The technology stories enterprise IT leaders actually read.
Volume 343, Issue 2 · 8 articles from 3 issues
Hands-on comparison of 10 AI text-to-speech tools, including ElevenLabs, Speechify, Murf and Amazon Polly, by use case and quality.
Techiexpert tested ten AI text-to-speech tools on voice quality, pronunciation, editing, language support and free plans: ElevenLabs, Speechify, Murf, NaturalReader, TTSMaker, PlayHT, Amazon Polly, Google Cloud TTS, Descript and WellSaid Labs.
ElevenLabs is rated best for overall voice quality with 70+ languages, Speechify for mobile reading, Murf for business teams and NaturalReader for documents.
TTSMaker is a free browser tool, Amazon Polly and Google Cloud TTS suit developers, Descript fits video and podcasts, and WellSaid Labs targets studio narration.
StorageReview ranks PCIe data center GPUs for enterprise inference using one vLLM container and model set.
StorageReview's buyer's guide covers add-in PCIe accelerators that fit the 2U and 4U servers most data centers already run, rather than HGX or OAM modules.
Every card was tested in the same NGC vLLM container and driver, mostly in a Dell PowerEdge R770 at the power that server supplies, at 1 to 32 concurrent streams across four models.
The key finding is that memory matters more than compute, as 24GB and 32GB cards run out of KV cache on long prompts.
Meanwhile, 48GB and 96GB cards support multi-tenant concurrency, and 94GB and 141GB HBM cards keep 70B to 120B models on one GPU.
Q2 2026 server revenue hit $166.3 billion, up 52% year-over-year, as GPU and XPU systems capture most of the spend.
Server market revenue reached $166.32 billion in Q2 2026, a 52% year-over-year jump, led by GPU-accelerated systems that generated 52.5% of revenue despite a 10.7% drop in unit shipments, signaling a shift to higher-end configurations.
XPU systems surged 3.4x to $27.5 billion as cloud providers push custom AI inference chips. Accelerated machines are just 16.1% of units shipped but capture 69.1% of revenue, underscoring the market's pivot to AI infrastructure.
Original design manufacturers earned $89.66 billion collectively, with Dell notably expanding share during the boom.
Physicists showed non-Abelian anyons can perform all operations needed for universal quantum computing.
Researchers from the University of Chicago, Harvard, Stony Brook, and Quantinuum created and tested non-Abelian anyons on a 54-qubit processor, combining braiding and fusion operations to achieve universal quantum computation.
This approach may bypass the costly magic state distillation process usually required for quantum error correction, potentially making fault-tolerant quantum computers more practical.
The team encoded information in topological qutrits and showed the method could directly produce magic states through topological operations.
It is a proof of principle; next steps involve integrating the technique with active error correction for larger-scale systems.
AI makes compute, memory, networking, power and cooling a tightly coupled system that must be planned together.
For decades, processors, memory, networking and facilities evolved on separate roadmaps and could be upgraded on different timelines.
This opinion piece argues that AI has ended that model, because accelerator performance depends on the interaction of memory, fabrics, packaging, power delivery and cooling.
Since training and inference are limited by data movement across the platform, idle accelerators mean stranded capital and wasted energy.
For colocation providers, facility roadmaps may increasingly follow customer accelerator deployments, requiring earlier investment in power distribution, liquid cooling and high-density racks.
Memory has eclipsed compute as the control point of the IT industry in the AI era, per Gartner chip revenue data.
The Next Platform's Timothy Prickett Morgan argues that for decades CPUs and later GPUs were the focal point of systems, with memory, storage and I/O treated as secondary and often skimped on.
Machine learning, generative AI and agentic AI have moved memory front and center, eclipsing compute as the control point in most of the IT sector with surprising speed.
Using Gartner's breakdown of memory versus non-memory chip revenue, he notes the semiconductor market nearly doubled between 2025 and 2026, with growth forecast to slow in 2027, and estimates the split between DRAM and NAND flash.
A spec comparison of AI accelerators from NVIDIA, AMD, Google, Intel, AWS and Cerebras covering compute, memory and interconnect.
Analytics Insight compares leading AI processors to help buyers match hardware to workloads instead of marketing claims.
NVIDIA's B200 has 208 billion transistors, 192GB of HBM3e at 8TB/s and 1.8TB/s NVLink, with a GB200 NVL72 rack reaching 720 petaFLOPS at FP4, while AMD's MI325X and MI355X lead on memory with up to 288GB of HBM3E.
Google's TPU v7 Ironwood delivers 4,614 TFLOPS at FP8 per chip and scales to 9,216-chip pods.
The guide also covers Intel, AWS Trainium and Cerebras, noting that memory capacity, bandwidth and interconnect matter as much as raw compute for large models.
A primer on how chiplet-based CPUs change server density, power efficiency and data center infrastructure planning.
CPUs have traditionally been single pieces of silicon, with servers and data centers designed around that model, whereas chiplets combine small modular dies for compute, I/O, cache, memory connectivity or acceleration into one package.
The article explains the benefits: higher manufacturing yields and lower costs, design freedom to tailor processors to workloads, potentially higher compute density without larger servers, and some supply chain flexibility.
It also notes that advanced packaging brings its own supply chain risks.
Data center planners should consider how chiplet designs affect density and power.
More from Technology · All IT section newsletters · This week