CES has long been a consumer electronics show at heart — a place for TVs, gadgets, and concept cars. But Jensen Huang’s two-hour keynote this week made it clear that NVIDIA has essentially turned CES into an AI infrastructure conference. The announcements were dense, ambitious, and consequential for anyone building or deploying AI systems.
The headline consumer product — the GeForce RTX 5090 and 5070 series based on the new Blackwell architecture — is impressive on its own. But what caught my attention were three announcements that matter more for the broader technology landscape: Project DIGITS, the Cosmos foundation models, and NVIDIA’s deepening push into physical AI and robotics.
Project DIGITS: A Personal AI Supercomputer#
The most surprising announcement, at least to me, was Project DIGITS — a desktop-sized device powered by the new GB10 Grace Blackwell Superchip that can run 200-billion-parameter AI models locally. It runs NVIDIA’s DGX OS (Linux-based), has 128GB of unified memory, up to 4TB of NVMe storage, and delivers a petaflop of AI computing performance. Two units can be linked for even larger models. Price: $3,000, shipping in May.
Let me put this in perspective. When I first started working with neural networks in the late ’90s, training even modest models required access to university computing clusters. Five years ago, running a large language model required cloud GPU instances costing thousands per month. Now NVIDIA is putting a petaflop of AI compute on your desk for the price of a high-end laptop.
This matters enormously for AI development workflows. Running inference and fine-tuning locally means faster iteration, no cloud costs, and no data leaving your premises. For startups, researchers, and independent developers, DIGITS could democratize access to the kind of AI experimentation that’s currently gatekept by cloud GPU availability and pricing.
The 200B parameter capacity is the sweet spot too — it covers running Meta’s Llama models, Mistral’s offerings, and many other open-weight models at full precision. You could realistically run a competitive AI application entirely on local hardware.
Cosmos: Foundation Models for Physical AI#
NVIDIA also announced Cosmos, a platform of world foundation models designed for physical AI development — robotics, autonomous vehicles, and industrial automation. These are generative models that understand physics and can generate synthetic environments for training robots and autonomous systems.
This is NVIDIA playing the long game. The current AI boom is primarily about language and image generation. But the next wave — the one NVIDIA is positioning for — is about AI that interacts with the physical world. Training a robot to navigate a warehouse or a vehicle to handle edge cases requires massive amounts of scenario data. Generating that synthetically is orders of magnitude cheaper and faster than collecting it in the real world.
The Cosmos models are being released as open-source under the NVIDIA Open Model License, which is a smart community-building move. By providing foundational world models freely, NVIDIA ensures that the entire physical AI ecosystem develops on their platform and hardware. The models are free; the GPUs to run them are not.
The Blackwell Architecture in Consumer GPUs#
The RTX 5090 brings the Blackwell architecture — previously reserved for data center GPUs — to consumer hardware. The standout feature is the fifth-generation Tensor Cores and a new neural rendering pipeline that NVIDIA calls “the biggest generational leap” in their history.
The interesting technical story here is how NVIDIA is using AI to make traditional computing faster. Their new DLSS 4 with Multi Frame Generation can generate up to three frames for every traditionally rendered frame, using AI models that run on the Tensor Cores. The GPU is essentially doing less traditional rasterization and more AI inference to produce the final image.
This approach — using AI acceleration to improve performance in non-AI workloads — is going to spread well beyond gaming. We’re already seeing it in video encoding, image processing, and audio. The pattern of “compute the answer approximately with traditional methods, then use AI to refine it” is becoming a general-purpose optimization strategy.
The Infrastructure Play#
Zooming out, what struck me most about the keynote was how thoroughly NVIDIA has positioned itself across every layer of the AI stack. They make the chips (Grace, Blackwell). They make the systems (DGX, HGX, DIGITS). They provide the software platform (CUDA, cuDNN, TensorRT, Triton). They’re building foundation models (Cosmos). They have the networking (Spectrum-X, NVLink). And they have the cloud partnerships (every major cloud provider).
This vertical integration is reminiscent of what made IBM dominant in enterprise computing for decades. The difference is that NVIDIA doesn’t lock you in the same way — CUDA notwithstanding — because the underlying models and many tools are open source. But the practical effect is similar: if you’re building AI, you’re almost certainly building on NVIDIA’s stack.
The competition isn’t standing still. AMD’s MI300X is gaining traction in data centers. Intel’s Gaudi accelerators are finding niches. Custom silicon from Google (TPUs), Amazon (Trainium), and others is real. But at CES, NVIDIA demonstrated why they remain the gravitational center of AI computing: they’re not just selling chips, they’re selling an ecosystem.
My Take#
I’ve attended CES presentations that felt like tech demos looking for a problem. This wasn’t one of those. Every announcement connected to a clear market need: cheaper local AI inference (DIGITS), better training data for robotics (Cosmos), and continued gaming dominance (RTX 50 series).
The DIGITS device is what I’m most excited about personally. At $3,000, it’s within reach for serious independent developers and small teams. If it delivers on the promise of running 200B parameter models locally with reasonable performance, it could shift a meaningful chunk of AI development away from cloud providers and back to local hardware. That has implications for privacy, cost, and the pace of experimentation.
Jensen’s keynote ran over two hours and could have run four. The density of announcements reflects a company that’s firing on all cylinders. Whether you’re building AI applications, deploying infrastructure, or just trying to understand where the industry is headed, NVIDIA’s CES showing is required viewing.