Read on for our intrepid engineer’s latest set of predictions for the year(s) to come.
As has been the case the last couple of years, we’re once again flip-flopping what might otherwise seemingly be the logical ordering of this and its companion 2025 look-back piece. I’m writing this 2026 look-ahead for December publication, with the 2025 revisit to follow, targeting a January 2026 EDN unveil. While a lot can happen between now and the end of 2025, potentially affecting my 2026 forecasting in the process, this reordering also means that my 2025 retrospective will be more comprehensive than might otherwise be the case.
Without any further ado, and as usual, ordered solely in the cadence in which they initially came out of my cranium…
AI-based engineering
Likely unsurprisingly, as will also be the case with the subsequent 2025 retrospective-to-come, AI-related topics dominate my forecast of the year(s) to come. Take “vibe coding”, which entered the engineering and broader public vernacular only in February and quickly caught fire. Here’s Wikipedia’s introduction to the associated article on the subject:
Vibe coding is an artificial intelligence-assisted software development technique popularized by Andrej Karpathy in February 2025. The term was listed on the Merriam-Webster website the following month as a “slang & trending” term. It was named Collins Dictionary‘s Word of the Year for 2025.
Vibe coding describes a chatbot-based approach to creating software where the developer describes a project or task to a large language model (LLM), which generates code based on the prompt. The developer does not review or edit the code, but solely uses tools and execution results to evaluate it and asks the LLM for improvements. Unlike traditional AI-assisted coding or pair programming, the human developer avoids examination of the code, accepts AI-suggested completions without human review, and focuses more on iterative experimentation than code correctness or structure.
Sounds great, at least in theory, right? Just tell the vibe coding service and underlying AI model what you need your software project to do; it’ll as-needed pull together the necessary code snippets from both open-source and company-proprietary repositories all by itself. If you’re already a software engineer, it enables you to crank out more code even quicker and easier than before.
And if you’re a software or higher-level corporate manager, you might even be able to lay off (or at least pay grade-downscale) some of those engineers in the process. Therein explaining the rapid rollout of vibe coding capabilities from both startups and established AI companies, along with evaluations and initial deployments that’ll undoubtedly expand dramatically in the coming year (and beyond). What could go wrong? Well…
Advocates of vibe coding say that it allows even amateur programmers to produce software without the extensive training and skills required for software engineering. Critics point out a lack of accountability, maintainability, and the increased risk of introducing security vulnerabilities in the resulting software.
Specifically, a growing number of companies are reportedly discovering that any upfront time-to-results benefits incurred by AI-generated code end up being counterbalanced by the need to then reactively weed out resulting bugs, such as those generated by hallucinated routines when the vibe coding service can’t find relevant pre-existing examples (assuming the platform hasn’t just flat-out deleted its work, that is).
To that point, I’ll note that vibe coding, wherein not reviewing the resultant software line-by-line is celebrated, is an extreme variant of the more general AI-assisted programming technology category.
But even if a human being combs through the resultant code instead of just compiling and running it to see what comes out the other end, there’s still no guarantee that the coding-assistance service won’t have tapped into buggy, out-of-date software repositories, for example. And there’s always also the inevitable edge and corner cases that won’t be comprehended upfront by programmers relying on AI engines instead of their own noggins.
That all said, AI-based programming is already having a negative impact on both the job prospects for university students in the computer science curriculum and the degree-selection and pursuit aspirations of those preparing to go to college, not to mention (as already alluded to) the ongoing employment fortunes of programmers already in the job market.
And for those of you who are instead focused on hardware, whether that be chip- or board-level design, don’t be smug. There’s a fundamental reason, after all, why a few hours before I started writing this section, NVIDIA announced a $2B investment in EDA toolset and IP provider Synopsys.
Leveraging AI to generate optimized routing layouts for the chips on a PCB or the functional blocks on an IC is one thing; conventional algorithms have already been handling this for a long time. But relying on AI to do the whole design? Call me cynical…but only cautiously so.
Memory (and associated system) supply and prices
Speaking of timely announcements, within minutes prior to starting to write this section (which, to be clear, was also already planned), I saw news that Micron Technology was phasing out its nearly 30-year old Crucial consumer memory brand so that it could redirect its not-unlimited fabrication capacity toward more lucrative HBM (high bandwidth memory) devices for “cloud” AI applications.
And just yesterday (again, as I’m writing these words), a piece at Gizmodo recommended to readers: “Don’t Build a PC Right Now. Just Don’t”. What’s going on?
Capacity constraints, that’s what. Remember a few years back, when the world went into a COVID-19 lockdown, and everyone suddenly needed to equip a home office, not to mention play computer games during off-hours?
Device sales, with many of them based on DRAM, mass storage (HDDs and/or SSDs), and GPUs, shot through the roof, and these system building blocks also then went into supply constraints, all of which led to high prices and availability limits.
Well, here we go again. Only this time, the root cause isn’t a pandemic; it’s AI. In the last few years’ worth of coverage on Apple, Google, and others’ device announcements, I’ve intentionally highlighted how much DRAM each smartphone, tablet, and computer contains, because it’s a key determinant of whether (and if so, how well) it can run on-device inference.
Now translate that analogy to a cloud server (the more common inference nexus) and multiply both the required amount and performance of memory by multiple orders of magnitude to estimate the demand here. See the issue? And see why, given the choice to prioritize either edge or datacenter customers, memory suppliers will understandably choose the latter due to the much higher revenues and profits for a given capacity of HBM versus conventional-interface DRAM?
Likely unsurprising to my readers, nonvolatile memory demand increases are pacing those of their volatile memory counterparts. Here again, speed is key, so flash memory is preferable, although to the degree that the average mass storage access profile can be organized as sequential versus random, the performance differential between SSDs and lower cost-per-bit HHDs (which, mind you, are also increasingly supply-constrained by ramping demand) can be minimized.
Another traditional workaround involves beefing up the amount of DRAM—acting as a fast cache—between the mass storage and processing subsystems, although per the prior paragraph it’s a particularly unappealing option this time around.
I’ve still got spare DRAM DIMMs and M.2 SSD modules, along with motherboards, cases, PSUs, CPUs, and graphics cards, and the like sitting around, left over from my last PC-build binge.
Beginning over the upcoming holidays, I plan to fire up my iFixit toolkits and start assembling ‘em again, because the various local charities I regularly work with are clearly going to be even more desperate than usual for hardware donations.
The same goes for smartphones and the like, and not just for fiscally downtrodden folks…brace yourselves to stick with the devices you’ve already got for the next few years. I suspect this particular constraint portion of the long-standing semiconductor boom-and-bust cadence will be with us even longer than usual.
Electricity rates and environmental impacts
Not a day seemingly goes by without me hearing about at least one (and usually multiple) new planned datacenter(s) for one of the big names in tech, either being built directly by that company or in partnership with others, and financed at least in part by tax breaks and other incentives from the municipalities in which they’ll be located (here’s one recent example).
And inevitably that very same day, I’ll also see public statements of worry coming from various local, state, and national government groups, along with public advocacy organizations, all concerned about the environmental and other degrading impacts of the substantial power and water needs demanded by this and other planned “cloud” facilities (ditto, ditto, and ditto).
Truth be told, I don’t entirely “get” the municipal appeal of having a massive AI server farm in one’s own back yard (and I’m not alone). Granted, there may be a short-duration uptick in local employment from construction activity.
The long-term increase in tax revenues coming from large, wealthy tech corporations is an equally enticing Siren’s Song (albeit counterbalanced by the aforementioned subsidies). And what politician can’t resist proudly touting the outcome of his or her efforts to bring Alphabet (Google)/Amazon/Apple/ Meta/Microsoft/[insert your favorite buzzy company name here] to his or her district?
Regarding environmental impacts, however, I’ll “showcase” (for lack of a better word) one particularly egregious example: Elon Musk’s xAI Colossus 1 and 2 data centers in Memphis, Tennessee.
The former, a repurposed Electrolux facility, went online in September 2024, only 122 days after construction began. The latter, for which construction started this March, is forecasted, when fully equipped, to be the “First Gigawatt Datacenter In The World”. Sounds impressive, right? Well, there’s also this, quoting from Wikipedia:
At the site of Colossus in South Memphis, the grid connection was only 8 MW, so xAI applied to temporarily set up more than a dozen gas turbines (Voltagrid’s 2.5 MW units and Solar Turbines’ 16 MW SMT-130s) which would steadily burn methane gas from a 16-inch natural gas main. However, according to advocacy groups, aerial imagery in April 2025 showed 35 gas turbines had been set up at a combined 422 MW. These turbines have been estimated to generate about “72 megawatts, which is approximately 3% of the (TVA) power grid”. According to the Southern Environmental Law Center (SELC), the higher number of gas turbines and the subsequent emissions requires xAI to have a ‘major source permit’, however, the emissions from the turbines are similar to the nearby large gas-powered utility plants.
In Memphis, xAI was able to sidestep some environmental rules in the construction of Colossus, such as operating without permits for the on-site methane gas turbines because they are “portable”. The Shelby County Health department told NPR that “it only regulates gas-burning generators if they’re in the same location for more than 364 days. In the neighborhood of South Memphis, poor air quality has given residents elevated asthma rates and lower life expectancy. A ProPublica report found that the cancer risk for those living in this area already have four times the risk of cancer than what the Environmental Protection Agency (EPA) considers to be an acceptable risk. In November 2024, the grid connection was upgraded to 150 MW, and some turbines were removed.
Along with high electricity needs, the expected water demand is over five million gallons of water per day in “… an area where arsenic pollution threatens the drinking water supply.” This is reported by the non-profit Protect Our Aquifer, a community organization founded to protect the drinking water in Memphis. While xAI has stated they plan to work with MLGW on a wastewater treatment facility and the installation of 50 megawatts of large battery storage facilities, there are currently no concrete plans in place aside from a one-page factsheet shared by MLGW.
Geothermal power
Speaking of the environment, the other night I watched a reality-calibrating episode of The Daily Show, wherein John Stewart interviewed Elizabeth Kolbert, Pulitzer Prize-winning author and staff writer at The New Yorker:
I say “calibrating” because it forced me to confront some uncomfortable realities regarding global warming. As regular readers may already realize, either to their encouragement or chagrin, I’m an unabashed believer in the following:
- Global warming is real, already here, and further worsening over time
- Its presence and trends are directly connected to human activity, and
- Those trends won’t automatically (or even quickly) stop, far from reversing course, even if that causational human activity ceases.
What I was compelled to accept after watching Stewart and Kolbert’s conversation, augmenting my existing opinion that human beings are notoriously short-sighted in their perspectives, frequently to their detriment (both near- and long-term), were conclusions such as the following:
- Expecting humans to willingly lower (or even flatline, versus constantly striving to upgrade) their existing standards of living for the long-term good of their species and the planet they inhabit is fruitless
- And given that the United States (where I live, therefore the innate perspective) is currently the world’s largest supplier of fossil fuel—specifically, petroleum and natural gas—energy sources, powerful lobbyists and other political forces will preclude serious consideration of and responses to global warming concerns, at least in the near term.
In one sense, those in the U.S. are not alone with their heads-in-the-sand stance. Ironically, albeit intentionally, the photo I included at the beginning of the prior section was of a coal-burning power plant in China.
That said, at the same time, China is also a renewable energy leader, rapidly becoming the world’s largest implementer of both wind and solar cell technology, both of which are now cheaper than fossil fuels for new power plant builds, even after factoring out subsidies. China also manufactures the bulk of the world’s lithium-based batteries, which enable energy storage for later use whenever the sun’s not shining and the wind’s not blowing.
To that latter point, though, while solar, wind, and many other renewable energy sources, such as tidal power, have various “green” attributes both in an absolute sense and versus carbon-based alternatives, they’re inconsistent in output over time. But there’s another renewable option, geothermal power, that doesn’t suffer from this impermanence, especially in its emerging “enhanced” variety. Traditional geothermal techniques were only limited-location relevant, with consequent challenges for broader transmission of any power generated, as Wikipedia explains:
The Earth’s heat content is about 1×1019 TJ (2.8×1015 TWh). This heat naturally flows to the surface by conduction at a rate of 44.2 TW and is replenished by radioactive decay at a rate of 30 TW. These power rates are more than double humanity’s current energy consumption from primary sources, but most of this power is too diffuse (approximately 0.1 W/m2 on average) to be recoverable. The Earth’s crust effectively acts as a thick insulating blanket which must be pierced by fluid conduits (of magma, water or other) to release the heat underneath.
Electricity generation requires high-temperature resources that can only come from deep underground. The heat must be carried to the surface by fluid circulation, either through magma conduits, hot springs, hydrothermal circulation, oil wells, drilled water wells, or a combination of these. This circulation sometimes exists naturally where the crust is thin: magma conduits bring heat close to the surface, and hot springs bring the heat to the surface.
To bolster the identification of such naturally geothermal-friendly locations (the photo at the beginning of this section was taken in Iceland, for example), companies such as Zanskar are (cue irony) using AI to locate previously unknown hidden sources. I’m admittedly also pleasantly surprised that the U.S. Department of Energy just announced geothermal development funding.
And, to even more broadly deploy the technology, other startups like Fervo Energy and Quaise Energy are prototyping ultra-deep drilling techniques first pioneered with (again, cue irony) fracking to pierce the crust and get to the constant-temperature, effectively unlimited energy below it, versus relying on the aforementioned natural conduit fractures. That it can be done doesn’t necessarily mean that it can be done cost-effectively, mind you, but I for one won’t ever underestimate the power of human ingenuity.
World models (and other LLM successors)
While the prior section focused on accepting the reality of ongoing AI technology adoption and evolution, suggesting one option (of several; don’t forget about nuclear fusion) for powering it in an efficient and environmentally responsible manner, this concluding chapter is in some sense a counterpoint. Each significant breakthrough to date in deep learning implementations, while on the one hand making notable improvements in accuracy and broader capabilities, has also demanded ever-beefier compute, memory, and other system resources to accomplish its objectives…all of which require more energy to power them, along with more water to remove the heat byproduct of this energy consumption. The AI breakthrough introduced in this section is no exception.
Yann LeCun, one of the “godfathers” of AI whom I’ve mentioned here at EDN numerous times before (including just one year ago), has publicly for several years now been highly critical of what he sees as the inherent AGI (artificial general intelligence) and other limitations of LLMs (large language models) and their transformer network foundations.
A recent interview with LeCun published in the Wall Street Journal echoed many of these longstanding criticisms, adding a specific call-out for world models as their likely successor. Here’s how NVIDIA defines world models, building on my earlier description of multimodel AI:
World models are neural networks that understand the dynamics of the real world, including physics and spatial properties. They can use input data, including text, image, video, and movement, to generate videos that simulate realistic physical environments. Physical AI developers use world models to generate custom synthetic data or downstream AI models for training robots and autonomous vehicles.
Granted, LeCun has no shortage of detractors, although much of the criticism I’ve seen is directed not at his ideas in and of themselves but at his claimed tendency to overemphasize his role in coming up with and developing them at the expense of other colleagues’ contributions.
And granted, too, he’s planning on departing Meta, where he’s managed Facebook’s Artificial Intelligence Research (FAIR) unit for more than a decade, for a world model-focused startup. That said, I’ll forever remember witnessing his decade-plus back live demonstration of early CNN (convolutional neural network)-based object recognition running on his presentation laptop and accelerated on a now-archaic NVIDIA graphics subsystem:
He was right then. And I’m personally betting on him again.
Happy holidays to all, and to all a good night
I wrote the following words a couple of years ago and, as was also the case last year, couldn’t think of anything better (or even different) to say this year, given my apparent constancy of emotion, thought, and resultant output. So, once agai,n with upfront apologies for the repetition, a reflection of my ongoing sentiment, not laziness:
I’ll close with a thank-you to all of you for your encouragement, candid feedback and other manifestations of support again this year, which have enabled me to once again derive an honest income from one of the most enjoyable hobbies I could imagine: playing with and writing about various tech “toys” and the foundation technologies on which they’re based. I hope that the end of 2025 finds you and yours in good health and happiness, and I wish you even more abundance in all its myriad forms in the year to come. Let there be Peace on Earth.
p.s…let me (and your fellow readers) know in the comments not only what you think of my prognostications but also what you expect to see in 2026 and beyond!
—Brian Dipert is the Principal at Sierra Media and a former technical editor at EDN Magazine, where he still regularly contributes as a freelancer.
Related Content
- 2025: A technology forecast for the year ahead
- 2024: A technology forecast for the year ahead
- 2023: A technology forecast for the year ahead
- 2022 tech themes: A look ahead
- A tech look back at 2022: We can’t go back (and why would we want to?)
- A 2021 technology retrospective: Strange days indeed
The post 2026: A technology forecast for AI’s ever-evolving bag of tricks appeared first on EDN.
