0.5 ExaFLOPS: What Deutsche Telekom's Setup Really Means for AI Training

0.5 ExaFLOPS is a strong number – but the precision behind it determines whether it translates to “AI autonomy in months” or “in years.”

The Telekom Industrial AI Cloud

Deutsche Telekom has launched its Industrial AI Cloud in Munich (Tucherpark): ~10,000 NVIDIA Blackwell GPUs (DGX B200 + RTX PRO Servers) delivering “up to 0.5 ExaFLOPS,” operated under German requirements for data protection, security, and availability.

The Key Interpretation

These 0.5 ExaFLOPS are very plausibly FP64 (HPC metric). A single B200 is rated at roughly ~37–40 TFLOPS FP64; multiply by 10,000 and you land exactly in the 0.4–0.5 EFLOPS range.

For LLM training, however, BF16/FP8 is what counts – and here the facility is in a completely different league, it’s just not being marketed that way.

How Fast Could a SOTA Training Run Be?

Let’s take Llama 3.1-405B as a reference with ~3.8×10²⁵ FLOPs. From “>1,000 DGX B200” we get ≥8,008 B200 GPUs (8 per DGX) – plus RTX PRO in the remaining mix, which typically wouldn’t be the compute bottleneck for pretraining at this scale.

🔹 Back-of-the-envelope (B200 share only, BF16 dense):

• Peak: ~8,000 × ~2.25 PFLOPS ≈ 18 EFLOPS (dense, no sparsity)
• With a realistic Model-FLOP-Utilization of 35–40%, that effectively yields ~6.3–7.2 EFLOPS
• Result: ~58–70 days of continuous training

Important: 58–70 days of training ≠ 58–70 calendar days. In practice, checkpointing, failures/restarts, evaluation, and tuning add up. A realistic buffer is ~1.5–2×, so roughly ~3.5–5 months wall-clock.

Cost

Public cloud rates for B200 sit roughly at ~$4.7–$8 per GPU-hour (depending on provider/commitment). As a “rental equivalent” for 8,000 GPUs, that means:

• ~$0.9–1.54M/day
• Over 58–70 days: ~$52–108M

Caveat: These are rental prices, not Telekom’s internal costs (depreciation, power, utilization effects, etc. are typically lower).

Bottom Line

This facility is less “just another cloud” and more a lever for European AI sovereignty – provided access, utilization, and programs are designed so that not only large corporations benefit.