Nvidia just unveiled Project Digits at CES 2025, a $3,000 personal AI supercomputer powered by its new GB10 Grace Blackwell Superchip. The compact workstation packs 128GB of memory, letting researchers and developers run models with up to 200 billion parameters right on their desks—or link two together for 405-billion-parameter beasts. It begins shipping in May.
NVIDIA just unveiled its GeForce RTX 50 series GPUs at CES, doubling generative AI performance with FP4 support. The new quantization cuts model memory needs in half, allowing models like FLUX.1 to run locally on under 10GB of VRAM in just five seconds. The flagship RTX 5090 launches January 30.
AMD just previewed its next-gen Radeon RX 9070 and 9070 XT GPUs at CES, built on a new 4nm RDNA 4 architecture. The biggest shift is FSR 4, which now relies on machine learning upscaling powered by dedicated on-die AI accelerators. Both cards drop in Q1, though specs and prices remain under wraps.
Qualcomm just expanded its Snapdragon X lineup with a cheaper Arm laptop chip, bringing Copilot Plus PCs down to around $600. It features an eight-core Oryon CPU and a 45 TOPS NPU to keep all of Microsoft's on-device AI features intact. Why it matters: Qualcomm is aggressively taking Windows on Arm mainstream to squeeze Intel where it hurts.
Intel just unveiled a massive PC chip refresh at CES 2025, led by its Core Ultra 200 series across desktop and mobile. The new silicon features dedicated NPUs, AI-based power management, and up to 3.3x faster local Llama 3 8B performance on the 285H. Systems begin shipping later this month.
Elon Musk just revealed that xAI's upcoming Grok 3 has finished pretraining, using 10 times more compute than Grok 2. He added that the new flagship model is coming soon. Why it matters: xAI is rapidly scaling its infrastructure to challenge OpenAI and Google at the very top of frontier AI.
xAI just closed a massive $6 billion Series C funding round, pushing its valuation to $45 billion. Backed by Nvidia, AMD, and Sequoia, the fresh cash will help double its 100,000-GPU supercluster and accelerate the rollout of its next-generation Grok models.
Google just made Trillium, its sixth-generation TPU, generally available on Google Cloud. The chip powers Google's flagship Gemini 2.0 model, delivering over 4x faster training and up to 3x higher inference throughput compared to the prior-gen TPU v5e. Why it matters: Google is ramping up the pressure on Nvidia by giving developers a massive, cost-effective alternative for training giant models.
IBM just unveiled a breakthrough in co-packaged optics (CPO), introducing the first polymer optical waveguide prototype to replace copper wires inside data centers. The tech can slash standard LLM training times from three months to three weeks while cutting power consumption fivefold. It has also passed all key manufacturing stress tests.
Microsoft just unveiled a next-generation datacenter design that consumes zero water for cooling, saving over 125 million liters per site each year. Built for heavy AI workloads, it relies on chip-level closed-loop liquid cooling. The architecture has been standard for all new builds since August 2024, with pilots launching in 2026 before coming online in late 2027.
Microsoft AI CEO Mustafa Suleyman just pushed back against claims that AGI is achievable on current hardware like Nvidia GB200s, estimating it will take two to five more chip generations—or five to ten years. He also distanced practical AGI from sci-fi superintelligence, arguing builders should focus on useful, accountable AI companions instead of chasing the singularity.
Hugging Face just made 83 popular open models available directly within Amazon Bedrock Marketplace. The models are managed via SageMaker JumpStart under the hood, allowing developers to plug open weights into Bedrock features like Agents, Knowledge Bases, and Guardrails with standard Bedrock API calls.
AMD just released and open-sourced Nitro Diffusion, a pair of one-step text-to-image models based on Stable Diffusion 2.1 and PixArt-Sigma. Trained on AMD Instinct MI250 accelerators, they match full-step visual quality in a single inference step for both data centers and edge PCs. Code and weights are now up on Hugging Face and GitHub.
Google just unveiled Willow, a state-of-the-art quantum chip that cuts errors exponentially as physical qubits scale up. It cracked a nearly 30-year challenge in quantum error correction and finished a benchmark in under five minutes that would take supercomputers 10 septillion years. Why it matters: It proves large-scale, fault-tolerant quantum computers are actually buildable.
The world’s second-fastest supercomputer, Frontier, just simulated the observable universe at an unprecedented scale using 9,000 AMD-powered nodes. By calculating dark matter and gas dynamics rather than simple gravity, the 1.4-exaFLOP system completed hydrodynamic modeling roughly 300 times faster than expected. Why it matters: Exascale computing is smashing scientific simulation ceilings, setting up massive compute gains for future AI and climate models.
Google DeepMind just published GenCast, an open-source diffusion ensemble model for 15-day global weather forecasts. Trained on ECMWF ERA5 data, it beat top operational system ENS on 97.2% of targets while running in just 8 minutes on a single TPU. It marks a huge leap for extreme weather and renewable energy forecasting.
Apple just confirmed it uses Amazon's Inferentia and Graviton chips for services like search, delivering a 40% efficiency boost. The company is now evaluating AWS's new Trainium2 chips to pretrain models like Apple Intelligence, eyeing up to a 50% efficiency gain and signaling alternatives to Nvidia's pricey dominance.
Amazon is teaming up with Anthropic to build Project Rainer, a massive AI supercomputer powered by hundreds of thousands of its custom Trainium 2 chips. AWS says the system is 30% to 40% cheaper than Nvidia setups and five times larger than Anthropic’s current training cluster. Why it matters: Amazon is proving Nvidia isn't the only viable game in town, giving frontier AI builders a much cheaper route to compute.
Researchers from the Allen Institute for AI and UC San Diego just introduced Spherical DYffusion, cutting a 10-year global climate simulation from 78 hours to under 3 hours. The model adapts diffusion architectures directly to spherical geometry to emulate the physics-based FV3GFS model, trading higher polar biases for massive compute and energy savings.
Apple just placed orders with TSMC for its upcoming M5 chip, targeting production development ahead of late 2025. Apple is skipping the pricier 2nm node in favor of advanced 3nm tech with 3D chip stacking. Crucially, Apple plans to deploy M5 across future hardware and its own cloud AI servers.
