Seven gigafactories in 2028. My clients need sovereign inference this quarter.

On 30 July the Commission opened the tender for up to seven AI gigafactories: around €30 billion mobilised, up to €10 billion in EU and national money meant to pull in at least €20 billion private, at least 100,000 AI chips per site, letters of intent with AMD, Nvidia and Qualcomm for hardware access. Bidding closes 12 November 2026. First sites operational in 2028.

I am genuinely glad this exists. Europe cannot argue about sovereignty forever without building the compute to back it. But I have now sat in four meetings where a CIO cited the gigafactory programme as a reason to defer a decision, and that is a misreading worth naming.

2028 is not a plan, it is a hope

Count the steps between a tender and usable capacity. Consortium formation, award, sites, grid connections, transformers, cooling, chip delivery against a global queue, commissioning. Any one of those slipping a year is unremarkable in infrastructure. "Operational by 2028" means first sites, best case, and it means capacity allocated by a process that does not yet exist to applicants who will queue.

Meanwhile the roadmap in front of you has a board date this quarter.

What is available right now

This is the part that gets lost. Sovereign inference is not blocked on the gigafactories, because inference is not frontier training. You do not need 100,000 chips to serve a bank. You need a handful of GPUs in a European datacentre owned by a European company, and those exist today at OVH, Hetzner, Scaleway and a dozen smaller providers.

That is the whole architecture I keep deploying: open weights on rented European bare metal, Proxmox for the hypervisor layer, vLLM and Ollama for serving, weights and logs and keys under EU jurisdiction rather than merely EU residency. Available in weeks. Priced in euros. At sustained scale it lands materially below hyperscaler GPU instance rent, which is the argument I made in more detail elsewhere on this site.

The gigafactories will matter for training frontier models in Europe. Almost none of my clients are training frontier models. They are serving them.

The open-weight side made the case stronger, not weaker

The last two months have been the busiest stretch in open-weight releases I can remember. Kimi K3 arrived in July with open weights at 1M context. GLM and Qwen shipped substantial jumps. DeepSeek's V4 Pro landed a large capability step and, notably, a sharp price increase with peak and off-peak billing attached.

That last detail is the one I would put in front of a CFO. If your three-year plan assumes today's API price per million tokens, you have built a forecast on someone else's pricing page. Open weights on metal you rent do not reprice on you. The capability curve keeps rising and you adopt it by pulling new weights, which is a deployment task rather than a renegotiation.

What I actually advise

• Do not defer sovereign inference to 2028. It is a procurement decision available now, at small scale, with a real exit.
• Do watch the tender, particularly the access model. If gigafactory capacity is genuinely reachable for SMEs and enterprises for fine-tuning, that is a useful option to hold in 2029. Plan for it as upside, not as a dependency.
• Separate training from serving in your roadmap. They have different hardware needs, different economics and different timelines. Conflating them is what makes people think they need to wait.
• Keep your inference layer portable. One OpenAI-compatible gateway in front of everything, cloud and self-hosted, so that swapping where a model runs is a routing change. That is the thing that makes both today's rented metal and 2028's gigafactory a config edit rather than a migration.

Industrial policy operates on decade horizons. Your regulator, your board and your users operate on this one.