and Brave Search (pvt)) check the internet on every search, while others either have their own decision mechanism (Gemini, Claude (Anthropic), ChatGPT, GROK (xAI), META (FB)) or users can request an internet search.  But going forward things get more complicated.  Some CE companies use their own AI infrastructure to run embedded chatbots, while others are based on ChatGPT or other existing infrastructure.  Apple (AAPL) runs Apple intelligence on its own servers, it does not ‘pay’ for searches, although the all-in cost of each search is amortized over the infrastructure cost.  Others, who might use ChatGPT  or Gemini as their AI infrastructure providers that support their chatbot, would have to pay for each search, as both ChatGPT and Google charge on a per search basis, so the advantage goes to those who can support their own chatbot/search infrastructure.
All of this comes down to the value that consumers see in ‘free’ chatbots.  If results from trained data are enough for most, chatbots will have a modest impact on search and a modest cost to chatbot owners.  If chatbot results are not sufficient for the average user, the cost to maintain ‘free’ chatbots will rise as search fees climb, unless the chatbot owner can convert users to paid plans. 
Search has been around for a while and chatbots are relatively new so we expect the impact from chatbots will be small over the next year, but as the general public gets more used to using chatbots, the competition between them will increase and we expect search will be an integral part of all chatbots, especially as the average consumer becomes more aware of chatbot data sources. 
We use eight different chatbots, typically querying at least two each time, especially if the query has a need for very current data.  We know which chatbots are search oriented and which are not, and which will cite sources, either trained or search related, but the average person will likely use what is easiest, and that is usually whatever is built in to the OS or specific applications, so it is up to the brand to decide what level of search is sufficient for their users.  As Ai is such a high-level feature now, the competition will continue to increase and that means more ‘free’ features  which is certainly  good for consumers

A new bill has been introduced in the House of Representatives that would require owners of operating systems with over 100m US users (such as IoS or Android) to  allow users to choose their own default applications or app stores, as opposed to having the choice made for them with pre-loaded applications.  It also requires that OS owners allow users to hide or delete pre-installed applications or app stores.  Developers must also be provided with access to the OS interfaces, hardware, and software  at the same level as the company or its 3rd party partners.  Under the bill OS companies cannot require that developers use the company’s in-app payment system as a condition for access to the OS or app store, and cannot require pricing terms that are more favorable than on other comparable sites, while they are also restricted from using non-public information collected from an application on their platform to compete with that application.
Currently a number of states have passed or are discussing similar legislation, but most are focused on age verification (Utah put it into law).  Florida has proposed a law that would force Apple to open access to outside app stores and payment systems and a number of other states have proposed legislation that allows sideloading and prohibits mandating OS owned applications.  However, this bill would set rules at the federal level and would prohibit states from enacting laws that would be counter to actions stated in the federal bill, which is the 2nd of its kind to be introduced to Congress.  Similar bills have been subject to aggressive lobbying, partisan politics, and challenges from companies and 1st Amendment supporters, and we expect this bill will face the same, but as politicians look for causes that are popular with consumers, we expect Apple and others will ultimate lose some of their proprietary ‘rights’ and face a more level competitive playing field.  Its hard to tell if the time is right for such a bill to pass, but bills that have gone before and failed have at least built some awareness, so if consumers get more involved in the conflict, the law will eventually be put into effect.  Whether it would be better as a federal law or a state law, depends on who you ask.  Full text of the bill here
​

The US government has been working toward providing high-speed internet (100Mbs) to all Americans, but while the FCC shows that only 1.5% of those living in urban areas don’t have such access, 22.5% in rural areas are without same and much of that population has 25Mbs or less.  The BEAD program, a $42.45b government funded program that provides subsidy capital to each state to expand high-speed internet access to rural communities is a major driver behind the efforts to broaden high-speed internet coverage, however while the program was signed into law on 11/15/21 and funded by Congress on 6/26/23, the progress has been slow with preparation for construction underway, but with few providers willing to step up before capital is allocated.
Here's the current BEAD status:

• 56 of 56 state and territory applications have been approved by the NTIA
• 47 of 56 entities have completed the state challenge process
• 32 of 56 entities have begun selecting service providers
• 4 of 56 entities have completed service provider selection
• 3 of 56 have released final proposal for public comment

In most cases the construction will be primarily the laying or stringing of fiber, a complex and costly process that entails gaining right-of-way access, local permits and the cost of trenching fiber, which can run, depending on the location, between $15,000 and $20,000/km and can take weeks to install.  In some cases it is just not feasible to bury fiber optic cable, so it must be strung, leading to higher annual maintenance costs, but there are alternatives.
Elon Musk’s Starlink (pvt) system, a group of LEO satellites that provide coverage in certain areas, presents an alternative.  However the service is relatively expensive at $300 - $400 for the hardware (basic) and between ~$80 and $120/ month for residential customers.  While Starlink has brought down the cost of placing satellites in orbit, it is expected that the current network of a few hundred satellites would likely have to be brought up to a much higher number to provide global service, and at a cost of ~$1m each (that’s the reduced cost) to launch  new Starlink satellites and a ~5 year life for existing ones, the cost of such a system would seem to make it an option that is not suitable for rural areas where much of the population is low-income.  Unless funded by the government Starlink would seem to be out or reach for many and  currently funding for such programs is quite difficult as Congress recently let the Affordable Connectivity Program, which supplemented the monthly cost of internet access for low-income families expire.
The is another alternative when the cost of fiber installation is prohibitive or virtually impossible and the customer economics do not make sense for satellite.  The technology is FSOC, aka Free Space Optical Communication, and is used commercially as both a last mile alternative in fiber systems, for disaster recovery when other means of communication are unavailable, and in military tactical situations as it provides a high bandwidth point-to-point link that is inexpensive and easy to deploy.  The technology is relatively simple to understand as it consists of:

• Optical Transceiver – This consists of a laser which carries the signal
• Modulator – Takes the incoming signal and encodes it into the laser beam using intensity or other optical properties.
• Optics – Lenses and mirrors that shape and direct the laser beam.
• PAT System (Pointing, Acquisition, & Tracking) – This system has movable platforms that can steer the laser beam.  Sensors and cameras detect the position of the incoming signal and control hardware and software that calculates position and adjusts the beam.
• Optical Amplifier – Increases the signal strength in longer range FSOC systems and compensates for atmospheric changes.

Depending on the installation, commercial FSOC systems have a maximum range of 20km. (~12 miles) on a direct line-of-sight basis, but has lower latency than RF (satellite) and very low installation cost (~$1,000/km).  Here’s how it stacks up against fiber and satellite:

As with all communication that travels through the atmosphere, different conditions can cause issues with key variables in optical systems  FSOC systems that are for short distances (campus, etc.) typically operate at 850nm, while longer distance systems, where the effects of weather are more pronounced, operate at 1310nm and 1550nm,which are less affected by rain or snow and are compatible with fiber operating wavelengths.  When the weather is poor (droplets reflect and absorb laser light) FSOC systems can adapt in a number of ways, the primary being by increasing power, similar to RF systems (satellite).  However RF systems are power limited so as not to overpower adjacent bands, while FSOC is limited only by laser (eye) safety requirements.
Other techniques, such as spatial diversity (similar to MIMO), essentially creating multiple optical paths (multiple lasers) or multiple frequency paths for data, adding repeaters (shortening the signal path), specialized modulation techniques that vary the data rate depending on weather, and predictive atmospheric modeling for proactive adjustments, all can contribute to offsetting the atmospheric issues that occur with atmospheric optical systems.  We note also that maintaining precise alignment with the sensors at the receiver end is a key function, as small variations can interrupt transmission, however FSOC systems have progressed substantially, allowing them to self-correct and maintain a solid line-of-sight connection in almost all circumstances, and new chip level systems reduce the lens and mirror count considerably, making the systems even more stable.
Interestingly, what brought FSOC technology back to our attention was the fact that Alphabet (GOOGL) recently spun out Taara (pvt), an FSOC company from its “Moonshot Incubator” (Where ‘Google Brain’ lab and Waymo (pvt) came from) retaining a minority stake along with initial VC funding, with the idea that the company can now look at other financing options.  We originally heard about Taara back years ago (2018) when Google (GOOG) mentioned a project (LOON) that was attempting to create a system of high-altitude (stratosphere) balloons that could deliver internet service to remote locations across the globe, an idea we thought foolish.  While the laser technology (Free-space optics) has been used to communicate with the International Spece Station the idea seemed a bit on the ridiculous side, and eventually the Loon project was cancelled due to regulatory issues about flying large ballons in public airspace.
That said, the laser technology continued to be developed and refined and the company has announced (2026 release) the development of a silicon photonics chip that will replace the mirrors, sensors, and hardware used to ‘steer’ the laser light with software, reducing the overall size of the Lightbridge (the company’s current product), which is roughly the size of a traffic light, to a single chip containing the emitters and all of the directional correcting mechanisms.  The new chip is the size of a fingernail and is expected to be incorporated in a new (much smaller) device.  This will allow the company to lower the hardware and installation price and create modules that are able to deployed indoors, such as on a factory floor (ceiling).

Taara has deployed its systems with a number of partners in a number of unusual instances.  One such was in Pipeline Estate, a low-income, densely populated community in Nairobi, Kenya.  Taara’s partner in Nairobi expected to see a large number of sign-ups when the service was connected, but after a few early new customer sign- ups things tailed off quickly.  After close examination it turned out those new users were consuming ~10 times the normal bandwidth and after a more detailed scrutiny they found that each of the new users were reselling pieces the large bandwidth now available.  At a cost of $10 to the primary user and a similar price to their shadow customers, they were making a substantial profit on the new service.  As this was not the model expected, the company created software that allowed ISPs and local residents to ‘micro-share’ the bandwidth and manage the usage.  Taara, after that experience, that capitalized on this situation and devised a system that allowed retail store owners to resell connectivity from the Taara system as an extra income source.
In India, Aitel (532454.IN) has used Taara’s system for FWA backhaul in conjunction with microwave links in a number of areas in India where physical issues (rivers, mountains, etc.) have limited capacity as a hybrid solution, and to provide service to locations where India’s well-known bureaucracy made it difficult or impossible to get microwave licensing within a reasonable amount of time. 
T-Mobile (TMUS) has used the technology to provide cell service at Coachella and the Albuquerque International Balloon Fiesta, where large crowds typically overloaded cell systems and laying fiber was definitely not cost effective.  As the Taara system could be set up quickly, they were able to provide 5G service for over a million calls over the nine day festivals at 99.9% uptime.  The Taara system was attached to a fiber link 2 km from the festival and paired with a T-Mobile Cell-on-Wheels station near the event and a second COW about a mile away.
There are others working on FSOC technology, but as can be seen below, most are looking at the technology from a military perspective, with few focused on backhaul as an adjunct to fiber.

• Raytheon (RTX) – has developed “NexGen Optix”, a FSOC tactical communication system for “challenging environments” to provide secure connectivity.

​Figure 5 - Mynaric - aerial communication module - Source: Mynaric

• fSona Networks (pvt) – FSOC back haul system but no active website.  Recent reverse merger
• CaiLabs (pvt) – Primarily aerospace and defense base stations and fiber lasers but also have TILBA-LOS FSOC point-to-point to be released this year.

Most of the FSOC development we have seen is oriented toward military applications and air-to-ground communication, likely the reason why Alphabet figures it’s time to a move Taara from the lab to a more potentially financially viable position.  This technology is relatively new to the back-haul market and seems to have become robust enough that it can have a place among other high bandwidth communication technologies, along with some more specific applications where laying fiber is not practical due to physical issues. 
While the more obvious cases are related to geophysical objects (rivers, offshore islands, and emergency disaster communication) short-haul links in urban areas where the cost and time to lay fiber is prohibitive, but line-of-site is available, does make sense if the technology proves to be as cost effective as Taara seems to indicate.  With the ability to eliminate much of the hardware needed to maintain alignment, the cost of base stations and repeaters should continue to drop, making the FSOC viable in more instances and certainly more competitive against regulated systems (microwave, RF).  While FSOC would not completely solve the need for high-speed, low-cost internet across the US, used in conjunction with fiber, it has the potential to make a dent, if the silicon proves to be as cost effective as it is promoted to be.   It’s a long way from the ballon network…
 
Please note that we do not receive any compensation from or have any connection to the companies or technology we write about.  We look for interesting products, companies, and technologies that we believe might be of interest to our readers.  While we might speak with some of these companies, we receive no proprietary information, and all opinions are our own.

Apple (AAPL) has changed its healthcare focus from a meager “Project Quartz” to a more meaningful and robust “Project Mulberry”, including AI agents to collect and process the data that Apple devices collect about you.  This is not the ‘secret stuff’ brands collect, like what OS you are using, what device you are on, your search results (if you let them), and almost everything about what you have bought[1], but more the data that you allow Apple to collect by using the Apple watch, your iPhone, your AirPods, and even some 3rd party applications.  This is ‘health’ data, that includes sleep patterns, steps, calories, heart rate, weight, and a variety of other metrics about your bodily functions.
The objective is to provide Apple users with information that will make them healthier and more fit, but Apple, even before the platform is available, has made the upgrade to AI agents and an integration with Apple Intelligence, to make that information more ‘real-time’, personal, and meaningful.  The agents are the scavengers that will poll your Apple devices for the health information they collect and bring it to Apple Intelligence for monitoring and evaluation.  It is thought that Apple will not only offer you evaluations of your nutritional and sleep habits but could even offer camera-based assessments of your workouts and access to educational videos, put together by internal and external health experts. 
While the range of detail is thought to delve into physical therapy, mental health, and even cardiology, the initial focus is thought to be nutritional, with monitoring and alerts leading to personalized health advice based on your data, although there has been talk of AI-based mental health counseling and chronic disease predictive analysis.  As one might expect, Apple’s focus seems to be on the ‘user experience’, the part of the Apple persona that allows them to charge a premium for their products, but Apple is certainly not the first to go in this direction in this new age of AI.  Google’s  (GOOG) Fit is a similar collector of personal health data through Android’s Health Connect.  This platform allows permitted 3rd party apps to supply and collect data that feed the Google Fit app, but is more a collector, aggregator, and visualizer than an advice tool, although Google is currently working to integrate that data into its other health related services, with a tie-in to reference ‘reputable sources’ on YouTube.
Amazon (AMZN) also has a health program, but its focus is more oriented toward B2B with the Amazon Pharmacy supplying information on medications and interactions and the Amazon Clinic and One Medical able to set up virtual video or text sessions with clinicians (some on staff) that can evaluate conditions, make diagnoses, and prescribe medication for relatively common illnesses.  There are also companies like Noom (pvt) or MyFitnessPal (pvt) that are more specific to food and calorie management but given the enthusiasm for Ai that seems rampant across the health sector, we expect almost every health related application to leverage AI to stay competitive.
There are a few caveats here, particularly HIPAA regulations which regulate any health information that is maintained or transferred.  Entities involved must encrypt health data, limit access, perform risk assessment, maintain audit trails, breach notifications, and take ‘reasonable steps’ to prevent access to or disclosure of patient information.  HIPAA is difficult enough to understand and maintain, but adding AI to the mix opens everything up to new legal questions, many of which have yet to reach the courts and as liability becomes a potential issue when health-related advice is being given, we expect many new court cases that will not only focus on the potential liability of poor or incorrect data, but will include questions of algorithmic bias, inadequate software testing, and the fact that Ai systems are essentially ‘black boxes’ that make it impossible to derive where or how an AI arrived at a particular diagnosis or conclusion. 
Smart lawyers will not only include site owners but also those who wrote the algos that run them, looking for biases that could cause hallucinations, errors in judgement, or flawed diagnoses based on poor human vetting.  When Ai developers are called into court to defend issues like what data was included in an AI’s training or what process was used to draw a conclusion, high level math will not be how they are judged by a jury, so while Apple jumps into the fray to provide a positive health experience through Project Mulberry and Apple Intelligence, its not like Wikipedia, where you take things with a grain of salt.  Healthcare decisions affect people’s lives, as some can be significantly influenced by the information given by Ai healthcare.  There are good and bad doctors, and sometimes doctors make mistakes, which is why malpractice insurance exists, but will there be malpractice insurance for an application that gives incorrect advice or misdiagnoses an ailment or mental condition?


[1] IP Address
Device Type & Model
Operating System
Device Identifiers (trackers like AAID, IMEI
Screen Resolution
Installed apps (some)
Browser type & version
Cookies (optional)
Browsing History (Optional)
Location Data (Optional)
Referring websites
App usage
Contacts & Calendar (Optional)
Photos & Videos (Optional)
In-app purchases
Search queries (Optional)
Social Media Activity
Shopping activity
Form submissions
Wi-Fi network name
Data usage
Bluetooth data
Sensor tracking
Accelerometer & Gyroscope data
Ambient Lighting data
‘like’ data
DNS lookups
…To name a few.

Alibaba (BABA) Cloud announced that it was lowering the price of its LLM model for the 3rd time to remain competitive in the Chinese AI market.  The model, known as Qwen-VL has a number of primary features, such as multi-modality (Can accept both text and image input), High-resolution processing (>1m pixels), enhanced image extraction, and multilingual support (Eng, Chinese, Japanese, Korean, Arabic, Vietnamese) and is closest to Google’s (GOOG) Gemini, which has similar features.  The model is part of Alibaba’s cloud-based AI chatbot family, which focuses on enterprise customers rather than the consumer market as a way to differentiate itself from Chinese and other AI competitors.
While much has been said about the competitive nature of Chinese companies, that rhetoric has been typically focused on manufacturing, however it seems that the AI market in China has spurred an even more intense competition to gain share in its own market.  In June of 2024 there were over 230 million AI product and services users in China, according to state-sponsored data, which grew to over 600 million by the end of October, with almost 200 LLM commercially available models to choose from.  While we believe the share of the potential user base that is using an AI on at least a weekly basis is higher currently in the US than in China, and generated more sales in 2023, expectations for industry growth over the next seven years are higher for China (25.6% CAGR for China v. 23.3% for the US[1] ), which is the impetus for the even more aggressive nature of Chinese AI Chatbot brands.
With this intense level of competition among AI Chatbot model providers, we were curious to not only to see if we could quantify the rate of price reductions but also compare those to model price reductions outside of China.  We note that this is an unscientific comparison, as each of the models has its own set of features and characteristics, and the availability of this data is, at best, poor, but we gathered as much data as possible, and converted the Chinese price data to US dollars for comparison.  Most notable is that the price of the most recent Tencent (700.HK) model, which has been available for roughly one month, is now the same as the Alibaba Qwen-VL model, which has been available for well over a year, and while the non-Chinese model prices have come down at a similar rate to the Chinese models, the current prices of the non-Chinese models are appreciably higher.  Overall, one might question the viability of the current business models behind commercial chatbots based on the data.
We note also that Baidu’s (BIDU) ERNIE model is now free, and Google’s BARD has morphed into Gemini.  We can find no specific data on how the MetaAI (FB) models are broken out pricewise and we note also that there are newer models for some (GPT-4 for example) that have much higher performance and cost, but this is as close to a comparison as we could make given the time involved.  The prices are for 1,000 tokens of input data in all cases.


[1] GrandView Research