Vitalik Buterin recently declared 2026 the year to “take back lost ground in computing self-sovereignty.” He shared the personal changes he’s made: swapping Google Docs for Fileverse, Gmail for Proton Mail, Telegram for Signal, and even experimenting with running large language models locally on his laptop instead of relying on the cloud.
The instinct makes sense. Centralised AI infrastructure is a real and growing problem. Three companies — Amazon, Microsoft, and Google — now control roughly two-thirds of global cloud infrastructure spending, a market that surpassed $100 billion in a single quarter last year. When nearly every AI prompt flows through infrastructure owned by a handful of corporations, users inevitably give up control over data that should arguably remain private. If you care about digital autonomy, that concentration should feel uncomfortable.
But the solution Vitalik points toward — running AI locally on personal hardware — creates a tradeoff that may not actually be necessary. For serious builders, it doesn’t offer a viable path forward.
The limits of local compute
Running models on your own machine is appealing. If the model stays on your laptop, your data does too. No third parties. No hidden surveillance. No dependence on hyperscale infrastructure. For lightweight use cases, this works. Developers experimenting with small models or individuals running basic inference can absolutely extract value this way. Vitalik acknowledges the current friction around performance and usability, suggesting these are temporary constraints that will improve over time.
But there’s a hard ceiling here.
Training models, running inference at scale, and deploying always-on agents require persistent GPU power that personal hardware simply cannot provide. Even a single AI agent operating overnight demands continuous compute. The idea of an “always-on” assistant collapses the moment you close your laptop.
At enterprise scale, the numbers escalate quickly. Thousands of GPU-hours per day are not unusual. A startup training a specialised model could consume more compute in a week than a high-end laptop could deliver in a year. Research teams often spend the majority of their budgets — sometimes 80% or more — just securing GPU capacity. That’s money that could otherwise fund talent, product development, or market expansion.
Large tech companies absorb these costs without blinking. Everyone else feels the squeeze.
Local hosting doesn’t resolve this tension. It quietly presents a binary: stay small and sovereign, or scale up and surrender your data to Amazon, Google, or Microsoft.
A false choice
But decentralisation was never about shrinking ambition in exchange for independence. It was about enabling both at once. Crypto’s core promise was not “do less so you can control more,” but “build systems where control and scale coexist.”
That principle applies to compute just as much as it applies to money.
Around the world, millions of GPUs sit underutilised — in regional data centres, enterprises, universities, and independent facilities. A growing number of decentralised compute networks are beginning to aggregate this fragmented capacity into elastic, on-demand infrastructure. Instead of relying on a single hyperscaler, developers can tap into distributed pools of independent operators.
The result isn’t theoretical. These networks now span over 100 countries and can offer enterprise-grade GPUs at significantly lower cost than traditional cloud providers. Pricing is dynamic. Access is flexible. Idle hardware becomes productive. Developers get scalable infrastructure without committing to a single corporate gatekeeper.
Who this actually helps
The implications go far beyond marginal cost savings.
Independent researchers can pursue ambitious experiments instead of downsizing ideas to fit hardware constraints. Startups in emerging markets can train models tailored to local languages, healthcare systems, or agriculture without first raising massive venture rounds just to secure cloud credits. Regional data centres can participate in global markets instead of being locked out by entrenched enterprise contracts.
That’s how you close the AI divide — not by asking builders to accept weaker tools, but by redesigning how compute is accessed and distributed.
Vitalik is right about the diagnosis: AI infrastructure is dangerously concentrated. But retreating to personal hardware isn’t the only alternative — and for most serious builders, it’s not a sustainable one.
The real test for crypto
If decentralisation is more than a slogan, this is where it matters.
Decentralised compute networks offer an opportunity to demonstrate that distributed systems can match — and even outperform — centralised incumbents. Lower costs. Wider access. No single point of failure or control.
The infrastructure is emerging. The real question is whether the industry will embrace it — or settle for a version of sovereignty that only works if you’re willing to stay small.
