A hardware teardown by iFixit has revealed that the Trump Mobile T1 smartphone is almost identical to the HTC U24 Pro. Internal components including the Snapdragon 7 Gen 3 processor, 12 GB of RAM, and 512 GB of storage match the HTC model exactly. Comparative CT scans and physical disassembly confirmed that the mainboards are interchangeable between the two devices.

Source

Meta is shifting its hardware focus toward smart glasses, betting that wearable devices are the natural successor to smartphones for interacting with AI. The primary advantage of glasses over other wearables is their ability to see and hear from the user's perspective, facilitating "agentic" behavior where the AI proactively handles tasks like scheduling or summarizing conversations. Unlike specialized AI pins or clips, smart glasses leverage an existing consumer habit, offering utility as traditional eyewear even when the digital features are inactive.

Source

This deal drops the Dell XPS 14 to $1,749.99 and gets you a 3.2K OLED touch display, an Intel Core Ultra 7 processor, 16GB of RAM, and 1TB of storage.

UGREEN hosts a 15% discount on its NASync DH2300, a NAS enclosure that's easy to use for beginners while boasting up to 64TB of compatible storage, 4GB of RAM, and 125MB/s transfer speeds.

Microsoft is expanding support for local AI capabilities in Windows 11 to include systems equipped with dedicated Nvidia graphics cards. Previously, these features were exclusive to Copilot+ PCs, which require a dedicated Neural Processing Unit (NPU) with at least 40 TOPS of performance. This shift allows the Windows 11 local Language Model APIs to leverage the raw processing power of traditional GPUs rather than relying solely on specialized AI chips.

Source

Amazon Web Services has announced the general availability of M9g and M9gd instances powered by the new Graviton5 processor. These instances utilize the sixth generation of the AWS Nitro System and offer up to 25 percent better compute performance than their predecessors. Early testing by production users shows significant improvements in database query speeds and overall throughput for observability workloads.

Source

AMD has released preliminary performance estimates for its upcoming EPYC Venice processors, which utilize the new Zen 6 architecture. The flagship model features 256 cores and is designed to compete directly with Nvidia's Vera CPU in high-density data center environments. These initial benchmarks focus on rack-scale performance within a fixed 100kW power budget rather than individual socket metrics.

Source

In an interview with the BBC, TSMC's CFO explained that price hikes for its foundry services seem almost inevitable due to inflation. Add that to shareholder pressure as the company maxes out production, and you have a recipe for almost all tech going up further in price.

The Surface Laptop Ultra shows what Microsoft can achieve at the high end, and several of its features could make the standard Surface Laptop a stronger option for everyday users.

Google has launched Gemma 4 12B, a mid-sized open-source AI model designed to run locally on enterprise laptops with 16GB of RAM. This release bridges the gap between the smaller edge-focused models and larger high-performance versions by offering advanced reasoning in a compact footprint. The model is released under an Apache 2.0 license, making it accessible for diverse development environments and local inference pipelines.

Source

Microsoft has established a dedicated Beta channel for Windows 11 version 26H1 to provide a more stable testing environment. This release is specifically engineered as a platform-focused update to support upcoming silicon, including Qualcomm Snapdragon X2 and NVIDIA RTX Spark processors. The new structure separates the Beta channel into the 28000 build series while moving Experimental devices to the 28100 series.

Source

Dell's new XPS 13 for 2026, announced at Computex, is the new best competition for the MacBook Neo. But how exactly do the two laptops compare? I break things down to help discover where the XPS fills in the Neo's blind spots.

Google has released Gemma 4 12B, an open-weight AI model designed to run locally on consumer hardware with at least 16GB of RAM. This 12-billion parameter model is unique for its size because it natively processes text, images, and audio without requiring separate encoder modules. By utilizing an integrated architecture, the model allows standard notebooks to function as multimodal AI agents without relying on cloud-based processing.

Source

The development of Linux kernel 7.1 has reached its final testing phase with the release of the seventh candidate. Stabilization is proceeding according to schedule, and the final version is expected to be available next week. Recent updates have focused heavily on graphics subsystems, networking fixes, and various driver improvements.

Source

I had a chance to test out new gaming handhelds featuring Intel's latest Arc G3 Extreme chips. Here's what I learned (and why these chips are the new standard for mobile gaming).

Geekom, one of the world's leading proprietors of mini-PCs, is offering people a chance to buy the Geekom A6 Mini-PC, which we gave a near-perfect review, at a 20% discount thanks to a limited-time promo code.

Global governments and private enterprises are investing billions to transition quantum computing from experimental science to industrial-scale technology. The United States and United Kingdom have earmarked significant funding to establish domestic quantum foundries and scale hardware production. This shift aims to integrate quantum processing units into standard data center architectures alongside existing CPUs and GPUs.

Source

Microsoft vient d'annoncer lors de son événement Build 2026 l'arrivée de sa Surface RTX Spark Dev Box, un petit boîtier qui se pose sur le bureau et qui fait tourner des modèles IA de 120 milliards de paramètres en local, sans rien envoyer dans le cloud.

Et bien sûr derrière le badge Surface, c'est NVIDIA qui se tape tout le boulot.

Dans cette boîte noire, vous avez donc la puce NVIDIA RTX Spark, qui rassemble un GPU Blackwell et un processeur Grace pour sortir environ 1 pétaflop de puissance IA et 128 Go de mémoire unifiée.

De quoi donc faire tourner un gros modèle avec une fenêtre de contexte d'un million de tokens, ou carrément affiner (fine-tuner) un modèle sans louer des GPU dans le cloud. Le tout dans un châssis en aluminium pensé pour servir de dissipateur, donc refroidi passivement. Et un malheur n'arrivant jamais seul (je plaisante ^^), Windows 11 Pro arrive préconfiguré dessus pour les devs, avec tous les outils qui vont bien déjà installés.

D'après le site de Microsoft, ce petit joujou sera donc dispo fin 2026, aux États-Unis d'abord.

Détails du châssis

Maintenant, le truc à bien capter, c'est que cette puce RTX Spark, c'est exactement la même famille que la DGX Spark , le mini-PC que NVIDIA vend depuis octobre dernier. Même architecture Grace Blackwell, même pétaflop, mêmes 128 Go unifiés.

Eh oui, Microsoft n'a pas conçu de puce maison pour cette box (ses puces Maia, c'est pour ses datacenters), mais a juste pris la plateforme d'NVIDIA et l'a habillée en Surface avec une image Windows maison. Ce qui n'est pas grave, hein, mais autant le savoir avant de croire à une révolution Microsoft.

Côté tarif, pas de chiffre officiel encore mais les estimations tournent autour de 3500 dollars. Pour vous donner une idée, la DGX Spark d'NVIDIA, sa cousine sous Linux, est passée de 3999 à 4699 dollars récemment, la faute à la flambée des prix de la mémoire. Donc, ce ne sera pas donné, mais vous vous en fichez parce que vous êtes probablement pété de thunes ^^.

Cela dit, même si c'est cher, l'idée de faire tourner un modèle costaud entièrement chez soi, ça reste sacrément séduisant. Vos données ne sortent jamais de la machine, y'a zéro facture d'API qui gonfle à chaque requête, et vous pouvez bidouiller un fine-tuning maison tranquillement. C'est une tendance qu'on voit monter depuis un petit moment maintenant avec par exemple des gens qui glissent un GPU de datacenter dans leur PC gaming juste pour s'affranchir du cloud ^^.

Après, vous n'avez pas besoin d'attendre cette box pour faire de l'IA locale. La DGX Spark existe déjà, un Mac avec assez de mémoire unifiée encaisse de gros modèles aussi, sans oublier qu'il y'a carrément moyen de remplacer l'API d'OpenAI par votre propre Mac . Sans parler des PC AMD Strix Halo...

Non, le vrai plus de Microsoft ici, c'est le combo refroidissement passif et image Windows dev clé en main, taillé pour le futur "Windows agentique" qu'ils nous préparent, et grâce auquel les agents IA tourneront en permanence sur nos machines pour taffer à notre place.

Bref, rien de dingue, c'est certain mais ça peut clairement dépanner ceux qui veulent un PC IA local sans avoir à bricoler. J'ai hâte de connaître le prix en tout cas !

Source

Google has launched Gemma 4 QAT, a series of open-weights models optimized for local execution on hardware with limited memory. These models utilize Quantization-Aware Training to simulate lower-precision calculations during the training phase, which preserves output quality even when the model files are heavily compressed. The release includes a 12B parameter version specifically designed to run on standard enterprise laptops equipped with 16GB of RAM or unified memory.

Source