Not everything published is worth reading. And if I do read it, it's probably not worth a whole book report. These are some of the stories, studies, stunts, and other stuff that I think data center types might actually want to spend a minute or two on, with a little bit of additional context from me. No spin, no press release rewrites and if I think something is shaky, I'll say so.
Is The Matrix Coming True? Maybe...
Scientists at Singapore's Yong Loo Lin School of Medicine have demo'd a prototype rack that uses living human neurons for processing, built with DayOne and Cortical Labs. Interesting Engineering describes 20 CL1 units, each with about 800,000 lab-grown cells, or roughly 16 million neurons in one rack. The hook for data center folks is power: the researchers say this could run on a fraction of the electricity silicon needs. That is a belief, not a measured result, and this is a lab prototype, not a product for the GPU aisle. Let's see some benchmarks, right?
Source: Interesting Engineering, August 21, 2026. Tags: Power, AI, Data Center Design
Average Rack Turned Up to 11
Uptime Institute's 16th annual data center survey shows average rack power draw at a record high of 11 kW, up from 9 kW last year. Other high points include worries about costs and the struggle to plan for both conventional CPU-driven workloads and newer AI applications.
The article makes it sound like users see those as two separate things. I don't buy it. Some sort of AI mechanism is going to get folded into every application where it makes sense, and lots where it doesn't.
The survey also seems to concentrate on model training, which is a big deal and needs lots of accelerators. But over time the bigger workload will be inference. That will still need accelerators in most cases, just at lower density than training. I also get the sense that the respondent base might be skewed toward hyperscalers and co-location providers.
The most interesting bit: users track server utilization, but they don't seem to do much to maximize it. Uptime files utilization under sustainability. I look at it as a critical economic and operational factor, and it's the difference between AI gear that pays for itself and AI gear that just sits there costing you money.
Source: Data Center Knowledge, July 31, 2026. Tags: Data Center Design, Power, Industry
Is Google Search Over? Yeah
Google Search as you know it is over, says TechCrunch, after Google I/O showed AI Overviews, follow-up questions in AI Mode, information agents, and generative UI landing in the search box. Google says AI Overviews already reach more than 2.5 billion monthly users and AI Mode more than 1 billion; those are the company's numbers. Links are not gone, which TechCrunch had to clarify after the headline did some damage, but they are no longer the priority for many queries. For data center folks, this means more inference in Google's plants, and a shift in how enterprises and vendors get found on the web. Worth watching, and if you're relying on Google search results to drive people to your site, be prepared to make a change.
Source: TechCrunch, May 19, 2026. Tags: AI, Industry
GigaIO Sells Composable Business to d-Matrix
GigaIO sold its SuperNODE platform and FabreX PCIe Gen 5 AI fabric to d-Matrix, along with key rack-scale engineering talent. d-Matrix says it will use that IP to build rack-scale, low-latency inference systems around its Corsair accelerators. No purchase price was disclosed, and claims like "the industry's most scalable AI inference platform" come from the companies, not an independent benchmark. GigaIO is pivoting to Gryf, a suitcase-sized edge system it calls the world's first carry-on AI supercomputer, which is a vendor slogan, not a measurement. If you were looking at GigaIO for composable GPU fabric in the data center, that product line now lives at d-Matrix.
Source: GigaIO, April 2, 2026. Tags: Accelerators, AI, Industry
Server market rockets higher in '25
Our pal TPM at The Next Platform breaks down the Q4 2025 server market with the kind of context you actually need. IDC put quarterly server sales at $125.3 billion, up 52.4 percent year on year. TPM calculates that $70.65 billion of that, or 56.4 percent, was GPU-accelerated gear, and he is careful to say some earlier chart gaps were filled with estimates. Dell led named OEMs at $12.65 billion and Supermicro had $11.7 billion, while ODM vendors as a group shipped $66.62 billion. TPM's warning is that if this AI buildout ever reverses, the hole will dwarf the old dot.com bust, which is a definitely something to keep in mind. All merry go rounds stop at some point, some abruptly, others gradually, but they all stop at some point.
Source: The Next Platform, March 26, 2026. Tags: Industry, AI, HPC
Ecolab Buys CoolIT
Water treatment company Ecolab agreed to pay $4.75 billion for liquid cooling mainstay CoolIT, in a deal that closed July 2, 2026, earlier than the "later this year" timing in the original announcement. Ecolab said at the time that CoolIT was likely to generate $550 million in sales over the next 12 months; that is the buyer's projection, not an audited result. The logic is obvious for high-density AI racks: Ecolab already sells water chemistry and monitoring, and CoolIT brings the direct liquid cooling hardware. Big industrial names keep buying cooling specialists, which tells you the market thinks liquid cooling is no longer optional. Just don't treat the sales forecast as a done deal.
Source: Data Center Dynamics, March 20, 2026. Tags: Cooling, Industry
Fake Science Racket Revealed!
Northwestern says fake science is spreading faster than real science, and the researchers say they brought receipts. The linked piece is a ScienceDaily writeup of a PNAS paper by Luis Amaral, Reese Richardson, and colleagues, who used retraction records, journal metadata, and other databases to map paper mills, brokers, and compromised journals. The "faster than legitimate publications" claim is the authors' conclusion, not something I independently checked against their methods. The warning that should matter to all of us: if this literature is getting poisoned, it will also train the next round of AI models, and vendor white papers and academic citations are not automatically clean. Science has to police itself better, or we will all be citing garbage.
Source: ScienceDaily, March 7, 2026. Tags: Industry, AI
Trane Adds LiquidStack
Trane agreed to buy LiquidStack, in a deal set to close in March. Trane already held a minority stake from 2023. LiquidStack brings direct-to-chip cooling plus single-phase and two-phase immersion. That combination is what what data centers need as rack densities climb: a path from the plant to the chip, not just another CRAC. LiquidStack claims that its two-phase gear is "future-proof" and "outperforms all other immersion cooling systems" is vendor language, not an independent test. But we agree that two-phase liquid is probably more efficient than single phase, would be great to have some numbers to prove that. The purchase price was not disclosed, and Joe Capes stayed on to lead the business.
Source: Facilities Dive, February 12, 2026. Tags: Cooling, Industry, Data Center Design
Eaton Buys Boyd Thermal
Power giant Eaton bought Boyd Thermal from Goldman Sachs Asset Management for $9.5 billion, a deal announced in November 2025 and closed March 12, 2026, ahead of the original Q2 target. Boyd forecast $1.7 billion in 2026 sales, with $1.5 billion of that in liquid cooling; those are Boyd's numbers, not audited results. The portfolio includes cold plates, loops, CDUs, and single- and two-phase systems, plus a 2.3 MW CDU Boyd said can cool more than ten Nvidia NVL72 racks. For operators, the point is simpler: a major power vendor paid a fortune to own the path from the grid to the chip, because cooling is now a first-order power problem.
Source: Data Center Dynamics, November 3, 2025. Tags: Cooling, Power, Industry
"Replace junior staff with AI? Dumbest thing....."
AWS CEO Matt Garman called replacing junior staff with AI "the dumbest thing I've ever heard," and I am with him on this one. His point is practical: juniors are often your cheapest people, they already use the tools, and if you stop hiring them you will have nobody who learned the work ten years out. He also called "percent of code written by AI" a silly metric, because more lines are not better lines. He said more than 80 percent of AWS developers use AI in some way; that is his figure, not an audited survey. For infrastructure teams, the same logic applies: you still need people who can decompose a power, cooling, or cluster problem, AI assistant or not.
Source: The Register, August 21, 2025. Tags: AI, Industry
'Shiny object syndrome' and AI
Good advice from US EPA CIO Carter Farmer: AI is not a wonder drug, and shiny object syndrome is how projects go off the rails. Speaking on a FedInsider webinar, he said teams jump in because they saw what someone else did, without a specific use case, and often the problem did not need AI. He wants people to reimagine the process, not lift and shift the old workflow, and he noted the "value add" can cost two to four times doing it manually. The Register also cited a survey that only one in four AI bets pay off; that figure is theirs, not Farmer's, so I am leaving it aside. Have a use case and clean data first, then talk about GPUs and gear.
Source: The Register, July 9, 2025. Tags: AI, Industry
Chip Dive: Grace Hopper, AMD MI300A, AWS Graviton
Chips and Cheese did a deep, rambly dive into Nvidia's GH200 after the Grace Hopper intro, with comparisons to AWS Graviton 4 and a shorter look at AMD's MI300A. Chester Lam's measurements are from mid-2024 cloud instances. The useful takeaway: GH200 is not a true APU, the CPU and GPU still have separate memory pools, and NVLink-C2C's headline 900 GB/sec comes with high latency and, in his tests, some ugly hangs. The title mentions MI300A, but that part is mostly a closing comparison, not a full teardown. Still worth reading if you want to understand why packaging CPU and GPU together is harder than the slideware suggests.
Source: Chips and Cheese, July 31, 2024. Tags: Accelerators, HPC, AI
Ethernet Gang Looks to Topple InfiniBand
Very informative Next Platform article on UEC vs. InfiniBand, and why Cornelis wants to make Omni-Path speak Ultra Ethernet. Timothy Prickett Morgan's June 2024 piece is worth your time: he notes that 200 and 400 Gb/sec InfiniBand fabrics often eat 20 percent or more of cluster cost, in markets used to spending under 10 percent on the network. UEC's bet is a multi-vendor Ethernet fabric with InfiniBand-style congestion control and adaptive routing, which matters as GPU racks get denser and more expensive. Just know the dates slipped. The 1.0 spec was due in Q3 2024 and landed in June 2025 (it is now at 1.1), and Cornelis still lists Ultra Ethernet on its 800 Gb/sec CN6000 as preview.
Source: The Next Platform, June 26, 2024. Tags: Networking, HPC, Industry
Wafer-Scale Triumph on Hardcore Science AND LLM Inference
Amazing results from Cerebras on two completely unrelated workloads, at least if you take the company's numbers at face value. IEEE Spectrum reported that a WSE-2 system ran an 800,000-atom molecular dynamics problem 179 times as fast as Frontier, work done with Sandia, Livermore, and Los Alamos. Separately, with Neural Magic, a sparse Llama 7B model (70 percent zeros, accuracy recovered after extra training) did inference in one-third the time and energy of the dense model, the 66 percent reduction I originally flagged. Those are Cerebras and collaborator results on WSE-2, from mid-2024, on specific workloads, not a general "we beat every supercomputer" claim. Worth a look, just keep the caveats attached.
Source: IEEE Spectrum, June 12, 2024. Tags: Accelerators, AI, HPC