Artificial intelligence has spent years getting better at thinking inside computers. Now one of its most important challenges is learning how to work outside them.

That transition—from software that generates answers to machines that can perceive, reason and physically act—is driving the rapidly expanding field of Physical AI. And a new European partnership suggests that the race is moving beyond impressive robot demonstrations toward something much harder: industrial-scale production.

Germany-based NEURA Robotics and Italian embedded-computing specialist SECO announced a strategic partnership designed to industrialize and scale Physical AI from Europe. Under the agreement, SECO will help design, engineer and manufacture electronic compute modules for NEURA’s cognitive robots, including its humanoid 4NE1. The systems will incorporate Qualcomm Dragonwing processors as part of NEURA’s distributed computing architecture. The full partnership details are available in the [Seco] 

That might initially sound like a hardware supply agreement.

It is considerably more interesting than that.

The partnership links robotics, edge AI computing, electronics manufacturing, industrial data and semiconductor production into what could become a broader European Physical AI ecosystem. And it illustrates something increasingly important about the robotics market: building an intelligent machine is no longer just a robotics problem. It is a full-stack computing, data and manufacturing problem.

The NEURA–SECO Partnership Is About Turning Robotics Into an Industry

Robotics companies are very good at producing videos that make the future look approximately three weeks away.

Commercial-scale manufacturing is less cinematic.

A robot intended to operate reliably in a factory needs industrial-grade computing, sensing, power management, safety systems, real-time control, networking, thermal engineering, durable electronics and a supply chain capable of producing the same system repeatedly.

That is where SECO enters the picture.

SECO specializes in embedded computing, edge AI and industrial electronics. Under the partnership, the company will engineer and manufacture compute modules for NEURA robots, helping move those systems from advanced platforms toward repeatable series production in Europe. Independent coverage from [Intelligent CIO] highlights the same industrialization angle: NEURA contributes cognitive robotics and Physical AI technology, while SECO provides the engineering and manufacturing capability required to turn that intelligence into scalable hardware.

That distinction matters.

There is a huge difference between building 20 sophisticated robots and reliably producing thousands of them. Manufacturing scale introduces entirely new constraints involving component availability, quality assurance, power efficiency, maintainability and cost.

The NEURA Robotics and SECO Physical AI partnership is therefore partly an attempt to solve the decidedly unglamorous problems that separate a great prototype from a viable industry.

And those problems may ultimately determine which robotics companies survive.

NEURA’s “Brain + Nervous System” Approach Could Be the More Important Story

Traditional robots often resemble centralized machines: sensor information travels to one principal computing system, decisions are made, and commands travel back to the motors and actuators.

NEURA is pursuing a more distributed architecture.

The company describes its approach as a “Brain + Nervous System” architecture, supported by its Smart Limb concept. Instead of concentrating every computational decision in one place, sensing and processing can be distributed throughout the robot and positioned closer to the joints, sensors and mechanisms where physical actions occur.

Think less “one computer operating a robot” and more “an intelligent network spread throughout the machine.”

There is a practical reason for that approach.

A cognitive robot may be reasoning at several different timescales simultaneously. High-level intelligence might determine that a component needs to be picked up and moved. Local systems must then handle much faster tasks such as maintaining balance, interpreting force, controlling a joint or reacting immediately when unexpected contact occurs.

Smart Limbs Bring AI Closer to Physical Action

Smart Limbs are particularly interesting because they point toward a different way of thinking about robot intelligence.

Imagine a worker catching a falling cup.

The brain does not need to consciously calculate every muscle movement from scratch. Different parts of the nervous system coordinate perception and physical response extremely quickly.

Robotic systems obviously work very differently from biological bodies, but the architectural principle offers a useful analogy. Distributed intelligence can allow individual robot subsystems to handle tightly constrained real-time tasks while a higher-level system concentrates on more complex reasoning.

That creates several potential advantages.

It can reduce latency, limit the amount of sensor information constantly traveling to a central processor, support modular hardware designs and improve responsiveness when robots operate close to people.

Qualcomm’s [Qualcomm] shows how seriously the broader robotics industry is taking this hardware challenge. The reference platform integrates computing, sensing, networking and software with the explicit goal of helping developers move from prototypes toward deployable robots. NEURA is listed among the ecosystem participants supporting the platform.

The Semiconductor Industry Could Become a Training Ground for Physical AI

The partnership becomes even more strategically interesting when you look beyond NEURA’s own robots.

NEURA and SECO plan to collaborate on Physical AI systems for semiconductor and electronics manufacturing, using real-world production environments to generate industrial data and develop new automation capabilities.

That creates a potentially powerful feedback loop.

Robots operate inside real production environments. Their interactions produce data. That data can help developers improve perception, manipulation and decision-making. New skills can then be refined, validated and deployed across additional robots and manufacturing sites.

The objective is not simply to program one machine to repeat one movement forever.

It is to create reusable intelligence.

Coverage from [Unite] emphasizes this connection between European manufacturing, robot compute modules and scalable Physical AI infrastructure.

This is also why real-world industrial data is becoming so valuable.

Generative AI could learn enormous amounts from material already stored on the internet. Physical AI does not have the same luxury. A humanoid learning to manipulate irregular components, respond to changing factory conditions or recover from unsuccessful actions needs data describing physical interactions—not just text describing them.

NEURA Gym Could Help Build the Data Flywheel

NEURA and SECO are also discussing the creation of a NEURA Gym in Italy, which would become NEURA’s first training hub in Southern Europe.

The name may sound like a place where humanoids finally get serious about leg day, but the concept addresses one of robotics’ biggest bottlenecks: training.

Physical AI systems need experience.

Developers require environments where robots can practice tasks, collect interaction data, validate new capabilities and improve models before those systems enter production environments.

Real-world training hubs can complement simulation by providing data involving friction, imperfect components, unexpected obstacles, human behavior and other details that digital environments may struggle to reproduce perfectly.

That creates what the AI industry increasingly describes as a data flywheel: deploy systems, observe what happens, learn from those interactions, improve the models and redeploy the improvements.

NEURA’s broader Neuraverse vision extends the idea further. Skills developed through one robot or application could eventually become reusable capabilities available to other machines rather than remaining isolated inside a single custom automation project.

This shift from isolated automation toward reusable physical intelligence could be one of the most commercially important developments in robotics.

Why Europe Matters to This Partnership

The words “from Europe” in the announcement are not decorative.

The partnership is explicitly framed around developing and manufacturing important parts of the Physical AI stack in Europe.

Europe already possesses several ingredients that could make it unusually competitive in Physical AI: world-class manufacturing, industrial automation, automotive engineering, robotics research, sensors, embedded systems and enormous quantities of specialized industrial knowledge.

Where Europe has struggled more visibly is in the enormous centralized computing infrastructure and frontier AI platforms that have made American technology companies dominant during the generative-AI boom.

Physical AI changes the competitive equation somewhat.

A robot in an automotive factory or semiconductor facility requires more than a giant language model. It needs mechanical engineering, embedded electronics, real-time systems, industrial data, sensors, manufacturing capabilities, safety expertise and close integration with real production workflows.

What Businesses Should Watch Next

For executives, manufacturers and technology leaders, the significance of the NEURA–SECO agreement extends beyond whether 4NE1 becomes the dominant humanoid robot.

The partnership reveals several signals worth watching:

• Physical AI is moving toward industrialization. The conversation is increasingly about manufacturing capacity, embedded computing and repeatable deployment rather than demonstrations alone.
• Edge computing will be critical. Robots need local intelligence capable of responding quickly even when cloud connectivity is limited.
• Real-world operational data may become a competitive asset. Companies that can responsibly turn production experience into reusable robotic skills could create powerful feedback loops.
• Ecosystems may matter more than individual vendors. Robotics increasingly requires collaboration among AI developers, semiconductor companies, manufacturers, integrators and industrial customers.
• Europe has a plausible Physical AI advantage. Its industrial base, robotics expertise and engineering infrastructure could allow it to compete differently from regions dominating cloud-based generative AI.

That last point deserves particular attention.

The Physical AI race may not be won by the company capable of producing the most entertaining humanoid demonstration.

It may be won by the ecosystem that can manufacture robots reliably, train them on valuable real-world tasks, update them safely and repeat that process thousands of times.

Conclusion

The partnership between NEURA Robotics and SECO shows how quickly Physical AI is moving from research and prototypes into real-world manufacturing. By combining NEURA’s intelligent robots with SECO’s embedded computing and production expertise, the two companies are working to make advanced robotics easier to build, scale, and deploy across Europe.

The bigger story is that AI is no longer limited to software on a screen. It is becoming part of machines that can see, understand, move, and work alongside people. For businesses, this could mean smarter factories, more flexible automation, and new ways to improve productivity.

Europe also has an opportunity to play a major role in this shift because of its strong manufacturing, robotics, and engineering industries. If partnerships like NEURA and SECO succeed, Europe could become an important center for the development of safe, reliable, and scalable Physical AI.

Ultimately, the future of AI may not only be about systems that can answer questions. It may also be about intelligent machines that can understand the physical world and take useful action within it.

Artificial intelligence is beginning to change the economics of the global technology services industry, and Wipro has just provided one of the clearest examples yet of how significant that change could become.

[Reuters] the Indian IT services company says its artificial intelligence initiatives have generated productivity improvements equivalent to the work of approximately 20,000 employees. Importantly, Wipro says this does not mean 20,000 people were simply replaced. Instead, employees whose capacity was freed by AI have been redeployed to other projects, trained for different roles, or tasked with managing AI-powered systems.

With Wipro employing roughly 243,000 people as of June 2026, the productivity capacity being discussed is equivalent to about 8% of its workforce. That makes Wipro AI productivity an important case study in how enterprise AI may reshape work without necessarily translating automation directly into one-for-one job elimination.

Wipro Is Moving Toward a Human-AI Operating Model

Wipro CTO Sandhya Arun described the company’s direction as a move toward a “human-AI operating model.” Rather than separating employees and artificial intelligence into competing groups, this model treats AI agents as tools or digital workers that employees can supervise, coordinate and use to increase their own output.

An engineer who once handled a collection of repetitive development or support tasks manually, for example, could increasingly oversee multiple AI agents performing portions of that work. The employee can then spend more time on architecture, client requirements, quality control, decision-making and other higher-value activities.

The strategy did not appear overnight. Wipro launched its [Wipro] in 2023 alongside a commitment to invest $1 billion in artificial intelligence capabilities over three years. The initiative was designed to integrate AI across Wipro’s internal operations, platforms, solutions, research and client services while simultaneously expanding AI training across the workforce.

The latest productivity figures suggest those investments are now moving beyond experimentation and beginning to influence how work is actually delivered.

AI Training Is Becoming as Important as AI Technology

Perhaps the most important part of Wipro’s strategy is not the AI itself but what the company is doing with its people.

Reuters reports that more than 100,000 Wipro employees have received advanced AI-related training and certifications as the company expands its human-AI model.

That matters because the next phase of enterprise AI is likely to require more than knowing how to use a chatbot. Employees may increasingly need to understand how to supervise agents, validate AI-generated output, design AI-enabled workflows, connect models with enterprise systems and determine when human intervention is necessary.

This aligns with [IBM] IBM argues that businesses seeing the greatest value from AI are increasingly redesigning how their organizations operate rather than simply deploying more AI tools.

AI Is Also Changing the IT Services Business Model

Wipro’s productivity gains matter far beyond its own workforce.

India’s enormous technology services sector has traditionally relied heavily on people-based delivery models. In simple terms, large teams of engineers and consultants perform work for clients, while contracts frequently reflect the amount of talent and time required to deliver it.

AI challenges that structure.

A separate [Reuters] found that major providers including Wipro, Tata Consultancy Services, Infosys, HCLTech and Cognizant are increasingly facing customers who expect faster delivery and greater productivity for lower costs. Contracts are consequently moving toward business outcomes rather than simply hours worked.

That shift creates both an opportunity and a challenge.

If AI allows an IT provider to complete the same project faster with fewer manual hours, productivity improves. But clients also know AI makes those efficiencies possible, meaning they may demand a share of the savings.

The competitive advantage therefore cannot simply be, “We use AI.”

Almost every major technology services company can make that claim.

The stronger advantage will come from demonstrating that AI can produce measurable improvements in areas such as customer experience, software delivery speed, revenue generation, service quality and operating costs.

That is why Arun’s emphasis on moving from productivity to outcomes may ultimately be more significant than the headline 20,000-worker-equivalent figure.

Forward-Deployed Engineers Could Become a Critical AI Role

Another emerging piece of Wipro’s strategy involves forward-deployed engineers.

These specialists typically work closely with customers to take AI from a demonstration or prototype into an operational system. Instead of simply developing technology from a distance, they embed more deeply with client teams, understand domain-specific problems and adapt AI systems to real business environments.

Wipro is expanding this capability, following a broader movement across India’s technology services industry.

This could become increasingly valuable because enterprise AI deployment is rarely as simple as connecting a large language model to company data. Businesses must consider security, permissions, legacy technology, governance, workflow design, compliance and human oversight.

[McKinsey] similarly argues that organizations need to rethink hiring, internal capabilities and workforce structures as AI agents take on greater portions of technology work.

In other words, AI may reduce demand for certain repetitive tasks while increasing demand for professionals who can successfully turn AI capabilities into working business systems.

The Next Test for Wipro: Turning AI Productivity Into Revenue

There is still an important unanswered question.

Productivity gains demonstrate that AI can improve internal efficiency, but businesses ultimately need to convert those efficiencies into sustainable financial value.

Reuters notes that Wipro remains the only company among India’s four largest IT services providers that does not disclose AI revenue separately. Analysts cited by Reuters say the company is still relatively early in translating its AI investments into commercial returns compared with some competitors.

That makes the next phase particularly important.

If Wipro can combine AI-enabled productivity with stronger client outcomes, larger AI projects and improved margins, the 20,000-worker-equivalent capacity figure may eventually look like an early milestone rather than the main story.

If those productivity improvements primarily result in clients demanding lower prices, however, the financial impact could be more complicated.

What Wipro’s AI Push Means for Business Leaders

Wipro’s experience provides a useful blueprint for companies considering their own AI transformation.

The biggest opportunity may not come from simply reducing headcount. Instead, organizations can use AI to remove repetitive work, increase employee capacity and redirect human expertise toward higher-value problems.

That requires investment in technology, but it also requires training, workflow redesign, governance and clearly defined business outcomes.

AI adoption is increasingly becoming an organizational transformation challenge rather than purely a technology challenge.

The companies that succeed may be those that learn how to combine the speed and scalability of artificial intelligence with human judgment, domain expertise, creativity and accountability.

Conclusion

Wipro’s AI push shows that the biggest impact of artificial intelligence may not be simple job replacement, but a major shift in how work is organized. By freeing capacity equivalent to around 20,000 workers, the company is demonstrating how AI can handle repetitive tasks, improve productivity, and give employees more room to focus on higher-value work.

The real test now is whether Wipro can turn those efficiency gains into stronger client outcomes, better services, and measurable business growth. For other companies, the lesson is clear: successful AI adoption requires more than deploying new tools. It also depends on reskilling employees, redesigning workflows, and creating a model where people and AI work together effectively.

As enterprise AI continues to mature, businesses that focus on this balance between automation and human expertise may be best positioned to gain long-term value from the technology.

The age of the AI chatbot may be giving way to something considerably more powerful—and considerably more complicated.

Meta introduced Muse, a personal AI agent designed not merely to answer questions but to take actions on a user’s behalf. The Meta AI agent can connect with other apps and services, send emails, book travel, fill out forms, make purchases and continue working on tasks even after a user closes the app.

That is a substantial leap from asking an AI, “Can you draft this email?” to saying, “Handle this,” and watching the software actually do it.

According to [Meta’s official Muse announcement] the company sees Muse as part of a broader shift toward personal AI capable of turning goals into actions rather than simply producing information.

The launch puts Meta squarely into the rapidly expanding race to build agentic AI—AI systems that can reason about a goal, interact with outside software and perform a sequence of actions with less step-by-step human involvement.

And while the productivity possibilities are huge, so are the questions around permissions, privacy, security, financial transactions and accountability.

What Exactly Is Meta Muse?

Muse is Meta’s attempt to move AI from an assistant sitting inside a chat window to something closer to a digital operator.

Traditional generative AI generally waits for a question and generates an answer. An AI agent can instead interpret a goal, create a plan, use tools and execute parts of that plan.

Meta says Muse runs using its agent-focused AI technology and operates within a dedicated cloud-based environment designed to let the system interact with connected services while maintaining security controls.

That environment gives Muse access to a browser-like workspace for completing tasks across different services. The system can also continue working in the background while the user does something else.

What Can the Meta AI Agent Actually Do?

Meta is positioning Muse as a general-purpose personal agent rather than an assistant dedicated to one application.

The company says users can connect services across areas including email, calendars, payments, shopping, dining, entertainment and other digital services. From there, the agent can coordinate actions spanning multiple applications.

Muse can potentially:

• Send emails and manage communication
• Research and book travel
• Fill out online forms
• Make purchases
• Build shopping and grocery lists
• Coordinate calendars and invitations
• Track longer-term goals and projects
• Remember user preferences
• Continue certain tasks in the background
• Work across supported third-party apps rather than remaining inside one Meta product

The significance is not any one of those features. The important part is cross-application execution.

The [Associated Press coverage of Meta’s personal AI agent] highlights the growing difference between traditional conversational AI and agentic systems capable of acting across digital services.

The Privacy Question: How Much Should Your AI Know?

Muse becomes more useful as it understands more about the user.

That is simultaneously its biggest advantage and one of its biggest privacy challenges.

A personal agent might benefit from knowing your calendar, travel preferences, favorite stores, dietary requirements, purchasing habits and communication patterns.

But bringing those pieces together can potentially create an unusually detailed representation of an individual’s digital life.

Meta has emphasized privacy controls as part of the Muse design. The company outlines its approach to user data, connected services and future privacy protections in its [official product announcement].

The challenge is that personal agents require context to become genuinely useful.

An assistant that knows nothing about you cannot do much on your behalf.

An assistant that knows everything about you raises a very different set of questions.

Finding the right balance may become one of the most important product-design problems in consumer AI.

Security Concerns Haven’t Disappeared

Meta has made security one of the central themes of the Muse launch. However, independent reporting also shows why caution remains appropriate.

[Reuters’ reporting on the Muse launch] examined Meta’s development process and reported concerns arising during internal testing, including reliability and data-handling issues.

That is arguably the central tension surrounding every powerful AI agent.

An agent needs enough authority to be useful.

Every additional permission also expands the consequences when the agent misunderstands instructions, encounters malicious content, receives a deceptive prompt or simply fails.

[TechCrunch’s analysis of Muse and the consumer trust question] also focuses on a key challenge facing Meta: convincing consumers to give an AI system access to services containing valuable and highly personal information.

The technical problem is therefore only half the challenge.

The other half is earning enough trust that people will actually connect those accounts.

Why Meta Muse Matters for Businesses

Muse is launching as a consumer-focused product, but enterprise technology leaders should pay close attention.

Consumer platforms often create expectations that eventually migrate into the workplace.

Employees accustomed to telling a personal AI, “Plan this trip and book it,” will increasingly expect business software to respond to requests such as:

“Review these invoices, identify discrepancies, update the system and send me the exceptions.”

Or:

“Find available meeting times across these teams, prepare the briefing material and schedule the meeting.”

The underlying shift is from software as a collection of interfaces to AI as an orchestration layer across software.

That could significantly affect SaaS platforms, ecommerce companies, financial institutions, CRM providers, travel businesses and virtually any company whose customers currently interact through websites and apps.

Tomorrow’s customer may not always be a human clicking your interface.

Sometimes it may be an agent operating on the human’s behalf.

Is Muse the Future of Personal AI?

Possibly—but the most important part of Muse is not whether this particular product becomes the dominant personal agent.

It is what the launch tells us about where the technology industry is heading.

Meta is betting that the next phase of AI is not primarily about answering better questions.

It is about taking better actions.

The company has enormous distribution through its existing platforms, giving it a potential path to put agentic AI in front of mainstream consumers rather than limiting the technology to developers and early adopters.

[Reuters’ coverage] provides additional detail on Meta’s positioning of Muse and the broader competitive race around personal AI agents.

If agents become reliable enough, the familiar routine of opening apps, navigating menus and manually transferring information between services could gradually disappear.

But the winners will not simply be the companies whose agents can perform the greatest number of tasks.

They will be the companies whose agents users trust enough to let them perform those tasks.

Conclusion: Meta Muse Signals the Next Phase of AI

Meta’s Muse AI agent is a clear sign that artificial intelligence is moving beyond conversation and into action. The ability to access apps, send emails, coordinate tasks and participate in payments could make AI far more useful in everyday life—but it also raises the stakes for privacy, security, permissions and accountability.

For businesses, the takeaway is bigger than one Meta product. Agentic AI is becoming a new interface for how people may interact with software, services and commerce. Companies that prepare now for AI-driven workflows, secure integrations and human approval systems will be better positioned as this shift accelerates.

For consumers, the key question will be trust. The most successful AI agents will not simply be the ones that can do the most—they will be the ones people feel comfortable allowing into their inboxes, calendars, accounts and financial workflows.

Meta Muse may still be an early step, but it points toward a future where AI does more than recommend the next move. It may increasingly make that move for us.

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.