Five Tech Trends Coming in 2025Five Tech Trends Coming in 2025

A quarter-century into the new millennia, the computing industry is undergoing change and advancement as significant as the development of the integrated circuit and microcontroller was back in 1975. 

Here, I’ve compiled overviews of five subjects to keep an eye on in 2025.

Data centers enter a new era in 2025 

In the early part of the 21^st century, data centers were largely all about, well… data. Early stages of the Internet showed that distributed data centers could improve search and data retrieval as well as add a layer of redundancy. Then came Bitcoin, and data centers became massive number-crunching warehouses full of graphics processing units (GPUs). Most recently, generative AI has stolen the spotlight, and data centers have added AI-oriented tensor and neural processors (TPUs and NPUs) to the processor mix.

The current growth of AI will impact data centers with required increases in electrical power, processor manufacturing, and technical labor at a level that simply doesn’t exist. It will need a diversity of resources and skills required, perhaps only comparable to the Manhattan Project during World War II. Sustaining the necessary growth is not possible within the current ecosystem model. This fact will become quite clear in 2025, with much press covering the coming data center capacity shortage. 

Related:The Embedded Systems Industry in 2024: Top Trends & What’s Next

The demand for AI large language model (LLM) processing and consumer generative AI services is growing so fast that there is talk of using nuclear power to keep up with the AI power demand. It remains to be seen if nukes can get regulatory approval, but my hunch is that we’ll see that approval starting late 2025.

One partial solution we’ll see more of in 2025 is a significant emphasis on edge AI. Edge computing holds many applications that can benefit from AI. However, with higher-performing, lower-power CPUs and AI-capable microcontrollers, many of those applications can now be served with edge AI. We saw the start of this trend in 2024, but getting AI out into the field will go full-steam in 2025. Anything AI that can be broken out and processed at the edge rather than at the data center will be done in the field. This will allow data centers to stay focused on the massively parallel processing needs.

The ultimate solution lies with data center class quantum computing, which is half a decade or more from large-scale commercial viability. In the interim, the industry needs a new data center architecture. The current model of more racks with more processors with more cores is reaching physical and logistical limits.

Related:Energy Efficiency Becomes Focus of Chip Research

AI reality drift stokes fear

Reality drift (RD) is a recursive phenomenon that occurs when generative AI output ends up back in the original data source for the large language model (LLM). The current generative AI systems are ripe for automated reality drift.

The best example of AI reality drift is when an AI LLM uses a large sampling of the Internet as its data source. Generative AI output from that LLM ends up back on the Internet to be used as source material. The result is an information equivalent to audio feedback. Either the absurdity levels of generated information will expand to ridiculous levels, or the highs and lows will be filtered off with information dumbed down to a lowest common denominator.

When a new technology shows incredible promise but doesn’t yet have adequate safeguards, the dark side of technology often weasels its way in. This will result in intentional reality drift (IRD). IRD occurs when the data source for the generative AI engine is deliberately manipulated. The manipulation may be performed for nefarious purposes, like William Randolf Hearst and his Spanish American War agenda. Many people have ulterior motives and will encourage reality drift, either through the injection of false information into the raw data sets or by adjusting the AI algorithms to suit their intended level of disinformation.

Related:Quantum Computing No Longer a Pipedream

Reality drift will also occur when AI suppliers start limiting the processing allocated to generative or interpretive requests based on cost (either in dollars of Kilowatts). Providers may offer differing levels of accuracy based on subscription fees. History shows that humanity most often selects low cost over high quality. That low quality material will be fed back into the source, ad infinitum.

Probability joins analog & digital

Anyone who has spent even the slightest amount of time around electronics understands the distinction between digital and analog electronics. Get ready for probability electronics to join the fold. Probability electronics already exist in super high-speed communications. These signals operate so fast that conventional digital electronics cannot read them. Multiple interpretations are summed together to get a probability-based result that is statistically equivalent to a digital answer.

Probability-based computing (PBC) will get more screen time as well. PBC is at the heart of LLM AI computing. Algorithms are run comparing an unknown model to a known model. The output is a probability of a match. AI systems rerun the algorithm thousands of times, with each repeated running increasing the probability of an accurate response. 

With the rise of probability-based computation, you will be hearing more about the probability threshold (PT). The PT is a value, above which the system is understood to be most likely to be correct, and below which there is doubt. Doubt increases the further you go below the PT and certainty increases as you get closer to and cross above the PT. As data center capacity shortage takes hold, AI providers are likely to start skimping on accuracy by lowering their accepted PT. Lowering the corporate PT is the information equivalent to lowering manufacturing quality standards.

General-purpose processors retake the stage

Processor technology has advanced rapidly over the last decade, evolving from monolithic structures to chiplet-based designs with high-speed interconnects and stacks of main system RAM on the same interconnect substrate. The general-purpose central processing unit (CPU) and graphics processing unit (GPU) have been joined by specialized neural and tensor processing units (NPU and TPU). Programmable logic array (PLA) fabric is even finding a home alongside the hard silicon processing cores. The CPUs have become less a key processing engine in the data center world and more the main arbiter between the GPUs, NPUs, TPUs, and PLAs.

Intel is surrounded by rumors of a potential takeover, with Qualcomm and Apple being the most prominent rumored suitors. The “good money” is on Intel’s demise. However, Intel has been in trouble before and worked through it. The general-purpose CPU may not be the media darling it once was. But it will be again. GPUs, NPUs, TPUs, and others as of yet unannounced “xPUs” get all of the attention now, but the CPU will be back. Nvidia is getting into the desktop CPU game, ARM is making inroads, RISC-V is nipping at heals, and the “demise of Intel” may be a premature announcement. 2025 will be the year of the reemergence of the CPU.

Wires make a comeback

High-speed WiFi and the latest in mobile data live up to their promise of nearly unlimited data, fast multi-device two-way streaming, and nearly lag-free top-gear gaming—if you live in an open field with no structures or other wireless users nearby. The newer standards use higher frequencies. At the corresponding shorter wavelengths, individual rooms start to look like Faraday cages. Mesh networking devices are easy to set up, easy to use, and can deliver great coverage across large houses or commercial settings. However, they often only really meet their potential if each unit is close enough to another unit to receive signal that has not degraded below the maximum speed capability. With remote work arguably being a permanent fixture in our economy, data bandwidth and speed across the home become mission critical rather than a luxury.

Security will also factor into the return of wires. Internet-connected cameras, locks, and alarms are quickly replacing older non-IP based systems. These systems need to be responsive, robust, and secure. Solar-powered wireless security devices are easy to install and connect. However, they are also easy to disable or jam. More and more homeowners want wired Power over Ethernet (PoE) security systems with a backup power supply.

The easiest solution is to run Ethernet wires to each room and hardwire mesh WiFi wherever it will be used. A hardwired mesh node is as easy to set up as a wireless one and as easy to connect to devices. However, each node starts out with the same maximum signal as is available at the router. Wireless mesh nodes still have to contend with signal degradation at each point in the daisy chain. New construction needs to accommodate wired networking, nodes in every room, battery backup, and node access at key security points around the house.

2025—An exciting year to come

AI got most of the headlines in 2024, but in the next year, we’ll see much more about what’s behind the AI curtain. Data center capacity shortages and AI reality drift will make big news, with reality drift becoming a pop-culture boogeyman in 2025. Probability-based electronics and computing make people think, and bring urgency to the advancement of quantum computing. General-purpose CPUs will jump out of the shadows and get some much-deserved attention, and the lowly data wire will be back in the construction world. Tipping points are always a challenge and this one, at the edge of a new quarter-century, is no different.