For years, cloud computing trained us to think of infrastructure as an abstraction.

You clicked a button, and a server appeared. A storage bucket appeared. A model endpoint appeared. The physical world stayed out of view, which was useful for developers and companies building fast. It was also incomplete.

AI is making that abstraction harder to sustain.

Every prompt runs somewhere. Every RAG pipeline, model endpoint, document search, agentic workflow, fine-tuning job, and batch generation task depends on chips, electricity, cooling, land, grid access, networking, contracts, maintenance, and the legal geography of where the workload lives.

The cloud did not become less physical because we stopped looking at it. The machinery was always there. AI is making it visible again.

The next cloud region will be shaped by more than latency, compliance, or a city name in a dropdown. It will be shaped by power: where it is produced, where it is available, how fast it can be connected, and whether it can support compute without turning every new AI product into another claim on an already strained grid.

That is the part of the AI boom that deserves more attention.

We have written before that AI’s next phase hinges on infrastructure, not algorithms, and that neoclouds are where the AI boom meets the real world. This is the next step in that argument: AI compute will increasingly follow power.

The old cloud map is getting out of date

A cloud region used to answer a simple question: where should the workload run so users get decent latency and the customer can satisfy a basic residency requirement?

That still matters. It will always matter. AI adds another layer, though. A region is no longer just a point on an infrastructure map. It is also a power question.

The International Energy Agency estimates that data centers consumed around 415 TWh of electricity in 2024, about 1.5% of global electricity consumption. In its base case, that figure roughly doubles to 945 TWh by 2030.

AI is not the only pressure on the grid, but it is one of the fastest-growing and most concentrated. Data centers do not spread evenly across a country like household appliances. They land in specific places, ask for large grid connections, and turn local energy planning into a cloud infrastructure problem.

Goldman Sachs Research expects US data center power demand to rise from 31 GW in 2025 to 66 GW in 2027. It also expects only about 50% to 60% of scheduled data center capacity over the next one to two years to come online on time, because permitting, construction, supply chains, labor, and grid access do not move at software speed.

Europe is facing the same physical constraint. Reuters reported that EU data center capacity is expected to grow from 12 GW to 28 GW by 2030, while the European Union prepares minimum energy-efficiency standards for data centers.

This is what the cloud interface hides. Capacity is a product, but it is also a negotiation with the physical world.

Inference makes the bottleneck permanent

Training gets the drama. It has the frontier labs, the giant clusters, the benchmark announcements, and the eye-watering capital expenditure.

Inference is quieter. It is also where AI becomes ordinary infrastructure.

A training run starts, finishes, and can be studied afterward as a large cost. Inference sits inside the product. It runs every time a customer asks a question, a support ticket gets classified, a document gets summarized, a legal file gets searched, a student receives feedback, a code assistant completes a function, or an internal agent routes a workflow.

That changes the economics.

Deloitte expects inference to account for roughly two-thirds of AI compute in 2026, up from one-third in 2023. That shift matters because inference compounds. A bad infrastructure choice does not stay in a lab budget. It becomes a tax on every useful interaction.

We made that argument recently in The H100 tax: many AI teams overpay because they treat every production prompt like a frontier workload. That mistake shows up first as a GPU bill. At scale, it becomes something larger: wasted capacity, higher energy pressure, and a stronger dependency on the same small group of providers.

Inference is where AI stops being a demo and becomes a utility-like load. That load needs infrastructure that is affordable enough to run every day, close enough to serve real users, governed enough for sensitive workloads, and efficient enough not to waste scarce power.

Power is moving into the architecture

The cloud industry spent years treating energy as a procurement issue. Data center operators secured power, built campuses, and customers saw the result as a region.

AI changes the order of questions.

When power is the bottleneck, compute cannot always wait for a massive new campus to win a grid connection. In many cases, compute should move toward available power.

That is the core idea behind the Antimatter approach. We deploy compute where renewable energy is produced. That solves the power bottleneck blocking hyperscalers.

It sounds obvious once stated plainly. If renewable energy sites already have land, grid connections, and local energy production, then AI infrastructure can be designed around those facts instead of assuming the only serious model is another giant centralized campus.

That does not mean every workload belongs next to a solar plant. It also does not mean hyperscalers will disappear or that one architecture will replace every other architecture. Infrastructure is rarely that tidy.

The important change is that energy stops being an invisible input and becomes part of the architecture.

Some workloads will still belong in large centralized regions. Some will run on specialized GPU clouds. Some will run through distributed micro data centers close to power, users, or regulatory boundaries. The cloud map starts to change because the constraints underneath it have changed.

We have already argued that the cloud was never supposed to live in five places. AI makes that argument less philosophical and more practical. Concentrated infrastructure is becoming more expensive, more politically exposed, and more limited by local power realities.

The next cloud will have to be more distributed because the physical world is distributed. Energy is distributed. Demand is distributed. Regulation is distributed. Users are distributed. Data is distributed.

The cloud should be built with that reality in mind.

A contract is not the story. The infrastructure shift is.

The recent Antimatter and CloudGrid Energy announcement matters because it points to that wider shift. (June 2026)

Antimatter signed a €580M framework contract with CloudGrid Energy to deploy 280 Policloud units and 29,000 GPUs across Europe by the end of 2027.

The planned deployment also represents 2,000,000 vCPUs and 35 MW of energy capacity, with 16 sites already secured in France, Germany, Italy, Spain, and Sweden. The first unit is already operational at the Bonne Voisine site in Aube, France, according to the official Policloud announcement.

The numbers show scale. The model behind the numbers matters more.

CloudGrid Energy develops and operates modular digital infrastructure units in renewable energy sites. Policloud brings modular computing infrastructure. Hivenet brings the cloud platform layer for AI, high-performance computing, storage, and secure file sharing.

In plain English: power, infrastructure, and software have to meet.

A data center alone is not a cloud. A GPU fleet alone is not a product. A renewable energy site becomes AI infrastructure only when workloads can run there through usable tools, predictable pricing, regional control, storage, transfer, support, and an interface developers can work with.

That is Hivenet’s role in the wider Antimatter group: to make distributed infrastructure usable for real workloads.

The distinction matters because AI teams do not buy an energy thesis. They need to run workloads.

They need inference that does not crush margins. They need storage near compute. They need secure transfer. They need support for open-source and open-weight models. They need region choices that mean something. They need to know where their data and workloads live. They need a path out if the relationship stops working.

That is what practical sovereignty looks like.

We have written before that Europe keeps talking about sovereign AI. The phrase can get vague quickly. It becomes a slogan, a policy panel, a procurement checkbox, or a regional label in a sales deck.

The useful version is simpler: can European companies run AI workloads on infrastructure they can access, afford, understand, and control?

If they cannot, sovereignty remains decorative.

SMEs and startups cannot build on invisible constraints

The largest AI companies can reserve huge clusters, negotiate power deals, hire infrastructure teams, and absorb waste for longer than anyone else.

Most companies cannot.

This infrastructure is built for European SMEs, startups, and entrepreneurs facing prohibitive inference costs and dependency on non-European players.

That sentence is not a side note. It is the market reality.

If AI infrastructure becomes available only to companies with hyperscaler contracts, giant cloud credits, or the patience to wait through capacity queues, then the next phase of AI will narrow instead of widen. Smaller teams will rent APIs from the companies that already control the stack. European companies will talk about autonomy while routing more of their intelligence through infrastructure they do not own, govern, or meaningfully influence.

That is dependency with a better interface.

The answer is not to romanticize smallness. AI infrastructure needs serious engineering, serious operations, and serious capacity. The better path is to build capacity in a shape that gives more teams a fair chance to use it.

That means right-sizing GPUs instead of overbuying prestige hardware. It means deploying compute closer to available energy. It means treating storage and data movement as part of the workload, rather than an afterthought. It means making regional control part of the product, rather than a promise stapled to the contract.

It also means being honest about cost.

Inference is no longer a one-time experiment. It is becoming a recurring operating expense for companies that use AI in real products. If that cost is too high or too unpredictable, many useful AI products will never leave the prototype stage.

The AI industry likes to talk about intelligence. Builders still have to pay the bill.

The better question for AI teams

The old cloud question was: which provider should we use?

The better AI infrastructure question is: where should this workload live?

That sounds simple, but it opens the right set of concerns.

• Where is the power coming from?
• Can the workload run in the region we need?
• Do we need the most expensive GPU, or the right GPU?
• What is the cost per useful inference when the product is busy?
• Where is the storage?
• What happens to egress?
• Can we move if the architecture changes?
• Who controls the infrastructure path?
• Are we building a product, or are we slowly building a dependency we will regret?

These questions decide who can build AI products sustainably, who can scale inference without burning margins, and who gets trapped inside infrastructure choices made too early.

That is why the “grid as cloud region” idea matters. It puts the physical constraint where it belongs: at the center of infrastructure strategy.

The cloud is becoming accountable again

The first cloud era won because it made infrastructure feel easy.

The next cloud era has to do something harder. It has to make infrastructure visible enough to trust without making it so complex that only specialists can use it.

That means being clearer about the material system behind AI: the chips, the power, the grid, the sites, the regions, the data paths, and the cost of every useful inference. It also means being clearer about sovereignty. Regional control cannot be treated as an add-on if the workload still depends on infrastructure that companies cannot access, price, move, or understand.

AI needs a different cloud because AI has changed the shape of demand.

Inference runs continuously. Power matters locally. Capacity comes online slowly. Europe needs infrastructure it can actually use. SMEs and startups need costs that make sense before they scale. The next wave of builders will care less about the elegance of a dashboard and more about whether the underlying system is available, efficient, affordable, and under their control.

The cloud region is becoming a grid question.

That may sound less glamorous than model benchmarks, but glamour was never the scarce resource. Capacity is.

At Hivenet, we think the future of AI infrastructure will be built closer to power, closer to users, closer to data, and closer to the real constraints that decide whether products can run at all.

The cloud is still here. AI is forcing more people to see the physical system behind it.

This time, we should build it as if the physical world matters.

Written by Hivenet Editorial
Hivenet is a distributed cloud platform for storage, file sharing, and compute. It helps teams run cloud and AI workloads across independent infrastructure, with clearer control over cost, resources, and data movement.

Discover more at hivenet.com.

Download the latest version of the Hivenet app

The AI industry has made one of its least rational habits sound like discipline: renting the most expensive GPU before proving the workload needs it.

For a while, access to top-end datacenter GPUs became a badge of seriousness. Saying “we’re running on H100s” sounded like ambition, technical maturity, and enterprise safety in one sentence. Investors understood it. Buyers recognized it. Engineers could defend it. In the first AI infrastructure scramble, that made sense. Frontier training was hungry, GPU supply was tight, and the companies building the largest models were fighting for every accelerator they could get.

That reflex is now leaking into places where it does not belong.

A lot of what AI businesses actually ship is inference: extracting fields from documents, classifying support tickets, summarizing calls, running RAG, answering internal questions, moderating content, generating first drafts, routing workflows, and powering narrow copilots. It runs repeatedly inside real products under cost pressure. It needs to be fast enough, reliable enough, close enough, and cheap enough to run all day.

It does not need every request to dress like a frontier lab.

We have written before that AI has made the physical reality of the cloud harder to ignore. Every prompt runs somewhere. That somewhere has power, cooling, chips, fiber, contracts, jurisdiction, and cost attached to it. The same logic applies inside the GPU decision. The useful question is simple: what does this workload actually need?

AI hardware has developed a quiet social hierarchy.

At the top sit H100s and H200s. Below them, A100s. Further down, RTX cards carry the unfair baggage of being called “consumer” GPUs. That word does more damage than people admit. It tells buyers the hardware is less serious before anyone has asked whether the workload fits.

“Consumer” describes the sales channel. It does not define the ceiling of useful computation.

Don’t get us wrong; H100s are extraordinary hardware. They earn their place for large models, large context windows, heavy batching, training, 80 GB HBM, NVLink, MIG, strict procurement rules, and workloads that need datacenter GPU features. But when a team uses that class of hardware for sub-30B open-weight inference that can run on one 24 GB or 32 GB card, the decision starts to look less like engineering and more like insurance.

Expensive insurance.

This is where the conversation becomes uncomfortable. The AI industry likes to talk about breakthroughs, benchmarks, and model capabilities. It spends far less time talking about overbuying infrastructure because overbuying infrastructure sounds prudent. Nobody gets criticized for choosing the expensive option. Plenty of people get questioned for choosing the cheaper one.

That is exactly why the H100 tax exists.

Compute with Hivenet currently offers RTX 4090 and RTX 5090 instances for AI workloads. The RTX 4090 provides 24 GB of VRAM, while the RTX 5090 provides 32 GB. Current pricing is €0.40/hour for a 4090 and €0.75/hour for a 5090, with fixed on-demand pricing, per-second billing, no egress fees, and regional control.

Run a single instance around the clock for 30 days, and the math is straightforward: roughly €288 for a 4090 or €540 for a 5090, before tax and any surrounding costs. Compare that with the H100 market, where on-demand rental prices vary widely but often land in the low single-digit euros or dollars per GPU-hour, with some significantly higher.

The exact comparison depends on the provider, region, contract, support model, and workload. That is why benchmarks matter. But the shape of the problem remains the same: if the model fits on a dedicated RTX-class card and the workload does not require datacenter-only features, the larger GPU may be buying comfort rather than value.

Training waste is painful. Inference waste compounds.

A training run happens, finishes, and becomes a cost you can examine. Inference is different. It sits inside the product. Every user action, support ticket, document upload, search query, or chatbot response can carry a small compute cost. When the product grows, that cost grows with it.

This is where poor GPU sizing becomes dangerous. It does not show up as a dramatic failure. It shows up as a margin that quietly disappears.

No one writes “we overpaid because the famous GPU felt safer” in the postmortem. The line item will say cloud compute. The cause will be vaguer: conservative architecture, unclear benchmarks, fear of underprovisioning, procurement comfort, or the desire to standardize too early around hardware that sounded serious in the board meeting.

Cloud teams already know this pattern. Companies have spent years overprovisioning CPUs, memory, databases, and managed services because bigger instances feel safer than right-sized ones. AI is repeating the same behavior with more expensive hardware and a stronger status script.

The H100 has become the new oversized instance.

RTX will not replace H100 across AI infrastructure. That would be a weak argument. The stronger argument, and the one we are making, is narrower: many everyday inference workloads should start with the smallest serious GPU that fits the model, latency target, context window, and reliability requirement. If that is a 24 GB or 32 GB card, start there. If the workload crosses the boundary, move up.

That boundary is practical. Larger models, long context windows, strict latency targets, heavy batching, multi-GPU memory requirements, formal enterprise procurement rules, or large-scale fine-tuning may genuinely require A100, H100, H200, or another datacenter-class option. Smaller models can often run comfortably on GPUs with 24–32 GB of VRAM.

A mature AI team knows where the boundary is.

An immature one asks for the biggest GPU before asking what the model needs.

There is another reason this conversation matters: overpaying is often social rather than technical.

Engineers like to believe infrastructure decisions are rational. Some are. Many are shaped by fear.

It is easy to defend H100. It is the famous answer. It signals seriousness. It reassures non-technical stakeholders. If something goes wrong, the team can say it chose the premium option.

Choosing an RTX-class GPU requires more explanation. You have to show the benchmark. You have to explain VRAM, quantization, throughput, batching, context length, and cost per useful token. You have to persuade people that “consumer” does not mean unserious.

That is why the H100 tax survives. It is procurement anxiety with a CUDA driver.

The irony is that careful sizing is the more serious decision. It forces the team to understand the workload rather than hide behind the accelerator's brand name. It asks whether the model actually needs 80 GB of HBM, whether the request pattern benefits from heavy batching, whether latency is strict or flexible, whether a single card can carry the service, and whether the extra cost yields a measurable gain.

In AI infrastructure, prestige is a poor substitute for measurement.

A broader systems question sits underneath the GPU bill.

AI infrastructure is already putting pressure on power systems. The International Energy Agency estimates that data centers consumed about 415 TWh of electricity in 2024, around 1.5% of global electricity use, and projects that figure could more than double to about 945 TWh by 2030.

That does not mean teams should feel guilty every time they run inference. Moral bookkeeping is a poor infrastructure strategy. The point is simpler: waste at scale becomes a systems problem.

When every workload defaults to the largest centralized hardware tier, the industry does not just spend more money. It also burns scarce capacity, strains grids, increases cooling demand, and reinforces the idea that AI infrastructure must be concentrated in a few giant places. We have argued before that the AI boom cannot keep pretending it runs on air. The same applies to GPU selection.

Efficiency is the discipline of matching the tool to the job.

There is a market structure hiding inside this discussion, too.

If serious AI products require datacenter GPU economics by default, then AI becomes easier for companies with giant contracts, deep cloud credits, and enough capital to absorb waste. Smaller teams become API renters. Researchers, agencies, public institutions, and startups get pushed toward platforms they do not control, in regions they may not choose, at prices that can change underneath them.

The dependency matters as much as the cost.

We have written before that compute is becoming a source of leverage in AI. That leverage can harden around a few centralized providers, or it can be spread across more practical infrastructure choices. The sovereignty debate makes the same point from another angle: control is not a speech, a label, or a clause in a sales deck. It depends on where workloads run, who owns the infrastructure, who controls the terms, and how hard it is to leave.

This is where right-sized inference becomes more interesting than a cheaper GPU bill.

If more everyday inference can run on dedicated RTX-class GPUs, in chosen regions, with predictable pricing and standard tools, more teams can build without asking permission from the same small group of infrastructure gatekeepers.

That is the practical side of distributed cloud: a better match between workload, geography, cost, and control.

As we argued in our piece on neoclouds and the real-world AI boom, AI infrastructure is becoming a real-world problem again: energy, capacity, regional deployment, software orchestration, and customer control all matter. Inference makes that especially clear because it runs repeatedly, close to users and data, inside ordinary applications.

The strongest argument against the H100 tax is empirical: Run the benchmark.

What model are you serving? What precision or quantization is acceptable? What context length do you need? How many concurrent users matter today, not in the fantasy roadmap? Can requests be batched? Is latency strict or forgiving? Does the workload need 80 GB memory, NVLink, ECC, MIG, or formal datacenter GPU procurement? Does the data need to stay in a specific region? What is the cost per useful token when the product is busy, not when the demo is quiet?

Only after those answers should the GPU tier enter the conversation.

If the workload fits comfortably on one 24 GB or 32 GB card, test it there. If it fails, move up. If it succeeds, do not apologize for using the right tool.

The serious AI teams in the next phase will not be the ones who rent the most prestigious accelerator out of habit. They will be the ones who understand their inference shape, measure their cost per useful token, and know exactly when an H100 earns its place.

At Hivenet, this is why we built Compute with Hivenet around dedicated RTX 4090 and RTX 5090 instances, fixed on-demand pricing, per-second billing, no egress fees, and regions teams can choose. The point is not that every workload belongs on RTX. The point is that many workloads do, and the industry has been strangely slow to admit it.

The H100 tax will not announce itself. It will hide inside defaults, procurement habits, and the comforting belief that the bigger GPU must be the safer one.

For frontier work, rent frontier hardware.

For everyday inference, make your math work.

Written by Hivenet Editorial
Hivenet is a distributed cloud platform for storage, file sharing, and compute. It helps teams run cloud and AI workloads across independent infrastructure, with clearer control over cost, resources, and data movement.

Discover more at hivenet.com.

Download the latest version of the Hivenet app

For a while, the AI story was told as if software had escaped the physical world.

Bigger models. Longer context windows. Smarter agents. Better benchmarks. A new demo every week, each one asking us to believe that intelligence was becoming lighter, faster, more abstract, almost weightless.

That was always the comforting version.

Every AI system runs somewhere. That somewhere needs chips, electricity, cooling, land, fiber, permits, contracts, technicians, and a grid that can absorb the load. The model may feel like software. The infrastructure does not.

This is the part of the AI boom that is becoming harder to ignore. The International Energy Agency estimates that data centers consumed about 415 TWh of electricity in 2024, around 1.5% of global electricity use, and projects that figure could rise to about 945 TWh by 2030. The same IEA analysis says data center electricity consumption is growing much faster than overall electricity demand.

In the United States, Goldman Sachs Research expects data center power demand to rise from 31 GW in 2025 to 66 GW in 2027, driven by AI infrastructure buildout. It also expects only about 50% to 60% of scheduled near-term data center capacity to come online on time, because the physical world still has delays, labor shortages, permitting, grid limits, and construction bottlenecks.

So perhaps the next useful AI question is not “Which model is smartest?”

It is: where will all of this intelligence actually run?

That is where neoclouds enter the conversation.

A neocloud is an AI-first cloud provider built around GPU compute rather than the broad, general-purpose cloud model that defined the last two decades. The term is awkward, but the category is useful. It names a real market shift: AI teams need faster access to GPUs, clearer pricing, and infrastructure designed for training, fine-tuning, inference, rendering, and other workloads that do not behave like a typical web app.

For the technical version, we have a full explainer on what a neocloud is. This piece is about the larger point.

Neoclouds exist because the old cloud story has run into a new kind of demand.

The first wave of cloud computing won because it made infrastructure feel invisible. You clicked a button, a server appeared, and everything underneath became someone else’s problem. That abstraction was useful. It helped millions of teams build without owning hardware.

AI has made that abstraction thinner.

A GPU is not just another virtual machine. A long training run is not a website. Inference at scale is not a side workload. A public institution running AI over sensitive data is not simply shopping for a cheaper instance. These use cases care about performance, but they also care about location, cost, energy, jurisdiction, resilience, and the buyer's ability to leave without rebuilding everything from scratch.

This is why we argued in “Why AI’s next phase hinges on infrastructure, not algorithms” that compute is becoming a real source of power in AI. Models still matter. They will keep improving. But models without available infrastructure are demos waiting for capacity.

The first neoclouds answered one part of the problem: GPU scarcity.

That matters. Uptime Institute described neoclouds as a cost-effective AI infrastructure alternative that emerged as hyperscalers struggled to meet GPU demand. Its analysis found that at the start of 2024, high-end GPU-backed instances from hyperscalers could cost more than $100 per hour when available, while neoclouds offered lower-cost access to comparable GPU infrastructure.

The market is growing fast. ABI Research expects global GPU-as-a-Service revenue from neocloud providers to grow from US$42 billion in 2025 to nearly US$250 billion by 2030, with inference revenue growing faster than training revenue as AI moves into daily applications.

But cheaper GPU access is not enough to make a durable cloud category.

A GPU rental page can be useful. It can also become a commodity. If two providers offer the same GPU in the same region, and the buyer sees no difference beyond price, the category quickly becomes a margin fight. McKinsey has made a similar point: neoclouds were born from GPU scarcity, but their long-term viability depends on moving beyond bare metal and into more differentiated AI-native infrastructure and services.

That is the fork in the road.

One version of the neocloud becomes a cheaper place to rent chips.

The better version becomes the infrastructure layer AI actually needs: GPU access, yes, but also storage near compute, predictable pricing, regional control, workload orchestration, migration paths, energy awareness, and software that makes distributed capacity usable without asking every customer to become an infrastructure engineer.

This is where the word “neocloud” starts to become interesting.

Not because it is new. Tech has enough new words already. We are practically buried under them.

It is interesting because it forces the cloud industry to admit that AI infrastructure is not placeless. It has a body. It consumes energy. It sits under the law. It creates dependencies. It affects local grids and public policy. It shapes who can build, who can compete, and who has to ask permission.

That last part matters in Europe.

We have written about this from a few angles already: in “Europe keeps talking about sovereign AI”, in “Large cloud providers won’t fix the sovereignty problem”, and in “What Greenland reveals about power, and what cloud hides”.

The thread is simple: infrastructure dependency is still dependency, even when the interface is elegant.

Sovereignty is often presented as a policy issue. That is only partly true. It is also an engineering issue. It depends on where workloads run, who owns the infrastructure, who controls the keys, who can change the terms, and how hard it is to leave.

A cloud that makes exit painful is not just providing infrastructure. It is collecting leverage.

The AI boom sharpens that problem because AI will not stay inside research labs. It is moving into schools, hospitals, public services, enterprise workflows, customer support systems, code tools, search, design, legal work, and daily decision-making. That means inference becomes ordinary infrastructure. It runs again and again, close to users and data, under cost pressure and regulatory pressure.

Training can often justify huge centralized clusters. Inference has a different rhythm. It is repeated, local, latency-sensitive, and tied to the messy world of real customers and real institutions.

That is why the energy question is not a side issue.

We wrote in “The AI boom is fueling a new energy crisis” that the cloud cannot keep pretending it runs on air. The point was not to shame people for using AI. That kind of moral bookkeeping gets boring quickly. The point was that infrastructure design has consequences.

Europe is already treating data centers as energy policy. Reuters reported on June 3, 2026, that the EU plans minimum energy-efficiency standards for data centers, with EU data center capacity expected to rise from 12 GW in 2025 to 28 GW by 2030.

A serious neocloud cannot ignore this.

It has to ask better questions. Can compute move toward available energy instead of forcing energy to move toward compute? Can infrastructure be deployed in months rather than years? Can smaller sites be coordinated as one cloud? Can workloads follow demand, price, capacity, and local constraints? Can sovereignty be enforced by architecture rather than promised in a contract?

This is where Antimatter changes the shape of the Hivenet story.

In “The cloud was never supposed to live in five places”, we explained why Hivenet is now part of Antimatter. The short version is that Hivenet brings the distributed cloud software layer. Antimatter brings that software together with energy and modular infrastructure.

Antimatter combines three operating businesses: Data Factory for energy, Policloud for modular micro data centers, and Hivenet for distributed cloud software. Its public materials frame the company as a vertically integrated neocloud for AI inference, built around energy access, modular deployment, and software orchestration.

That matters because a data center is not a cloud.

A GPU fleet is not a cloud.

A collection of sites is not a cloud.

A cloud is what happens when capacity becomes usable: APIs, billing, orchestration, workload placement, storage, support, security, observability, migration, and the dull practical details that make infrastructure safe to depend on.

Hivenet’s role inside Antimatter is to make distributed capacity feel like something people can build on. Antimatter’s “How it works” page explains that Hivenet software treats distributed capacity as a single logical cloud and delivers compute, storage, and file transfer via APIs across distributed sovereign infrastructure.

That is the point many neocloud conversations miss.

AI infrastructure will not be fixed by throwing more GPUs into the same old shape. The old shape is part of the constraint. Bigger campuses will still matter. Hyperscalers will still matter. Some workloads belong there. But the idea that every serious workload should flow into a few massive regions, under a few providers, with power and jurisdiction treated as background details, looks weaker every month.

The next cloud market will be mixed:

Large campuses where they make sense.

Specialized GPU clouds where fast access matters.

Distributed micro data centers where energy, latency, and sovereignty matter.

Software layers that make all of it coherent enough to use.

This is why decentralization is finally the point. Not decentralization as a slogan. Not the old ideological fog that made everything sound like a manifesto written inside a Telegram channel. Practical decentralization: spreading compute, control, and value because concentrated infrastructure is becoming too expensive, too brittle, and too politically exposed.

Hivenet has always argued for that shift. Antimatter gives it a larger industrial base.

That does not make the work easy. Distributed infrastructure has to be disciplined. It has to be measurable, secure, orchestrated, and boring in the best sense. If it feels improvised, it fails. If developers cannot use it, it fails. If buyers cannot understand the price, it fails. If sovereignty is just a claim on a website, it fails.

The neocloud category will have the same problem every infrastructure category has: the word will get cheaper as more companies use it.

Some neoclouds will be useful GPU shops.

Some will become acquisition targets.

Some will become thin wrappers around familiar infrastructure.

A few may become something more durable: AI infrastructure companies that understand the full chain from power to hardware to software to workload.

That is the bet behind Hivenet and Antimatter.

The AI boom is not running out of intelligence. It is running into the cost of making intelligence useful.

That cost is measured in GPUs, but also in megawatts, construction timelines, energy contracts, grid queues, data residency, customer trust, and whether teams can build without surrendering control to the same handful of platforms.

Neoclouds are not the whole answer. No single architecture is.

But they are a sign that the old cloud story has reached its limit. The next phase of AI needs infrastructure that is closer to energy, closer to users, closer to data, and honest about the physical world it depends on.

That is the cloud worth building now.

Not another black box.

A system people can understand, choose, leave, and rely on.

That is where the AI boom becomes real.

Recommended links:

• The neocloud business model explained
• What is a neocloud?
• Neocloud vs hyperscalers
• The economics of the neocloud
• Sustainability in the neocloud era
• The future of cloud sovereignty
• When to use a neocloud

Written by Hivenet Editorial
Hivenet is a distributed cloud platform for storage, file sharing, and compute. It helps teams run cloud and AI workloads across independent infrastructure, with clearer control over cost, resources, and data movement.

Discover more at hivenet.com.

Download the latest version of the Hivenet app

You clicked a button, a server appeared, and the physical world stayed politely out of view. Power, land, water, cooling, chips, fiber, contracts, jurisdiction, and construction time were somebody else’s problem. The cloud became an abstraction, and the abstraction worked well enough that many people forgot what it was abstracting.

AI has ended that illusion.

Every prompt, agent, document search, model endpoint, training run, and workflow now depends on physical infrastructure that cannot be wished into existence. The question is no longer just whether there are enough GPUs. It is where the power comes from, how fast capacity can be deployed, where the workload runs, who controls the path, and what happens when a small number of regions or providers become the default answer to every technical question.

The cloud was never supposed to live in five places.

The number is not literal. It is a shorthand for a pattern we all recognize: a small set of providers, a small set of regions, a small set of campuses, and a small set of contracts quietly shaping the digital lives of everyone else. That concentration made sense for a long time. It gave the industry standard interfaces, reliable capacity, and economies of scale. It also created a dependency we now have to look at directly.

AI inference changes the logic.

Training large models can (probably) justify huge centralized clusters. The jury is still out on that. But inference is different. It runs inside real products, customer workflows, copilots, agents, search tools, internal systems, and public services. It runs repeatedly, close to users and data, under cost pressure, and often under regulatory pressure. It needs capacity that is available, governable, efficient, and close enough to the work it serves.

A bigger campus will still have a place. It cannot be the only answer.

This is why Hivenet is now part of Antimatter.

What Antimatter is

Antimatter is an AI infrastructure company built from three operating businesses: Data Factory, Policloud, and Hivenet.

Data Factory brings flexible power. Policloud brings modular micro data centers. Hivenet brings the distributed cloud software layer that turns capacity across many places into products people can actually use.

In simpler terms, Antimatter moves compute toward energy, then uses software to make that distributed infrastructure behave like one cloud.

That last part matters. A data center is not a cloud. A GPU fleet is not a product. A collection of sites is not an answer until customers can deploy workloads, store data, enforce residency, manage costs, and get predictable service through a coherent interface.

Antimatter starts from a physical fact: AI infrastructure is no longer only a data center problem. It is an energy problem, a deployment problem, a coordination problem, and a service problem. The companies that solve only one part of that stack will still matter, but they will inherit constraints from the parts they do not control.

That is the reason for bringing these three businesses together.

Power decides where compute can live. Modular infrastructure decides how fast capacity can come online. Software decides whether any of it becomes usable by developers, startups, enterprises, public-sector teams, and ordinary users.

Hivenet belongs in Antimatter because the distributed cloud thesis has reached its industrial stage.

Why Hivenet belongs here

Hivenet has always rejected an unnecessary assumption: that useful cloud infrastructure has to sit inside a handful of centralized hyperscale campuses.

The point was never to make infrastructure scattered, casual, or improvised. Distributed cloud only works when it is coordinated. It has to be programmable, resilient, secure, measurable, and understandable. Otherwise, it is just capacity with a prettier diagram.

Hivenet’s job inside Antimatter is to make distributed infrastructure usable.

That means turning physical capacity into services: compute, storage, file transfer, and AI services. It means giving customers a clear interface to infrastructure that may be spread across regions, sites, and power assets. It means enforcing residency through architecture rather than asking people to accept a policy promise. It means making distributed capacity feel boring in the best sense: available, predictable, and ready to build on.

The old cloud abstraction told customers not to care where things ran. That was convenient, and sometimes it was useful. It also hid control.

Hivenet takes a different view. Location, energy source, jurisdiction, and deployment model are not minor implementation details. They shape cost, trust, resilience, and the user’s ability to leave.

That is why Hivenet’s role is larger inside Antimatter, not smaller. Without Hivenet, Antimatter would have power and modular infrastructure. With Hivenet, those assets can become cloud services.

Distributed does not mean undisciplined

There is a fair question here.

Hivenet has argued for distributed cloud. Antimatter talks about vertical integration. Are those ideas in conflict?

They are not. They solve different parts of the problem.

Distribution is about where compute lives. Vertical integration is about how capacity becomes reliable. A distributed cloud with no control over energy or hardware becomes a clever software layer sitting on somebody else’s constraints. A hardware network with no orchestration becomes a set of boxes. The future needs both: distributed placement and serious operational control.

The misconception is that distributed means lightweight. It does not.

The more distributed a system becomes, the more disciplined the coordination has to be. Workloads have to be placed. Capacity has to be measured. Data has to stay where it was supposed to stay. Failures have to be contained. Customers need billing, APIs, observability, support, and clear migration paths. The system cannot ask every user to understand the physical topology underneath it.

That is Hivenet’s work inside Antimatter.

It is also why the story has to move beyond “GPU rental.”

A GPU is a component. Customers need useful work: throughput they can plan around, storage near compute, no surprise egress, predictable pricing, data residency, and tools that fit how teams already build. The better story is not “here is a cheaper machine.” The better story is “here is a more practical way to run AI workloads without giving up control.”

Why this matters now

The AI market is moving from model fascination into operational reality.

Models still matter. They will keep improving. But the real pressure is shifting toward deployment. How do you run inference every day? How do you serve users without sending sensitive data into a black box? How do you keep costs from becoming a tax on every product interaction? How do you build in a region that matters for your customers, your regulators, or your own risk model?

A model answering a customer support question, summarizing a legal file, searching a company knowledge base, grading a student’s response, routing a logistics workflow, or helping an analyst write code is not a research workload. It is business infrastructure.

Business infrastructure has to be close enough, affordable enough, governed enough, and reliable enough to become ordinary.

That kind of AI does not want the cloud to be invisible. It wants the cloud to be accountable.

This is where Antimatter’s structure matters. Data Factory addresses the power side. Policloud addresses the physical deployment side. Hivenet addresses the software and product side. Each layer is useful alone, but the pressure of AI inference makes them stronger together.

Compute cannot escape energy. Energy cannot become AI infrastructure without hardware. Hardware cannot become cloud without software.

What changes for Hivenet customers

For Hivenet customers, Antimatter should not feel like a detour. It should make the roadmap more credible.

It means Hivenet is backed by a physical infrastructure thesis, not only a software thesis. It means Compute, Storage, Send, and AI services can sit on a stronger base: energy availability, modular deployment, and a coordinated cloud layer. It means Hivenet can keep building for the people who already understand why sovereignty, price clarity, and non-hyperscaler options matter, while giving production AI teams a more serious path forward.

Most customers should not have to think about Antimatter every day. They should care that workloads can run where they choose. They should care that bills are understandable. They should care that familiar tools still work. They should care that migration does not become a trap. They should care that their infrastructure does not quietly route through a third party they never approved.

The practical promise is simple.

You should be able to run AI and storage workloads without making the hyperscaler the default middleman.

You should be able to pick a region and have that choice mean something.

You should be able to use standard tooling rather than adopt a private universe of proprietary services.

You should be able to leave.

That last point matters. A cloud that works by making exit painful is not just infrastructure. It is leverage.

Sovereignty is architecture, not paperwork

Sovereignty is often marketed badly.

It becomes a banner over conventional infrastructure. A compliance line in a sales deck. A regional checkbox. A contract clause on top of a dependency that has not really changed.

That is not enough.

Sovereignty is weakest when it is presented as something someone promises. It is strongest when it is part of how the system is built.

For Hivenet, sovereignty means architecture should enforce where workloads and data live. The selected region should not be a label. It should be a boundary that the system respects.

It also means sovereignty should not be the only story.

Nobody wants a sovereign cloud they have to apologize for. Performance matters first because teams need infrastructure that works. Affordability matters because better governance cannot require waste. Sustainability matters because AI cannot pretend energy is free.

This is why Antimatter’s energy-first model matters. If infrastructure follows available power rather than forcing power to follow infrastructure, deployment can be faster, less wasteful, and more honest about physical limits.

That does not make the problem easy. It makes the architecture more grounded.

The future will be mixed

The next cloud will not be one giant thing replacing another giant thing.

Hyperscale data centers will remain important. Large training clusters will still exist. Some workloads benefit from enormous centralized systems, and that will not change because a new category appears.

The problem is the monopoly of architecture.

For too long, cloud infrastructure has been treated as if one shape can serve every need: bigger campus, bigger region, bigger contract, bigger dependency. AI inference exposes the limits of that shape.

The right future is mixed.

Large campuses where they make sense. Modular sites where deployment speed and energy access matter. Distributed software to coordinate capacity across places. Customer choice that is real enough to survive procurement, migration, and production workloads.

In that future, Hivenet is the layer that makes distributed capacity practical. Antimatter is the structure that gives that layer the physical base it needs.

Together, the idea is simple to describe and hard to build: bring compute closer to energy, closer to users, closer to data, and make it available as one coordinated cloud.

What we still believe

Hivenet’s role has changed. The underlying belief has not.

We still believe the cloud should be less concentrated.

We still believe users should have more control over where their data lives.

We still believe efficiency matters, not as a slogan, but because waste becomes cost and cost becomes exclusion.

We still believe developer experience has to be practical. A noble architecture that developers cannot use is just a diagram.

We also believe the next cloud has to stop pretending it floats above the world. It sits in the world. It draws power. It occupies space. It depends on grids, chips, contracts, teams, and software.

The best infrastructure companies of the AI era will be the ones honest enough to build from that fact rather than hide it behind abstraction.

That is why Hivenet is part of Antimatter.

The original distributed cloud thesis did not fail. AI made it more urgent. Software alone cannot solve the physical limits of AI infrastructure. Physical infrastructure alone cannot become cloud without software. These layers now have to meet.

The cloud was never supposed to live in five places. It was supposed to connect capacity, people, and work in a way that made computing more available.

AI gives us a chance to correct the shape of the system before the next decade hardens around the same dependencies.

Hivenet’s job is to make the alternative usable.

Antimatter gives it the power, hardware, and deployment base to do that at scale.

Suggested final line if you want it slightly warmer and less declarative:

That is the work ahead: make distributed AI infrastructure feel practical, accountable, and ordinary enough for people to build on it.

Written by Hivenet Editorial
Hivenet is a distributed cloud platform for storage, file sharing, and compute. It helps teams run cloud and AI workloads across independent infrastructure, with clearer control over cost, resources, and data movement.

Discover more at hivenet.com.

Download the latest version of the Hivenet app

We need to talk about infrastructure dependency and European risk after 2020

Europe can argue about Greenland for weeks and still miss the point.

Greenland is not a metaphor. It is a real place with real people, land, energy routes, and strategic weight. When it enters political conversation, it does so because geography still matters. Power still anchors itself somewhere. The lesson worth paying attention to is not rhetorical escalation, but what it exposes: control tends to sit where the infrastructure sits (and a %^&load of rare minerals).

That matters far beyond maps.

Right now, Europe’s digital life runs on infrastructure it does not control. Not at the edges, but at the center. Applications, identity, data storage, backups, monitoring, internal tools. The boring, load-bearing layer. Most of it lives on platforms operated by a small number of U.S. companies such as Amazon Web Services and Microsoft Azure. This is rarely visible day to day. It becomes visible when politics intrudes, prices move, or access changes.

The problem is that cloud dependence is usually treated as a technical preference. Teams choose what ships fast. Buyers choose what looks safe. That logic might have held when the cloud felt like neutral plumbing. It is not that anymore, and, since 2020, it has been harder to defend. Cloud has become entangled with jurisdiction, sanctions, export controls, and policy shifts. The same dependency that saved time early on can quietly turn into a constraint later.

For founders and infrastructure buyers, this is not abstract. It shows up in contracts, invoices, compliance reviews, and architectural trade-offs.

But what can you do? Well, to be honest, a couple of things come to mind:

Start with jurisdiction. If your provider is a U.S. company, U.S. law can reach your data, even when servers are in Europe. Europe has spent years trying to stabilize this through legal frameworks for transatlantic data transfers. The collapse of Privacy Shield in 2020 forced many companies into a permanent state of legal uncertainty. Assessments, clauses, safeguards. The large GDPR fine against Meta in 2023 made the risk concrete. Teams stopped treating this as a lawyer’s problem and started treating it as a business one.

That uncertainty tracks political cycles. Executive orders change. Oversight bodies weaken. Agreements wobble. Most startups do not want their compliance posture to depend on the next administration’s priorities. Infrastructure buyers do not enjoy betting their risk profile on foreign courts and temporary fixes.

Sanctions show the harder edge of the same dynamic. Sanctions are serious tools, often used for serious reasons. The point here is not whether they are justified, but where their consequences land.

When U.S. sanctions reportedly led to the suspension of Microsoft-hosted email accounts tied to the International Criminal Court, the mechanism became clear. A political decision in one country created an obligation for a provider, and a critical service in Europe stopped working. Cloud accounts are not neutral utilities. They are dependencies that can be terminated by forces outside your market and outside your control.

The same pattern appeared after Russia’s invasion of Ukraine. Cloud services were restricted or withdrawn under sanctions. Many agreed with the policy. Many also learned how exposed their operations were when infrastructure followed foreign policy in real time. You can support the decision and still recognize the structural risk.

Export controls are moving in a similar direction. Restrictions that once applied to physical hardware are extending toward remote access to advanced computing. If access to GPUs becomes governed like exports, the enforcement point will sit with providers. European companies working across borders, even for ordinary commercial reasons, may find themselves asked to justify projects or counterparties. This is not hypothetical. It is already being discussed in policy circles.

Pricing is the quieter version of power. It rarely triggers headlines, but it shapes outcomes.

When cloud providers adjust European pricing to reflect currency shifts or internal strategy, customers adapt. That is not abuse by default. It becomes a problem when exit is impractical. Complex pricing models, bundled services, and high data transfer costs discourage movement. Over time, teams design systems that only one supplier can comfortably host. Then, cost control becomes guesswork. And right after, negotiation becomes asymmetric.

Europe has started to respond. The EU Data Act aims to reduce switching barriers and phase out certain cloud exit fees by 2027. That is an acknowledgment that lock-in is both inconvenient and structural. A market where leaving is punished is not resilient, no matter how advanced the product.

Reliability adds another layer. Hyperscalers invest heavily in uptime. They still fail. When they do, the blast radius is wide because the concentration is wide. Outages cascade across sectors at once. Banking apps, messaging tools, internal systems. A single provider event becomes a multi-industry event. That is not a failure of engineering effort but a consequence of dependence.

All of these points bring us to a conclusion that is less dramatic than the sovereignty rhetoric suggests:

Europe does not need a cloud revolt. Europe needs options.

Options change incentives. Options change negotiation. Options change how risk is priced. They reduce the chance that one policy decision, one pricing shift, one outage, or one legal reversal becomes existential.

This is where many sovereignty debates go wrong. They jump from dependence to replacement. But that is not how buyers behave, and not how systems evolve. Replacements are rare, and second rails are common.

A practical approach does not require ideology. It starts with questions. Which workloads can move without breaking the product? Background jobs often can. Storage can, if planned. Internal tools can, if tested. Which data categories create the highest legal exposure? Where are the real lock-in points? Identity systems and proprietary managed databases often are. What is the true exit cost, measured in time and operational disruption, not just money?

Optionality is not resistance. It is engineering discipline.

Greenland reminds Europe that power still has geography. Cloud reminds founders that power also has contracts.

If a supplier can change your operating reality without your consent, it deserves competition in your stack. Not as a statement, but as a safeguard.

We are looking at the wrong issue. We think that Europe just needs to replace hyperscalers to regain balance. What it really needs is to stop designing its future as if dependence carries no cost.

Written by Hivenet Editorial
Hivenet is a distributed cloud platform built on everyday devices, not data centers. It powers storage, compute, and file transfers through a global network of contributors — faster, fairer, and radically more sustainable.

Discover more at compute.hivenet.com and hivenet.com.

Download the latest version of the Hivenet app

There’s a peculiar mood in Europe right now. You can feel it in conferences, policy panels, and the way executives lower their voices when the word “sovereignty” comes up. It’s half pride, half anxiety, like someone trying to convince themselves that the door isn’t already closing. It’s difficult to admit there is a gap between what we want and what we can achieve.

In other words: Everyone wants sovereign AI. No one seems sure what that means.

On paper, Europe knows exactly what it wants: a future that isn’t dictated by foreign platforms and foreign governments. A future where its companies, hospitals, courts, and ministries don’t depend on another country’s political climate or corporate roadmap. A future where decisions made in Washington or San Francisco don’t silently ripple into European institutions overnight.

Yet look at today’s infrastructure. Look at the contracts, the partnerships, the so-called “European AI” solutions built on someone else’s stack. For all the speeches, Europe still entrusts the heart of its digital life to providers it cannot influence. The imbalance is obvious, and the continent feels it. It’s a quiet but sure humiliation: wanting independence while sending another check to the same three vendors. They must be secretly laughing in their American-based underground bunkers.

The mistake, however, is not ambition. The mistake is where that ambition points. Europe keeps staring at the wrong finish line. People talk about building a European model to rival the giants, as if sovereignty will sprout from a bigger dataset and a better benchmark score. It sounds noble. It’s also a distraction. Sovereignty was never about who builds the tallest model. It’s about who can keep the lights on when someone else decides they shouldn’t.

France understood this before anyone was ready to admit it. Instead of buying the story that Europe needed a moonshot, it built something far less glamorous and far more important: places where intelligence can live without asking permission. Real infrastructure. Real sites. A few micro data centers anchored in French soil started to appear, expanding outward like roots rather than monuments. That quiet, stubborn effort is meant to cross borders. The strategy is spreading, not through headlines, but through hardware.

France was the place where Hivenet first appeared — not as a slogan, not as a flag, but as a different kind of answer. A cloud that doesn’t look like a fortress, but like a fabric. It doesn’t worship scale for its own sake. It uses what already exists. It builds what doesn’t. It treats Europe not as a market to be captured, but as a continent full of untapped capacity — grid power, homes, offices, greenhouses, and edges where energy and computing power wait for someone to give it purpose.

There’s something honest about that approach. It doesn’t promise a clean break from the United States or a return to some digital Versailles. It offers something more grounded: the ability to stay standing if someone else tries to pull the plug. We have to understand that sovereignty is bargaining power. It’s the freedom to say no without fear of collapse.

Sovereignty is the freedom to say no without fear of collapse.

Critics, usually from a comfortable distance, insist Europe is too far behind to matter. They mistake sovereignty for spectacle. Real power rarely looks spectacular. It looks like local infrastructure that keeps running when policies shift. It looks like companies that can move workloads without rewriting their identity. It looks like a future that can’t be revoked by an email from another jurisdiction.

Europe has spent a decade learning that privacy without ownership is theater. It is now learning that intelligence without control is perilous. Renting AI capacity is not a strategy. Not a viable one, at least. It’s deferral. It feels fine until the dependency starts to itch, until the terms change, until someone else’s election nudges your infrastructure into instability.

The path ahead should be humble but stubborn. It should be European in the best way: pragmatic, skeptical of monopolies, resistant to being cornered. The question isn’t whether Europe can beat Silicon Valley at its own game. The question is why Europe keeps agreeing to play it.

France answered by doing something quieter: it built places for intelligence to stay. Hivenet is betting more countries will follow — not out of pride, but out of self-respect. Europe doesn’t need another AI mascot. It needs a spine.

Sovereignty is freedom. It is refusing to let your future depend on someone else’s weather. And once you see it that way, the choice stops being ideological. It becomes personal.

Europe doesn’t have to shout. It just has to stop asking permission.

Written by Hivenet Editorial
Hivenet is a distributed cloud platform built on everyday devices, not data centers. It powers storage, compute, and file transfers through a global network of contributors — faster, fairer, and radically more sustainable.

Discover more at compute.hivenet.com or hivenet.com.

Download the latest version of the Hivenet app

Big Tech built AI on centralization. Hivenet shows why the future depends on distributed compute, shared value, and real digital sovereignty.

For years, Big Tech told us that scale was the answer. Bigger clusters. Bigger models. Bigger promises. Every problem could be solved by stacking more GPUs, feeding more data, and burning more power. But something about that story is beginning to crack.

As PYMNTS recently put it, decentralization is emerging as a real test of Big Tech’s AI power. The question is changing: It’s not anymore about who builds the best models… It’s who gets to benefit from them. Who owns the compute, who controls the data, and who profits from the intelligence that everyone helps to create.

In general, we are trying to steer clear from any ideological pushes, especially when it comes to things like “freedom tech.”But it’s a practical correction. Centralization seems to work until it doesn’t — until cloud outages cripple critical services, until supply chains buckle, until regulators wake up and realize most of the world’s AI runs in a handful of American data centers.

At Hivenet, we’ve taken a different route from the start. We built a distributed cloud that runs on everyday devices, not data centers. It’s the opposite of vertical integration: it’s horizontal cooperation. Thousands of small contributions add up to something much more resilient, efficient, and fair. (If you’re curious how that works, here’s our deep dive into distributed cloud computing’s next wave).

Decentralized AI isn’t a new concept, but it’s finally finding its moment. Researchers at MIT Media Lab talk about five major hurdles: privacy, verifiability, incentives, orchestration, and user experience. They’re right. None of this works if it isn’t verifiable, rewarding, and simple enough for normal people to use. But the direction is clear: AI doesn’t have to live behind the walls of hyperscale infrastructure. It can breathe out into the world again.

You can already see the shift. The Linux Foundation found that nearly all organizations experimenting with generative AI rely on open-source software for at least part of their stack. Many critics say that’s a fringe movement, but they are wrong: that’s the foundation of a new ecosystem. Sahara AI, for instance, raised fresh funding on the promise that users and data contributors should share the economic upside of the AI they help create. Even California’s new Frontier AI Act pushes for transparency in large-model development. Regulation is catching up to reality.

In Europe, the conversation is becoming existential. The EU AI Act and its upcoming Action Plan both point to the same bottleneck: dependence. Europe relies heavily on foreign data centers and clouds to power its AI ambitions, which means sovereignty is, at best, conditional. There’s growing interest in spreading that load across smaller, local, and more sustainable providers — precisely the kind of distributed infrastructure that makes sense in a continent built on diversity and collaboration. We’ve written about this shift in The future of cloud computing: trends and the pivotal role of the distributed cloud.

Distribution changes everything. It brings compute closer to where data lives, keeps value circulating among those who create it, and makes systems more fault-tolerant by design. It’s not just greener; it’s fairer. When power is shared, innovation follows. You can read more about how that plays out in our article on the cost-efficiency of the distributed cloud.

At Hivenet, that idea is running today. Our network stores, computes, and transfers data across a mesh of idle devices, owned by real people who can choose how and when to contribute. It’s practical, efficient, and more aligned with how the world actually works: decentralized, messy, but cooperative when given the right structure.

The future of AI won’t be decided by who trains the biggest model. It’ll be decided by who builds the fairest infrastructure and who dares to question whether centralization ever made sense in the first place.

We’ve had decades of “cloud” that looks suspiciously like the old mainframes with better marketing. Maybe it’s time to remember what the cloud was supposed to mean: something everywhere, something shared, something we all have a stake in.

Sources & further reading

• PYMNTS: Decentralization Emerges as a Test of Big Tech’s AI Power
• MIT Media Lab: Five Challenges for Decentralized AI
• Linux Foundation: State of Open Source in Generative AI 2024
• Reuters: Sahara AI Raises Funding to Reward Data Contributors
• California SB 53: Transparency in Frontier Artificial Intelligence Act
• European Commission: EU AI Act Summary
• European Commission: AI Action Plan & Sovereignty Strategy

Written by Hivenet Editorial
Hivenet is a distributed cloud platform built on everyday devices, not data centers. It powers storage, compute, and file transfers through a global network of contributors — faster, fairer, and radically more sustainable.

Discover more at compute.hivenet.com or hivenet.com.

Download the latest version of the Hivenet app

At the RAISE Summit in Paris — a sleek event tucked beneath the Louvre — former Google CEO Eric Schmidt casually dropped a phrase that’s still echoing in our minds. He called it the “San Francisco Consensus”: a tight circle of Bay Area insiders convinced that AI is about to trigger the most significant transformation in human history. And soon. As in, the next two to four years.

It would be easy to dismiss this as another tech guru hyping his latest obsession. There’ve been a lot of that. Yet Schmidt isn’t exactly new to dramatic tech shifts. He oversaw Google’s pivot into AI at a time when most of us thought machine learning was still science fiction. This claim feels weighty, urgent, even credible.

What exactly does this supposed consensus predict?

Three things:

First, autonomous AI agents that don’t just chat — they execute tasks end-to-end. Imagine software that doesn’t merely answer questions, but manages entire businesses, writes code, plans projects, or executes deals, almost independently. If you think this is far-fetched, think again. Startups and research labs are already quietly reshaping workflows.

Second, large language models (LLMs) are getting smarter — scarily smart. They’re starting to reason and remember, building something eerily close to genuine understanding. GPT-4o, Google’s Gemini, and Anthropic’s Claude 3 have already shown what improved memory and chain-of-thought reasoning mean in practice. Soon, those models will remember more than just your prompts. They’ll grasp context across multiple interactions, elevating their capabilities exponentially.

Third, and perhaps most intriguing: recursive self-improvement. AI systems are learning to enhance themselves at a rate faster than humans ever could. Theoretically, this creates a feedback loop that could spiral rapidly out of control—or toward unprecedented innovation, depending on one's perspective. If this loop becomes real, even modest advancements could yield staggering results.

Of course, skeptics aren’t convinced. And skepticism is healthy … tech has a long history of promising revolutions that never materialized. But let’s assume Schmidt’s insiders are only partly correct. Even if just one of these predictions pans out, we’re facing a fundamental shift in the market dynamics around AI.

And here’s where things get interesting.

CapEx becomes the new moat

Look at the balance sheets of Alphabet, Amazon, Meta, and Microsoft. You’ll quickly see where they’re betting the house: GPUs, datacenters, and compute infrastructure. Billions are pouring into Nvidia’s coffers and into data warehouses sprawling across deserts and tundras. You could say that this sounds like a teenager’s impulsive spending, and if you are thinking of Zuck, you may have been right. But this feels strategic. Something is happening.

In Schmidt’s words, “CapEx becomes the new moat.” Forget data or even proprietary models. The real competitive advantage now lies in raw computing power. He who controls the GPUs, controls the future. Models can be replicated, open-sourced, tweaked. But infrastructure — especially at scale — can’t be easily copied.

But centralized clouds have limits

Centralized computing isn’t unlimited or infallible. We’ve already glimpsed the cracks: GPU shortages, geopolitical disruptions, rising electricity demands, even something as mundane as the limited bandwidth in certain regions. And let’s not overlook sovereignty laws, such as GDPR and Schrems II, which have sent multinational companies scrambling to build local clouds practically overnight.

Yet, despite these constraints, Big Tech continues to double down, betting on massive, centralized infrastructure they hope will be too big to fail.

That’s a risky bet.

At Hivenet, our view is different.

A decentralized answer

Our thesis is simple: decentralized compute changes the entire CapEx equation. Instead of sinking billions into monolithic datacenters, you leverage millions of idle devices spread across the globe. Rather than enormous facilities that guzzle power and strain local grids, you tap into existing hardware, closer to the data, closer to the users, and infinitely more resilient.

Through our partnership with Policloud, we’re quietly building precisely that: compute infrastructure that is distributed, sustainable, secure, and scalable. It’s not flashy — it’s infrastructure, after all. But it matters immensely. Because if Schmidt and his San Francisco insiders are even partially right, soon compute won’t just be fuel; it’ll be leverage. Control it, and you control the future. Democratize it, and you transform the future into something far more robust and far less monopolized.

What builders should do now

If you’re developing a product, scaling a startup, or leading tech teams, you need to act now — not in two to four years. That means auditing future compute needs today, diversifying infrastructure early, and treating GPUs like the strategic inventory they are. It means designing your workflows with anticipation of agent-based AI systems and preparing your business model to run securely under increasingly stringent sovereignty regulations.

Hivenet isn’t your only option, but it’s a ready path, and one already aligned with where the world is heading. Infrastructure, not flashy models, will decide who thrives in the next decade.

A closing reflection

When Schmidt stepped off the Paris stage, he left an open question lingering in the air: Is the coming shift two years away? Four years away? Or are these insiders blinded by their own optimism?

The truth: it doesn’t matter much. The direction is clear. Compute wins, infrastructure defines leverage, and the companies or communities holding that leverage will shape the next decade.

Don’t wait to find out if Schmidt nailed the timing. Start building your foundations now.

When the shift hits — and it will — you’ll want to be ready.

Download the latest version of the Hivenet app

It began like so many internet storms do, with one late‑night screenshot. A video editor in Berlin scrolled through WeTransfer’s revised terms and spotted a sentence that sounded harmless until the middle clause swerved:

“You hereby grant us a perpetual, worldwide, non‑exclusive, royalty‑free, transferable, sub‑licensable license … including to improve performance of machine‑learning models that enhance our content‑moderation process.” (from Reddit)

The line landed on X (Twitter) at 02:11 CET, and by dawn, the tag #WeTrustNoMore was trending in half a dozen languages. Artists, agencies, and journalists (many of the platform’s 80 million users in 190 countries ) asked the same blunt question: If the clause kicks in, what stops my work from becoming AI training data?

The backlash travelled fast. The Independent called it “global outrage,” noting that entire studios froze deliveries until clients signed off on safer transfer tools. The Next Web framed the episode as proof that “trust in centralised file‑sharing is cracking.” Even the Writers’ Guild of Great Britain weighed in, welcoming any climb‑down but warning that members’ scripts must never feed an algorithm without permission.

Forty‑eight hours after the fire started, WeTransfer pulled the clause, issued a mea culpa, and rewrote the licence to a boiler‑plate line about “operating, developing and improving the service.” “We don’t use machine learning or any form of AI to process content shared via WeTransfer,” the statement added, describing the wording as “poorly phrased.”

Of course, they would. They are not stupid. They are a business. Relief rippled across feeds, like a silent victory, but then stalled. Trust, once bent, rarely springs back. What stops them from trying to do this again?

The updated wording lets WeTransfer “use” content to operate and improve the service. That’s normal, but it’s also open‑ended. If you handle NDA‑bound or embargoed material, assume nothing and encrypt before upload. And when it comes to AI, WeTransfer now says any future moderation AI would run on synthetic test data, not customer files. That doesn’t mean that AI couldn’t actually enter through a back door. \

Broken trust, though, brings unease. And that unease has history. In June 2019 the service mis‑sent user files to unintended recipients for two days and could not explain why, forcing password resets system‑wide. Last year its new owner, Bending Spoons, cut 75 percent of staff within weeks of acquisition, the kind of belt‑tightening that leaves lawyers juggling more code than clauses. Against that backdrop, a sweeping AI licence felt less like a typo and more like a stress test to see what users would tolerate.

There is no way that clause was a ‘mistake’. A ‘typo’. ‘Poor wording’. We have heard all the buzzwords until now. Without the backlash, they would have gone through with it.

The deeper problem is structural. Centralised clouds copy every file, keep every key, and can widen their license overnight. One paragraph can flip private art into corporate data, and the user learns about it only when the social feed explodes. Change the architecture (shard each file, encrypt it client-side, and spread it across volunteer nodes), and that power shift never happens because no single platform holds a complete copy or the keys. That’s how we build Hivenet’s network: tools stay quiet, keys stay local, and licences remain narrow. It isn’t publicity; it’s physics.

None of this means you must abandon the service on principle (even if there are better alternatives to WeTransfer). It does mean you must read the policy and terms changes text line by line. Ask yourself who holds the keys, whether “machine learning” or “derivative works” sneaks into the licence, and whether the provider can rewrite the deal while you sleep. If any answer feels vague, wrap the file in your own encryption or choose a service that never touches your keys in the first place.

Trust is not a marketing slogan. It is custody, expressed in code and contracts. The next edit, wherever it happens, will be easier to spot if we keep the habit of reading the small print and holding our tools to it.

Download the latest version of the Hivenet app

AI is reshaping the internet — and if we don’t rethink how it’s powered, the planet will pay the price.

You’d be forgiven for thinking this whole digital world runs on air. Tap a screen, ask a question, generate a song, store your photos — and it all feels invisible. But that convenience has a physical cost, one most of us never see.

A new report from the International Energy Agency just made it harder to ignore. It warns that by 2030, energy use from data centers could quadruple. If that sounds abstract, here’s a clearer picture: AI, cryptocurrencies, and cloud systems could soon consume more electricity than some countries.

This is a warning light if we choose to stop at red.

So what’s behind the surge? A big part of it is scale. Running large models doesn’t just mean more code. It means more hardware. Bigger machines. Racks of graphics processors that run day and night, producing heat, demanding cooling, and drawing power nonstop. Even asking a model to write a sentence or find a file often takes more electricity than a standard web search. And when billions of queries hit every day, those numbers grow fast.

What makes this even trickier is how hidden it all feels. There’s no smoke. No server room humming in the background of your phone. But the emissions? Very real. Some countries, like Ireland, are already feeling the pressure. Data centers there eat up nearly 20% of the national electricity, and that number is climbing. The Netherlands and Germany are seeing similar debates. Who gets priority: homes or servers? Farms or training models?

This isn’t just about what powers the cloud. It’s about how we’ve built it. Most of today’s infrastructure is centralized — massive warehouses in a handful of places, owned by a handful of companies. That concentration doesn’t just shape policy. It concentrates risk. Local power failures can ripple globally. Local politics can shift how the internet works for everyone.

The environmental impact is just one layer of the issue. There’s also the question of control. When so much of our digital lives run through a few corporate systems, what happens to choice? What happens to resilience? You don’t need to be a privacy advocate or a climate scientist to feel uneasy about how top-heavy this all feels. We explored this trade-off more in our post on what it means to own your data in a connected world.

To be fair, some companies are trying. They’re experimenting with renewable energy, smarter cooling systems, and more efficient chips. It’s a start. But let’s be honest: we can’t keep stacking servers and hope solar panels fix everything. We can’t efficiency our way out of overconsumption. There’s a deeper redesign needed — not just in how we power our infrastructure, but in how we think about it.

One alternative? Stop thinking in terms of mega-centers and start thinking about networks. Distributed models don’t require you to build yet another warehouse — they tap into what’s already out there. Idle laptops. Dormant workstations. Machines with plenty of capacity just sitting there, waiting to help. That’s what Hivenet is built around. A way to turn the excess we already have into something useful and secure. It’s not about building more. It’s about using better. If you’re curious how this works in practice, check out our recent breakdown on why the future of cloud storage needs to change — and what that means for us.

There’s a deeper idea here worth sitting with. It’s tempting to chase speed, scale, and novelty. But that’s how we got into this mess. The more urgent question now is: how do we make all this sustainable — not just in terms of carbon, but in terms of trust and longevity? We talk a lot about energy efficiency, and yes, it matters. But resilience might matter more. Can we build systems that flex with demand instead of buckling under it? Systems that work with the environment, not against it? Systems people can understand, contribute to, and benefit from directly?

We’ve written before about why the future isn’t nuclear — it’s distributed and shared — and this is exactly why. Sustainability is about systems that adapt and invite participation, not just systems that are technically impressive. And resilience means keeping things human-scaled.

The good news is we still have time. This isn’t a doomsday scenario. It’s a fork in the road. The convenience of this digital age doesn’t have to come at the cost of the planet. The tools to build something better already exist — we just have to be willing to use them.

At Hivenet, we’re betting on a different kind of future. One where the cloud is made up of people, not just machines. Where your data lives closer to home, powered by real communities. Where sustainability isn’t a feature — it’s the foundation.

Because if we’re going to build something lasting, it shouldn’t be another power-hungry black box. It should be something transparent, adaptable, and shared.

And it’s not too late to start.

Download the latest version of the Hivenet app