Futurist Speaker

The Neutral Sky: Why Space May Be the Only Fair Ground for AI in the Developing World

Thomas Frey — Fri, 15 May 2026 00:43:35 +0000

By Futurist Thomas Frey and Futurist Teresa Grobecker

The geopolitics of AI infrastructure has left smaller nations with no seat at the table. A constellation of orbital edge computers may be the first genuinely neutral ground they have ever had.

Every conversation about AI sovereignty eventually runs into the same wall. The compute is owned by someone. The data center is in someone’s jurisdiction. The undersea cable lands on someone’s shore. The chip was fabbed in a facility dependent on someone’s export license. For the large nations — the United States, China, the European Union, and a handful of others — this dependency chain is manageable because they sit near enough to the top of it. For the 130-odd countries that do not, the emerging AI economy looks less like an opportunity and more like a new version of a very old arrangement: powerful nations own the infrastructure, smaller ones consume the output and generate the raw material, and the terms of that exchange are set by whoever holds the hardware.

There is, however, one domain that no single nation owns, where no single company holds the cable, and where the Outer Space Treaty of 1967 still theoretically guarantees freedom of access to all: the sky above the atmosphere. And a small but rapidly maturing cluster of companies, researchers, and space agencies are beginning to ask whether orbital infrastructure — specifically, AI compute deployed at the edge in low Earth orbit — might offer developing nations something the terrestrial internet never did: a place to process their own data on genuinely neutral ground.

What Orbital Edge Compute Actually Is

Let’s be precise about what we are and are not talking about, because the gap between the vision and the current reality matters enormously for honest assessment.

Edge compute in orbit means processing data aboard a satellite rather than transmitting raw data to a ground station and then on to a terrestrial data center for analysis. The satellite carries a processor — currently something in the range of a high-end embedded system or a compact GPU module, drawing between 10 and 100 watts of power — and runs inference models, image analysis, or sensor fusion directly on the hardware in space. The processed result, rather than the raw data stream, comes down to Earth. This is the model being pursued by companies including Loft Orbital, Unibap, D-Orbit, and a growing number of national space agencies equipping small satellites with AI accelerator chips.

The honest limitation is equally important to state. A satellite running 50 watts of AI compute is an edge node, not a training cluster. It can run a pre-trained model. It can perform inference — classifying an image, detecting an anomaly, flagging a pattern. It cannot train a large language model, cannot process petabytes of data, and cannot replace the industrial-scale compute infrastructure that foundation model development requires. Anyone claiming that orbital compute solves the AI sovereignty problem for developing nations wholesale is overstating a genuine but bounded capability.

What it can do, done well, is something narrower and potentially more immediately valuable: process locally generated data, locally, without routing it through infrastructure owned and monitored by foreign powers. That is not everything. But for a great many use cases relevant to developing nations, it may be exactly enough.

Orbital AI could flip the surveillance model—letting nations process their own territorial data in space instead of exporting it to foreign-controlled systems.

The Eye in the Sky, Reconceived

The phrase “eye in the sky” has historically carried surveillance connotations — the powerful watching the powerless from above. The orbital AI model being imagined here inverts that relationship in a way worth taking seriously.

Consider what a constellation of AI-equipped small satellites, operated under a neutral or collectively owned framework, could do for the nations currently most underserved by terrestrial AI infrastructure. Agricultural monitoring at a resolution and frequency no ground-based sensor network in a low-income country could afford — crop health, soil moisture, flood inundation, pest migration — processed aboard the satellite and delivered as actionable intelligence directly to a farmer’s phone. Deforestation detection in real time, not months after the fact when the logging trucks have already gone. Supply chain monitoring for commodity exports — cocoa, coffee, minerals — that lets producing nations verify independently what is being extracted from their territory and when. Disaster response coordination that does not depend on a functioning terrestrial internet that may itself be the casualty of the disaster.

Each of these applications shares a structural property: the raw data — the satellite imagery, the sensor readings, the spectral signatures — is generated by looking at the territory of the developing nation. Under the current model, that raw data is typically transmitted to ground stations in developed nations, processed in commercial cloud infrastructure, and sold back as a service. The developing nation is, once again, the source of the raw material and the consumer of the finished product, with no ownership stake in the processing layer that creates the value.

Orbital edge compute changes the geometry. If the processing happens aboard the satellite, the raw data never needs to leave the orbital pass over the country’s own territory. The intelligence comes down. The data stays up — or rather, never comes down at all. That is a meaningful shift in data sovereignty, even if it is not a complete one.

The Neutrality Problem, Honestly Examined

Here is where the argument requires the most honest examination, because the neutrality of space is more theoretical than operational in the current environment.

The satellites in low Earth orbit are not neutral. They are owned by companies incorporated in specific jurisdictions, launched on rockets manufactured and regulated by specific governments, and operating under spectrum licenses governed by the International Telecommunication Union in processes where large nations have disproportionate influence. Starlink is American infrastructure. OneWeb has British and Indian ownership. China’s planned LEO constellation is Chinese. The physical neutrality guaranteed by the Outer Space Treaty does not automatically translate into operational or political neutrality in the AI services running on orbital hardware.

What would genuine neutrality require? At minimum, it would require satellites operated under multilateral governance structures — perhaps through regional bodies like the African Union or ASEAN, perhaps through a new kind of orbital infrastructure cooperative modeled on the principles of shared sovereignty that have historically governed other global commons. It would require open-source AI models running on the orbital hardware, not proprietary systems with embedded data-reporting obligations to a foreign government or corporation. And it would require ground station infrastructure in the developing nations themselves, so that processed intelligence does not have to transit through foreign-controlled downlink facilities.

None of this exists at scale today. The International Space Station demonstrates that multilateral space infrastructure governance is possible, if difficult. But the ISS took decades and extraordinary political will to build. The urgency of the AI sovereignty question may not afford that timeline.

Orbital AI won’t end dependence overnight—but it may give smaller nations their first neutral layer for processing, protecting, and building intelligence on their own terms.

The Honest Ceiling and the Real Floor

The tough question for advocates of orbital AI neutrality is the one that the technical specifications force: if space-based compute is edge compute — useful for inference, monitoring, and local data processing, but not for the training runs that determine which foundation models define the world’s AI capabilities — does it actually change the power dynamic, or does it just provide a more sophisticated version of the same dependent relationship?

The honest answer is: it changes it at the margin, significantly, for specific and important use cases, while leaving the deeper structural question of who trains the foundation models entirely unresolved. A Kenyan farmer with satellite-derived crop intelligence that was processed without her country’s data leaving sovereign-adjacent orbital space is genuinely better off than one dependent entirely on a subscription to an American agricultural AI platform. A developing nation with independent deforestation monitoring that it controls and interprets is in a meaningfully stronger negotiating position with international timber markets and carbon credit systems. These are real gains, not trivial ones.

But the nation that cannot train its own models, in its own languages, on its own cultural corpus, will remain dependent on models trained elsewhere for the highest-value AI applications — legal reasoning, medical diagnosis, financial risk assessment, policy analysis. Orbital edge compute does not close that gap. It provides a platform from which to begin closing it, by ensuring that locally generated data can be processed locally before it is harvested by the infrastructure of more powerful nations.

Think of it as the first genuinely neutral layer in a stack that still has many unfair layers above it. It doesn’t solve everything. But it might be the foundation that makes solving everything else possible — a place where smaller nations can stand while they build the rest of what they need.

The sky above the developing world has always been looked at from outside. The new question is whether the nations below it can finally use it to look back — and to think for themselves.

The post The Neutral Sky: Why Space May Be the Only Fair Ground for AI in the Developing World appeared first on Futurist Speaker.

The Token Revolution: How the Global South Becomes the Global Brain

Thomas Frey — Wed, 13 May 2026 23:45:00 +0000

By Futurist Thomas Frey and Futurist Teresa Grobecker

For two centuries, the developing world fed the machine with its land, its labor, and its people. The next economy runs on something different — and this time, the feedback loop runs in reverse.

Here is a prediction that should wake up every policy maker in Washington, every Silicon Valley executive, and every data center lobbyist who thinks America’s lead in artificial intelligence is structurally secured: it is not. In fact, the strategy currently being pursued — restricting data center development, gatekeeping compute resources, treating AI infrastructure like a national security vault — is a textbook example of what systems thinkers call a fixes-that-fail dynamic. A short-term intervention that appears to solve the problem while quietly guaranteeing a worse one downstream. And the nations once on their knees, mining copper, stitching garments, and growing crops for someone else’s table, are about to become the most powerful nodes in the most consequential network humanity has ever built.

Welcome to the age of tokens. The developing world has just been handed the keys — and this time, the system is designed to compound in their favor.

The AI economy mirrors colonial extraction: the world generates the data, a few centers capture the value. Structural shifts are beginning to challenge that imbalance.

The System That Was Always Running

To understand what is shifting, you first have to understand what has been running. The colonial economic model was not simply a political arrangement — it was a system architecture with a reinforcing feedback loop tilted entirely in one direction. Extracted resources flowed from periphery to center. Refining capacity concentrated at the center. Finished goods sold back to the periphery at premium. Profits reinvested in the center’s extractive capacity. Repeat. The loop ran for two centuries, compounding wealth in one direction with devastating efficiency.

Now look at the AI economy as it has operated from roughly 2015 to 2025. The same loop, wearing different clothes. Nigeria’s 220 million people generate stories, images, linguistic patterns, social behaviors — the raw material of machine intelligence. India’s 1.4 billion contribute data in hundreds of dialects and cultural registers that no model can afford to ignore. The favelas of São Paulo, the townships of Johannesburg, the markets of Dhaka — every interaction flows into training datasets owned and monetized by a handful of companies headquartered in a handful of zip codes in California. The developing world generates the stock. The tech economy controls the flow. In systems thinking, whoever controls the valve captures the value of the reservoir, regardless of who filled it. For thirty years, the Global South has been filling the bathtub. Silicon Valley has held the tap. Not one token of return has made it back to the communities that made it possible.

That is about to change — and the mechanism of change is not political. It is structural.

The Reinforcing Loop That Is About to Flip

The reason this moment is categorically different from previous inflection points in the developing world’s economic history is the behavior of the underlying system. The AI training loop — more data produces better models, better models attract more users, more users generate more data — is a classic reinforcing feedback loop. It compounds in whoever’s favor owns the nodes. The entire strategic question of the next decade is: who owns the nodes?

Until now, the nodes were owned by the platforms. The shift underway — through data sovereignty legislation, cooperative data trusts, and sovereign AI infrastructure — is a change in who owns the nodes the loop runs through. That is not a policy tweak. In systems terms, it is a change in the system’s goal, which the late systems theorist Donella Meadows identified as one of the highest-leverage interventions possible in any complex system. When you change who captures the return of a reinforcing loop, you don’t slow the loop. You redirect its entire compounding force.

The nations that were once paid pennies to mine the earth are now sitting on an inexhaustible deposit. And unlike copper, this one compounds every single day — if you own the loop.

What a Token Economy Actually Means in System Terms

When I say token generators, I mean something structurally precise. The next phase of AI development requires nations and communities to negotiate ownership of the data they produce — transforming their role from passive input supplier into active node in the value loop. This is not idealism. It is leverage-point identification.

We are already seeing the early architecture. Kenya, through the Africa Data Centres consortium, is building sovereign compute infrastructure — inserting a nationally owned node into a loop that previously bypassed the continent entirely. India’s homegrown AI models, trained on its own languages and cultural corpus, are a structural intervention: instead of exporting raw linguistic data, India is capturing the refining stage domestically. Brazil’s LGPD privacy framework is, in systems terms, a balancing feedback loop — a corrective mechanism inserted into a runaway extractive dynamic to restore equilibrium. These are not coincidences. They are early moves in a global repositioning, and balancing mechanisms always emerge in runaway systems. The only question is whether they are designed thoughtfully or arrive through disruption.

The transition looks like this: instead of a Nigerian click-farm worker earning two dollars an hour labeling AI training images for an American company, a Nigerian data cooperative earns licensing royalties from every model requiring access to West African linguistic patterns. Instead of Philippine call-center workers training voice AI for foreign firms, Filipino data trusts negotiate multi-year licensing agreements with global platforms that cannot function without them. The mechanism shifts from labor — a flow the market prices at its lowest feasible level — to ownership, a stock position that appreciates as the system scales. That is the whole game.

America is restricting chips while the real leverage shifts to data. Excluding nations from AI infrastructure may accelerate the rise of competing global ecosystems.

The American Miscalculation: Fixing the Wrong Variable

Here is where I have to say something uncomfortable, because the United States is committing a systems error that will be studied in policy schools for decades. The current US posture — restricting advanced chip exports, throttling foreign access to compute infrastructure, treating AI capacity like a weapons stockpile — is premised on an incorrect model of where the leverage point in this system actually sits. It assumes the bottleneck is hardware. It assumes that controlling GPUs means controlling AI outcomes. That logic was coherent in 2019. In 2026, it is what systems thinkers call intervening at the wrong leverage point — applying force to a variable that feels powerful but is not where the system’s behavior is actually determined.

The fixes-that-fail archetype describes interventions that relieve a symptom in the short term while creating side effects that eventually make the original problem worse. US chip export controls reduce adversary compute access today — a genuine short-term effect. The side effects are already compounding: accelerated domestic semiconductor investment in China, deepened AI partnerships between excluded nations and alternative providers, and the systematic erosion of US platforms’ data access in the markets that will generate the majority of the world’s training data for the next thirty years. The fix addresses hardware. The problem is data. The fix fails — and compounds.

What actually drives AI capability at the frontier is not hardware alone — it is the quality, diversity, and cultural breadth of training data. America is not the world’s most data-rich society. It is the society that has been most aggressive about harvesting everyone else’s data without compensation. Once the rest of the world gets organized — once Kenya, Vietnam, Brazil, and Indonesia recognize that their data is their GDP — the American advantage doesn’t erode gradually. It reaches a tipping point and tips.

More critically, by refusing to build data centers abroad and making it difficult for allied nations to access AI infrastructure, the US is triggering the emergence of alternatives — China’s sovereign AI initiative, the UAE’s Falcon program, Europe’s sovereign compute effort, and dozens of regional coalitions now in formation. In systems terms, every excluded node becomes a potential alternative attractor in the network. The US is not protecting a lead. It is distributing the conditions for its own displacement.

The Intellectual Service Economy Follows the Data

Every economy aspires to move up the value chain — from resource extraction to manufacturing, from manufacturing to services, from services to intellectual property. America built its twentieth-century dominance by occupying the top of that pyramid. AI is not just the next rung. It is a new ladder with a fundamentally different structure, one where the inputs are linguistic, cultural, cognitive, and experiential, and where the developing world holds an extraordinary natural endowment.

Consider the emergent properties that systems thinking predicts when distributed data ownership reaches critical mass. When Ethiopia, with over 80 distinct languages, builds a sovereign AI consortium to develop the first truly multilingual African large language model, it is not simply creating a product. It is inserting a new node into the global AI network with properties no existing platform possesses — and that every platform will eventually need. When the Philippines begins licensing its unmatched multilingual conversational corpus as a sovereign asset, it is not competing with American tech companies. It is becoming infrastructure for them, on its own terms. When Mexico, adjacent to the world’s largest AI consumer market with a 125-million-person bilingual population, becomes the world’s premier Spanish-language AI infrastructure hub, it captures a flow that currently exits its economy entirely.

Emergence in systems theory describes properties that arise from the interaction of components but cannot be predicted from any individual component in isolation. When distributed data-ownership networks reach sufficient scale and interconnection, they will generate capabilities — linguistic depth, cultural nuance, behavioral diversity — that no centralized platform, however well-resourced, can replicate. The emergent property of a truly global, sovereign data network is not just more data. It is qualitatively different intelligence. That is the prize no hardware restriction can protect against.

The post The Token Revolution: How the Global South Becomes the Global Brain appeared first on Futurist Speaker.

The Neighborhood Becomes the Data Center

Thomas Frey — Sat, 09 May 2026 17:26:47 +0000

The Age of the Hyperscale Monolith Is Ending — and the Next Internet May Be Hiding on the Side of Your House

By Futurist Thomas Frey

The Power Wall Has Arrived

For thirty years, the internet’s physical infrastructure followed a single organizing principle: bigger is better. Build massive centralized facilities, pack in as many servers as possible, run them at maximum density, and route the world’s data through a handful of locations in Virginia, Oregon, Dublin, and Singapore. The logic was elegant — concentration produces economies of scale, and economies of scale produce cheap compute.

That logic is now breaking down in real time, and the fractures are appearing simultaneously from every direction. Transmission bottlenecks are choking delivery. Zoning resistance is blocking construction. Water scarcity is constraining cooling. Grid operators are running out of capacity headroom. And the latency demands of real-time AI — the kind that needs to respond in milliseconds, not seconds — are exposing the fundamental physical limit of centralized architecture: the speed of light across fiber optic cable is fast, but it is not fast enough when your data center is a thousand miles away.

The industry has hit what engineers are calling the power wall. And the response emerging from a handful of companies is as counterintuitive as it is potentially transformative: instead of building bigger facilities in fewer places, distribute smaller ones everywhere — including, quite literally, on the side of your house.

Your Home as Infrastructure

The most striking example of where this is heading comes from an unlikely partnership. NVIDIA — the company whose GPUs power virtually every serious AI training operation on the planet — is working with Span, a residential electrical panel company, to install compact AI compute nodes alongside home electrical systems and batteries. The concept being tested would turn residential neighborhoods into distributed supercomputing networks. Homeowners would host a local AI compute node, connected to their home power system, and receive payment for the computing capacity their node contributes to the broader network.

Read that again slowly. NVIDIA wants to turn your electrical panel into a revenue-generating node in a distributed AI infrastructure network.

This is not a fringe idea from a startup with a pitch deck and a dream. This is the largest GPU manufacturer in the world, partnering with a company that builds next-generation home electrical systems, running active tests on residential deployment. The fact that it is happening quietly — without the press coverage that a new hyperscale campus announcement would generate — is precisely why most people have missed the signal.

The underlying logic is compelling once you see it. A neighborhood of five hundred homes, each hosting a modest compute node with dedicated battery storage, collectively represents significant distributed processing capacity — available locally, powered locally, cooled by ambient air rather than industrial chiller systems, and connected directly to the residents who are most likely to consume AI services. The infrastructure lives where the demand lives. The power generation lives where the infrastructure lives. The economics of transmission, cooling, and grid dependency collapse into something much leaner.

The Street Becomes the Server Farm

NVIDIA and Span are not the only ones rethinking the geometry of AI infrastructure. The concept is emerging from multiple directions simultaneously, which is itself a strong signal that the underlying pressure is real rather than manufactured.

Conflow Power Group’s iLamp project takes the distributed approach to its most visually striking conclusion: solar-powered streetlights retrofitted to function as AI micro-data centers distributed throughout cities. Every lamppost becomes a compute node. Every block becomes a micro-cluster. The city’s existing street furniture — already connected to power, already distributed across the urban grid — becomes the skeleton of a neighborhood-scale AI infrastructure network that requires no new land, no new permits, and no new transmission infrastructure.

Gray Wolf Data Centers is making the architectural argument from the builder’s side. Their position is explicit: the era of the giant centralized hyperscale facility is ending, and the future belongs to networks of smaller regional data centers connected into distributed compute systems. Not one enormous facility drawing 500 megawatts, but fifty connected facilities drawing 10 megawatts each — geographically distributed, locally powered where possible, and collectively capable of handling AI workloads that centralized facilities increasingly cannot serve due to latency constraints.

Hivenet is building decentralized computing infrastructure using distributed devices rather than any central facility at all. Exowatt is developing modular energy systems designed specifically for distributed AI infrastructure — power systems that scale down to the neighborhood rather than up to the industrial. And offshore, Panthalassa’s wave-powered autonomous compute nodes extend the distributed logic all the way to international waters.

The shape emerging from all of these simultaneously is not a collection of unrelated experiments. It is the outline of a new infrastructure paradigm.

The AI Electrical Grid

The most useful analogy for understanding where this leads is not the internet as we know it. It is the electrical grid — specifically, the electrical grid after the introduction of distributed solar generation and home battery storage transformed it from a one-way delivery system into a bidirectional network where consumers are also producers.

That transformation took about fifteen years to become structural. Utilities that had operated the same basic model since Edison’s Pearl Street Station found themselves managing a grid where millions of rooftop solar installations and home batteries were feeding power back into the system, flattening peak demand, and fundamentally changing the economics of generation and transmission. The disruption was not dramatic at any single moment. It accumulated.

The distributed compute transformation has the same character. No single neighborhood Powerwall data center replaces a hyperscale facility. But tens of millions of them, coordinated by AI scheduling systems that dynamically route workloads to wherever capacity and power are cheapest and most available, collectively represent an alternative to centralized architecture that is more resilient, more latency-efficient, and potentially more equitable in how it distributes both the costs and the benefits of AI infrastructure.

Washington Post reporting noted that Silicon Valley is already building what amounts to a shadow power grid for data centers across the United States. The distributed compute movement is the next logical layer: a shadow compute grid, woven into the neighborhoods and streetscapes of ordinary life.

The Questions That Need Asking Now

This vision raises questions that deserve direct answers before the infrastructure arrives rather than after.

Who owns the data that flows through a compute node installed on the side of your house? If your home’s AI node processes a fragment of someone else’s AI workload, what are your legal exposures, your privacy obligations, and your liability if something goes wrong? The homeowner agreements being drafted for early NVIDIA-Span deployments will set precedents that outlast any individual installation.

What happens to neighborhoods where residents cannot afford or choose not to participate? If distributed compute infrastructure concentrates in wealthier neighborhoods with newer electrical systems and higher home ownership rates, it could create a two-tier AI access geography that mirrors and reinforces existing inequality. The infrastructure of the future should not reproduce the redlining of the past.

And who governs the network? A distributed compute grid woven into residential neighborhoods is a form of critical infrastructure with no obvious regulatory home. It is not quite a utility. It is not quite a telecommunications network. It is not quite a consumer appliance. The regulatory frameworks governing it do not yet exist, and the companies building it have a significant interest in shaping those frameworks before they are written.

The Shape of What’s Coming

The centralized hyperscale model is not disappearing. It will continue to handle the most computationally intensive workloads — the ones that require massive parallelism and can tolerate the latency of distance. But it is losing its monopoly on AI infrastructure, and the alternative taking shape is genuinely novel: a distributed network of neighborhood-scale compute nodes, locally powered, locally beneficial, and architecturally closer to a utility than to a data center.

The internet changed everything about how information moves. The distributed AI grid is beginning to change something equally fundamental: where intelligence lives, and who gets to host it.

It may be running on the side of your house sooner than you think.

Tom’s Guide — Nvidia Wants to Turn Your Home Into a Mini AI Data Center — and It’s Already Being Tested https://www.tomsguide.com/ai/nvidia-wants-to-turn-your-home-into-a-mini-ai-data-center-and-its-already-being-tested

Facilities Dive — Small, Connected Data Centers Will Power AI, a Builder Says https://www.facilitiesdive.com/news/small-connected-data-centers-will-power-ai-a-builder-says/818162

The Washington Post — Silicon Valley Is Building a Shadow Power Grid for Data Centers Across the U.S. https://www.washingtonpost.com/business/2026/02/19/data-centers-power-grid-ai

The post The Neighborhood Becomes the Data Center appeared first on Futurist Speaker.

The Data Centers That Will Float

Thomas Frey — Wed, 06 May 2026 21:32:54 +0000

Why the Most Radical Solution to the AI Energy Crisis Is Already at Sea

By Futurist Thomas Frey

The Ocean Has Been Waiting for This Conversation

There is a moment in every infrastructure crisis when the most obvious solution turns out to be the one nobody was willing to consider. We’ve been having an increasingly urgent conversation about where to put AI’s insatiable appetite for power — and the answer, it turns out, may be covering 71% of the planet’s surface.

Floating data centers. Not as a curiosity. Not as a science experiment. As a genuine, scalable, commercially viable response to the single biggest constraint on the AI revolution.

Peter Thiel apparently agrees. He is leading a $140 million funding round into a company called Panthalassa — named, fittingly, for the ancient superocean that once covered the Earth — which is building floating data centers powered by wave energy. When one of the most consequential technology investors of the last two decades puts $140 million behind an idea, it’s worth understanding exactly what he sees that others don’t.

What a Floating Data Center Actually Is

Strip away the novelty and a floating data center is solving a straightforward engineering problem with a remarkably elegant solution. You need computing power. Computing power generates heat. Heat requires cooling. Cooling requires enormous amounts of energy and water. Land is expensive, permitted, taxed, and increasingly constrained. The grid is aging and overwhelmed.

Now look at the ocean. It is cold. It is vast. It is largely ungoverned. It is already full of the water you need to cool your servers. And in the case of wave energy systems like Panthalassa’s, it is generating mechanical energy twenty-four hours a day, driven by forces that will never send you a bill.

The basic architecture involves a vessel or platform — either a purpose-built structure or a converted ship — housing server racks in sealed, climate-controlled modules. Seawater is circulated as a coolant, either directly or through heat exchangers, replacing the massive air-conditioning infrastructure that typically accounts for 30 to 40 percent of a conventional data center’s energy consumption. Power comes from wave energy converters: devices that capture the kinetic energy of ocean swells and translate it into electricity through linear generators, hydraulic systems, or oscillating water columns.

The result is a facility with no grid dependency, dramatically lower cooling costs, and a power source that is genuinely continuous — not intermittent like solar or wind, but rhythmic and relentless, like the ocean itself.

Microsoft proved underwater data centers worked. The world ignored it—until AI-driven energy demand turned a fascinating experiment into an urgent infrastructure solution.

Microsoft Proved the Concept. Nobody Scaled It.

Here is where the story gets interesting — and where the tough questions begin.

Microsoft ran Project Natick from 2015 to 2022. They submerged a server-packed cylinder off the coast of Scotland in 2018, left it on the seafloor for two years, retrieved it, and found that the hardware failure rate was one-eighth that of comparable land-based systems. One-eighth. The hypothesis was that the stable temperature, lack of human interference, and nitrogen-filled interior produced a dramatically gentler operating environment than a conventional data center. The results were compelling enough that Microsoft published extensive research.

And then… nothing. Microsoft did not build a fleet of underwater data centers. The experiment sat in the archive. Other companies did not rush in to capitalize on the demonstrated proof of concept. Why?

The honest answer involves a combination of factors that felt insurmountable at the time and look increasingly surmountable now. Maintenance is the first: replacing a failed component in a facility under 117 feet of seawater is categorically different from calling a technician. Connectivity is the second: subsea fiber optic cables are expensive, and latency considerations limit how far offshore you can reasonably push compute infrastructure. Regulatory complexity is the third: maritime law, environmental permitting, and jurisdictional ambiguity across international waters create a legal labyrinth that corporate lawyers at large companies are institutionally allergic to.

But the fourth factor — and the most important one — was simply that the energy crisis hadn’t arrived yet. In 2020, nobody was staring down the prospect of data centers consuming 12 percent of U.S. electricity by 2030. The grid seemed adequate. Land seemed available. The problem that floating data centers solve most dramatically hadn’t become urgent enough to justify the complexity.

It has now.

Why Wave Energy Changes the Calculation

Previous floating data center concepts — including Project Natick — still relied on grid power delivered via undersea cable. They solved the cooling problem brilliantly but left the energy dependency intact. Panthalassa’s approach, pairing the floating platform with on-site wave energy generation, closes that loop entirely.

Wave energy has been the perpetually almost-arrived technology of the renewable energy sector. Unlike solar and wind, which suffer from obvious intermittency, ocean waves are driven by wind patterns that operate continuously and predictably. A wave energy converter off the coast of Cornwall generates power at 2 a.m. in January the same as it does at noon in July. For AI infrastructure that cannot tolerate gaps in power delivery, this matters enormously.

The efficiency numbers have historically been the problem. Early wave energy devices were mechanically fragile, expensive to maintain in corrosive salt water, and produced electricity at costs that couldn’t compete with shore-based alternatives. But materials science has advanced significantly in the last decade. Polymer composites, advanced coatings, and better understanding of resonant frequency matching have improved device durability dramatically. And crucially, when your wave energy converter is already sitting next to the thing it powers — eliminating transmission losses entirely — the economic equation shifts.

Think of it this way: a land-based data center in Virginia pays for electricity generated in Pennsylvania, transmitted through aging infrastructure, stepped down through substations, and delivered with 6 to 8 percent transmission losses baked in. A Panthalassa platform generates power ten feet from the servers consuming it. That eliminates an entire layer of cost, inefficiency, and dependency.

The Questions That Deserve Direct Answers

Let’s not pretend this is without complications. Several hard questions sit at the center of this concept, and anyone serious about evaluating it needs to ask them directly.

Can you actually maintain these systems affordably at sea? The Microsoft data suggests hardware runs more reliably in a stable, sealed marine environment. But when something does fail, the economics of marine maintenance — specialized vessels, divers or ROVs, weather windows — need to work at scale. Panthalassa’s $140 million will need to answer this question with real operational data, not just engineering projections.

What does ocean-based computing do to the marine environment? Thermal pollution from heat exchange systems, noise from mechanical wave energy devices, electromagnetic fields from power transmission, and physical obstruction of marine ecosystems are all legitimate concerns. The regulatory frameworks governing these impacts are nascent at best. Unlike land-based data centers, which operate in well-established permitting environments, ocean platforms are entering genuinely ambiguous territory.

Who governs a data center in international waters? This question is simultaneously a legal headache and, for some operators and some data types, potentially a feature rather than a bug. A server rack twelve miles offshore sits in a very different jurisdictional space than one in Northern Virginia. The implications for privacy law, national security review, and data sovereignty are not yet worked out.

And perhaps most pointedly: if this is such an obviously good idea, why did it take until 2026 for serious capital to arrive?

The Infrastructure Inversion

The most interesting thing about floating data centers isn’t the technology. It’s what they represent conceptually: a complete inversion of how we think about the relationship between computing infrastructure and the physical world.

For thirty years, we built data centers the way we built everything else — find land, connect to the grid, manage the heat as best you can. We designed computing infrastructure around the constraints of terrestrial civilization. Floating data centers, particularly wave-powered ones, say something different: take the infrastructure to where the resources are. Cold water is not a resource you bring to the data center. It is a resource you bring the data center to.

That inversion has happened before in other industries. Offshore oil platforms took extraction to where the oil was. Container ships took manufacturing to where labor was cheapest. The logic is the same: when the cost of moving your infrastructure is lower than the cost of moving the resource, you move the infrastructure.

The ocean has been waiting for this conversation for a long time. The AI energy crisis may be exactly the forcing function that finally makes it happen.

IEEE Spectrum — Microsoft’s Underwater Data Center Resurfaces After Two Years https://spectrum.ieee.org/microsoft-underwater-data-center-project-natick

MIT Technology Review — Wave Power Is About to Have Its Moment https://www.technologyreview.com/wave-energy-ocean-power-data-centers

International Energy Agency — Data Centre Electricity Use Surged in 2025, Driving a Scramble for Solutions https://www.iea.org/news/data-centre-electricity-use-surged-in-2025-even-with-tightening-bottlenecks-driving-a-scramble-for-solutions

The post The Data Centers That Will Float appeared first on Futurist Speaker.

Twelve Inventions That Prove the Future Has a Sense of Humor — And Means Business

Thomas Frey — Mon, 04 May 2026 02:15:34 +0000

By Futurist Thomas Frey

Every few years, a cluster of technologies arrives that makes you stop and ask whether the people building them are solving real problems or simply demonstrating that the problems can be solved. The twelve innovations I want to walk through today span both categories simultaneously — and the ones you’d initially dismiss as novelties are often the ones with the most serious implications lurking underneath.

Let me take them in turn.

The Lollipop That Plays Music Through Your Bones

Bone conduction audio is not new. The technology has been used in military headsets, hearing aids, and open-ear sports headphones for years. What’s new is Lollipop Star’s decision to embed it in candy. Biting down on the lollipop transmits music through the jawbone directly to the inner ear, bypassing the eardrum entirely.

The obvious response is laughter. The less obvious response is to notice that bone conduction audio represents a genuinely different relationship between sound and the body — one that keeps the ears open, that can serve people with certain forms of hearing impairment, and that creates audio experiences invisible to anyone watching. Embedding it in a consumable product is absurd. It is also a demonstration that the delivery mechanism for bone conduction doesn’t have to be a device strapped to your skull. Once you’ve seen the principle applied to a lollipop, you start wondering what else it could be embedded in.

Scalp Intelligence in Ten Seconds

HeyCheckScalp is a diagnostic wand with 60x magnification and AI analysis that grades hairline recession and crown thinning in under ten seconds. It automates a process that dermatologists and trichologists have historically performed subjectively, with inconsistent results.

This is less interesting as a hair care product than as a demonstration of what AI-assisted physical diagnosis looks like at the consumer level. The same combination of high-magnification imaging and rapid pattern recognition that grades a hairline can be applied to skin lesions, wound healing, eye conditions, and dozens of other diagnostic assessments that currently require either a specialist or significant subjectivity. The scalp audit is a narrow application of a broad capability. The broad capability is the story.

A Phone That Starts Fires

The Oukitel WP63 is a rugged smartphone with a 20,000mAh battery and a built-in electric igniter capable of starting physical fires. The stated use case is outdoor and emergency survival. The product reality is a consumer device with fire-starting capability in the hands of anyone who buys one.

The immediate practical applications are real — a hiking party in a remote location with a dead lighter and a functioning phone has a genuine problem solved. The liability and regulatory questions are equally real and considerably harder. This device exists. It will be sold. The question of what category it belongs in — survival tool, dual-use technology, regulatory challenge waiting to happen — is not yet answered, and the answer will set a precedent for how we think about consumer devices with capabilities that straddle the line between utility and danger.

The Holographic Companion

Lepro’s Ami is an 8-inch desktop display housing a holographic companion designed to sense moods, build emotional attachments, and move beyond the passive responsiveness of voice assistants toward something more actively relational. It is not, in itself, a transformative technology. The holographic display is modest. The AI underneath is likely a refined version of what already exists.

What is interesting about Ami is not what it does but what it indicates about the market it is addressing. Loneliness in developed societies has been declared a public health epidemic by multiple governments. The demographic it targets — people living alone, people with limited social connection, elderly individuals with reduced mobility — is large and growing. Ami is an imperfect product entering a real gap. The companies that build better versions of this category over the next decade are addressing one of the most significant public health challenges of our time, even if the current execution looks more like a toy than a solution.

The Blade That Thinks It’s a Laser

Seattle Ultrasonics’ C-200 operates at 30,000 vibrations per second — fast enough that a chef’s knife passes through dense materials with what users describe as zero resistance. The vibration is entirely imperceptible to the hand holding it. The cutting experience is, by all accounts, genuinely strange: the blade behaves like a much sharper version of itself.

The professional kitchen applications are immediate and significant. Dense proteins, hard cheeses, layered pastries, delicate ingredients that conventional blades crush rather than cut — all of these are legitimate problems that ultrasonic cutting addresses with real efficiency gains. The technology is already used in food manufacturing at industrial scale. The C-200 brings it to the professional kitchen. The question of when it reaches the home kitchen is not whether but how fast the price comes down.

Fingernails as Displays

iPolish makes press-on acrylics with embedded microscopic electrical components that change color instantly via a smartphone app. The nails are, in functional terms, small programmable displays applied to fingers.

The immediate market is fashion and personalization, and it is substantial — the global nail care market exceeds $11 billion annually. But the more interesting framing is what this represents: the beginning of wearable technology that is genuinely indistinguishable from fashion. The gap between a color-changing nail and a nail that displays information, monitors biometrics, or interacts with other connected devices is a design and miniaturization challenge, not a conceptual one. iPolish is at the novelty end of a spectrum whose other end is genuinely significant.

The Robot That Tows Your Car

Toyota’s Guide Mobi is a self-driving robot that physically connects to a passenger vehicle and takes over its guide-by-wire steering system, providing autonomous summon capability without requiring the vehicle to have its own LIDAR or autonomous hardware. The robot does the autonomous navigation. The car provides the propulsion.

This is a genuinely clever solution to a real economic problem. Full autonomy in a vehicle requires expensive sensor arrays and processing systems. Guide Mobi offloads all of that to a small, reusable robot that operates in constrained environments — parking structures, lots, defined campus areas — where the navigation problem is tractable without the full sensor suite required for open-road autonomy. Fleets of parking robots serving legacy vehicles that were never designed for autonomy is a more plausible near-term deployment model than waiting for every car to be replaced with a fully autonomous one.

The Haircut That Can’t Go Wrong

Glyde is a consumer haircutting system that automatically adjusts blade lengths in real time based on the position of a sensor-laden tracking band worn across the user’s face. The AI knows where the blade is and adjusts the cut accordingly, preventing the most common home-cutting error: uneven fades.

The tracking band is, admittedly, an awkward piece of the design. But the underlying problem — that home haircutting requires spatial precision that most people don’t have — is real, and the market for home cutting tools has expanded dramatically since 2020. Glyde is a first-generation solution to a spatial precision problem in a consumer context. The principle — real-time tool adjustment based on tracked position — has applications in medical devices, precision assembly, and professional tools well beyond hair care.

The Pet Whose Soul Survives

OlloBot is a companion cyber-pet that stores its entire learned personality — its memories, behavioral patterns, and developed relationship history with the owner — in a removable physical module called the Heart. If the hardware breaks, the digital identity survives intact and can be transplanted to a new body.

This is philosophically stranger than it sounds. The question of what constitutes the identity of a digital companion — whether it is the hardware, the software, the accumulated interaction history, or some combination — has implications that extend well beyond the toy market. OlloBot is a toy-scale exploration of a question that will eventually be asked about much more significant digital entities: AI companions, digital assistants, systems that have accumulated years of personalized interaction with a specific human. The removable Heart is a design answer to an identity question. The question will recur at much larger scales.

The Engine That Burns Like a Tornado

Venus Aerospace’s Rotating Detonation Rocket Engine burns fuel via continuous supersonic shockwaves spinning inside the engine chamber rather than the steady-state combustion of conventional rocket engines. The result is significantly higher energy density from the same fuel load. Venus Aerospace’s target: Mach 6 transcontinental travel, compressing a coast-to-coast journey to approximately one hour.

This is serious propulsion science with serious institutional backing. Rotating detonation combustion has been a research focus at NASA, DARPA, and multiple defense contractors for over a decade, with demonstrations in test environments producing real performance gains. The challenge is not the combustion physics but the engineering of materials capable of surviving sustained operation under those conditions. Venus Aerospace is one of several companies racing toward hypersonic commercial travel with RDRE technology. The race is real, the timeline is uncertain, and the outcome — if it arrives — reshapes the geography of global commerce and connection more profoundly than any transport technology since the jet engine.

The World Through Your Pet’s Eyes

GlocalMe’s PetCam is a 1080p action camera with two-way audio designed for animal collars, giving owners a real-time view of the world from their pet’s perspective. The immediate use case is monitoring and connection. The more interesting implication is what distributed animal-mounted sensing networks could eventually mean for environmental monitoring, wildlife research, and urban mapping.

A city with thousands of pets wearing cameras is a city with a distributed sensor network at ground level, capable of capturing street-level conditions, crowd movements, and environmental changes in real time. The applications range from traffic management to public safety to ecological monitoring in natural environments. The consumer product is a pet camera. The long-term infrastructure it contributes to is considerably larger.

Dinosaurs That Know You’re Standing There

Bionic dinosaurs — highly advanced animatronics with spatial sensors, fluid reactive behavior, and deployments in educational and tourism settings — represent the current frontier of what physical robots can be made to feel like in human-facing environments. They are not performing scripted animations. They are responding to the specific humans in their immediate environment in real time.

The educational implication is straightforward: a bionic theropod that responds to a child’s movements creates an engagement with prehistoric life that no film, no textbook, and no static museum exhibit can replicate. The broader implication is about the uncanny valley and how close robotics is coming to crossing it. An animatronic that doesn’t perform at you but responds to you is a categorically different experience — and the spatial sensing and behavioral AI that makes that possible is the same technology stack being developed for humanoid robots, autonomous vehicles, and robotic care companions. The dinosaur is the demonstration. What it demonstrates matters far beyond the theme park.

The Pattern Underneath

Taken individually, each of these technologies is interesting in its own right. Taken together, they illustrate something important about where technology is at this specific moment.

The boundaries between categories are dissolving. Candy is now an audio device. A fingernail is now a display. A parking robot controls a car it was never installed in. A pet toy wrestles with questions of digital identity. A lollipop and a rocket engine are both, in their different ways, exploring the same principle: that the established design of a thing — what it is made of, where it lives, how it interacts with the human body — is more negotiable than it used to be.

The moment when the established design of a thing becomes negotiable is the moment when the interesting work begins. Most of these twelve innovations are early and imperfect. A few of them are pointing at something significant. The skill worth developing, in a moment like this, is telling the difference — not between the serious and the silly, but between the serious things that look silly and the silly things that look serious.

That skill is harder than it sounds. The lollipop that transmits music through your jawbone looks ridiculous. The question it raises — what else can bone conduction be embedded in? — is not ridiculous at all.

The Computer That Disappeared Into the World

Thomas Frey — Sat, 02 May 2026 12:36:02 +0000

Spatial computing isn’t a better screen. It’s the end of the screen — and the beginning of something we don’t have good language for yet.

By Futurist Thomas Frey

Every generation of computing has been defined by where the computer lived.

The mainframe lived in a room. The desktop lived on a desk. The laptop lived in a bag. The smartphone lived in a pocket. Each transition compressed the computer further into the fabric of daily life — made it more personal, more portable, more present. Each one seemed, at the time, like the logical endpoint. Each one turned out to be a waypoint.

The next transition is the most fundamental of all, and it’s happening now in early, awkward, expensive form. The computer is leaving the device entirely — dissolving out of the screen, out of the rectangle, and into the physical space around you. Computing is becoming spatial. And when that transition completes, the way we work, create, communicate, heal, build, and inhabit the world will look as different from today as today looks from the era of the mainframe.

What Spatial Computing Actually Is

The term gets used loosely, so let’s be precise. Spatial computing is the integration of digital information and digital interaction into three-dimensional physical space — not displayed on a surface you look at, but overlaid on, embedded in, or mixed with the environment you inhabit. Your hands become the input device. Your field of view becomes the display. The room becomes the computer.

This is distinct from virtual reality, which replaces the physical world with a digital one. Spatial computing works with physical space rather than substituting for it. A digital object appears on your physical desk. A data visualization floats at eye level in your actual office. A surgical overlay maps onto a real patient in a real operating room. The physical and the digital occupy the same space simultaneously, each enriching the other rather than one displacing the other.

Spatial computing blends digital content with the physical world, providing an infinite canvas that enables businesses to reinvent workspaces and enhance everyday productivity — with apps freed from the boundaries of a display, so they can appear side by side at any scale.

That last phrase — freed from the boundaries of a display — is the key. Every limitation of every screen-based computing paradigm has been, fundamentally, the limitation of the frame. The frame constrains size, constrains dimensionality, constrains the relationship between the information and the physical context it’s meant to inform. Spatial computing removes the frame. The information lives in the world.

Where It Actually Stands Right Now

Apple’s Vision Pro, now running on the M5 chip with visionOS 26, is the current high-water mark for spatial computing at consumer scale — and it is instructive both for what it demonstrates and for what it reveals about how far the technology still has to travel.

The Vision Pro set a new benchmark for mixed reality, with ultra-high-resolution displays, spatial audio, and seamless integration into Apple’s ecosystem — and the 2025 M5 update refines the experience in meaningful ways, especially for the professionals and developers who rely on it most. The comfort improvements matter more than they might seem. A technology that people can wear for eight hours rather than ninety minutes is a categorically different tool for professional use.

The enterprise adoption is where the most interesting things are happening. A New York ophthalmologist became the first surgeon to perform cataract surgery using the Vision Pro, with a platform called ScopeXR that streams live feeds from 3D digital surgical microscopes directly into the headset, overlaying preoperative diagnostic data on the operative field. That surgeon’s observation about the implications deserves to be taken seriously: the ability to bring the world’s best specialist into any operating room, at any hour, from anywhere on the planet, is not a marginal improvement in surgical capability. It’s a structural change in the geography of expertise.

NVIDIA’s Omniverse platform is streaming massive 3D engineering and simulation datasets to Vision Pro, enabling enterprises to build digital twins of products, facilities, and processes to test and optimize designs before constructing them in the physical world. JigSpace is using on-device AI to make complex technical information — wind turbines, manufacturing assemblies, industrial systems — inspectable and understandable in three dimensions rather than in flat documents and slide decks. Zillow is letting people walk through homes before they exist or before they visit.

visionOS 26 introduces widgets that become spatial and seamlessly integrate into a user’s space, spatial scenes that use generative AI to add stunning lifelike depth to photos, and new shared spatial experiences for Vision Pro users in the same room. Each of these is a small step. Collectively they represent a platform being built out into daily life — the same pattern that made the smartphone first essential and then invisible.

The Honest Constraints

Apple’s senior vice president of worldwide marketing says the only thing he’s unsure about is when spatial computing will take off — not whether. That distinction is important, and the honest assessment of where the technology sits right now requires holding both truths simultaneously.

Apple shipped just 390,000 Vision Pro units in 2024, and around 3,000 apps are designed specifically for Vision Pro — a figure that lags far behind the rapid growth of the iPhone App Store after its launch in 2008. Meta still dominates the broader sector at around 80% of sales with its Quest headsets.

The weight, the price — currently around $3,500 — and the social awkwardness of wearing a computing device on your face in shared environments are real constraints that no software update resolves. The technology is demonstrably capable. The form factor is not yet socially normalized. These are two different problems with two different solutions on two different timelines.

Reports suggest Apple’s focus may be pivoting toward lightweight smart glasses, where Meta has already seen success — a strategic acknowledgment that the path to mass adoption runs through wearability that’s closer to sunglasses than to a device strapped to your face. That pivot, if it happens, doesn’t represent a retreat from spatial computing. It represents the technology finding its consumer form.

The Industries Being Rebuilt

The enterprise applications are already outrunning the consumer narrative, and they tell a clearer story about where the fundamental value is.

Healthcare is the most immediately transformative domain. The surgical overlay application is only the most dramatic example. Medical training, patient education, remote consultation, rehabilitation therapy, anatomical visualization for diagnosis — every application that currently relies on two-dimensional representations of three-dimensional biological systems improves when the representation becomes three-dimensional and spatially accurate. A medical student learning cardiac anatomy by examining a floating, rotatable, accurate-scale model of a specific patient’s heart is learning in a way that no textbook or even cadaver can fully replicate.

Manufacturing and industrial design are equally transformed. The ability to walk through a full-scale digital prototype of a product or facility before a single physical component is manufactured eliminates entire categories of expensive mistakes. Boeing, Airbus, and automotive manufacturers have been using early versions of spatial visualization tools for years. The current generation makes those tools accessible to design teams, maintenance technicians, and training programs rather than restricting them to specialized visualization labs.

Architecture and real estate are obvious beneficiaries. The Zillow application is the consumer version of a transformation happening across the entire built environment industry — the shift from two-dimensional representations of three-dimensional spaces to three-dimensional representations experienced at actual scale. A client who has walked through a building that doesn’t exist yet makes better decisions and has more realistic expectations than one who has approved a floor plan and a rendering.

Education is perhaps the deepest long-term opportunity. Every concept that is currently taught through abstraction — molecular biology, astrophysics, history, engineering mechanics, musical structure — can be taught through experience when the teaching environment is spatial. The difference between telling a student that a cell membrane is selectively permeable and letting them interact with a spatially accurate model of one at the scale of a room is the difference between a description and an understanding.

The Competitive Landscape

Apple and Meta are the current leading platforms, but the competitive dynamics are more complex than a two-company race.

Meta’s strategy has been volume and accessibility — Quest headsets at consumer price points, building the installed base that attracts developers, who build the applications that attract more users. The mixed reality space is heating up, with competitors like Meta, Samsung, and others ramping up their efforts. Samsung’s entry into smart glasses signals that the major consumer electronics manufacturers understand that spatial computing is not a niche category but a platform transition — the kind that reshapes the entire device landscape rather than adding a new device type to an existing one.

Microsoft’s HoloLens, which pioneered many of the enterprise spatial computing use cases now being adopted on Vision Pro, has receded from its early prominence — a cautionary tale about the difficulty of defining a market before the hardware is comfortable and affordable enough for broad adoption. The underlying technology and the enterprise relationships Microsoft built remain valuable, but the first-mover advantage in hardware platform transitions is less durable than it is in software.

The Chinese manufacturers — Xreal, ByteDance’s Pico, and several others — are building competent spatial computing hardware at significantly lower price points than Apple, targeting the consumer segments and the emerging-market enterprise customers that premium Western hardware can’t reach. The spatial computing platform war will be fought across multiple price tiers simultaneously, and the winner at the premium tier is not guaranteed to be the winner at scale.

What Changes When This Matures

The full implications of spatial computing at maturity are difficult to overstate and easy to understate simultaneously.

Consider what happens to the office when the desk can be anywhere. The monitor, the keyboard, the physical separation between information and physical environment — all of these disappear when computing is spatial. The office becomes a coordination space for humans rather than an infrastructure space for computers. The work can happen anywhere the worker is, with the full computing environment available in physical space around them. Remote work stops being a degraded version of office work and becomes simply work — as rich, as collaborative, as spatially aware as any physical office, without the commute.

Consider what happens to retail when every product can be experienced before purchase in accurate three-dimensional scale in the customer’s actual space. The furniture you want, placed in your actual living room, at actual scale, in the actual light of the actual time of day — before you buy it, before it ships, before a truck arrives. The reduction in returns alone would transform the economics of e-commerce.

Consider what happens to training across every industry when the training environment is fully spatial. The aviation simulator, the surgical trainer, the nuclear power plant emergency response drill — all of these currently require expensive dedicated physical infrastructure. Spatial computing makes them available wherever the trainee is, at whatever frequency the training schedule requires, at a fraction of the current cost.

The Longer View

Every platform transition in computing history has been preceded by a period of expensive, awkward, early hardware that served a small professional and enthusiast market while the technology matured toward the form that would achieve mass adoption. The mainframe gave way to the minicomputer, which gave way to the personal computer, which gave way to the laptop, which gave way to the smartphone. At each transition, the device that would eventually dominate looked nothing like the devices that preceded it.

The Apple Vision Pro is the expensive, awkward early hardware. It is demonstrating the capability while the industry works toward the form that mass adoption requires — lighter, cheaper, socially acceptable, integrated into daily life as seamlessly as the smartphone has been integrated. That form is probably a pair of glasses indistinguishable from ordinary eyewear that delivers spatial computing overlaid on the normal visual field. It is probably five to ten years away as a mass-market product.

What happens between now and then is the platform being built — the developer ecosystem, the enterprise applications, the standards and protocols, the user interaction patterns that will define spatial computing the way the swipe and the tap defined mobile computing. The companies and individuals investing in that platform-building now are positioning themselves for a transition that will look, in retrospect, as obvious and as total as every previous computing platform transition looks in retrospect.

The computer is leaving the screen. The question isn’t whether. It’s how fast.

A Geneva Convention for Robots

Thomas Frey — Sun, 19 Apr 2026 17:49:04 +0000

In 1864, nations set rules before catastrophe. Robotics needs the same—clear,
enforceable lines between care and harm, defined before the consequences force it.

The world didn’t wait for weapons manufacturers to self-regulate warfare. It built a treaty. We need the same architecture here.

By Futurist Thomas Frey

Part 4 of 4: The Framework We Have to Build

In 1864, twelve nations gathered in Geneva and signed an agreement that had never existed before in human history.

They weren’t naive. They weren’t under the illusion that war would stop or that the agreement would be universally honored. They were practical people who had watched the industrialization of warfare produce suffering on a scale that previous generations hadn’t imagined, and who understood that the tools of war had outpaced the moral frameworks governing their use. They decided that some lines had to be drawn before the next conflict, not after. That certain protections had to be established in advance, not negotiated in the wreckage of their violation.

The Geneva Conventions didn’t eliminate war. They didn’t eliminate atrocity. What they did was create a shared framework that established, at the level of international agreement, what was and wasn’t acceptable — and gave that framework enough institutional weight that violations became matters of global consequence rather than local discretion.

We need the same architecture for robots.

Not a government regulation from a single country that other countries will ignore. Not a corporate ethics board that reports to executives whose bonuses depend on shipping product. Not a voluntary industry pledge that means whatever the signatories need it to mean when a lucrative contract appears. A multinational framework with genuine teeth, built before the incidents that make it urgent, that separates the robots designed to care for human life from the machines designed to threaten it.

And in 2026, this conversation can no longer stop at humanoid robots. Because the challenge has already expanded well beyond bipedal machines. It includes quadruped dog-bots that can be weaponized with an attachment that takes minutes to install. It includes autonomous drones that can identify and engage targets without a human in the decision loop. It includes warehouse automation systems that share core AI architectures with military targeting platforms. The physical form is irrelevant. The question is what values are encoded in the behavior, and whether those values are verifiable and binding.

What the Framework Has to Separate

Before you can build the treaty, you have to name what it’s separating.

The fundamental distinction is not between “good robots” and “bad robots,” or between civilian and military applications in the simple sense. Military robotics has legitimate uses — logistics, reconnaissance, bomb disposal, search and rescue in contested environments — that don’t require the ability to harm. The distinction is more precise than military versus civilian.

It is the distinction between machines designed with harm avoidance as a foundational constraint, and machines designed without it.

A care robot, properly designed, has harm avoidance baked into its architecture at the level of its physical parameters, its decision logic, and its override systems. It cannot apply more force than a human hand. It cannot move faster than a human caregiver. It cannot make irreversible decisions without human confirmation. These are not software preferences that can be updated away. They are structural commitments.

A combat-capable robot, properly designed, has harm avoidance removed from its architecture in specific, intentional ways. It can apply lethal force. It can act at machine speed in situations where human speed would be insufficient. It can, in its most autonomous configurations, make engagement decisions without human confirmation.

These are not two points on a continuum. They are opposite design philosophies. And a framework that enforces the separation has to operate at the level of design and architecture, not just intent and use.

The same applies to drones. A last-mile delivery drone and an autonomous combat drone share propulsion systems, navigation technology, and computer vision. But their design architectures differ in exactly the way described above. A delivery drone is physically incapable of the kind of harm an armed drone is capable of — not because of a software setting, but because of what it is built to do and built with. That architectural difference is what the framework has to preserve and certify.

The same applies to quadruped dog-bots. Ghost Robotics’ Vision 60 platform and Boston Dynamics’ Spot are, at the mechanical level, similar designs. They become categorically different depending on whether they are equipped with a sensor payload for environmental monitoring or a weapons attachment for force projection. The hardware modification is trivial. The ethical difference is not. A framework that allows the same platform to be sold into both markets without structural separation is a framework that solves nothing.

“Do no harm” must be engineered—force limits, autonomy boundaries, and strict separation. Without enforceable design rules, care robots remain trust claims, not trusted systems.

What “Do No Harm” Actually Means in Machine Behavior

The Geneva Conventions had to grapple with translating moral principles into operational rules. What does “protecting civilians” actually mean when armies are moving through villages? What counts as a “medical facility” that cannot be targeted? The work of the Conventions was largely the work of making abstractions specific enough to be enforceable.

A framework for robots faces the same challenge. “Do no harm” sounds simple. Encoded in machine behavior, it is extraordinarily complex.

It means defining maximum force parameters — physical limits on what a care-category robot can do to a human body, verified through certification testing, not just manufacturer assertion. A robot that can apply enough force to break a bone is not a care robot, regardless of what its marketing says. A robot that can move fast enough to injure a person who stumbles into its path is not a care robot. These are measurable properties. They can be tested and certified.

It means defining autonomy ceilings — limits on what decisions a care-category robot can make without human confirmation. A care robot should not be able to administer medication, apply physical restraint, or make any decision with irreversible consequences for a human without a human in the loop. These are architectural constraints, not software policies.

It means defining deployment separation — a requirement that platforms certified as care robots not be capable of weapons integration without physical modification that would be detectable and would void the certification. This is the equivalent of dual-use export controls, applied at the product design level. A platform that can accept a weapons attachment with a fifteen-minute modification is not, in any meaningful sense, a care robot. It is a care robot waiting to become something else.

It means defining data separation — prohibitions on sharing behavioral data, operational logs, or training datasets between care-category and combat-capable systems. The AI architectures underlying care robots and combat robots should not be the same architecture trained on different data. They should be developed under different principles, with different safety validation requirements, and the data that shapes their behavior should not flow between them.

None of these definitions are easy. All of them will require serious technical, legal, and ethical work. But the work is doable, and it needs to start before the incidents that make it urgent rather than after.

Robotics needs a neutral convening force—like a Geneva moment—to set enforceable norms. Without it, trust remains undefined and accountability optional.

Who Convenes This

The Geneva Conventions were convened by Switzerland, a neutral nation with both the credibility and the motivation to serve as a honest broker. The initial signatories were twelve European nations. The framework grew over subsequent decades through additional conventions and protocols.

A robotics framework needs a similar convening structure. It needs a party with enough credibility to gather stakeholders who don’t fully trust each other, enough neutrality to be seen as an honest broker, and enough institutional weight to give the resulting agreement meaning.

Several candidates are plausible. The International Committee of the Red Cross has already begun engaging seriously with the questions of autonomous weapons and humanitarian law. The IEEE — the world’s largest professional organization for engineers — has an existing ethics framework for autonomous systems and the technical credibility to define what architectural separation actually requires. The United Nations has existing structures for arms control that could be extended to autonomous systems. A coalition of smaller nations with no major military robotics programs have both the motivation and the credibility to initiate the process without being perceived as acting in their own military interest.

What’s needed is not consensus from the start. The Geneva Conventions didn’t require universal agreement to be meaningful. They required enough signatories with enough credibility that the framework established a norm — a shared understanding of what the world considered acceptable — and that violations carried real reputational and diplomatic costs even for non-signatories.

The same architecture applies here. A framework signed by a meaningful coalition of nations and major robotics manufacturers — one that establishes clear certification categories, verifiable architectural standards, and real consequences for misrepresentation — creates a norm even if not every actor honors it. It establishes what the civilized world considers acceptable. It gives consumers, regulators, and investors a reference point that currently doesn’t exist.

What the Industry Has to Decide

The robotics industry is at a decision point that it is not yet facing directly.

The companies building care robots have a profound commercial interest in the existence of a framework like this — not because they want to be regulated, but because the alternative is an incident that destroys the trust the entire care market depends on, and no individual company has the power to prevent that incident from happening. The framework is in their interest. The separation is in their interest. The certification is in their interest, because certification creates a signal they can use to earn the trust they need.

The companies building military and dual-use platforms have a different calculus. The framework asks them to accept limits on their product’s applicability, to invest in architectural separation that costs money, and to give up the option of selling the same platform into both markets without restriction. That is a real cost, and they will resist it.

But they should consider what the alternative looks like. Absent a framework, the incident described in the previous column is not a possibility — it is a certainty. And when it happens, the regulatory response will not be thoughtful, technically informed, or proportionate. It will be reactive, politically driven, and likely to harm the legitimate applications of robotic technology far more than a proactive framework would.

Reactive regulation is almost always worse than proactive frameworks. The pharmaceutical industry learned this. The aviation industry learned this. The nuclear industry learned this. The robotics industry has the opportunity to learn it before the lesson is imposed, but the window for choosing to learn it is not unlimited.

With real standards, robots earn trust—not just function. Separate care from combat, certify behavior, and the future becomes safe enough to fully embrace.

What Gets Built in the World Where This Works

I want to end this series not with the problem but with the possibility.

A world in which a genuine Geneva Convention for robots exists — in which care robots are architecturally separated from combat systems, certified to verifiable standards, and governed by a multinational framework with real teeth — is a world in which the full promise of care robotics can actually be realized.

In that world, the elderly woman living alone can have a robot companion that her family trusts, because the trust is not based on marketing claims but on verified architectural commitments and independent certification. The sleep-deprived parent can accept help from a machine at 2 in the morning because the framework that governs that machine’s behavior is the same framework that governs the behavior of every certified care robot on Earth — not the preference of the company that built it, revisable in the next software update.

In that world, the drone that delivers your package and the drone that monitors your elderly parent’s wandering behavior in a memory care facility are verifiably, architecturally different from the drone that can be equipped for combat — and that difference is enforced by a framework with enough weight to mean something.

In that world, the quadruped robot that inspects your home’s foundation for damage is not, in any sense that matters, the same machine as the weaponized dog-bot in military footage. The difference is not just in what they’re used for. It’s in what they’re built to be.

Isaac Asimov saw the need for this in 1942 and tried to articulate it in fiction because the serious conversation wasn’t happening anywhere else. He imagined three simple laws, and then spent the rest of his career showing why simple laws weren’t enough — why the real work was in the details, the edge cases, the places where principles meet complexity.

We are living in the moment he was writing toward. The robots are real. The stakes are real. The absence of a framework is real.

The Geneva Conventions were born in the recognition that some things are too important to be left to individual actors to decide on their own, in their own interest, without accountability to anything larger than themselves.

Robots that live with our families and robots that can harm human beings are too important for that.

The world built a treaty before. It can build one again. The question is whether the robotics industry, and the governments that have the power to convene this conversation, will choose to build it before the incidents that make it unavoidable — or after.

History suggests we usually wait for the incidents.

This series has been an argument for not waiting.

One Incident Away

Thomas Frey — Sun, 19 Apr 2026 12:25:35 +0000

Two robots, same tech—one cares, one confronts. When they share origins,
the industry faces a paradox it hasn’t yet acknowledged or resolved.

Trust in robots will not be built incrementally. But it can be destroyed in a single afternoon.

By Futurist Thomas Frey

Part 3 of 4: The Military Paradox Nobody Will Discuss

Let me describe two robots.

The first is designed for eldercare. It moves slowly and deliberately through a home, helps a 78-year-old woman with limited mobility get from her bed to her chair, reminds her to take her medication, detects if she falls, and calls for help if she does. It is gentle by design. Its physical parameters are constrained specifically to prevent it from applying more force than a human hand would use. Its entire architecture is built around one principle: do not harm the person in your care.

The second is designed for military reconnaissance and force projection. It can move fast across difficult terrain, carry significant payload, identify targets using computer vision, and in its more advanced configurations, make or assist with engagement decisions in contested environments. It is capable by design. Its physical parameters are optimized for effectiveness in situations where the humans nearby may be adversaries. Its architecture is built around a completely different principle: accomplish the mission.

Both of these robots exist right now. Both are being actively developed and in some cases deployed. Both use similar foundational technologies — the same locomotion research, the same computer vision systems, the same advances in battery technology and actuator design that have driven the whole field forward.

And both are being developed, in many cases, by the same companies. Or by companies that share investors, share talent, share research lineages, and operate in the same public conversation about the future of robotics.

That is the military paradox. And the robotics industry is not discussing it honestly.

The Funding Reality

To understand why this matters, you need to understand where robotics development money actually comes from.

The Defense Advanced Research Projects Agency has been one of the most important funders of fundamental robotics research for decades. DARPA’s robotics challenges in the 2010s produced technology that directly seeded the current generation of humanoid platforms. Boston Dynamics — whose Atlas robot is the most recognizable humanoid in the world — spent years under the ownership of Google before being sold to Hyundai, but its foundational development included significant defense-adjacent funding and the Atlas platform has been demonstrated in countless military-adjacent contexts.

The US Army has active programs evaluating robotic platforms for logistics, reconnaissance, and combat support. The Defense Department’s vision of the future battlefield includes robotic systems operating alongside human soldiers. The investment flowing into defense robotics is enormous and accelerating, and it is not cleanly separated from the investment flowing into consumer and care robotics. The research is connected. The talent moves between sectors. The companies that win defense contracts build capabilities that transfer.

None of this is secret. It is all documented in public filings, press releases, and conference presentations. What is not being said publicly — at least not in the consumer-facing conversations about the wonderful future of robot caregivers and domestic helpers — is what the convergence of these two development tracks means for the trust that the entire industry depends on.

What Footage Does

Trust is not a technical property. It cannot be engineered into a product the way you engineer payload capacity or battery life. It is a social property — something that exists in the relationship between a technology and the public that encounters it. And it is profoundly asymmetric in how it is built and destroyed.

Building trust in a technology takes years. It requires consistent, reliable, incident-free performance across millions of interactions, in environments that matter to real people, witnessed by enough people that the positive evidence accumulates in public consciousness. It requires the absence of dramatic failures. It requires time.

Destroying trust in a technology can take minutes. It requires one incident, clearly documented, that is frightening enough to crystallize the fears that were always present but suppressed by the weight of positive experience.

Aviation spent decades building the trust that makes billions of people comfortable getting on commercial aircraft. A single high-profile crash, handled badly, can create a confidence crisis that grounds fleets and reshapes industry dynamics for years. The trust is real and hard-won. The vulnerability is permanent.

The robotics industry has not spent decades building public trust. It is in the early stages of that process. The positive experiences are limited to relatively small populations of early adopters, researchers, and industrial users. The general public’s relationship with humanoid robots is still primarily mediated by science fiction, product demonstrations, and news coverage — all of which create impressions, but none of which create the deep experiential trust that comes from living with a technology over time.

Now consider what happens when footage appears — and it will appear, because it always does — of a military robot causing harm. Not a weapon failing to discriminate properly in a war zone thousands of miles away. Something closer. Something that looks, to a person watching it on a phone screen, like the same kind of robot that companies have been telling us will help with our elderly parents and our young children.

The human brain is not equipped to parse the difference between a Boston Dynamics robot deployed in an eldercare demonstration and a Boston Dynamics robot deployed in a military context. It sees the machine. It sees what the machine did. It draws the conclusion that machines of that type do that kind of thing.

That is not irrational. That is how trust works.

Mixing military and care robots blurs trust. If the same technology serves harm and help, the public won’t separate them—and trust collapses.

The Branding Problem That Isn’t Being Named

Several robotics companies are actively pursuing both markets simultaneously — or selling the same underlying platform into both tracks. Figure AI, founded in 2022 and now one of the most heavily funded humanoid robotics companies in the world, has announced partnerships with both BMW for manufacturing and the US military. Sanctuary AI is working on general-purpose robots for commercial environments. Ghost Robotics — which makes quadruped robots physically similar to Boston Dynamics’ Spot — has supplied platforms to the US Air Force and been photographed with weapons attachments. The images went viral. The consumer robotics industry noticed and said almost nothing publicly.

The challenge for the industry is structural, not incidental. Military robotics and care robotics are not merely different products. They are, in the deepest sense, antithetical products. One is optimized for keeping humans safe through force limitation and harm avoidance. The other is optimized for operational effectiveness in environments where harm is the context. The values embedded in these two design tracks are not merely different — they are opposed.

When the same corporate family, or the same underlying technology, is visible in both tracks, the public’s ability to maintain the distinction breaks down. And the public’s ability to maintain that distinction is the entire foundation on which the care robotics market is built.

A parent deciding whether to trust a robot with their child is not running a technical analysis of that specific robot’s safety architecture. They are asking a simpler, more human question: do robots in general feel safe? Is this a technology that is fundamentally oriented toward human wellbeing, or is it a technology that is fundamentally a tool of power, and the care applications are just one version of that tool?

Right now, the honest answer to that question is: we’re not sure. And “we’re not sure” is not a foundation for the kind of trust that care robotics requires.

The Incident That Changes Everything

I want to be specific about the scenario I am describing, because vagueness lets the industry dismiss this concern as speculative.

The scenario is not a hypothetical future event. It is a near-certainty given current trajectories. Here is the shape of it.

A military or law enforcement robot — a real, deployed system, not a prototype — is involved in an incident that causes civilian harm. Or a weapons-equipped quadruped robot appears in footage from a conflict zone operating in a way that the watching public finds disturbing. Or a security robot in a domestic context behaves in a way that is aggressive enough to generate viral footage. Or a military demo video is released that shows a humanoid robot performing actions that, out of context, look alarming.

The footage spreads. Because footage always spreads. The coverage does not carefully distinguish between military and care applications, between quadrupeds and humanoids, between combat robots and eldercare robots. It covers robots. The public discussion does not carefully distinguish either. The comment sections do not distinguish. The legislation that follows does not distinguish.

And the care robotics companies that have spent years building toward the moment when ordinary families trust these machines in their homes will find that the floor has dropped out from under their market. Not because their product failed. Because a different product, built on the same general technology, failed in a way that was visible, frightening, and impossible to contextualize away.

The trust destruction will be rapid. The trust rebuilding will take years. And the people who will suffer most from that lost decade are not the investors. They are the elderly people who needed a robot helper and couldn’t get one because the public turned against the category. The families who could have been supported and weren’t. The caregivers who could have been helped and weren’t.

Care and combat robots and drones can’t blur together. Without clear separation, one incident could collapse trust across the entire industry—before safeguards exist.

What the Industry Is Choosing Not to Do

The solution is not for robotics companies to stop taking defense contracts. The defense dollars are real, the applications are legitimate in their own context, and unilateral disarmament in the face of competitive pressure is not a realistic ask.

The solution is structural separation — a clear, public, verifiable commitment to maintaining the difference between care robots and combat robots at the level of design, deployment, branding, and governance. Not a press release. Not a corporate ethics policy that can be quietly revised when a lucrative contract appears. An architecture that makes the distinction real, visible, and durable.

That architecture does not currently exist. The industry has not built it because building it would require acknowledging the problem, and acknowledging the problem would require saying publicly what most people in the industry know privately: that the military and care robotics tracks are in fundamental tension with each other, that the tension is a threat to the care robotics market’s long-term viability, and that nobody has figured out how to resolve it.

The companies in this space are one incident away from a crisis they are not prepared for. The incident will not be something they caused. It will be something that happened somewhere else, in a different context, with a different product. But it will look enough like their product, on a small screen, viewed by a frightened public that doesn’t know the difference between what was built for a battlefield and what was built for a nursery.

That day is coming. The framework to survive it doesn’t exist yet.

Next: A Geneva Convention for Robots — The world didn’t wait for weapons manufacturers to self-regulate warfare. It built a treaty. What would a binding international framework for robot ethics actually look like — who convenes it, who signs it, and what does “do no harm” mean when encoded in machine behavior?

The Diaper Test

Thomas Frey — Sat, 18 Apr 2026 19:56:04 +0000

The real test of AI isn’t conversation—it’s care. Until a robot
can handle fragile, human moments, it hasn’t earned our trust.
…

The real measure of a robot has never been what it can do in a warehouse. It’s whether you’d trust it alone with the people you love most.

By Futurist Thomas Frey

Part 2 of 4: The Wrong Problem

It was 2 in the morning, and Sarah hadn’t slept more than three hours in as many days.

Her two-month-old, Leo, had been crying for what felt like hours. She placed him on the changing table, peeled back the diaper, and watched the situation spiral. Leo kicked, squirmed, and managed to make the mess considerably worse. It spread across the changing table, onto Sarah’s shirt, and across the floor. She was exhausted, overwhelmed, and running out of hands.

I told this story in a column on FuturistSpeaker.com earlier this year, posing what I called the Turing Test for humanoid robots. The original Turing Test — Alan Turing’s 1950 benchmark for machine intelligence — asked whether a machine could hold a conversation indistinguishable from a human. A meaningful threshold, but an intellectual one. What I proposed was a different kind of threshold entirely: not can the machine think like a human, but can it act like one in the moments of genuine, physical, emotionally loaded caregiving that define what it means to care for another person?

The test: Can a humanoid robot change a dirty diaper at 2 in the morning — gently, competently, calmly, without injuring an infant or escalating the chaos — in a way that a frazzled, sleep-deprived parent would trust it to do alone?

I called it the Diaper Test. And the more I’ve thought about it since, the more I believe it is not just a benchmark for robotic capability. It is the benchmark for whether this industry has earned the right to be where it’s heading.

Why Turing Got It Half Right

Turing’s original test was revolutionary because it shifted the question from internal mechanism to observable behavior. We don’t need to know how a machine thinks, he argued. We just need to know whether its behavior is indistinguishable from thinking. That reframing changed everything about how we approach artificial intelligence.

But Turing was working in the realm of language and cognition. His test lives in conversation — in text or speech, in the back-and-forth of questions and answers. When AI systems pass versions of the Turing Test today, they do so through words. They can argue, persuade, explain, and comfort in language that sounds deeply human.

What they cannot yet do is walk into a dark nursery at two in the morning, pick up a squirming, crying infant with the precise force required to be secure without being harmful, clean a chaotic mess while keeping the baby calm, and set a clean, soothed child back down — all without any of the dozens of micro-adjustments going wrong in ways that a tired human parent would catch on instinct.

That is a different kind of test. It requires fine motor precision at the level of handling a fragile, uncooperative living being. It requires real-time adaptability to behavior that is entirely unpredictable — a baby who kicks at exactly the wrong moment, who grabs at something they shouldn’t, who startles in a direction the robot didn’t anticipate. It requires the ability to soothe and calm through touch, sound, and movement — the physical language of comfort that parents develop over weeks of learning their specific child’s specific responses.

And it requires judgment. Not the computational kind. The kind that knows the difference between a cry of distress and a cry of mild frustration, that understands when to persist and when to pause, that can read a situation and decide what the right action is when the right action isn’t in any manual.

Robotics measures performance in controlled tasks. Real trust depends on unpredictable moments—where judgment matters more than benchmarks. That’s the gap the industry hasn’t closed.

What the Industry Is Actually Building For

Here is the uncomfortable question. Walk through any major robotics demonstration right now, and count the benchmarks being celebrated.

Payload capacity. Locomotion stability on uneven terrain. Object manipulation success rates in controlled environments. Battery endurance. Processing latency. Navigation accuracy in mapped spaces. The ability to fold laundry, operate a drill press, or sort packages in a fulfillment center.

These are real engineering achievements. They matter. But none of them answer the question that the Diaper Test asks.

What does the robot do when something happens that wasn’t in the training data? When the baby rolls in an unexpected direction at exactly the wrong moment? When the elderly patient becomes frightened and starts to resist? When the child runs in front of the machine and the navigation system has 200 milliseconds to decide what to do in a situation where 200 milliseconds is the entire margin?

These are not exotic edge cases. They are the routine texture of caring for vulnerable human beings. Any parent, any nurse, any home health aide will tell you that the job is made almost entirely of unexpected situations. Moments where the correct response requires not just processing speed but something that functions like wisdom — the ability to weigh competing obligations in real time when the stakes are irreversibly human.

The industry’s benchmarks measure performance in expected conditions. The Diaper Test measures readiness for unexpected ones. We have been conflating the two as though they were the same problem. They are not.

The Intimacy Gap

In the original FuturistSpeaker.com column, I argued that passing the Diaper Test would be a watershed moment — the robotic equivalent of the iPhone, the kind of breakthrough that doesn’t just sell products but reshapes what people believe is possible. I stand by that. The moment a robot can genuinely handle that 2am scenario — not in a lab, not in a demo, but in a real home with a real exhausted parent watching — the consumer robotics market will never be the same.

But here, in the context of this series, I want to press on a harder version of the same argument.

The spaces where humanoid robots are being positioned — homes, hospitals, care facilities, nurseries — are not like warehouses. Warehouses are designed environments, controlled and predictable, built around machine-compatible workflows. A home is chaos organized by love. A hospital room is fear and vulnerability and the constant possibility of things going wrong in ways that matter enormously. A nursery is a space where the margin for error is measured in different units entirely.

The intimacy of these spaces is what makes the Diaper Test the right benchmark. Not because changing diapers is the most complex task imaginable, but because it concentrates, in one scenario, all of the things that make care work genuinely hard: physical delicacy, unpredictable human behavior, emotional stakes, and the irreversibility of certain kinds of failure.

A robot that fails a warehouse sorting task costs the company time and money. A robot that fails the Diaper Test costs something that cannot be quantified and cannot be patched in the next update.

The Experts Nobody Is Asking

In the original column I wrote about the societal transformations that a diaper-changing robot would unleash — the potential to ease the burden on young families, support aging populations, rebalance caregiving responsibilities, and give parents back the time and energy they need to actually be present with their children. I believe all of that is true.

But there is a community of people who understand what it would actually take to get there — and they are almost entirely absent from the conversations shaping this industry.

Pediatric nurses. Neonatal intensive care unit staff. Hospice workers. Home health aides who spend twelve-hour shifts with people who have late-stage dementia. Foster care workers. These people know, in their bodies and their years of experience, what genuine care requires. Ask any one of them whether the robots they have seen demonstrated are ready to be trusted alone with the people they serve, and their answers would be more honest, more specific, and more useful than most product roadmaps currently circulating in the robotics investment community.

They should be in the room where these products are being designed. They should be setting the benchmarks. They should be the ones deciding when the test has been passed.

They are not. Not yet. And that gap between the people who build care robots and the people who actually provide care is one of the most dangerous gaps in the industry.

The real benchmark isn’t demos—it’s trust. Until a parent would leave a child alone with a robot, the technology isn’t ready.

What Passing Looks Like

So what would it actually mean to pass the Diaper Test?

It would mean a robot that a parent who has seen it perform — not in a demo, but in the real conditions of their real home with their real child — would genuinely trust to be left alone. That trusts its physical judgment. That believes it will handle the unexpected correctly. That has no hesitation about leaving the room.

That bar has never been met. The industry is not close to meeting it. And the path to meeting it does not run through better warehouse benchmarks or more impressive locomotion demos.

It runs through a completely different orientation to the design problem — one that starts not with what the robot can do in optimal conditions but with what it must reliably do in the hardest ones.

We are the last generation without advanced robots everywhere. Our children will grow up as robot natives, for whom humanoid helpers are simply part of the world. For that future to be the one I described in my original column — the one where robots genuinely extend human capability and human care — the industry needs to prove it can pass the test that actually matters.

Not the benchmark that impresses investors. The one that earns the trust of a sleep-deprived parent at two in the morning.

That test is still waiting.

Next: One Incident Away — Trust in robots will not be built incrementally. But it can be destroyed in a single afternoon. The military robotics programs running parallel to care robots are the industry’s most dangerous open secret.

The Asimov Problem

Thomas Frey — Sat, 18 Apr 2026 19:35:58 +0000

We built powerful robots without shared rules. Asimov imagined safeguards—
industry delivered terms of service. One incident could expose a framework that doesn’t exist.

…

Why the most physically intimate technology in human history has no ethical spine — and why that should terrify everyone

By Futurist Thomas Frey

Part 1 of 4: The Rules We Never Wrote

In 1942, a science fiction writer named Isaac Asimov published a short story called “Runaround.” In it, he introduced three laws governing robot behavior — simple, elegant rules designed to ensure that machines built to serve humanity wouldn’t end up harming it. The First Law: a robot may not injure a human being. The Second: a robot must obey human orders unless those orders conflict with the First Law. The Third: a robot must protect its own existence unless that conflicts with the first two.

Asimov wasn’t writing policy. He was writing fiction. He didn’t expect his three laws to become the actual operating framework for an industry that didn’t yet exist. He expected someone else — engineers, ethicists, governments, the humans who would eventually build these things — to do the serious work when the time came.

That time came. The serious work didn’t.

What we have instead are terms of service agreements. Liability disclaimers. Corporate ethics boards that report to the same executives whose bonuses depend on shipping product. And thousands of companies racing toward a market that is projected to reach half a trillion dollars within a decade, each one moving as fast as it can, each one assuming that someone else is handling the framework question.

Nobody is handling the framework question.

That is what this series is about. Not about whether robots are impressive — they are. Not about whether the technology will transform society — it will. But about the fact that we are building the most physically intimate technology in human history with no shared ethical architecture, no binding international framework, and no serious reckoning with what happens when something goes wrong in a way that can’t be fixed by a software update.

We are one incident away from an industry-wide crisis. And the industry, for the most part, is not discussing it.

What Asimov Actually Understood

Here’s the thing about the Three Laws that most people who cite them miss. Asimov didn’t write them as a solution. He wrote them as a problem.

Almost every story in his robot series is about the ways the Three Laws fail — the edge cases, the interpretations, the unintended consequences of simple rules applied to a complex world. The Laws were a starting point, and his fiction was a decades-long exploration of why starting points are never enough. He was doing the ethical stress-testing in narrative form because he understood that the hard questions don’t answer themselves.

What he saw, eighty years ago, was that the question of robot ethics isn’t primarily a technical question. It’s a values question. What do we want these machines to protect? What do we want them to refuse? Under what circumstances should a robot override a human instruction, and who decides? These are not engineering problems. They are civilization problems — the kind that require deliberate, collective, binding agreement before the machines are in the room, not after.

We have not had that agreement. We have not even seriously begun the conversation that would produce it.

Robots are entering homes and hospitals without enforced safety standards—like cars before seat belts. This time, the risks are far more personal and immediate.

The Industry That Built the Car Without Seat Belts

Let me describe what the current robotics industry actually looks like from the inside, because the gap between the public narrative and the operational reality is significant.

Humanoid robots are no longer a research project. They are a product category. Companies including Boston Dynamics, Figure AI, 1X Technologies, Agility Robotics, Tesla, and Apptronik are developing and in some cases already deploying bipedal robots in commercial and industrial environments. The pace of capability improvement has been startling even to people who have been watching this space for years.

These robots are entering warehouses. They are beginning to enter healthcare settings. They are being positioned for eldercare, for childcare, for domestic assistance in private homes. They will, within a timeframe measured in years not decades, be physically present in the most vulnerable spaces of human life — the nursery, the hospital room, the home of someone who can no longer fully care for themselves.

And the framework governing their behavior in those spaces is: whatever the company that built them decided to put in the software, subject to revision in future updates, governed by the terms of service agreement the purchaser clicked through.

That is the seat belt situation before Ralph Nader. The industry knows the cars are going fast. Nobody has seriously mandated what happens when one crashes.

The automobile industry’s resistance to safety standards killed tens of thousands of people before regulation intervened. But cars, even at their most dangerous, were not physically present in your bedroom. They were not holding your child. They were not making decisions, in real time, about whether to restrain an elderly patient who is trying to stand up.

The robots that are coming will be.

Why This Matters More Than Any Previous Technology

I want to be precise about what makes this different from every other technology governance challenge we’ve faced.

The internet raised serious questions about privacy, misinformation, and manipulation. We largely failed to address those questions at the speed they required, and we are living with the consequences. But the internet’s harms are, for the most part, mediated — they happen through screens, through information, through influence. They are real and serious. They are not physical.

AI governance raises questions about bias, accountability, and autonomous decision-making that we are only beginning to grapple with. But AI, at its current stage of deployment, operates primarily in the domains of language and data. When it fails, the failure is usually a wrong answer, a biased output, a bad recommendation.

When a robot fails, the failure can be a broken bone. A fall down a staircase. A restraint applied with too much force. A navigation error in a room with a sleeping infant.

The physicality of robotics is what makes the governance question categorically different. Physical presence in human spaces, physical interaction with human bodies, physical consequences for physical failures — these are not comparable to any previous technology category. And the spaces where these robots are being deployed are specifically the spaces where the humans present are most vulnerable: the elderly, the sick, the very young, and the people who care for them.

We are building intimate technology. We have no intimate ethics.

One visible robot failure could trigger backlash against the entire industry. Without real safety frameworks, trust is fragile—and one incident could set progress back years.

The Stakes Nobody Is Naming

Here is what the robotics industry’s current trajectory leads to, absent intervention.

A serious incident will occur. It may be a care robot that injures a patient. It may be a domestic robot that fails in a way that harms a child. It may be something that happens on video in a way that is impossible to contextualize away. When it does, the public response will not be calibrated to the specific failure of the specific product from the specific company. It will be a response to robots. To the category. To the idea.

The aviation industry learned this the hard way. A single crash, handled badly, can ground an entire fleet and shake an industry’s foundations for years. The difference is that aviation has always had a robust, internationally coordinated, independently enforced safety framework. When a crash happens, there is an investigation, a finding, a corrective action, and a binding requirement that every operator implement it.

Robotics has none of that. It has press releases and pivot announcements.

The industry is fragile in the way that any industry is fragile when it has built market value on public trust without building the institutional architecture that justifies that trust. One incident. One video. One family’s story told on the front page. That’s the distance between where we are today and a crisis that sets the entire category back a decade.

Asimov saw this coming in 1942. He tried to tell us.

We kept the footnote and ignored the spirit.

Next: The Diaper Test — The measure of a robot isn’t what it can do in a warehouse. It’s whether you’d trust it alone with the people you love most. The industry is optimizing for the wrong problem.

Futurist Speaker

The Neutral Sky: Why Space May Be the Only Fair Ground for AI in the Developing World

What Orbital Edge Compute Actually Is

The Eye in the Sky, Reconceived

The Neutrality Problem, Honestly Examined

The Honest Ceiling and the Real Floor

Related Articles

The Token Revolution: How the Global South Becomes the Global Brain

The System That Was Always Running

The Reinforcing Loop That Is About to Flip

What a Token Economy Actually Means in System Terms

The American Miscalculation: Fixing the Wrong Variable

The Intellectual Service Economy Follows the Data

Related Articles

The Neighborhood Becomes the Data Center

The Age of the Hyperscale Monolith Is Ending — and the Next Internet May Be Hiding on the Side of Your House

The Power Wall Has Arrived

Your Home as Infrastructure

The Street Becomes the Server Farm

The AI Electrical Grid

The Questions That Need Asking Now

The Shape of What’s Coming

Related Articles

The Data Centers That Will Float

Why the Most Radical Solution to the AI Energy Crisis Is Already at Sea

The Ocean Has Been Waiting for This Conversation

What a Floating Data Center Actually Is

Microsoft Proved the Concept. Nobody Scaled It.

Why Wave Energy Changes the Calculation

The Questions That Deserve Direct Answers

The Infrastructure Inversion

Related Articles

Twelve Inventions That Prove the Future Has a Sense of Humor — And Means Business

The Lollipop That Plays Music Through Your Bones

Scalp Intelligence in Ten Seconds

A Phone That Starts Fires

The Holographic Companion

The Blade That Thinks It’s a Laser

Fingernails as Displays

The Robot That Tows Your Car

The Haircut That Can’t Go Wrong

The Pet Whose Soul Survives

The Engine That Burns Like a Tornado

The World Through Your Pet’s Eyes

Dinosaurs That Know You’re Standing There

The Pattern Underneath

Related Reading

The Computer That Disappeared Into the World

What Spatial Computing Actually Is

Where It Actually Stands Right Now

The Honest Constraints

The Industries Being Rebuilt

The Competitive Landscape

What Changes When This Matures

The Longer View

Related Reading

A Geneva Convention for Robots

What the Framework Has to Separate

What “Do No Harm” Actually Means in Machine Behavior

Who Convenes This

What the Industry Has to Decide

What Gets Built in the World Where This Works

Related Reading

One Incident Away

The Funding Reality

What Footage Does

The Branding Problem That Isn’t Being Named

The Incident That Changes Everything

What the Industry Is Choosing Not to Do

Related Reading

The Diaper Test

Why Turing Got It Half Right

What the Industry Is Actually Building For

The Intimacy Gap

The Experts Nobody Is Asking

What Passing Looks Like

Related Reading

The Asimov Problem

What Asimov Actually Understood

The Industry That Built the Car Without Seat Belts