World Models and AI Chips: Why Physical AI Needs a Different Kind of Speed

There's a huge problem in AI, and it comes down to speed.

ChatGPT can take anywhere from two seconds to multiple hours to answer or run an action. Those answers are rarely life or death. A slow reply is an annoyance. Put that same lag in a robot or a self-driving car and the stakes flip. The car has to see, remember, predict, and act in real time under a strict power budget, because batteries only hold so much. A two-second delay while avoiding a crash is the definition of life and death.

The chips built for that job do not exist yet. Not for ever more capable self-driving cars, humanoid robots, drones, and the rest of physical AI. That is about to change. Most people outside technical circles have no idea how big the shift is.

Today's AI was shaped by large language models: ChatGPT, Claude, Grok, Gemini. They receive a sequence of tokens and predict what comes next. The basic job is still next-symbol prediction inside a stream of text.

A world model has a harder assignment. Picture a robot in a kitchen. It sees a cup on the counter. A person walks in front of the camera and the cup disappears for a second. The robot still needs to remember the cup exists. It needs to know the counter is solid, the cup can fall, the person can move unpredictably, and its own arm will change the scene if it reaches forward. Then it has to ask: what happens if I move left? What if I grab the cup from this angle? What if the person steps back into my path?

The useful output is a prediction of how the world will change, then an action, then another observation. That loop may need to run dozens of times every second without losing track of what happened five seconds or five minutes ago.

The data problem is brutal. One minute of raw uncompressed 1080p color video at 30 frames per second is roughly 11 gigabytes. One minute of ordinary spoken text written down is usually less than one kilobyte. Video is about 10 million times larger. Models compress aggressively. Even so, the physical world dumps ridiculous amounts of information before you add audio, depth, touch, joint positions, or candidate actions.

Throwing more chips at it is not enough. An AI chip is like a city. Arithmetic units are factories. Memory is warehouses. Connections are roads. Double the factories and you still get stuck if every truck is in traffic and the warehouses are full. Expensive machinery sits idle waiting for data.

That bottleneck already shows up in language models. Nvidia said in the technical description of Rubin that the generation stage of modern AI inference is fundamentally constrained by the memory system. World models add time as a hard variable. The model may need a memory of the room, a history of motion, object representations, and several possible futures.

There is no single settled architecture. Google DeepMind's Genie 3 is a generative world model. DeepMind says it can create an interactive 720p environment at 24 frames per second while keeping consistency for a few minutes. Return to a place after a minute and the system still needs relevant information from the earlier visit while responding to new input in real time.

Meta's V-JEPA 2 has 1.2 billion parameters and was trained on more than one million hours of video. Instead of redrawing every pixel of the next frame, it predicts a compressed embedding, an internal summary of what matters in the scene. Meta then used about 62 hours of robot data so those predictions respond to actions. The model could plan basic robot movements in environments and with objects it had not seen in training.

Nvidia's Cosmos takes another route. It combines language, image, video, audio, and actions in one system using a mix of prediction and diffusion. Diffusion starts from noise and refines it into a coherent result.

So we already have at least three versions of the future. One generates the visible world. One predicts a compact representation. One tries to connect many information types and actions in a shared system. They do not even agree on the basic unit of intelligence. Is it a pixel, a video token, an object, a persistent three-dimensional map? If you cannot decide the unit, what should a chip accelerate? That disagreement is why the most important chip does not exist yet. Specializing silicon only pays off when the workload becomes predictable. We are not there.

GPUs stay valuable in this phase because they are flexible. Data centers will still train world models, generate synthetic data, run simulations, and help discover which architecture works. Once a workload is large, stable, and repetitive, a purpose-built chip can drop hardware the job does not need and spend more of its power budget on the parts it uses constantly. That is when a general GPU can lose on cost, power, or latency.

We already saw the pattern in the language-model boom. Google built TPUs. Amazon built Trainium and Inferentia. Meta built its own accelerators. Tesla designed inference chips for vehicles. Broadcom helps major AI companies create custom silicon. World models create the same incentive.

The end state likely needs two different computers. The first lives in the data center. It trains on enormous video datasets, generates synthetic environments, and simulates millions of situations. It needs huge memory capacity, huge bandwidth, fast chip-to-chip links, and flexibility for architectures that keep changing. The second lives inside the machine: car, humanoid, factory robot, drone, glasses. It must absorb streams from multiple sensors, update its internal model, choose an action with predictable latency, and stay inside strict power and heat limits.

Nvidia's products already show how different those jobs are. Rubin is a data-center monster: 288 GB of HBM4 stacked close to the processor, up to 22 terabytes per second of bandwidth, plus a serious high-speed fabric to other GPUs. Twenty-two terabytes per second is roughly 2,000 times the data in one full minute of that uncompressed 1080p video, processed every second. On the on-device side, Nvidia's Jetson Thor has up to 128 GB of low-power memory, about 273 GB per second of bandwidth, high-speed sensor connections, and a power range of 40 to 130 watts. Far less memory and bandwidth, because robots do not have data-center power.

Memory pressure keeps rising. The jump from Nvidia Blackwell's roughly 8 TB per second of memory bandwidth to Rubin's 22 TB per second is about 2.8 times in one generation. AMD's MI455X has 432 GB of HBM4 and up to 23 terabytes per second. Google's Ironwood TPU connects thousands of chips while giving each one 192 GB of HBM and 7.4 TB per second. Moving data is now a central constraint.

On the edge, consistent response time, low power, sensor interfaces, safety checks, and enough local memory when the internet is gone all matter. Tesla is a live case study. Its Full Self-Driving stack now supports coast-to-coast drives without interventions for millions of customers in the United States. Public materials stress performance per watt, high-frequency sensor data, memory latency, and work split across multiple processors. Tesla has also said AI5 and AI6 inference chips for humanoid robots and AI compute satellites, with production planned for 2027 and 2028, are progressing. We do not know if those chips become the best world-model processors. We do know Tesla has an incentive to co-design the model, the chip, the cameras, the vehicle, and the robot as one product.

Some companies that barely show up in AI chatter may capture a lot of value. Samsung, SK Hynix, and Micron make the high-bandwidth memory these systems need. Samsung began shipping HBM4 in 2026 and is working with AMD on memory for the MI455X. Micron is producing HBM4 for Nvidia's Rubin. SK Hynix says its HBM4 doubles bandwidth and improves power efficiency versus the prior generation. TSMC matters just as much, because the processor and memory have to sit close together through complex packaging.

From where I sit, and this is not investment advice, the winners fall into a few groups. Full-stack incumbents that can adapt before the architecture settles, with Nvidia as the clearest example. Vertically integrated companies that own both the use case and custom silicon, with Google and Tesla as the obvious names. Memory, manufacturing, packaging, and networking suppliers that get paid no matter which model architecture wins. Keep an eye on startups like Cerebras chasing wafer-scale chips and local memory. That could matter if world models reward low communication overhead.

Today's chips are the expensive laboratory where the architecture gets discovered. The specialized winner comes later, when researchers know what must be remembered, what can be thrown away, and which operations repeat often enough to hardwire.

When these models move into machines, the economic scale gets absurd: self-driving cars that run for years at a fraction of the cost of human drivers, humanoid robots that execute human tasks faster and cheaper. It can feel like AI is nearing an end if you only watch chatbots. The physical wave is just getting started.

Check the video here.