Neil Rimer isn’t one for alarmism. When the co-founder of Index Ventures – one of the funds that shaped the European tech ecosystem – says that the wealth AI is generating in Silicon Valley will have to be redistributed, by choice or by force, he’s not making a political prediction. He’s reading the economic balance of power that AI is consolidating, and which he believes is unsustainable in the long run.

The starting point is well known: a handful of companies – mostly cloud hyperscalers and their strategic partners – today control the entire AI stack, from models to inference hardware. An unprecedented concentration of capital, which has generated stellar valuations and even more stellar margins for those selling centralized computing power. Rimer argues that this accumulation cannot remain confined to the usual players. If AI wealth isn’t redistributed voluntarily, external forces – regulators, markets, geopolitical tensions – will step in to force the issue.

Beyond the cloud: the game is played on local infrastructure

Rimer’s thesis has an immediate corollary for those working on AI deployment. If value must flow from large computing centers to a broader fabric, on-premise and self-hosted inference hardware ceases to be a niche for control obsessives and becomes a strategic asset. It’s not just about data sovereignty – although GDPR and similar regulations continue to push in that direction – but pure economics: keeping the value of AI workloads within one’s own infrastructure means reducing dependence on ever-rising cloud fees and, potentially, participating in the redistribution Rimer speaks of as active beneficiaries.

Those producing GPUs and private cloud solutions (from NVIDIA DGX to inference platforms like vLLM) are in a privileged position. But also companies that are currently investing in local computing resources – servers with abundant VRAM, modular architectures for fine-tuning and quantization – may find themselves with assets the market will reward, because they lower total cost of ownership (TCO) over the medium term, as cloud providers’ margins start to be challenged.

It’s no coincidence that many organizations are already shifting inference workloads to self-hosted infrastructure: latency drops, control over the pipeline is total, and, most importantly, operational costs are no longer tied to a per-token rate decided day by day by someone thousands of kilometers away. If Rimer is right, this will no longer be just a technical choice for a few, but a structural necessity.

Winners and losers in a less concentrated AI

The flip side is equally clear. Large cloud AI providers, which today capitalize on hardware scarcity and the network effect of centralized models, could see their entrenched rents erode. They won’t disappear, but they’ll have to accept reduced margins and a less dominant role, in an ecosystem where computing power becomes a distributed commodity. Model-as-a-service vendors could also suffer if companies start preferring LLMs optimized to run locally – with aggressive quantization and frameworks designed for on-premise – instead of paying for every API call.

On the opposite side, those developing tooling for on-premise deployment, from orchestration managers to fine-tuning toolkits, will find fertile ground. And organizations that invested early in skills and hardware suited for AI workloads will see a return that goes beyond immediate savings: they’ll be able to actively participate in redistributing value, rather than endure it.

An open question touches the public sector and critical infrastructure: if forced redistribution of AI wealth became a political goal, with incentives or regulatory constraints, on-premise hardware could become the lever to ensure that AI works for a distributed ecosystem and not for a few monopolies. In such a scenario, technological sovereignty and local computing capacity would no longer be just a competitive advantage, but a prerequisite for existing in the new AI market.