Europe expands its AI factories; Cohere's new AI model runs on one GPU; CEOs and government ministers turn to chatbots for advice; a startup tries to replace the US national weather service

The quest for scientific superintelligence; the AI talent race is reshaping the job market; "vibe coding" AI-generated video games are coming; DeepSeek founder focuses on research rather than sales

Mar 14, 2025

∙ Paid

The European Commission announced this week that Austria, Bulgaria, France, Germany, Poland, and Slovenia will be part of an expansion of the AI Factories program, supported by national and EU investments of around €485 million.

AI Factories are a network of supercomputers built as part of the EuroHPC Joint Undertaking; they are meant to link universities and industry to supercomputing centers, with the hope that competitive generative AI models will be produced as a result.

However, a closer look at what’s available today reveals that many of these systems are built upon aging architectures. Take, for instance, Portugal's Deucalion supercomputer. Inaugurated in September 2023, it claims a peak performance of 5 petaflops. Yet, its architecture relies heavily on Fujitsu A64FX CPUs and Nvidia Ampere GPUs—components that, while great in their time, are now vastly outperformed by more advanced technologies. Similarly, Italy's Leonardo supercomputer, operational since November 2022, incorporates nearly 14,000 Nvidia A100 GPUs. While that was a pretty decent setup two years ago, the A100 has since been surpassed by Nvidia's newer offerings, such as the H100 and the latest Blackwell series. In fact, some of the EuroHPC supercomputers don’t have any GPUs at all such as Bulgaria’s Discoverer cluster.

If the above trend continues for the newer EU supercomputers, this reliance on older hardware will present a significant challenge for European AI startups. Training large-scale AI models demands more computational power and memory bandwidth, and newer reasoning models have even higher requirements as they use inference-time scaling which eats up GPU resources. Modern GPUs, like Nvidia's GB200 NVL72, are specifically designed to handle such intensive training and inference workloads, offering substantial improvements in speed, efficiency, and scalability. Without integrating these components, Europe's supercomputing infrastructure may inadvertently place its universities and companies at a disadvantage compared to their counterparts in North America and Asia.

The global landscape underscores this disparity. In the United States, Big Tech has gone on a spending spree, investing heavily in advanced AI infrastructure. A notable example is Meta’s $200 billion data center which would be several times larger than the 2GW, 4 million square foot facility under planning in Louisiana (a $200 billion data center would require up to 7 GW of power, according to sources). Meanwhile, it took 122 days for xAI to build a new 750,000 square foot Colossus supercomputer which has 100,000 Nvidia H100 GPUs. Such massive investments will ensure that United States AI ventures have access to the latest and most potent computational resources. Similarly, in Asia, companies like Foxconn are assembling servers equipped with Nvidia's Blackwell AI architecture, further widening the infrastructure gap.

To put Europe’s AI factories in perspective, MareNostrum 5 (the most advanced cluster from the list of existing EuroHPC JU clusters) has around 20 MW of capacity.

In my view, Europe must recalibrate its approach to supercomputing and aim for a bolder, interconnected data center approach rather than isolated HPC clusters. Encouragingly, some European entities are recognizing this need. Italian startup iGenius, in collaboration with Nvidia, is constructing one of the world's largest AI systems, dubbed Colosseum, in southern Italy. This data center will house approximately 80 Nvidia GB200 NVL72 servers, each equipped with 72 Blackwell chips.

But it’s not just about raw computing power: there are concerns around accessibility and ease of use of EuroHPC systems that AI startups in particular have to deal with.

Access to EuroHPC’s AI supercomputers is generally obtained through proposal-based allocations or special programs. Startups in EU member or associated states can apply for computing time on EuroHPC pre-exascale systems and the onboarding typically involves an application (e.g. via EuroHPC access calls) and an initial wait for approval. I’m sure resource availability is excellent once approved, but startups will likely face queue times during peak periods since these are shared national resources. That’s a lot of bureaucracy and time wasted at a time when generative AI is moving very fast.

In contrast, cloud providers like AWS, Azure, and Google Cloud make compute resources available on-demand to virtually anyone. There are no special eligibility requirements: a startup simply needs an account (and a payment method) to get started. This immediacy means no waiting: a developer can log in and launch GPU instances or AI services within minutes. Cloud platforms are globally accessible, which is great for distributed teams. Another plus is the abundance of startup incentives as major clouds offer generous credits to new startups to ease costs. For example, AWS’s Activate program provides up to $100,000 in credits for eligible early-stage startups (and even up to $300k for those in certain AI programs), and Azure and GCP have similar startup credit offerings. These credits and free tiers make initial access extremely affordable, effectively lowering the entry barrier.

Then there’s ease of use. Traditional HPC environments have a reputation for complexity because they typically use batch job schedulers, command-line interfaces, and require users to manage modules or container environments for their software. Even though some EuroHPC JU clusters offer technical assistance, building in a supercomputer requires some specialized adaptation: startups need to containerize their AI applications or adjust code to run efficiently on the underlying hardware. There may be stricter environment controls (for security, many supercomputers restrict direct internet access from compute nodes, etc.).

In general, once the workflow is established, running on HPC is straightforward, but getting to that point demands more initial effort and HPC-specific knowledge than using typical cloud services.

In contrast, a startup can launch a ready-to-use AI environment with a few clicks (for example, by spinning up a pre-configured GPU VM). Many AI frameworks and tools come pre-installed on cloud marketplace images or container offerings, saving time on setup. The cloud also excels in providing high-level AI services: APIs for vision, speech, or natural language processing that can be consumed without understanding the underlying ML models.

Secondly, the HPC model of computing is batch-oriented, which can slow interactive prototyping. Startups may need to queue jobs and wait for scheduling, which makes iterative debugging or trying quick changes a bit cumbersome compared to the instant run-stop cycle in a cloud notebook. In cloud computing, resources are available on-demand, so a startup can go from an idea to a running experiment in minutes.

Need to test a new model architecture? A developer can quickly spin up a high-performance GPU instance, run training for a few hours, then shut it down – paying only for that usage. This flexibility encourages experimentation: you can launch and terminate dozens of test runs in a single day without long setup, which hugely accelerates the build-measure-learn cycle. Clouds also offer managed ML services that handle a lot of the boilerplate (setting up distributed training, hyperparameter tuning, data pipelines, etc.), further speeding up development.

For companies developing commercial AI applications, the choice between EuroHPC supercomputers and cloud platforms depends on their needs and stage of development. I’m sure some startups will still find EuroHPC AI supercomputers useful, especially if they’re building in the areas of life sciences.

On February 10, at the AI Action Summit in Paris, the European Commission president Ursula der Leyen unveiled a plan to spend €20 billion to build four AI gigafactories aimed at training the most complex AI models. While ambitious, it’s worth remembering that it’s still 25 times less than what OpenAI alone plans to spend with the Stargate Project.

And now, here are the week’s news:

❤️Computer loves

Our top news picks for the week - your essential reading from the world of AI

Keep reading with a 7-day free trial

Subscribe to Computerspeak by Alexandru Voica to keep reading this post and get 7 days of free access to the full post archives.