We have officially entered the era where “I didn’t know it was going to rain” is becoming a much harder excuse to defend.

Google DeepMind and Google Research have introduced WeatherNext 3, Google’s latest AI weather model and its most advanced global forecasting system to date. Rather than simply making an existing forecast algorithm a little faster, WeatherNext 3 represents a more fundamental change in how artificial intelligence can observe the atmosphere, interpret rapidly changing conditions and turn that information into forecasts people can actually use.

According to Google’s detailed [WeatherNext 3 announcement] the model combines live global satellite observations with machine learning to produce updated forecasts every hour. Selected surface variables can be modeled at resolutions as fine as approximately five kilometers, giving the system a much more detailed view of local conditions than its predecessor. Google says WeatherNext 3 is roughly five times sharper overall than WeatherNext 2, which generally operated on a 25-kilometer grid with six-hour forecast intervals.

That might sound like an impressive collection of meteorological statistics. But the practical implication is much simpler: Google wants AI weather forecasting to become more local, more current and much better at answering the question everyone actually asks.

Is it going to rain where I am?

What Exactly Is WeatherNext 3?

Traditional numerical weather prediction is one of the most computationally demanding scientific tasks on Earth. Meteorological agencies feed observations from satellites, weather stations, aircraft, radar systems, ocean buoys and other instruments into enormous physics-based models. Supercomputers then solve equations describing how pressure, moisture, temperature and wind are expected to evolve.

These systems remain essential to modern meteorology, but they can require substantial computing resources and time.

AI weather models approach the challenge differently. Instead of explicitly calculating every atmospheric interaction from first-principles physics, machine-learning systems learn patterns from enormous quantities of historical and observational weather data.

WeatherNext 3 pushes that idea further by incorporating a mosaic of live geostationary satellite observations into its forecasting process. That matters because many earlier AI forecasting systems depended heavily on numerical weather prediction analyses that could arrive with several hours of delay.

Weather has the inconvenient habit of changing while your computer is calculating it.

WeatherNext 3 can instead incorporate newer observations and generate another forecast every hour, allowing the model to respond more quickly when storms, fronts or rainfall systems begin developing.

The Big Breakthrough: Better Rain and Snow Forecasts

Ask a meteorologist what is particularly difficult to forecast and precipitation will quickly enter the conversation.

Rainfall and snowfall depend on small-scale cloud processes, atmospheric instability, moisture and rapidly changing local conditions. Global models frequently have to represent these processes across relatively large grid cells, which can make precipitation appear blurred or place rainfall slightly away from where it eventually occurs.

WeatherNext 3 was designed specifically to improve this weakness.

Google trained parts of the system using high-quality precipitation datasets including NASA’s Integrated Multi-satellite Retrievals for GPM, better known as IMERG, alongside Google’s own satellite-radar precipitation reanalysis.

Google reports significant improvements across several precipitation evaluation metrics. In medium-range testing, the company reports improvements in Continuous Ranked Probability Score of up to 60% against IMERG observations, alongside improvements measured against U.S. radar and rain-gauge datasets. For consumer-facing forecasts made at least a day ahead, Google says people could see precipitation forecasts that are up to 50% more accurate in some circumstances.

The words “up to” matter. Weather is chaotic, forecast accuracy varies by region, lead time and measurement method, and no model suddenly makes precipitation perfectly predictable.

Still, better rain forecasting could have a surprisingly large impact.

[TechCrunch’s report on WeatherNext 3] notes that rain prediction, coarse spatial resolution and dependence on processed government datasets have historically been persistent weaknesses for AI weather models. WeatherNext 3 is designed to attack all three problems at once.

So yes, the umbrella joke is funny. The underlying technology is considerably more serious.

Five-Kilometer Detail Changes What a “Global Forecast” Can Mean

WeatherNext 3 also increases forecasting resolution.

Selected surface measurements, including station-calibrated temperature and dew point, can reach roughly five-kilometer resolution. Core gridded surface variables can operate closer to approximately ten kilometers, while other atmospheric variables are modeled at broader resolutions. In other words, WeatherNext 3 is not universally a five-kilometer forecasting system, but it can produce dramatically finer detail for important variables than WeatherNext 2.

Why does that matter?

Consider a coastal city, mountain valley or densely populated metropolitan region. Conditions separated by only a few kilometers can be surprisingly different. Mountains alter wind and precipitation. Coastlines influence temperature and humidity. Urban areas produce their own heat effects.

A 25-kilometer grid can smooth away some of those differences.

A five-kilometer-scale prediction can reveal much more of them.

[The Verge’s coverage of WeatherNext 3] highlights this combination of higher resolution and hourly satellite-informed updates, particularly for regions that historically have had fewer ground-based weather instruments.

That could make the technology especially important across parts of Africa, Latin America and Asia-Pacific, where building dense radar and meteorological infrastructure can be extremely expensive. Google says observationally driven AI forecasting could help bring higher-quality localized weather information to areas that have traditionally had less forecasting infrastructure.

WeatherNext 3 Is Coming Directly to Google Products

WeatherNext 3 is not being kept inside a research lab.

Google says its forecasts are being incorporated into Google Search, Gemini, Google Maps, Google Maps Platform and Google Earth Engine, while developers and researchers can also work with WeatherNext capabilities through Google’s cloud ecosystem.

That distribution may ultimately matter almost as much as the underlying model.

A meteorological breakthrough does not help many ordinary users when accessing it requires specialist software and atmospheric-science expertise. Putting AI forecasts inside Search, Maps and Gemini potentially puts the output in front of billions of users through interfaces they already understand.

Imagine asking Gemini whether an outdoor meeting should be moved indoors, checking Maps before driving into an approaching storm, or searching the weather for an upcoming weekend trip and receiving a more accurate precipitation outlook.

AI Is Also Getting Better at Forecasting Hurricanes

WeatherNext 3 did not appear from nowhere.

Google DeepMind has spent years developing AI systems for different parts of weather forecasting, including GraphCast, GenCast and specialized cyclone-prediction capabilities.

In August 2026, Google DeepMind published additional research showing how its WeatherNext technology could improve predictions of tropical-cyclone tracks, intensity and wind structure. The company said the system could potentially provide forecasters with roughly an additional day of useful warning in some circumstances. Readers interested in the severe-weather side of Google’s research can explore DeepMind’s

Extra warning time is not merely a more convenient forecast.

For hurricanes and typhoons, additional preparation time can influence evacuation planning, emergency logistics, transportation, business continuity and disaster response.

This is where AI forecasting moves from useful consumer technology into potentially life-saving infrastructure.

But Physics-Based Weather Models Are Not Going Away

It is tempting to frame AI weather prediction as an “AI versus traditional meteorologists” competition.

Reality is more interesting.

WeatherNext 3 still exists within a forecasting ecosystem built on decades of atmospheric science, observational infrastructure and numerical modeling. Weather agencies also rarely rely on a single forecast. Professional meteorologists compare multiple models, ensembles, observations and expert analyses before issuing official warnings.

Google itself explicitly tells users to rely on their local meteorological agency or national weather service for official forecasts, severe-weather warnings and public-safety advisories.

The likely future, therefore, is not AI replacing physics.

It is AI working alongside physics-based forecasting, observations and human expertise.

Google Research’s work on [NeuralGCM and AI-enhanced global precipitation modeling] illustrates this hybrid direction particularly well. NeuralGCM combines machine learning with physical modeling rather than treating the two approaches as mutually exclusive.

The most powerful future forecasting systems may be the ones that know when to use physics, when to use learned patterns and how to combine both.

Weather AI Also Raises Questions About Trust and Governance

As AI becomes integrated into systems that influence public safety and major economic decisions, accuracy cannot be the only consideration.

Organizations using AI-generated environmental predictions may also need to think about model transparency, uncertainty, validation, accountability and human oversight.

A consumer deciding whether to bring an umbrella can tolerate a wrong prediction.

An emergency-management agency, electric-grid operator or agricultural business may face significantly greater consequences.

That distinction makes responsible deployment essential. Forecasts should communicate uncertainty rather than creating false confidence, critical decisions should maintain human review, and organizations should understand the limitations of the data and models they rely on.

Conclusion

Google’s latest AI weather model, WeatherNext 3, shows how artificial intelligence is becoming useful in everyday life, not just in chatbots and business tools. By using live satellite data, updating forecasts more often and improving predictions for rain, snow and severe weather, the model could help people and businesses make better decisions.

The biggest benefit is simple: more accurate and more local weather forecasts can help us plan ahead. That could mean remembering an umbrella, changing travel plans, protecting crops, managing energy use or preparing earlier for dangerous storms.

WeatherNext 3 is not a replacement for meteorologists or official weather agencies. Instead, it is another powerful tool that can support experts and improve the information available to the public.

Artificial intelligence may live in the cloud, but keeping that cloud alive involves an enormous amount of very physical work.

Behind every AI model, recommendation engine, advertising system and digital assistant are rows of servers connected by thousands of cables, power systems, cooling equipment and networking components. When something goes wrong, a human technician may need to locate the problem, replace a cable, reset a server, inspect equipment or physically move hardware.

Meta now wants robots to help.

According to a recent WIRED investigation, Meta is testing several robotic systems inside its data centers that could eventually perform tasks ranging from plugging in networking cables and restarting servers to moving equipment and conducting inspections. The experiments offer an intriguing glimpse into the next stage of data center automation, where artificial intelligence is no longer simply running inside the servers—it is beginning to operate the physical infrastructure around them. [WIRED]

For businesses following the development of physical AI, robotics and hyperscale computing, Meta’s experiments matter for a much bigger reason. They could provide an early model for how intelligent machines will be introduced into some of the world’s most valuable and technically demanding workplaces.

Why Meta Is Experimenting With Robots Now

Timing is everything.

The AI boom is forcing major technology companies to build computing infrastructure at a scale that would have sounded extravagant only a few years ago. Meta has been aggressively expanding its AI infrastructure, including servers, specialized accelerators and new data centers. In 2026, the company raised its capital-spending outlook to roughly $130 billion to $145 billion as its AI ambitions continued to increase.

Meta itself says that it has broken ground on ten data centers during the previous 24 months as it expands its fleet of AI-optimized facilities. Its useful explainer on how these facilities work provides context for just how much physical infrastructure is required to support seemingly intangible digital services.

The larger the infrastructure footprint becomes, the larger the operational challenge becomes as well.

A hyperscale data center must be monitored constantly. Components fail. Cables loosen. Servers need to be reset. Hardware must be moved. Inventory has to be tracked. Problems can happen at 3 a.m. just as easily as 3 p.m.

That makes data centers an attractive proving ground for robots.

What Are Meta’s Data Center Robots Actually Doing?

This is not yet a scene from a science-fiction movie where humanoid machines independently run an entire facility.

Meta appears to be taking a more pragmatic approach: automate individual physical tasks where robotics can create measurable operational value.

WIRED reports that Meta has tested technology from Watney Robotics, Kinova and ABB. One experiment involves evaluating whether a Kinova Gen3 robotic arm could perform server power-cycling tasks. Another robotic system is being tested for replacing networking cables. Meta has also deployed simpler mechanisms capable of physically pressing buttons on machines when remotely instructed.

The company already has experience with less sophisticated automation. Meta has used self-driving tugger robots to transport heavy server racks and wheeled robots equipped with barcode readers to help track inventory. Those systems are reportedly operating at multiple Meta data centers, including facilities in Iowa and Virginia.

At Meta’s Altoona, Iowa campus, the company has reportedly been testing two dual-armed Watney robots for cabling work. The machines remain supervised by people and are not yet as fast as human technicians, illustrating an important distinction between an impressive robotics demonstration and a production-ready autonomous workforce.

Meanwhile, Meta is reportedly testing ABB-built mobile robots at its Prometheus data center campus in New Albany, Ohio. These machines combine a wheeled platform, an adjustable lifting mechanism and a robotic arm and are being evaluated for jobs such as reseating components.

The direction is clear: rather than building one magical robot capable of everything, Meta appears to be exploring a collection of robotic systems designed for narrowly defined operational tasks.

Physical AI Is Making These Robots More Capable

Robotic arms have existed for decades. So why is this happening now?

Artificial intelligence is changing what robots can understand and how easily they can adapt.

Traditional industrial robots typically perform carefully programmed motions inside highly structured environments. Modern physical AI systems can combine cameras, sensors, machine learning and increasingly sophisticated reasoning models to understand their surroundings and select actions.

NVIDIA describes this next generation of physical AI as systems that need to perceive the world, reason about what is happening, predict what may happen next and generate an appropriate physical action. Its recent Cosmos 3 work demonstrates how world models are increasingly being designed to support robotics, warehouse automation and other physical-AI applications. [NVIDIA]

But Data Centers Are Surprisingly Difficult Places for Robots

There is a catch.

Data centers were designed for people—not robots.

A cable connector that takes an experienced technician seconds to manipulate may be extremely challenging for a robot. Small variations in cable position, resistance, rack design and component layout can create difficult perception and dexterity problems.

WIRED reports that Meta’s existing inventory robot has encountered basic obstacles. Its grayscale camera reportedly cannot distinguish between certain red and green equipment indicators, meaning people are still required for some inspections. The machine can also struggle with corners and cables lying in its path. Battery charging creates additional downtime.

More advanced AI hardware can make the physical challenge even tougher. Dense GPU systems contain enormous amounts of cabling, connectors and tightly packed components. Meta has reportedly acknowledged internally that current robots are not ready for some of the intensive cabling associated with systems such as NVIDIA’s GB300 infrastructure.

This problem is known throughout the robotics industry as the gap between controlled demonstrations and messy reality.

ABB has written extensively about the “sim-to-real” gap—the difficulty of ensuring a robot trained or validated in simulation behaves reliably once it enters a real physical environment. [ABB]

AWS makes a similar argument in its Physical AI Blog. A 2026 simulation-first robotics demonstration showed how organizations can use digital environments to validate automated workflows before deploying them onto physical machines, reducing the cost and risk associated with learning exclusively in the real world.

For Meta, digital twins and high-quality simulation could eventually become nearly as important as the robots themselves.

The Real Opportunity: Faster Maintenance and Less Downtime

The most compelling argument for Meta data center robots may not simply be lower labor costs.

It is uptime.

Modern AI infrastructure is enormously expensive. When valuable servers, GPUs or networking systems are unavailable, those assets are not producing useful computing capacity.

Imagine an automated system detecting that a server has stopped responding. Instead of waiting for a technician to receive an alert, travel to the correct building, find the rack and inspect it, a nearby robot could potentially arrive automatically, visually inspect the hardware, perform an approved reset and report the result.

If the problem were more complicated, it could escalate the issue to a human.

That model turns robots into physical extensions of automated infrastructure-management software.

AWS has previously explained how cloud-connected robots can continuously collect operational data, feed monitoring systems, receive updates and interact with other machines and software. That architecture could become particularly valuable in data centers, where practically every physical action can be connected to telemetry and asset-management systems.

The future data-center technician may therefore work alongside a fleet-management platform rather than simply carrying a toolbox.

Dexterity Could Be the Breakthrough That Changes Everything

Moving equipment is comparatively easy.

Manipulating delicate components is where robotics becomes genuinely interesting.

Plugging a cable into the correct port requires perception, positioning, force control and precise hand-eye coordination. A robot has to identify the right connector, align it correctly, apply enough force to make the connection—but not enough to damage expensive equipment—and verify that the task actually succeeded.

Research across physical AI is increasingly focused on solving exactly these kinds of problems.

AWS recently highlighted how robotics developers are training systems using large quantities of real operational data so that machines can perform increasingly dexterous manipulation tasks.

What About Safety and Cybersecurity?

Putting intelligent machines beside multimillion-dollar racks of computing equipment creates another obvious requirement: robots cannot be allowed to improvise recklessly.

A mistaken physical action could disconnect the wrong server, damage networking equipment or interrupt critical infrastructure.

That means physical AI needs multiple layers of safety.

Robots will require collision detection, safe operating zones, authenticated instructions, access controls, detailed activity logs and mechanisms allowing people to stop or override an operation.

NVIDIA’s work on functional safety for robotics demonstrates how seriously the industry is beginning to treat this issue. Its Halos robotics architecture combines hardware and software safety mechanisms designed for autonomous machines working in environments shared with people. [NVIDIA]

Cybersecurity becomes equally important. A robot capable of physically interacting with servers effectively becomes another privileged endpoint on the network. Organizations must determine who can instruct it, what systems it can access, how commands are authenticated and what happens when connectivity is lost.

In other words, the robot may have arms—but enterprises should govern it more like critical infrastructure.

Will Robots Replace Data Center Workers?

This is likely to become the most closely watched part of Meta’s experiment.

One Meta worker cited by WIRED estimated that a successful cable-swapping robot could potentially replace as much as 80 percent of some workers’ workloads. That is an individual estimate, not a confirmed company projection, and widespread autonomous deployment remains far from guaranteed.

Meta, for its part, argues that it still needs more skilled workers rather than fewer and has launched programs intended to train people for electrical, mechanical and plumbing jobs associated with its infrastructure expansion.

Both developments can be true simultaneously.

Data-center employment may continue growing because the industry itself is expanding rapidly, while the composition of individual jobs changes because robots assume some repetitive tasks.

Technicians may increasingly supervise robotic fleets, perform complicated troubleshooting, maintain automation systems, analyze anomalies and handle tasks machines cannot reliably complete.

The smarter question is therefore not simply, “Will robots take jobs?”

It is: Which tasks will be automated, which jobs will change, and what new skills will workers need?

That distinction is especially important for communities considering subsidies or incentives for massive data-center projects. If employment benefits are part of the public justification for those investments, governments and companies will need increasingly transparent conversations about how automation affects long-term workforce requirements.

Conclusion: Meta’s Robot Experiment Could Redefine Data Center Operations

Meta’s push to bring robots into data centers is about much more than automating a few repetitive maintenance tasks. It signals a broader shift toward AI-powered infrastructure where software intelligence and physical automation work together to keep critical computing systems running efficiently.

For now, human technicians remain essential. Today’s robots still face real limitations in dexterity, navigation, reliability, safety and their ability to handle unexpected situations. But as physical AI improves, these machines could increasingly take responsibility for routine inspections, hardware movement, cable replacement, server resets and other predictable tasks that currently require manual intervention.

The bigger transformation may come when data centers themselves are redesigned around automation. Robot-friendly racks, standardized connectors, digital twins, autonomous monitoring systems and AI-driven maintenance platforms could dramatically change how future facilities are built and operated.

That makes Meta’s experiments worth watching far beyond the walls of its own data centers. They offer an early look at how robotics could reshape AI infrastructure, workforce skills and enterprise operations across the technology industry.

The irony is hard to miss: AI is creating unprecedented demand for data centers, and AI-powered robots may eventually help maintain the very infrastructure that keeps AI running.

As investment in artificial intelligence continues to accelerate, the next major breakthrough may not simply be a smarter model. It could be a smarter physical environment capable of monitoring, maintaining and increasingly managing itself.

Artificial intelligence has spent the last few years getting remarkably good at living inside computers. It can write code, analyze documents, search databases, operate software and increasingly complete multi-step digital tasks with limited supervision.

Anthropic now wants AI agents to reach beyond the screen.

Anthropic unveiled the Model Hardware Standard (MHS), a new framework designed to let AI agents communicate with and operate programmable physical equipment. The early research preview targets scientific laboratories and advanced manufacturing environments, where an AI agent could potentially coordinate microscopes, liquid-handling systems, robotic arms and other instruments as part of one autonomous workflow.

That may sound like a small change in plumbing. It is not.

The Anthropic Model Hardware Standard could represent an important bridge between agentic AI and physical AI—turning models that can reason about a task into systems capable of initiating actions in laboratories, factories and other instrument-heavy environments.

Anthropic describes MHS in its [ANTHTROPIC] as a shared specification for safely connecting AI agents with physical devices. [Reuters] similarly notes that the framework is initially being tested with selected partners before a planned open-source release.

What Is Anthropic’s Model Hardware Standard?

The easiest way to understand MHS is to think about the integration problem facing modern laboratories and manufacturing facilities.

A laboratory may contain instruments from several manufacturers, each with its own software, commands, APIs and communication methods. Connecting those devices into one automated workflow often requires specialized engineering and custom integrations.

Anthropic says those integrations can take weeks or even months. MHS is designed to reduce some of that work to hours or minutes by creating a more standardized way for software agents to understand what a device can do and communicate with it.

It is essentially an interoperability layer.

Instead of building a completely different AI integration for every microscope, robotic arm or liquid-handling machine, developers could potentially expose a standardized collection of capabilities that an agent can discover and use.

This matters because the future of physical AI will not be built around one magical robot doing everything. It will likely involve dozens—or thousands—of specialized machines coordinating with intelligent software.

AI Agents Could Coordinate Multiple Physical Devices at Once

One of the more significant features of MHS is that it is not simply about making Claude press a button on a machine.

Anthropic says AI agents can use the framework to operate multiple instruments in parallel.

Consider a research workflow involving a liquid handler, microscope, imaging system and robotic arm. Traditionally, scientists may need separate software platforms, custom scripts and manual intervention to move samples and data between stages.

With an agentic system, the workflow could become more dynamic.

An AI agent could reason through an experimental procedure, instruct one machine to perform a task, inspect the resulting data, modify parameters and then trigger the next instrument.

Anthropic says possible applications range from routine drug-discovery experiments to the calibration of lasers used in quantum computing.[WIRED] explains how the standard is intended to connect AI agents with equipment ranging from microscopes and liquid handlers to quantum hardware and robotic systems.

This is where agentic AI becomes considerably more interesting.

A chatbot answers a question.

An AI agent decides which tools it needs to answer the question.

A physical AI agent could potentially decide which machines it needs, operate them, analyze what happened and determine the next action.

That is a much bigger leap.

Autonomous Laboratories Could Run Around the Clock

Another major promise behind the Anthropic Model Hardware Standard is the possibility of more autonomous experimentation.

Anthropic says MHS could help researchers build round-the-clock experimental workflows in which AI agents reason through stages of an experiment, change settings based on results and, in some circumstances, recover from hardware problems without requiring immediate human intervention.

The attraction is easy to understand.

Scientific equipment is expensive. Many instruments spend significant amounts of time waiting for operators, scheduled processes or downstream analysis. If intelligently automated systems can safely keep experiments progressing overnight or across weekends, organizations could potentially increase equipment utilization while reducing repetitive manual work.

The broader scientific community is already exploring this direction. Research published in Communications Materials, for example, has examined multi-agent approaches to managing autonomous materials laboratories, while other work is exploring self-correcting AI architectures for automated experimentation.

MHS Is Model-Agnostic—and That May Be Its Most Strategic Feature

You might reasonably assume that something created by Anthropic would be designed exclusively for Claude.

Surprisingly, MHS is intended to be model-agnostic.

Anthropic says the standard can work with different AI models and agent frameworks rather than locking organizations into Claude. [report on Anthropic’s first major move into physical AI] notes that this approach is intended in part to reduce vendor lock-in for scientists and engineers.

That decision could be extremely important.

Standards become much more valuable when multiple vendors can adopt them.

The web benefited from open protocols. USB simplified hardware connections. APIs transformed software integration. Anthropic’s earlier Model Context Protocol (MCP) has pursued a similar concept for connecting AI models with software tools and data sources.

MHS appears to extend that thinking toward hardware.

In fact, Anthropic says agent systems can access MHS through standard protocols such as MCP.

So the emerging architecture starts to look something like this:

An AI model provides reasoning.

An agent framework manages goals and actions.

MCP connects the agent to software, services and information.

MHS connects the agent to programmable physical devices.

If that ecosystem gains adoption, developers could build AI agents capable of navigating both the digital and physical parts of a workflow without creating bespoke integrations for every individual tool.

Manufacturing Could Be Just as Important as Scientific Research

Although laboratories provide some of the most eye-catching examples, manufacturing may ultimately become one of the largest commercial opportunities.

Modern factories already contain extensive automation. The problem is that much of it remains highly specialized.

Traditional industrial automation works beautifully when conditions are predictable: move this component from Point A to Point B, tighten this bolt to a predetermined setting, inspect this component using a predefined rule.

AI agents potentially introduce greater flexibility.

An agent might interpret production data, inspect the status of several machines, identify a bottleneck and coordinate equipment differently based on changing conditions.

Anthropic already signaled its interest in this area before launching MHS. In July 2026, the company announced work with UST to bring Claude into engineering environments spanning semiconductors, automotive systems, manufacturing and connected devices.

Anthropic Is Not Alone in the Physical AI Race

Anthropic’s announcement arrives during a broader industry push toward physical AI.

NVIDIA has been building an extensive ecosystem involving robotics foundation models, simulation, edge computing and autonomous-system development. Its [Physical AI research and agent-skills announcement] highlights how developers are increasingly combining simulation, synthetic data, robot policies and autonomous agents into complete physical-AI workflows.

Google DeepMind is moving in a similar direction. Its [Gemini Robotics 2 announcement] focuses on models capable of whole-body robotic control, dexterity and cooperation between machines.

The approaches are different.

NVIDIA is building much of the computing, simulation and robotics infrastructure.

Google DeepMind is developing models specifically designed to reason about and control robots.

Anthropic’s MHS focuses on interoperability—helping intelligent agents communicate consistently with physical equipment.

Those strategies could eventually complement one another rather than compete directly.

A future factory could theoretically contain hardware powered by NVIDIA edge systems, robots using specialized embodied-AI models and higher-level agents communicating with equipment through standards such as MHS.

Physical AI is starting to look less like a single product category and more like an entire technology stack.

Why the Model Hardware Standard Matters for Businesses

For technology leaders, MHS is worth watching even if your company does not own a microscope or robotic arm.

Standards often matter before the applications built on them become obvious.

If MHS or a similar framework becomes widely adopted, organizations may eventually be able to connect AI agents with physical infrastructure using reusable integrations rather than building automation systems from scratch.

That could lower development costs and speed up experimentation across industries including manufacturing, biotechnology, semiconductor production, logistics, electronics and advanced research.

Businesses exploring this area should therefore watch three things closely: how many hardware manufacturers implement MHS-compatible interfaces, whether competing AI providers support the standard, and what safety requirements emerge from the research preview.

Adoption will determine whether MHS becomes foundational infrastructure or simply another interesting protocol in an already crowded AI ecosystem.

Conclusion

Anthropic’s Model Hardware Standard shows where AI is heading next: from working only on screens to helping control real-world machines and equipment.

By giving AI agents a common way to communicate with physical devices, MHS could make it easier to automate laboratories, manufacturing systems, robotics, and other technical environments. Instead of building a separate connection for every machine, companies may eventually be able to use one standard to connect AI with many different types of hardware.

At the same time, safety will be extremely important. When AI starts controlling physical equipment, businesses need clear rules, human oversight, permissions, monitoring, and emergency controls.

MHS is still in an early stage, but the idea behind it is important. AI is becoming more than a tool that gives answers—it is becoming a system that can take actions in the real world.

Artificial intelligence may look like software, but the latest AI race is getting increasingly physical.

At Hot Chips 2026, OpenAI pulled back the curtain on Jalapeño, its first custom AI inference accelerator, offering a much deeper look at a chip that could become an important part of the infrastructure behind ChatGPT, Codex, AI agents, and future OpenAI services.

The OpenAI Jalapeno Custom AI ASIC at Hot Chips 2026 story matters for more than semiconductor enthusiasts. It is a window into where the economics of artificial intelligence are heading: away from a world where every major AI workload automatically lands on a general-purpose GPU and toward one where hyperscalers and AI labs increasingly design specialized silicon around their own models, workloads, networking requirements, latency targets, and power constraints.

And OpenAI is not exactly dipping a toe in the water.

Jalapeño is intended to become the first generation of a multi-generation OpenAI compute platform.

What Is OpenAI Jalapeño?

Jalapeño is an application-specific integrated circuit, or ASIC, designed primarily for large-language-model inference.

That distinction is important.

Training is the computationally enormous process used to build or update a model. Inference is what happens after training, when millions of users ask that model questions, generate code, run AI agents, summarize documents, create content, or interact with AI-powered applications.

OpenAI formally introduced Jalapeño on June 24, 2026, through its announcement, [OpenAI and Broadcom unveil LLM-optimiezed inference chip] OpenAI said the accelerator was designed from scratch around modern LLM inference rather than adapting an architecture originally intended for a broader range of computing jobs. Broadcom is contributing silicon implementation and networking technology, while Celestica is involved in board, rack, and system integration.

That makes Jalapeño less of an attempt to create “another GPU” and more of an attempt to answer a very OpenAI-specific question:

What would the ideal computer look like if you could design the hardware, software, model-serving system, networking, memory architecture, and AI products together?

For more context on why specialized accelerators are becoming so important, Quantilus recently explored the same industry shift in [A New Era of AI Hardware Built for Training and Inference]

What OpenAI Revealed at Hot Chips 2026

Hot Chips 2026 ran from August 23–25 at Stanford University, with OpenAI presenting during the conference’s final AI session on August 25. Richard Ho, Ravi Narayanaswami, and Chris Leary delivered a wonderfully understated presentation titled “You Can Just Build Things … Chips.” The official Hot Chips program confirms OpenAI shared the AI 2 session with Google and SambaNova.

What followed was considerably more technical than the original June announcement.

Detailed coverage from [ServeTheHome’s OpenAI Jalapeno Custom AI ASIC at Hot Chips 2026 analysis] describes a full inference platform rather than merely a silicon die: accelerator hardware, memory, networking, racks, software, kernels, and a programming model designed to operate together.

That full-stack approach may ultimately prove more important than any individual specification.

Jalapeño’s Hardware Specifications Are Serious

According to the architecture details presented at Hot Chips, a Jalapeño accelerator delivers approximately 13.4 PFLOP/s of MXFP4 matrix compute, paired with 216 GiB of HBM4 memory and roughly 15.4 TB/s of HBM bandwidth.

The package has a rated thermal design power of 700 watts, although OpenAI says sustained power during the workloads it tested remained at or below approximately 550 watts.

That is particularly interesting because AI infrastructure is increasingly power-constrained.

More GPUs do not magically create more electricity. Data-center operators face fixed utility connections, cooling capacity, power-distribution infrastructure, construction timelines, transformers, backup systems, and local grid limitations. As a result, how much useful AI work can be produced from every kilowatt is becoming nearly as important as raw chip performance.

Jalapeño vs. Nvidia GB200 and GB300

Now for the spicy part. Yes, apparently the chip name has made pepper jokes legally mandatory.

OpenAI benchmarked Jalapeño using InferenceX, a public inference benchmarking framework developed by SemiAnalysis.

The tests included OpenAI’s GPT-OSS 120B, DeepSeek R1 670B, and Kimi K2.5 1T. According to OpenAI’s reported results, Jalapeño reached the performance-per-watt and latency Pareto frontier across those workloads.

On GPT-OSS 120B, OpenAI reported roughly 1.9 times higher peak throughput per kilowatt and approximately 1.7 times lower end-to-end latency at matched operating points against the comparison system.

For DeepSeek R1, reported improvements reached roughly 1.7 times higher peak performance per watt and 3.6 times lower end-to-end latency in one comparison.

On the enormous Kimi K2.5 workload, OpenAI reported approximately 1.5 times higher peak performance per watt and 3.4 times lower end-to-end latency.

Independent technical analysis is available in [SemiAnalysis and InferenceX’s deep dive into OpenAI Jalapeno] SemiAnalysis says it benchmarked the system with OpenAI engineers and describes Jalapeño as an unusually competitive first-generation custom accelerator.

There is, however, an important caveat.

SemiAnalysis argues that comparing Jalapeño primarily with Nvidia’s Blackwell-generation GB200 and GB300 does not tell the complete competitive story because newer Nvidia Rubin-generation hardware is the more forward-looking comparison.

That nuance matters. Benchmark headlines saying “OpenAI beats Nvidia” are irresistible click candy, but AI accelerator performance varies dramatically by model, batching strategy, precision, software stack, networking configuration, latency target, and system scale.

The sensible conclusion is not that Nvidia has suddenly been dethroned.

It is that OpenAI has produced a first-generation ASIC credible enough to enter the conversation.

That alone is significant.

AI Helped OpenAI Build the AI Chip

Perhaps the most fascinating part of the Jalapeño story is how quickly OpenAI says it was developed.

OpenAI says the core project moved from initial design to manufacturing tape-out in roughly nine months, with OpenAI models helping engineers during design and optimization.

There is some useful nuance around that figure: SemiAnalysis describes a longer period of roughly 16 months when counting from early team formation to tapeout, while OpenAI’s nine-month claim refers more specifically to the design-to-tapeout phase.

At Hot Chips, OpenAI also described using AI to optimize hardware kernels. ServeTheHome reported that AI-optimized attention and mixture-of-experts kernels achieved roughly 1.5x to 1.8x higher performance than existing expert-written implementations in the examples shown.

This creates a fascinating feedback loop.

OpenAI builds AI models.

Those AI models help engineers design and program better chips.

Those chips run future AI models more efficiently.

The improved infrastructure makes larger and more capable AI services economically viable.

Those services can then assist with the next generation of infrastructure.

Software eating the world was apparently just the appetizer.

Does Jalapeño Mean OpenAI Is Leaving Nvidia?

No.

OpenAI has been unusually explicit on this point.

The company says it expects to continue deploying accelerators from Nvidia and other partners for both training and inference, even as Jalapeño enters its infrastructure fleet.

That makes sense.

AI compute demand is expanding so rapidly that OpenAI’s problem is unlikely to be choosing one chip supplier and abandoning everyone else. The larger challenge is securing enough economically viable compute from multiple sources.

Jalapeño therefore looks less like an Nvidia replacement and more like a strategic new layer in OpenAI’s supply portfolio.

It gives OpenAI another source of compute, greater control over inference economics, tighter hardware-software optimization, and potentially more negotiating leverage throughout the semiconductor ecosystem.

Why Custom Silicon Could Change the Economics of AI

Inference is where an AI business repeatedly pays the computational bill.

A frontier model may require an enormous one-time training investment, but serving that model to hundreds of millions of requests creates an ongoing infrastructure cost.

Every reduction in energy per token, latency per request, server count, or networking overhead can therefore multiply across massive volumes.

For OpenAI, successful custom silicon could eventually mean lower operating costs, greater capacity, more predictable infrastructure supply, and improved performance for ChatGPT, Codex, API users, and AI agents.

For enterprises, those infrastructure gains could eventually surface as faster products, higher rate limits, more capable agentic workflows, or lower costs.

The Sustainability Angle: Performance Per Watt Is No Longer Optional

Jalapeño also highlights an increasingly important environmental and infrastructure issue.

AI’s energy requirements are growing rapidly, and large data-center deployments can create pressure on electricity grids, water systems, land use, and local communities.

That means performance per watt is not merely an engineering benchmark or cost-saving trick.

It is becoming part of responsible AI infrastructure design.

If a system can perform substantially more useful AI inference within the same power envelope, operators may be able to serve more users without requiring an equivalent increase in electricity consumption.

Of course, efficiency has a catch: cheaper and faster compute can encourage dramatically more AI usage. Better efficiency per request does not automatically guarantee lower total energy consumption.

This is why sustainability reporting should increasingly include both efficiency metrics and absolute consumption.

Policymakers, enterprises, and infrastructure providers will also need better transparency around electricity sources, grid impact, water consumption, carbon intensity, semiconductor supply chains, and lifecycle emissions.

The hardware may be getting smarter. Governance still needs to keep up.

What Jalapeño Means for the AI Chip Market

The bigger competitive story extends far beyond OpenAI and Nvidia.

Google has spent years building TPUs. Amazon has developed Trainium and Inferentia. Microsoft has its Maia accelerator program. Meta is developing custom AI silicon. Meanwhile, Nvidia, AMD, Cerebras, SambaNova, and other semiconductor companies continue pushing specialized architectures of their own.

Hot Chips 2026 effectively showcased this fragmentation in real time, with presentations from Meta, Microsoft, Google, Nvidia, Cerebras, SambaNova, AMD and OpenAI across the program.

The future of AI computing therefore probably will not be one universal processor.

It is increasingly likely to be a heterogeneous world where organizations select hardware according to the workload: training, high-throughput inference, interactive inference, long-context reasoning, agentic workloads, multimodal applications, or specialized enterprise deployment.

That competition should ultimately benefit AI users—provided increasingly powerful infrastructure is deployed responsibly and access does not become concentrated among only a handful of companies.

Conclusion

OpenAI’s Jalapeño custom AI ASIC is more than a new chip—it is a sign that the AI industry is entering a new phase where software, models, networking, memory, power efficiency, and silicon are being designed together as one integrated system.

At Hot Chips 2026, OpenAI showed that it is no longer relying solely on third-party accelerators to shape the performance and economics of its AI services. By building specialized hardware around real inference workloads, OpenAI is aiming to reduce latency, improve performance per watt, increase infrastructure efficiency, and support the growing demands of AI agents and large-scale model serving.

That does not mean Nvidia or other accelerator vendors are suddenly irrelevant. Far from it. The more likely future is a diverse AI infrastructure ecosystem in which custom ASICs, GPUs, TPUs, and other specialized processors coexist, each optimized for different workloads and deployment needs.

What makes Jalapeño especially important is the precedent it sets. If OpenAI can successfully combine custom silicon with its models, software stack, networking architecture, and data-center strategy, it could gain far greater control over the cost and performance of running AI at scale. That could influence everything from API pricing and response times to the capabilities of future autonomous agents.

The competitive race will also extend beyond raw speed. Energy efficiency, sustainability, supply-chain resilience, responsible infrastructure expansion, and transparency around resource consumption are becoming critical measures of success.

Ultimately, the OpenAI Jalapeno Custom AI ASIC at Hot Chips 2026 story is not simply about whether one processor can outperform another. It is about who can build the most efficient, scalable, and intelligent computing platform for the next generation of AI.

And if Jalapeño is only OpenAI’s first step into custom silicon, the real impact may become clearer with the generations that follow.

Humanoid robots just crossed another line that would have sounded like science fiction only a few years ago.

China’s Tiangong Ultra humanoid robot completed 100 metres in just 8.86 seconds during a semifinal at the 2026 World Humanoid Robot Games in Beijing, dramatically lowering the 9.39-second time it recorded earlier in the competition. Reuters reported that the result came in the large-size robot division on August 25 and put Tiangong more than seven-tenths of a second inside the famous 9.58-second human benchmark set by Usain Bolt in 2009. [Reuters]

That headline is spectacular. The technology behind it is arguably even more interesting.

Tiangong’s run is not simply a story about creating a machine that moves quickly. It demonstrates how rapidly physical AI, motion-control algorithms, actuators, autonomous navigation, mechanical engineering and real-time sensor processing are converging.

And those technologies could ultimately matter far more in factories, warehouses and infrastructure projects than they do on a running track.

From 21.50 Seconds to 8.86: A Remarkable One-Year Improvement

The scale of Tiangong’s improvement deserves some perspective.

At the inaugural World Humanoid Robot Games in 2025, Tiangong Ultra won the 100 metres in 21.50 seconds. One year later, its reported semifinal time is 8.86 seconds.

That means the robot has shaved 12.64 seconds—nearly 59%—off its 100-metre time in roughly a year.

At 8.86 seconds, its average speed across 100 metres works out to roughly 40.6 kilometres per hour, or about 25.3 mph.

That is quite a jump for a machine that needs to continuously balance a tall, articulated body while transferring force through two legs.

There is an important distinction, however: Tiangong has not replaced Usain Bolt as the World Athletics record holder. It has run the same distance in a lower recorded time, but robots and human athletes compete under entirely different categories, engineering constraints and rule systems.

So, no, Bolt does not need to return his medals to a robot.

But as a robotics benchmark, the result is extraordinary.

What Changed Inside Tiangong Ultra?

The biggest question is how a humanoid robot could improve so dramatically in such a short period.

According to technical details provided by the Beijing Humanoid Robot Innovation Center, Tiangong Ultra received upgrades to both its hardware and software. The robot reportedly uses improved joints, a lighter structure and a more streamlined physical design, while its control and navigation algorithms have also been optimized for high-speed running. The Center also described a move from simple track-line following toward map-based positioning, allowing the robot to navigate at speed without depending exclusively on painted lane markings. [Global Times]

That combination matters.

Running is an unforgiving robotics problem. Every stride involves balancing the robot on effectively one leg for part of the gait cycle while constantly adjusting joint torque, foot placement, body angle and momentum.

The robot needs to process sensor data, predict its next movement and correct small errors before they become large ones.

At sprinting speeds, it has very little time to think about it.

This is exactly why the broader development of physical AI is becoming such an important area of technology. Unlike a chatbot, which operates almost entirely in the digital world, physical AI must turn decisions into actions while dealing with gravity, friction, unpredictable environments and hardware limitations.

The World Humanoid Robot Games Are Becoming a Serious Technology Testbed

The setting is also important.

The second World Humanoid Robot Games, held at Beijing’s National Speed Skating Oval from August 22–26, have expanded dramatically.

According to official Beijing government information, the 2026 competition includes 2,056 robots from 666 teams, participating in 51 events and 1,301 competition sessions. The program includes 30 competitive events and 21 scenario-based challenges.

Those challenges go well beyond sprinting.

Robots are being tested in football, martial arts, dancesport and athletics, but also in environments intended to resemble factories, hotels, logistics facilities and other workplaces.

That shift is crucial.

Robot racing is entertaining. A robot that can reliably manipulate unfamiliar objects, identify equipment, navigate a changing building and recover from mistakes could create real economic value.

Being Faster Than Bolt Is Impressive. Plugging in a Cable May Be More Important.

There is a delightful irony in the current generation of humanoid robots.

Some can apparently sprint 100 metres faster than any human in recorded history.

Then they can struggle with a cable.

Reuters highlighted precisely this contrast in its broader reporting from the Games, noting that scenario-based competitions are testing robots on tasks such as connecting cables, charging electric vehicles, handling restaurant work and performing industrial operations. More than 40% of the relevant tests reportedly require full autonomy. [Reuters]

That is a useful reminder of how robotics progress should be measured.

A humanoid robot that runs extremely quickly demonstrates advanced locomotion, balance and mechanical performance. But a commercially useful robot also needs perception, dexterity, adaptability, reliability and safe decision-making.

The difference is similar to the difference between passing one difficult exam and being good at an entire profession.

Running is one skill.

Working in a messy human environment is thousands of skills stitched together.

Why Humanoid Form Factors Matter

Why build robots with two legs at all?

After all, wheels are generally simpler, more stable and more energy efficient.

The answer is largely that our world was built for humans.

Factories have stairs. Offices have doors. Warehouses have shelves at human heights. Homes contain handles, switches, tools, furniture and appliances designed around human bodies.

A capable humanoid robot could potentially operate within existing infrastructure without businesses rebuilding every workplace around a specialized machine.

That explains why so much investment is flowing into humanoid robotics.

The long-term objective is not necessarily to build machines that look impressive while sprinting around stadiums. It is to develop machines that can walk into environments designed for people and perform useful work there.

Speed Is Only Half the Story—Stopping Matters Too

There is also an amusing but important caveat hiding behind some of these spectacular sprint times.

Fast humanoid robots do not always stop particularly elegantly.

Coverage of the earlier 9.39-second performance showed robots relying on padded areas after races, highlighting the challenge of rapidly dissipating momentum while maintaining balance.

That matters because the ability to accelerate is only one component of useful mobility.

Real-world robots must also decelerate, change direction, avoid people, respond to unexpected obstacles and recover safely when something goes wrong.

In a warehouse, hospital or crowded workplace, smashing a speed record before smashing into the nearest wall would be a fairly serious design flaw.

The next major robotics milestone therefore may not be another tenth of a second removed from a sprint time. It may be achieving similar speed while demonstrating much more sophisticated control before, during and after the run.

What Tiangong’s Sprint Could Mean for Industry

The commercial implications extend well beyond competitive robotics.

Technologies refined through high-speed locomotion could eventually contribute to robots performing tasks such as infrastructure inspection, industrial maintenance, logistics and work in environments that are difficult for people to navigate.

The Beijing Humanoid Robot Innovation Center itself has connected technology developed through racing with potential applications including long-distance inspection, logistics and emergency response in complex environments.

That makes robot competitions surprisingly useful engineering laboratories.

A 100-metre race pushes actuators, joints, control systems, batteries and navigation algorithms toward their limits. Those stresses reveal weaknesses quickly.

Failure becomes data.

Engineers can then take what they learn about durability, balance and control and apply it to less glamorous but much more commercially valuable applications.

The Bigger Story: China’s Physical AI Strategy

Tiangong also illustrates how aggressively China is developing the humanoid robotics ecosystem.

The scale of the World Humanoid Robot Games—more than 2,000 robots in 2026 compared with just over 500 at the inaugural event—demonstrates how quickly participation is expanding.

Competitions create more than publicity.

They establish common benchmarks, generate data, expose hardware weaknesses, encourage universities and companies to compete against one another and demonstrate new technologies to investors and potential customers.

This can accelerate an entire ecosystem.

The result is an emerging global competition not simply over who builds the best humanoid robot, but over AI models, actuators, sensors, batteries, edge processors, simulation platforms, manufacturing capacity and robotics standards.

Safety, Ethics and Regulation Cannot Trail Behind the Robots

There is another reason Tiangong’s 8.86-second run matters.

Powerful robots are becoming genuinely powerful.

As machines become faster, stronger and increasingly autonomous, safety engineering needs to progress at the same rate.

Future commercial humanoids will require clearly defined operational boundaries, fail-safe mechanisms, collision detection, reliable emergency stops, cybersecurity protections, human oversight and transparent testing standards.

Organizations adopting physical AI will also need to determine who is responsible when an autonomous machine makes an incorrect decision.

Is responsibility carried by the robot manufacturer, AI-model developer, system integrator, operator or organization deploying it?

Those questions become much less theoretical when machines can move at more than 40 km/h.

Responsible robotics therefore cannot be separated from responsible AI. Performance benchmarks are exciting, but reliability, explainability, safety and accountability will ultimately determine whether humanoids earn public trust and large-scale commercial adoption.

What Comes Next for Humanoid Robots?

The most striking part of Tiangong’s achievement may not actually be the number 8.86.

It is the speed of improvement.

Going from 21.50 seconds in 2025 to below nine seconds in 2026 shows how quickly robotics can advance when better mechanical systems, AI control algorithms, sensing and intense real-world testing arrive simultaneously.

The next frontier will be combining those athletic capabilities with something considerably harder: general-purpose competence.

Can the same machine run quickly, slow down safely, climb stairs, manipulate fragile objects, follow natural-language instructions, recognize unexpected hazards and complete hours of useful work without constant intervention?

That is the benchmark businesses should watch.

Because the humanoid robot revolution will not be decided by which machine wins a medal.

It will be decided by which machines can reliably show up for work.

Conclusion

An 8.86-second 100-metre sprint is more than a flashy robotics milestone. It shows just how quickly physical AI is moving from controlled experiments into real-world capability.

The bigger story is not whether robots can outrun people. It is how rapidly machines are improving at balance, movement, navigation, sensing and decision-making — the exact capabilities needed for useful work in factories, warehouses, infrastructure, logistics and other human-designed environments.

As these systems become faster, stronger and more autonomous, the conversation must also expand beyond performance. Safety, reliability, accountability and responsible deployment will be just as important as speed.

The race for the future of robotics is clearly accelerating. And if an 8.86-second sprint tells us anything, it is that progress in physical AI may be arriving much faster than many expected.

AI coding has spent the last few years becoming remarkably good at helping one person build software faster. Slack now wants to answer a more interesting question: what happens when AI coding becomes a team activity instead of a private conversation between a developer and an agent?

That is the idea behind Slack Code, a new collaborative environment that brings AI coding agents into dedicated Slack channels where developers, product managers, designers, and even less-technical colleagues can participate in building software together.

Slack describes the concept as turning agentic coding into a “multiplayer” experience. Instead of an engineer disappearing into a terminal, IDE, or private AI chat and returning later with a pull request, a team can watch an agent work, provide context, review code changes, preview results, redirect the agent, and approve the final output from a shared workspace. Salesforce’s official introduction to Slack Code describes these spaces as project-specific code channels designed to keep the agent, conversation, code, documentation, and decision-making together. [Read Salesforce’s introduction to Slack Code]

What Is Slack Code?

Slack Code is essentially a dedicated workspace for a specific AI-assisted development task.

Imagine a product manager notices that a button is broken on a customer-facing page. Traditionally, the issue might become a ticket, get assigned to an engineer, enter a backlog, move into an IDE, generate a pull request, and eventually return to product or design for review.

With Slack Code, a teammate can bring a supported coding agent into the conversation. A dedicated code channel can then be created around the task, bringing relevant people and context together. Within that space, participants can see the agent’s plan, review code diffs, examine live HTML previews, provide feedback in natural language, and approve the work before deployment. Completed channels can be automatically archived while retaining the history as an audit record.

Computerworld’s coverage of the launch highlights the central idea particularly well: Slack is attempting to make coding agents available to multiple coworkers rather than keeping each agent interaction confined to an individual developer. [Read Computerworld’s Slack Code analysis]

This is a small interface change with potentially large organizational consequences.

From Single-Player AI Coding to Multiplayer Development

Tools such as Claude Code, GitHub Copilot, Devin, Codex, Cursor, and other coding agents have already changed how developers approach programming. They can help explain unfamiliar repositories, generate code, investigate bugs, draft tests, and implement features.

But there has been a structural weakness in many AI coding workflows: the agent’s working context can remain invisible to everyone except the person operating it.

A developer may spend 45 minutes prompting an AI agent, rejecting approaches, clarifying requirements, fixing assumptions, and refining the solution. When the resulting pull request appears, everyone else sees the outcome—but not necessarily the reasoning that produced it.

Slack Code attempts to expose more of that process.

The Verge describes the product as a collection of collaborative “vibe-coding” channels in which teams can compare changes and preview HTML output before something ships. [Read The Verge’s coverage of Slack Code]

That distinction matters because software development is rarely just about writing syntax. Good software emerges from discussions about customer needs, design trade-offs, deadlines, security, accessibility, architecture, business rules, and countless small decisions. Much of that information already lives in Slack.

Putting an AI coding agent closer to those conversations could reduce one of the biggest headaches in AI-assisted development: context loss between the business conversation and the technical implementation.

Developers Are No Longer the Only People in the Coding Conversation

Perhaps the most interesting part of Slack Code is not what it does for programmers. It is what it could do for everyone around them.

Product managers can clarify requirements while implementation is happening. Designers can examine a live preview instead of waiting for a completed build. Marketing teams can flag incorrect messaging. QA specialists can raise edge cases. Accessibility experts can identify issues earlier. Engineers can inspect the actual code diff before approving a change.

Non-technical contributors do not suddenly become senior software engineers—and businesses should not treat AI agents as a shortcut around engineering expertise. What changes is their ability to participate directly in the creation loop.

VentureBeat characterizes Slack’s strategy as moving AI coding “out of the terminal and into the group chat,” with teams able to watch, steer, review, and ship alongside AI coding agents. [Read VentureBeat’s analysis of collaborative AI coding]

That could shorten the distance between “here is what the customer needs” and “here is the software we built.”

Code Diffs and Live Previews Make AI Work More Visible

One of Slack Code’s strongest ideas is making the output of an AI agent inspectable.

Code channels can surface code diffs, allowing participants to see what lines changed rather than simply accepting an agent’s claim that it “fixed the problem.” For web-related work, teams can also view live HTML previews and check what the proposed change actually looks like.

This is important because natural-language interfaces can create a dangerous illusion of simplicity. A request such as “fix the checkout button” sounds trivial. The underlying code change may not be.

Making diffs, plans, previews, comments, and approvals visible in the same workspace encourages teams to treat AI-generated code as something to review, not something to blindly trust.

Slack Code Creates an Ecosystem Rather Than Another Coding Model

Slack is not trying to win the AI model race by creating yet another coding model. Instead, it is positioning itself as the collaborative layer where multiple coding agents meet teams.

Salesforce says Slack Code is being developed with an ecosystem that includes Anthropic, Cognition, GitHub, OpenAI, and Vercel. Its launch materials cite agents including Claude, Devin, GitHub Copilot, ChatGPT, and Vercel’s agents. Availability and individual partner requirements can vary, so organizations will still need the appropriate access to whichever agent they intend to use.

That ecosystem approach is strategically significant.

Enterprises increasingly use more than one AI model or agent. A company might prefer one model for development, another for research, and another for specialized internal workflows. If Slack can become the interface through which employees collaborate with those agents, Slack becomes less of a messaging app and more of an orchestration layer for human-AI work.

Governance Could Be Slack Code’s Real Enterprise Advantage

AI coding demos tend to focus on speed: describe an application, watch it appear, celebrate.

Businesses have less glamorous questions.

Who authorized the change? What data could the agent access? Which repository was modified? Who reviewed the code? What permissions did the agent inherit? Who approved production deployment? Can we reconstruct what happened later?

Slack Code is notable because the product is being framed around visibility, permissions, human approval, and auditability, not merely generation.

Salesforce says code-channel agents inherit Slack’s existing security model, permissions, and administrative controls. It also states that higher-stakes actions—such as pushing code into production—can require expert human sign-off. Completed code channels preserve a record of the work rather than simply disappearing.

Computerworld also notes an important caveat: permissions become especially complicated in multiplayer agent environments, because organizations must carefully decide whose authorization an agent inherits and what data or systems it can access.

That means Slack Code should not be interpreted as permission to abandon normal engineering controls.

Organizations adopting collaborative AI coding should still maintain clear repository permissions, branch protection, automated testing, security scanning, code review, deployment controls, observability, data-governance policies, and named human accountability.

What Slack Code Could Mean for the Future of Work

Slack Code may eventually matter for much more than software development.

Once organizations become comfortable placing AI agents into shared, auditable workspaces, the same interaction model can extend to other knowledge-intensive projects.

Slack has already suggested broader possibilities involving activities such as marketing projects and document review. The pattern is the important part: create a temporary workspace around a goal, give an AI agent controlled access to relevant context, allow multiple people to steer its work, surface its outputs for review, require appropriate approvals, and retain the final activity history.

That is essentially multiplayer agentic work.

A marketing team could collaborate with an agent on a campaign. A finance team might work with one on a reporting workflow. A legal team could review document revisions. Operations groups could coordinate agents working across connected systems.

In other words, Slack Code may be less about turning Slack into an IDE and more about testing a new model for how organizations interact with autonomous software.

Conclusion

Slack Code shows how AI coding is moving from a tool used by one developer to a shared team experience. Instead of keeping AI-generated code hidden inside a private chat or coding environment, teams can work together, review changes, give feedback, and approve updates in one place.

The biggest benefit is not simply faster coding. It is better collaboration between developers, designers, product managers, and other team members. Everyone can understand what the AI is doing and contribute before changes are released.

At the same time, human review remains important. AI coding agents can speed up development, but organizations still need strong security, testing, permissions, and approval processes.

Overall, Slack Code represents a broader shift in the future of work: AI is becoming a teammate rather than just a personal assistant. If this approach becomes widely adopted, software development could become faster, more transparent, and more collaborative—while still keeping people in control.

Artificial intelligence has taught robots how to see, recognize, reason and increasingly plan. But there is still one surprisingly human capability that machines struggle to reproduce: knowing what something feels like.

That gap may be closing.

Touch-sensitive electronic skin, or e-skin, is emerging as one of the most promising technologies in robotics. Flexible sensor systems can give machines the ability to detect pressure, contact, force, slipping and—in increasingly sophisticated designs—multiple types of physical stimuli across large areas of a robot’s body.

That matters because vision alone cannot tell a robot everything it needs to know.

Imagine a robotic hand picking up a strawberry. A camera can identify the fruit and estimate its location. But once the fingers make contact, the robot needs to determine whether the strawberry is secure, whether it is starting to slip, and whether additional grip force will turn lunch into jam.

This is where tactile intelligence enters the picture.

Touchlab CEO Zaki Hussein recently described the commercialization challenge succinctly in a [TechRadar discussion of touch-sensitive e-skin for robotics] “Groundbreaking technology often enters the market at a premium before scaling.” Touchlab’s goal is to make tactile sensing sufficiently practical and affordable that it eventually becomes a standard component of commercial robots rather than an exotic research feature.

That transition could become an important chapter in the rise of physical AI.

What Is Touch-Sensitive E-Skin?

Electronic skin is a broad category of thin, flexible or stretchable sensor technology designed to reproduce some of the sensory capabilities associated with biological skin.

Rather than installing a handful of rigid force sensors into a robotic gripper, engineers can potentially cover fingers, palms, arms or even much of a robot’s body with distributed sensing surfaces.

Touchlab takes another approach. Its technology uses quantum tunneling-based sensing to measure forces through an extremely thin material. According to the company in its TechRadar interview, the technology can support very rapid responses while measuring characteristics such as pressure, force and direction.

The bigger idea is simple: rather than merely telling a robot that contact happened, advanced e-skin can provide richer information about how that contact is changing.

For robots working around unpredictable objects and people, that difference is enormous.

E-Skin Could Make Robotic Hands Dramatically More Useful

Dexterity is becoming one of the biggest competitive frontiers in humanoid and general-purpose robotics.

Walking looks impressive in a product demonstration. Manipulating unpredictable everyday objects reliably is often much harder.

Plastic bags deform. Fruit bruises. Fabric folds. Tools can rotate inside a hand. Bottles become slippery. Electrical components can be damaged by excess pressure.

A sophisticated tactile system could help a robot recognize these changes as they happen.

Research is already demonstrating the value of combining tactile sensing with more capable hands. A 2025 Nature Communications paper, [Soft robotic hand with tactile palm-finger coordination] describes a robotic hand integrating high-density tactile sensing in the palm with soft fingers and coordinated manipulation strategies.

That is important because useful robotic manipulation will likely require more than increasingly sophisticated AI models. It requires a complete feedback loop:

Perceive → decide → act → feel → adjust.

The “feel” stage has historically been much weaker than the others.

E-skin could change that.

Edge Computing Could Give Robots Faster Physical Reflexes

Tactile sensing also creates an interesting computing challenge.

A robot covered in thousands of sensing points could potentially generate an enormous amount of information. Sending every tiny change to a centralized computer—or worse, across a network to a cloud service—could introduce unnecessary bandwidth requirements and latency.

A more practical approach is to process some tactile information at the edge, close to where it is generated.

Touchlab describes a shared-autonomy architecture in which low-level tactile reflexes, such as preventing an object from slipping, can be handled locally while a human operator or higher-level AI concentrates on broader decisions.

That resembles the division of labor in biological systems. You do not consciously calculate every tiny pressure adjustment required to keep a coffee mug from sliding through your fingers. Many responses occur automatically.

Giving robots a comparable low-level feedback layer could reduce both computational overhead and reaction time.

Manufacturing E-Skin at Scale Is the Real Test

Producing an impressive sensor in a laboratory is one thing.

Producing thousands or millions of reliable, affordable sensors that survive industrial use is something else entirely.

E-skin must potentially withstand bending, impact, repeated contact, abrasion, temperature changes and thousands—or millions—of manipulation cycles.

It also has to conform to irregular robotic surfaces while maintaining consistent measurements.

Manufacturing complexity is therefore one of the biggest barriers between advanced tactile sensing and widespread deployment.

A 2026 study in npj Flexible Electronics, [Scalable in-situ fabrication of multimodal electronic skin for intelligent interactive systems] specifically highlights scalability and customization as persistent limitations for many multimodal e-skin platforms because existing approaches can depend on complicated fabrication and assembly processes.

That helps explain why the “premium before scaling” argument matters.

Early commercial deployments do not necessarily need maximum possible sensor density across every square centimeter of a robot.

A manufacturer might instead place high-performance tactile sensing where it creates the greatest operational value: fingertips, palms, grippers or surfaces likely to encounter humans.

As production improves and costs fall, coverage can expand.

This is a familiar technology curve. What begins as specialized hardware can eventually become a commodity component once manufacturing processes, supply chains and standards mature.

The Business Case Goes Beyond Humanoid Robots

Humanoid robots may generate the headlines, but e-skin’s commercial potential is considerably broader.

Manufacturing

Industrial robots equipped with distributed tactile sensing could become better at handling irregular, fragile or variable components instead of relying exclusively on perfectly controlled production environments.

Healthcare and Assisted Living

Robots working near patients need exceptionally careful force control.

A machine helping someone reposition an arm, handing over an object or supporting mobility cannot behave like an industrial machine behind a safety cage.

Robotic skin could provide another layer of awareness when physical contact occurs.

Warehousing and Logistics

Tactile sensing could help robots grip oddly shaped packaging, identify unstable loads and handle objects whose physical properties vary even when their visual appearance is similar.

Home Robotics

Homes may ultimately be one of the strongest arguments for tactile robotics.

Unlike factories, homes are almost completely uncontrolled environments. Objects move constantly. Surfaces vary. Children, adults and pets may occupy the same space.

A robot expected to fold laundry, clear a table or assist an older adult needs more than object recognition. It needs continuously adaptive physical behavior.

Safety, Privacy and Standards Cannot Be an Afterthought

As physical AI becomes more capable, governance becomes more important.

A robot making decisions in software can generate an incorrect output. A robot making decisions in the physical world can move machinery, touch a patient, carry equipment or interact directly with a person.

Tactile sensing may improve safety by helping machines detect unexpected contact and regulate force, but organizations still need clear rules for testing, accountability, data governance and system failure.

Touchlab has also emphasized that tactile data can exist alongside video, audio, LiDAR and other sensor streams. Although tactile readings themselves may not identify a person, combining multiple streams can create broader privacy and security concerns.

Businesses deploying intelligent robotics should therefore treat security and governance as system-level requirements rather than individual sensor problems.

Standards will matter too. Comparable metrics for tactile sensitivity, spatial resolution, response time, durability and force detection could make it easier for robotics manufacturers to evaluate competing technologies.

Standardization is often less exciting than a humanoid demonstration—but it is frequently what transforms an emerging technology into infrastructure.

From Premium Technology to Standard Robotic Equipment

The most important question may therefore not be whether electronic skin works.

There is already significant research showing that flexible tactile sensing can provide remarkable sensitivity and enable increasingly sophisticated robotic interactions. A 2025 review of e-skin technologies describes the field as a convergence of advanced materials, sensors, AI and edge computing with major potential in robotics and human-machine interaction.

The question is whether the technology can become sufficiently durable, standardized, manufacturable and inexpensive to appear routinely in commercial machines.

That is why Hussein’s premium-to-scale observation is so relevant.

The first generation of advanced e-skin robots may be expensive. They may appear primarily in high-value environments where better manipulation or safer human interaction justifies the cost.

But commercialization rarely ends where it starts.

Production volumes increase. Manufacturing improves. Electronics become cheaper. Software becomes reusable. Standards emerge. Suppliers compete.

Eventually, something that once looked experimental can become simply another component on the bill of materials.

The Next Era of Physical AI May Be About Feeling

AI’s transition from digital assistants to machines operating in the physical world fundamentally changes what intelligence requires.

A chatbot needs to understand words.

A robot needs to understand consequences in three-dimensional space.

Vision tells it what an object looks like. AI reasoning tells it what the object may be and what should happen next. Motors let it act.

Touch tells it what is happening right now.

That final feedback loop could be one of the technologies that makes the difference between robots that perform carefully staged demonstrations and robots capable of working reliably beside people.

Touch-sensitive e-skin will not become commonplace simply because researchers can make impressive sensors. It will happen when robotics companies can deliver the right amount of tactile intelligence at the right price, package it into durable hardware, process its data efficiently and demonstrate that it improves real-world performance.

If that happens, electronic skin could follow the trajectory of cameras, accelerometers and other technologies that began as premium capabilities before becoming ordinary components.

And when that transition arrives, the smartest robots may no longer be defined only by what they can see or think.

They may be defined by what they can feel.

Conclusion

Touch-sensitive e-skin could be a major step forward for robotics because it gives machines something they have largely been missing: the ability to feel and respond to physical contact.

With e-skin, robots could handle fragile objects more carefully, work more safely around people, react faster when something slips, and perform more complex tasks in factories, hospitals, warehouses, and homes.

The technology is still relatively expensive and difficult to manufacture at scale, but that is common with new innovations. As production improves, costs may fall and e-skin could become a standard feature in future robots.

The bigger takeaway is simple: robots are becoming more useful because they are learning to combine sight, intelligence, movement, and touch. If tactile sensing continues to improve, the next generation of robots may not just see and understand the world around them—they may be able to feel it too.

Nvidia has already become one of the defining companies of the generative AI boom. Now it wants artificial intelligence to leave the screen, grow arms and legs, and start interacting with the physical world.

That ambition is becoming increasingly important to the Nvidia investment story.

Nvidia shares traded around $225 as investors digested another expansion of the company’s robotics strategy. Barron’s reported that the stock had gained more than 12% during August and highlighted Nvidia’s new agreement with South Korea’s LG Group to jointly develop humanoid robots. You can read the original [Barron’s report on Nvidia’s robotics bet here].

But the more interesting story is not whether robots can give Nvidia stock another good week.

It is whether Nvidia can do for robotics what it has already done for artificial intelligence: create the computing platform that an entire new technology industry builds on.

Nvidia and LG Are Taking the Robotics Partnership to the Next Level

The headline development is Nvidia’s expanding relationship with LG Group.

LG plans to unveil a next-generation bipedal humanoid robot in the first quarter of 2027. The robot is expected to use Nvidia’s Isaac GR00T foundation model for robotics together with the Jetson Thor computing platform. The two companies are also collaborating on wheel-based robots, AI infrastructure and advanced computing for future vehicles.

This is not a partnership that appeared overnight. In June, [Reuters reported that Jensen Huang and LG were already working together on humanoid robots and future data centers]. Huang specifically pointed to collaboration involving motor technology and mechanical systems—two areas that matter enormously when software intelligence must translate into reliable physical movement.

Nvidia and LG are approaching the problem from complementary directions. Nvidia brings AI models, accelerated computing, simulation technology and edge-processing hardware. LG brings manufacturing experience, electronics, sensing systems, batteries, actuators and access to real-world environments where robots can eventually be tested and deployed.

Why Nvidia Is Really Betting on “Physical AI”

The phrase to remember is physical AI.

Traditional artificial intelligence operates mainly in digital environments. It generates text, analyzes data, creates images or makes recommendations. Physical AI adds another challenge: the system must understand the physical environment and take actions inside it.

A robot needs to identify objects, understand spatial relationships, interpret instructions, plan movements, adjust to changing conditions and avoid creating unsafe situations. Doing that reliably requires far more than putting a language model inside a machine.

That is where Nvidia sees an opportunity.

At GTC 2026, Nvidia described an increasingly broad physical-AI ecosystem involving companies such as ABB Robotics, Agility Robotics, FANUC, Figure, KUKA, Universal Robots and Yaskawa. Its strategy combines Cosmos world models, Isaac robotics frameworks, GR00T models, Jetson edge hardware and simulation technologies. Nvidia’s [official physical AI announcement offers a useful look at the scale of this ecosystem].

The strategy resembles Nvidia’s approach to generative AI. Instead of trying to build every application itself, Nvidia provides much of the underlying infrastructure used by developers and manufacturers.

In other words, Nvidia does not necessarily need to win the humanoid robot race.

It could benefit if many different robot manufacturers win while using Nvidia technology.

Isaac GR00T Could Become a Key Part of the Robot “Brain”

One of the most interesting pieces of Nvidia’s strategy is Isaac GR00T.

GR00T belongs to a growing category of models designed to help robots perceive their environment, reason about what they see and translate instructions into physical actions. Nvidia’s robotics platform also combines these models with simulation tools so developers can train and validate systems before sending them into factories, warehouses, commercial spaces or homes.

That simulation layer is especially important.

Training a chatbot on another million examples is mostly a compute problem. Training a real robot by allowing it to make millions of physical mistakes can become expensive, slow and unsafe.

Simulation offers another route.

Nvidia Isaac Sim and related tools allow developers to reproduce environments digitally, generate training data and test robot behavior before deployment. Nvidia and LG, for example, have described plans to use Isaac Sim and Isaac Lab to train and validate robotic systems in physically accurate virtual environments.

The Bigger Opportunity Is an Entire Robotics Stack

This is where Nvidia’s robotics strategy becomes particularly interesting from a business perspective.

Think of a modern intelligent robot as a stack.

It needs powerful models to reason. It needs simulation environments for training. It needs synthetic data. It needs sensors. It needs local computing hardware capable of running AI with low latency. It needs software connecting perception with movement. And organizations need infrastructure capable of training, updating and monitoring those machines at scale.

Nvidia increasingly wants to participate across this entire workflow.

The company’s [official announcement detailing its AI factory collaboration with LG] describes an integrated system covering AI model development, physical-AI data generation, robot simulation, training, edge deployment and digital twins.

That matters because the most durable business opportunity may not come from selling a single robot. It could come from providing the technology required to develop and operate thousands of different robot models.

The parallel with cloud computing is useful. The biggest cloud businesses did not have to predict every successful software application. They benefited from supplying infrastructure to almost all of them.

Robots Are Moving Beyond the Demo Stage—But There Is Still Plenty to Prove

The excitement surrounding humanoid robots can make it seem as though general-purpose robotic workers are about to appear everywhere.

Reality is more complicated.

Commercial robotics is progressing, particularly in factories, logistics and structured industrial environments. But truly general-purpose humanoid robots still face substantial challenges involving dexterity, reliability, cost, batteries, safety and autonomous decision-making.

Even advanced systems can struggle when confronted with unfamiliar physical situations. Unlike a factory robot repeating the same movement in a controlled cell, a general-purpose robot may need to interact with irregular objects, unpredictable people and constantly changing environments.

That distinction is important for anyone evaluating the business opportunity.

A compelling demonstration is not the same thing as a commercially reliable deployment.

Recent reporting from the [Associated Press on Nvidia’s work with Fujitsu, Fanuc, Yaskawa and Kawasaki Heavy Industries] shows why industrial applications may be among the strongest early markets. Factories, hospitals and other organizations can use physical AI to address specific operational challenges while maintaining more controlled environments than the average home.

Why Robotics Could Matter to Nvidia Stock

Nvidia’s current business is still overwhelmingly tied to the AI computing infrastructure boom rather than humanoid robots. Robotics should therefore be viewed as a longer-term opportunity rather than the primary explanation for Nvidia’s present financial scale.

But markets tend to price companies partly on expectations about future growth.

For Nvidia, robotics adds another potentially enormous market on top of data-center AI, enterprise AI, autonomous vehicles and accelerated computing.

Jensen Huang has repeatedly characterized humanoid robotics as a potentially multitrillion-dollar economic opportunity. Barron’s cited an RBC projection envisioning approximately 350 million robots sold annually at roughly $25,000 each by 2050—a scenario that would imply a market approaching $9 trillion per year. Such long-range forecasts should be treated as scenarios rather than guarantees, but they demonstrate why Wall Street is paying attention.

There is another reason the opportunity is strategically attractive.

Nvidia does not necessarily need robotics revenue to come from one product.

Revenue could eventually be distributed across training infrastructure, simulation, chips, edge computing, software and the wider development ecosystem. That gives Nvidia multiple potential ways to participate if physical AI scales.

Nvidia Is Not Betting on Robots Alone—it Is Betting on an Ecosystem

Another important piece of Nvidia’s strategy is diversification.

LG is only one robotics partner.

At GTC 2026, Nvidia highlighted integrations and partnerships spanning major industrial robot manufacturers, humanoid developers and companies building generalized robotic intelligence. Established industrial companies collectively operate enormous fleets of robots, making integrations with firms such as FANUC, ABB, Yaskawa and KUKA strategically significant.

The company has also worked with emerging humanoid-robot developers. Nvidia previously invested in Figure AI, and its newer robotics initiatives have included collaboration across an expanding ecosystem of robot developers.

This helps explain why Nvidia’s robotics push may be more defensible than simply betting on one futuristic machine.

It is the classic platform strategy: provide useful technology to as many participants as possible.

What Businesses Should Take Away From Nvidia’s Robotics Push

For business leaders, Nvidia’s strategy is a signal that physical AI deserves a place on the technology roadmap—even if buying hundreds of humanoids tomorrow would be premature.

Organizations can begin by identifying workflows where intelligent machines could provide measurable value. Manufacturing inspection, material movement, warehouse operations, repetitive production processes and work in difficult environments are logical areas to evaluate because automation already has an established business case in many of them.

Companies should also think beyond the robot itself.

The organizations most prepared for physical AI will need clean operational data, digital representations of facilities, edge infrastructure, cybersecurity controls, integration with enterprise software and workers capable of supervising intelligent machines.

This is why developments such as Nvidia’s partnership with LG matter even to companies that have no intention of manufacturing a humanoid robot.

The industry is beginning to build the infrastructure layer for intelligent machines.

Conclusion

Nvidia’s push into robotics signals that the company is thinking well beyond the current generative AI boom. By combining advanced AI models, simulation platforms, edge computing and robotics hardware, Nvidia is positioning itself as a foundational technology provider for the emerging physical AI economy.

The company’s partnerships with LG and major robotics manufacturers show that this strategy is moving beyond research demonstrations and toward real-world industrial applications. While humanoid robots still face significant challenges around cost, reliability, safety and large-scale deployment, the long-term opportunity could be substantial if intelligent machines become commonplace across manufacturing, logistics, healthcare and other industries.

For investors, the key takeaway is that Nvidia’s robotics strategy is less about building one successful robot and more about supplying the technology stack that many robotics companies may depend on. That platform approach helped Nvidia become central to the generative AI revolution, and it could give the company another powerful growth opportunity as AI moves from digital environments into the physical world.

Businesses should watch this transition closely. Physical AI may still be in its early stages, but organizations that begin exploring automation, simulation, robotics infrastructure and AI governance today could be better prepared as the technology matures.

Ultimately, Nvidia’s robot bet is about something much bigger than a short-term rise in its stock price. It is a wager that the next major chapter of artificial intelligence will involve machines that can see, reason, move and work alongside people in the real world. If that future develops as Nvidia expects, robotics could become one of the company’s most important growth stories of the coming decade.

Artificial intelligence has spent the past several years getting bigger. Bigger models. Bigger GPU clusters. Bigger data centers. Bigger electricity bills.

Researchers at Northwestern University are exploring a very different idea: What if smarter AI hardware could simply do less work?

A new brain-inspired AI chip modeled on the human cerebellum has demonstrated the ability to identify unexpected events while requiring roughly 10,000 times fewer calculations than a conventional AI comparison system. Instead of continuously processing every incoming signal with the same intensity, the experimental hardware is designed to largely ignore what is predictable and concentrate computing resources on what changes.

That is a deceptively simple idea—and potentially a very important one.

The underlying research, published in [Nature Communications] describes a cerebellum-inspired memtransistor architecture for hardware-efficient novelty detection. The researchers reported more than 98% accuracy when identifying abnormal electrocardiogram patterns, with their approach detecting an irregular heartbeat before the heartbeat had finished.

There is one important clarification behind the headline. The researchers did not demonstrate a robot completing a physical motor-control benchmark with 10,000 times fewer operations. Rather, the hardware takes inspiration from the cerebellum, a brain region deeply involved in coordination, timing, balance and rapid responses, and applies similar principles to an ECG novelty-detection task. Coverage from [Live Science] likewise reports that the experimental system required around 10,000 times fewer calculations than the conventional AI approach used for comparison.

That distinction does not make the research less interesting. If anything, it helps reveal why brain-inspired computing could become such an important part of the next generation of edge AI, robotics, healthcare devices and autonomous systems.

Why AI Researchers Are Looking at the Cerebellum

Most conversations about artificial intelligence understandably focus on the parts of the brain associated with learning, reasoning, language and memory. But the Northwestern researchers turned their attention to something different: the cerebellum.

The cerebellum is exceptionally good at dealing with patterns that repeat over and over again. When an event behaves as expected, the brain does not need to treat it as a five-alarm emergency every single time. Instead, neural circuits can devote greater attention to deviations from the expected pattern.

Northwestern Engineering describes the approach in its article [AI Gets a Cerebellum]. Researchers engineered the device so that complementary excitatory and inhibitory behaviors work together to distinguish ordinary signals from genuinely novel ones.

Think of it like a security guard watching an empty hallway.

A traditional system might repeatedly analyze every frame: floor, wall, door, floor, wall, door—thousands upon thousands of times.

A novelty-focused system effectively asks: Did anything important change?

If the hallway looks exactly as expected, there may be little reason to throw maximum computing power at it. When the door suddenly opens, however, the system springs into action.

For always-on AI systems, that shift could be enormously valuable.

How the Brain-Inspired AI Chip Achieved 10,000x Fewer Calculations

To test the concept, researchers used electrocardiogram data containing both regular heart rhythms and arrhythmias.

Normal heartbeats represented predictable events. Arrhythmias represented novelty.

Rather than extensively analyzing every normal heartbeat, the experimental architecture focused its computational response on deviations from the expected pattern.

The result was striking.

The system achieved more than 98% accuracy, detected an abnormal heartbeat within roughly one-fifth of a heartbeat and operated more than twice as quickly as the conventional AI comparison, according to Northwestern and subsequent reporting. The simulated cerebellum-inspired network also required approximately 10,000 times fewer computational operations than the transformer-based comparison used by the researchers.

But that number needs context.

It does not mean that this experimental chip suddenly makes every AI workload 10,000 times more efficient. It was demonstrated on a specialized novelty-detection problem, and real-world performance will depend on hardware scaling, manufacturing, system integration, workload characteristics and other factors. The researchers have also noted that memtransistors have not yet been scaled to commercial silicon-chip dimensions.

So, no—the GPU is not packing its desk just yet.

What the result demonstrates is something potentially more fundamental: AI systems designed around event-driven, selective computation may be able to avoid huge amounts of unnecessary work.

Why This Matters for Edge AI

The most obvious opportunity may be edge computing.

Today, sophisticated AI workloads frequently rely on remote cloud infrastructure. A device gathers information, transmits some of that information to a server or data center, waits for computation and receives a response.

That model works well for many applications.

It is less appealing when a system needs to react almost instantly.

A wearable heart monitor should not need a hyperscale data center to decide whether a heartbeat looks unusual. A robot attempting to maintain balance cannot afford unpredictable network latency every time its environment changes. An autonomous machine encountering an unexpected obstacle may need to react immediately.

Local, energy-efficient inference could therefore become an important complement to cloud AI.

Wearable Healthcare Could Be an Early Opportunity

Healthcare monitoring is another compelling possibility because biological signals frequently involve long stretches of routine activity interrupted by events that matter.

A wearable ECG monitor, for example, may collect enormous amounts of normal heartbeat data before encountering an abnormal rhythm.

Instead of applying the same computational effort to every heartbeat, a future event-driven device could reserve additional processing for signals that depart from an expected pattern.

That could potentially support devices that are smaller, faster and less dependent on constant cloud connectivity.

Local processing may offer another benefit: privacy.

Keeping sensitive health information on a device can reduce the amount of personal data that needs to be continuously transmitted to external infrastructure. It does not automatically make a product private or secure—developers still need strong security, access controls and governance—but edge processing can reduce unnecessary data movement.

The research therefore sits at the intersection of several growing AI priorities: efficiency, latency, privacy and continuous monitoring.

Brain-Inspired Hardware Could Give Robotics Faster Reflexes

Robotics may be where the cerebellum analogy becomes especially intuitive.

A sophisticated robot might use a powerful general AI model for planning while simultaneously relying on smaller, specialized systems for rapid local responses.

The large model might answer:

“What should I do?”

A cerebellum-inspired subsystem could answer:

“Something unexpected just happened—react now.”

Imagine a warehouse robot carrying a container when an object unexpectedly falls into its path. Continuously running a massive model at maximum intensity would be computational overkill for every millisecond of operation.

A specialized novelty detector could instead monitor routine sensor patterns and activate more computational resources only when something changes.

Smaller Computation Does Not Automatically Mean Smaller Responsibility

More efficient AI hardware introduces an interesting policy question.

If AI becomes dramatically cheaper to run, it can spread into many more devices.

Wearables. Cars. Robots. Cameras. Industrial equipment. Medical sensors. Consumer electronics.

That makes responsible deployment even more important.

For healthcare applications, developers would need extensive validation before relying on novelty-detection hardware for clinical decisions. For autonomous machines, unusual-event detection would require rigorous safety testing because an event wrongly classified as “routine” could be more consequential than one mistakenly flagged as unusual.

Organizations would also need to think carefully about transparency.

If an event-driven system intentionally ignores predictable information, designers should understand what determines “predictable,” how those expectations change over time and whether important minority or edge-case events could be overlooked.

Energy-efficient AI is desirable, but efficiency should complement reliability, auditability and safety—not replace them.

The Biggest Challenge: Turning a Research Breakthrough Into a Commercial Chip

This technology remains experimental.

The Northwestern researchers have demonstrated device behavior and simulated networks based on measured memtransistor properties, but commercialization would require substantial additional progress.

Manufacturers would need reliable processes for producing large numbers of devices. Engineers would need to integrate them into practical computing architectures. Developers would need tools for programming them. Researchers would need to determine which workloads benefit most from the cerebellum-inspired approach.

And perhaps most importantly, future versions need to become more adaptive.

The human cerebellum does not permanently treat every unusual event as unusual. When something that was once surprising happens repeatedly, biological systems can update their expectations.

Northwestern’s researchers have identified this adaptive behavior as an area for further study. The eventual goal is not simply a device that detects novelty, but one that can learn what deserves attention as its environment changes.

That could turn novelty detection from a clever hardware feature into something closer to an artificial reflex system.

What This Brain-Inspired AI Chip Means for the Future of AI

For years, much of the AI industry has operated under a straightforward philosophy: when better performance is needed, add more computation.

This research suggests another path.

Do not calculate everything faster.

Calculate fewer things.

The new brain-inspired AI chip shows how hardware modeled on the cerebellum could distinguish routine information from unexpected events and concentrate computing resources where they matter most.

The roughly 10,000x reduction in calculations demonstrated in the ECG experiment should not be interpreted as a universal replacement for GPUs, transformers or conventional processors. This is early-stage hardware research focused on a specialized problem.

But specialized problems are everywhere.

Wearables constantly monitor biological signals. Robots constantly process sensor streams. Industrial systems constantly watch machinery. Autonomous devices constantly evaluate their environments.

Much of that data is ordinary.

If future AI systems can learn when not to compute, the impact could extend far beyond one experimental chip.

For businesses, researchers and technology leaders, the lesson is worth watching: the next major AI breakthrough may not come from building the biggest model or the largest data center.

It may come from borrowing one of biology’s oldest tricks—pay attention only when something changes.

Conclusion

The new brain-inspired AI chip shows that the future of AI may depend not only on more computing power, but on smarter, more selective computing. By dramatically reducing unnecessary calculations, cerebellum-inspired hardware could help make AI faster, more energy-efficient, and better suited for edge devices, robotics, and healthcare applications. While the technology is still experimental, it offers a promising glimpse of how biology could inspire the next generation of intelligent hardware.

Artificial intelligence may live in the cloud, but the infrastructure supporting it is becoming very, very physical.

Behind every generative AI assistant, enterprise copilot and increasingly sophisticated AI model sits an expanding network of GPUs, servers, networking equipment, cooling systems, electricity connections and enormous data centres. Big Tech is racing to secure that capacity before competitors do—and the financial commitments are becoming extraordinary.

According to an [August 2026 Reuters analysis] Microsoft, Meta Platforms, Oracle, Amazon and Alphabet had collectively committed approximately $1.09 trillion in payments under leases that had not yet commenced, mostly associated with data centres required for the AI expansion. That compares with roughly $285 billion of lease liabilities already recognized on their balance sheets.

And the number is still moving. Reuters reported that Meta subsequently signed another approximately $68 billion of data-centre leases in July, pushing the five companies’ known pipeline to roughly $1.16 trillion when those later agreements are included.

The AI arms race is therefore becoming more than a contest over who develops the smartest model. It is becoming one of the largest infrastructure and financing competitions in modern technology.

Why Big Tech Is Locking Up AI Data Centres Years in Advance

The logic behind the spending is straightforward: AI computing capacity cannot simply be switched on overnight.

Modern AI facilities require land, high-capacity grid connections, sophisticated cooling technology, advanced semiconductors, high-bandwidth memory, fiber connectivity, networking equipment and enormous amounts of electricity. Designing, permitting and connecting a major facility can take years.

That creates an unusual strategic problem. If technology companies wait until AI demand is obvious, the power, property and equipment they need may already be spoken for.

The spending numbers reinforce that view. Earlier in 2026, a [Reuters report on Bridgewater Associates’ analysis] estimated Alphabet, Amazon, Meta and Microsoft could collectively invest around $650 billion in AI-related infrastructure during 2026, compared with approximately $410 billion in 2025. Bridgewater argued that demand for compute continued to outstrip supply, encouraging hyperscalers to invest aggressively.

That is a remarkable transformation for companies traditionally celebrated for asset-light, high-margin software economics.

AI is turning portions of Big Tech into something closer to a digital infrastructure industry.

The $1 Trillion Number Needs an Important Accounting Footnote

Calling the entire $1.09 trillion figure “debt” would be misleading.

According to Reuters, signed leases generally do not become recognized lease liabilities until the relevant facility is available for use. Until then, future payment commitments are disclosed within notes to financial statements. The $1.09 trillion figure also represents largely undiscounted payments spread across many future years, while recognized lease liabilities reflect present values.

So investors should not simply add $1 trillion to Big Tech’s reported debt.

But they should not ignore the number either.

Once a company commits to a long-duration facility, it has effectively made a major bet on future demand. A data-centre lease that lasts 15 or 20 years represents a very different kind of commitment from ordering additional cloud capacity for a few months.

This is where the AI infrastructure race becomes financially interesting: today’s leases are being signed against forecasts of computing demand that may extend deep into the 2030s.

Microsoft Leads the Disclosed Lease Pipeline

Microsoft had the largest disclosed pipeline covered by Reuters, at approximately $329.1 billion in uncommenced leases, compared with $88.52 billion in recognized lease liabilities.

There is an obvious reason Microsoft is willing to keep spending: demand has been substantial. In its latest results, Microsoft disclosed a $678 billion cloud backlog, while Azure continued posting strong growth. Reuters reported that Azure revenue grew 43% in Microsoft’s fiscal fourth quarter of 2026.

Read the [Reuters analysis of Microsoft’s latest cloud results and AI spending] and the strategic calculation becomes clearer: Microsoft is spending aggressively because management believes insufficient capacity could constrain future AI and cloud revenue.

That does not make every investment automatically profitable. It does, however, explain why doing too little may look nearly as dangerous to Microsoft as doing too much.

Meta Is Building for an AI-First Future

Meta disclosed approximately $278.99 billion in uncommenced operating and finance lease payments, before signing a further roughly $68 billion of data-centre leases in July.

Unlike Microsoft, Amazon and Alphabet, Meta does not have a hyperscale public-cloud business comparable with Azure, AWS or Google Cloud. Its return on AI infrastructure therefore depends heavily on improvements in advertising, recommendations, engagement, AI products and whatever new business models emerge around increasingly capable AI systems.

Meta’s Q2 2026 capital expenditures, including principal payments on finance leases, reached $31.08 billion, illustrating the scale at which infrastructure is already flowing through its financial results. The company’s official [Q2 2026 earnings release] provides additional detail.

The question for Meta is no longer whether AI can improve its existing products. It is whether those improvements—and future AI businesses—can generate enough economic value to justify infrastructure commitments measured in hundreds of billions of dollars.

Oracle Shows Why Lease Duration Matters

Oracle may represent the clearest illustration of the financial risk.

The company disclosed roughly $260 billion in uncommenced commitments, nearly seven times its recognized lease liabilities of approximately $37.89 billion. Many of those data-centre leases are expected to begin between fiscal 2027 and fiscal 2029 and generally run for 15 to 19 years.

That duration matters because customers may not commit for anything close to the same period.

A separate [Reuters examination of Oracle’s AI strategy] highlighted the potential mismatch between long-term data-centre leases and customer agreements that may last only several years. Oracle’s leverage and infrastructure spending have consequently attracted particularly close attention from ratings agencies.

Think of it like renting a giant building for 19 years because a customer promises to use several floors for five. If that customer leaves after year five, the building does not magically disappear from the lease agreement.

That mismatch is one of the biggest financial questions surrounding the AI infrastructure boom.

The Bear Case: What Happens If AI Demand Falls Short?

The danger is that infrastructure commitments are long-lived while technology trends can change extremely quickly.

Several things could disrupt today’s assumptions.

More efficient models could require less computing power per task. Specialized chips could substantially lower inference costs. Customers could resist premium AI pricing. Competition could compress cloud margins. Enterprises could slow deployments after discovering that some AI projects do not produce sufficient returns.

There is also the risk of hardware obsolescence.

An expensive data centre can remain useful for decades, but the processors inside it may become economically outdated much faster. As newer chips produce more AI output for the same amount of electricity, operators may need to refresh hardware more aggressively to remain competitive.

That means success depends not simply on keeping facilities full, but on generating enough revenue and margin from them before the economics change.

Sustainability and AI Governance Can No Longer Be Separated From Infrastructure

The environmental debate around AI increasingly extends beyond whether an individual model is efficient.

The larger question is what happens when millions of people and businesses use AI continuously.

More computation means more servers. More servers require more electricity and cooling. And dramatically expanding power demand can affect surrounding communities as well as corporate sustainability goals.

The IEA expects renewables to supply a significant portion of additional data-centre electricity demand, while natural gas and other sources also contribute.

That means responsible AI strategies should increasingly include infrastructure governance: energy efficiency, water consumption, siting decisions, grid effects, lifecycle emissions and transparent reporting.

In other words, responsible AI is not only about what a model says. It is also about the physical systems required to keep that model running.

What the $1 Trillion AI Lease Burden Means for Business Leaders

For enterprise leaders, Big Tech’s spending spree delivers both an opportunity and a warning.

The opportunity is greater access to AI infrastructure. More data centres and compute capacity should support increasingly powerful cloud services and potentially lower the cost of running some AI workloads over time.

The warning is that enormous infrastructure spending does not guarantee that every AI implementation produces a positive return.

Businesses should continue evaluating AI projects against measurable outcomes: hours saved, costs reduced, revenue created, response times improved, errors prevented or processes automated.

Organizations should also preserve architectural flexibility. Depending entirely on one provider may create unnecessary exposure as model performance, cloud pricing and AI infrastructure economics evolve.

That is particularly important when hyperscalers themselves are making long-duration bets based on forecasts that remain uncertain.

Conclusion

The AI data-centre race and its $1 trillion lease burden show just how dramatically artificial intelligence is reshaping the economics of Big Tech. Microsoft, Meta, Oracle, Amazon, and Alphabet are no longer competing only on models, software, or cloud services. They are competing for land, electricity, GPUs, cooling capacity, and long-term access to the infrastructure required to keep AI systems running at global scale.

For now, strong cloud growth and rising enterprise AI adoption provide a credible case for aggressive investment. But the scale and duration of these commitments also introduce new risks. Long-term leases, rapidly changing hardware, uncertain customer demand, and growing pressure on energy grids mean that infrastructure strategy will be just as important as AI innovation itself.

The winners of the next phase of AI may therefore be the companies that balance ambition with discipline. Building more capacity is important, but so is ensuring that every dollar invested can eventually translate into sustainable revenue, efficient operations, and measurable customer value.

For business leaders and investors, the message is clear: the AI boom should not be judged only by how quickly companies can build data centres. It should be judged by how effectively they can turn that infrastructure into lasting economic value.

As AI continues to evolve, the biggest competitive advantage may not simply be having the most powerful model. It may be having the right combination of compute, energy, capital, customers, and long-term strategic discipline.