Beyond NVIDIA: AMD, ASICs and GPU Diversification in European Datacenters

NVIDIA dominates AI acceleration as few suppliers have ever dominated a market. Yet diversification is advancing: AMD is gaining ground, hyperscalers are deploying their own chips, and startups like Tenstorrent — which just raised 1.5 billion dollars — are funding alternatives. For datacenters, the challenge is no longer picking a side, but being able to host both.

NVIDIA's dominance rests on software as much as silicon: the CUDA ecosystem, built over fifteen years, is the real barrier to entry. European datacenters have massively built their AI offerings around this de facto standard — racks, networking, cooling and skills included.

But signals of diversification are multiplying: AMD is progressing with its Instinct range, Google and AWS deploy their ASICs at scale, and the silicon startup market is attracting billions. Just last week, research firm SemiAnalysis reported that neoclouds which had added AMD GPUs or Google TPUs to their offering were denied access to GTC Berlin — proof that competition has become a sensitive subject.

Why Customers Want an Alternative

The motivations are rational: securing supply amid delivery tensions, creating leverage in price negotiations, and meeting the digital sovereignty requirements of some European public customers. A cloud offering both architectures captures demand that a single-vendor site cannot serve.

Friction remains real: migrating a workload from CUDA to ROCm or an ASIC costs engineering time, and training performance for the largest models remains NVIDIA's territory. Diversification first concerns inference and new workloads, which are less locked in.

What This Changes for Investors

For a datacenter, hardware is the customer's choice, not the building's. What matters: sufficiently dense racks, liquid cooling compatible with several chip generations, an open network. An asset able to host NVIDIA, AMD or ASIC indifferently is structurally more resilient — its potential demand does not depend on a single supplier.

The discipline: do not price in a NVIDIA collapse — implausible — but price optionality. In contracts, dual-source clauses and multi-architecture commitments from cloud operators deserve to be read as an indicator of future demand strength.

How We Apply This at FINXIA

In the TITAN DC AI strategy, we design infrastructure as agnostic: power, cooling and connectivity sized for the roadmaps of all major vendors, not just one. The white room must outlast silicon cycles — it lasts twenty years, GPUs three.

This is a further argument for brown-to-green: a transformed site can integrate, from the renovation stage, power distribution and cooling compatible with any architecture, without paying the premium of a later conversion.

Key Takeaways

NVIDIA will remain dominant for a long time, but the market is diversifying — AMD, hyperscaler ASICs, well-funded startups. The winning datacenters will be those hosting all camps: an AI asset's resilience is measured by its hardware agnosticism.

Worth watching: orders for the next AMD Instinct generation by European neoclouds, and announcements at GTC Berlin this week — where the competition question will not be on the margins, but at the center.