Two Opportunities for 2025: Rethinking AI and Techno-miniaturization

As we enter 2025, I can’t help but think that the blush is off the AI rose.? I don’t expect sharp declines in NVIDIA stock or an end to the urge of consultants to reflexively recommend business solutions based on machine learning (ML).? In fact, as an active day trader I am personally betting that disasters are not in AI’s immediate future. Nonetheless, some rethink will be needed in the AI sphere soon and this rethink may itself lead to next thing in tech.? In this edition of From My Desk, I explain.

?What’s Wrong with AI:? Winter is Coming?

The truth is that problems are emerging for AI.? The hype and glory days for AI are ending.? Consider, for a start, that “AI hallucination” thing.? According to IBM (https://www.ibm.com/topics/ai-hallucinations),? “AI hallucination is a phenomenon wherein an LLM. . .perceives patterns or objects that are nonexistent or imperceptible to human observers, creating outputs that are nonsensical or altogether inaccurate.” In other words, AIs are liars.? In the words of Dr. House (https://en.wikipedia.org/wiki/House_(TV_series)), “Everyone lies!”? One of my wife’s students almost failed a class the other week because he inserted ChatGPT-generated errors into his work.

?More worrying to her husband and presumably to others of us who have hung out our shingles as industry analysts are the memories of “AI winters” past when AI projects were ruthlessly defunded as AI proved itself unable to meet (A) the expectation of customers and (B) the promises made in proposals to management and investors.? As this happened, the powers-that-be pulled the plug on finance and moved on to other things.

?At CIR, we would not be too surprised if – perhaps short of an AI winter – some sense of realism about AI begins to hit in 2025. Perhaps Elon will even make 2025 the year that he holds off predicting that his cars will become fully autonomous, “real soon now.” As an aside, one must realize that part of the silliness now surrounding AI is because a lot of people currently commenting on the AI and quantum networking scene apparently haven’t a clue.? (I hesitate to identify anyone in particular, but if you search for videos on YouTube under “Google Willow chip” you will be either amused or depressed or both.)

?But I digress.? Objectively speaking the AI market just seems to me to be overstated.? According to Statista the AI market will reach $827 Billion by 2030 (https://www.statista.com/forecasts/1474143/global-ai-market-size).? This is quite close to the current GDPs of Taiwan or Switzerland, which for me makes these numbers difficult for me to believe. This is because AI is proving to be (1) harder to implement and (2) of lower quality than was prophesized a year or two back.? I fear sleight of hand is at work by some market analysts.

?As far as (1) is concerned we note that both the high-speed (1.6T) Ethernet and Small Modular (nuclear) Reactors that will be required for fully functional AI data centers are still in development.? As far as (2) is concerned, we must ask ourselves whether the SIRI of a few years ago is really all that different to the SIRI of today?

?With these considerations in mind the projections for AI in our new report “Networks and Power Requirements for AI Data Centers:? A Ten-Year Market Forecast and Technology Assessment” (https://cirinc.wpenginepowered.com/reports-3/) are more realistic.? For 2030 we have just over $100 Billion.?

?That’s still a lot of money and much of it will be made by fixing the problems with AI that we have just identified.? That’s the first of the two opportunities that I am identifying in this piece.? The second is something I am calling techno-miniaturization.

?Opportunities Will Be Getting Smaller, But in a Good Way

While AI is being “fixed” investors, along with analysts like me, will be busy looking for the proverbial “next big thing.”? By this what I really mean is a big idea that lets you tap into the latest zeitgeist, and which can propel your thoughts and proposals towards acceptance and (especially) towards funding.??

?Good examples of such “next big things” from the recent past include “Sustainability” and “Quantum.”? They rise to prominence as the result of a perceived societal or market need, excite the imagination of entrepreneurs, investors, government and the media for a while and then problems surrounding them appear when reality sets in.? And then it’s on to the next “next big thing,” while the problems get sorted.? That’s where I think we are with AI right now.

?The problems that inevitably emerge from “next big things” are typically solvable in the medium term. Eventually, people will stop saying those silly things about AI, while AI hallucinations will stop saying silly things about almost everything!? For now, that’s all I have to say about all this.? (https://www.reddit.com/r/shittymoviedetails/comments/115h4kj/in_forrest_gump_1994_forrest_says_and_thats_all_i/)

?What I do want to say though is when we come to a little bump in the road such as I think we are going to see in 2025 with AI, the bump is enough to set the adventurous on a quest for the new again. And sometimes the “new” can be related to that older zeitgeist that we are beginning to find problems with.

?This is what I expect to happen in 2025. AI will become less of a gold rush and more of a business.? But I also expect entrepreneurs, strategists and corporate philosophers to see something potentially profitable in how we are solving AI’s problems.? I am going to call the how in this case “techno-miniaturization.” ?OK, there must be a better name, but it’s the best I can do for the moment.?

??Techno-Miniaturization Manifestations: Three Examples

My point here is this. AI so far is a huge AI resource hog.? It is threatening to drain resources to a frightening degree. Miniaturization preserves resources and here we are talking in the context of technology, hence “techno-miniaturization.” You will know techno-miniaturization by its features which may manifest themselves by as increased density, integration, localization and enhanced efficiency as well as actual miniaturization.

?Actual miniaturization: Miniaturization will directly impact AI industry will be through the arrival of Small Modular (nuclear) Reactors (SMRs), without AI’s great expectations may never be met.? Most of the hyperscalers are now exploring this power option.?

For example, Google has signed a deal with Kairos Power to build a fleet of SMRs?to power its data centers. These SMRs, are scheduled to be completed between 2030 and 2035 and will produce 500 megawatts of power.?However, we also note that, according to some, SMRs have been built in as little as three years.? On the other hand, as has always been the case with fission reactors nobody quite knows what to do with the radioactive waste.? Bury it or shoot it into space?

?Densification: Increased density is miniaturization in that more servers and racks are crammed into less space.? This is already a core strategy enabling AI data centers. In Northern Virginia, near where I live, there is one of the largest concentrations of data centers in the US, but with a vacancy rate of just 1 percent in 2024. In a few cases, ten- year leases are being signed and the hyperscalers and data-center developers often bank land up to ten years before they start to build. The shortage of land for data centers makes densification an imperative

?AI data centers need to be located optimally close to both power generation sources and (for latency-related reasons) to customers. This makes land suitable for AI data centers a wise investment.? As Mark Twain said, “Buy land, they're not making it anymore,” although I am pretty sure he never mentioned AI data centers specifically. Since densification means more heat – especially in AI data centers which typically use a lot of power another – another excellent investment is likely to be in the liquid cooling technologies which are emerging to serve the needs of data centers.

?Integration: ?Nothing new about integration, which is logically and historically tied to miniaturization.? And with AI there is so much to integrate.? There is, for example, a huge

opportunity to integrate AI MPUs/CPUs, accelerators, memory, security and communications chips and more.? Perhaps the AI data center will turn out to be the application that optical integration has always been seeking. I notice that there is currently a bit of a boom in chiplets and optical interconnects, all designed with the AI data center in mind. One recent innovation in this space is NVIDIA’s Grace-Blackwell product that contains two GPUs with one CPU to make a new “superchip. Ayar’s chiplet product with the scary name TeraPHY (which is pronounced “terrify”!) is another example of integration in the service of AI.

??Coda

So, what I am saying here is that “techno-miniaturization” is the coming buzzword of power to impress and close the sale, but the “features” mentioned above draw in the audience and show the ways in which techno-miniaturization can be achieved.? IMHO I think that techno-miniaturization or something like it may become an inspiration to both business and R&D activity and may even be a thing that spreads beyond AI and IT more generally.? ??

But this is a very big topic and a very small newsletter.