In the current landscape of 2026, training a frontier model is no longer just a research challenge; it is a massive capital allocation exercise. We have seen teams at European startups burn through seven-figure compute credits in weeks, only to end up with a model that underperforms because they optimized for the wrong variables. The sticker shock of a cloud bill often stems from a fundamental misunderstanding of how hardware actually interacts with transformer architectures. At Lyceum, we believe in radical transparency. Understanding the math behind your training run is the first step toward reclaiming sovereignty over your AI infrastructure and ensuring your compute budget translates into actual model performance.
Pricing Hourly Rates 6 min read
The Cost Per Training Run Calculator: A Guide for ML Engineers
Most AI teams realize their cloud bill is unsustainable only after the training run finishes. We break down the physics of compute costs and why Model Flops Utilization (MFU) is the only metric that actually matters for your bottom line.
Justus Amen
January 9, 2026 · GTM at Lyceum Technology
Last updated August 3, 2026
The Physics of Training Costs: The 6NP Formula
To build an accurate cost per training run calculator, you must start with the fundamental physics of the transformer architecture. For a standard dense model, the total number of floating-point operations (FLOPs) required for a single training pass is approximately 6 * N * P, where N is the number of parameters and P is the number of tokens in your dataset. This constant of six accounts for the forward pass (2 operations) and the backward pass (4 operations, including gradients and activations).
According to Epoch AI's June 2024 analysis of frontier training costs, the amortized hardware and energy cost of a final training run has grown by roughly 2.4x per year since 2016 (95% confidence interval 2.0x to 3.1x) [1]. This growth is driven by the 'Chinchilla scaling laws,' which suggest that for every doubling of model parameters, you should ideally double your token count to maintain optimal performance. If you are training a 70B parameter model on 2 trillion tokens, your raw compute requirement is roughly 8.4e23 FLOPs. However, translating this into a dollar amount requires factoring in the efficiency of your hardware.
- Parameters (N): The size of your model (e.g., 8B, 70B, 400B).
- Tokens (P): The total number of data points in your training set.
- Hardware Throughput: The theoretical peak TFLOPS of your GPU (e.g., 989 TFLOPS for an H100 at dense BF16).
Without accounting for efficiency, you might assume your GPUs are running at 100% capacity. In reality, most clusters operate far below their theoretical peak. This gap is where budgets are broken.
The MFU Trap: Why GPU Utilization is a Lie
One of the most common mistakes we see ML engineers make is relying on nvidia-smi to gauge efficiency. High GPU utilization metrics often mask deep architectural bottlenecks. You can hit 100% GPU utilization just by moving data in and out of memory without performing a single useful calculation. This is why we champion Model Flops Utilization (MFU) as the gold standard for cost estimation.
MFU measures the ratio of the observed model FLOPs throughput relative to the theoretical maximum FLOPs throughput of the system. The best public reference point is Meta's Llama 3 405B pre-training run, which reported an overall BF16 MFU of 38% to 43% across its 8,192 and 16,384 GPU configurations [2]. That is one of the most optimized runs published, so treat the high 30s to low 40s as a planning band, not a floor.
- Communication Overhead: As you scale to hundreds of GPUs, the time spent synchronizing gradients across the network increases, dragging down MFU.
- Memory Bottlenecks: If your batch size is too small, your GPUs spend more time waiting for data than processing it.
- Software Inefficiency: Unoptimized kernels or poor integration with frameworks like PyTorch can lead to significant 'dark compute' where the hardware is active but unproductive.
When using a cost per training run calculator, always input a realistic MFU. If you assume 60% but achieve 30%, your final bill will be exactly double your estimate. At Lyceum, our scheduling product uses memory and runtime prediction to help push MFU higher by optimizing batch sizes and parallelization strategies for European sovereign clusters.
Hardware Economics: H100 vs B200 in 2026
The arrival of the NVIDIA Blackwell architecture in 2025 fundamentally shifted the economics of training. On Lyceum's published on-demand rates, a B200 GPU VM costs a little over twice as much per hour as an H100, yet the total cost per training run is often lower on the newer hardware. This is the 'Price Paradox' of AI infrastructure.
NVIDIA rates a DGX B200 system at up to 3x the training performance of a DGX H100 [4], and on large-scale models that headroom can more than offset the higher hourly rate. The B200 also carries 180GB of HBM3e per GPU, or 1,440GB across the eight GPUs in a DGX B200 [4], which allows for much larger batch sizes and reduces the total number of GPUs required to fit the model in memory. Whether that nets out cheaper depends on the MFU you actually reach on your own workload.
However, for smaller models (under 10B parameters), the H100 remains a highly cost-effective workhorse. As H100 prices have matured and dropped in early 2026, they offer a stable platform for fine-tuning and smaller-scale pre-training where the massive memory of the Blackwell series isn't fully utilized.
The Hidden 20%: Egress, Checkpointing, and Idle Time
Your cost per training run calculator is incomplete if it only looks at GPU hours. There are three 'silent killers' of AI budgets that often account for 20% or more of the final invoice. First is data egress fees. AWS charges $0.09 per GB for the first 10TB per month of data transferred out to the internet from eu-central-1 (Frankfurt), tiering down to $0.05 per GB above 150TB per month, read from Amazon's public price feed on 3 August 2026 [5]. AI training also doesn't read data once; it pulls the same data repeatedly across epochs and distributed workers. Moving a 100TB dataset out of that region costs roughly $8,000 at those tiered rates, and a job that re-reads it every epoch pays that line again.
Second is activation checkpointing. To save memory and avoid Out-of-Memory (OOM) errors, engineers often recompute certain activations during the backward pass. While this saves VRAM, full recomputation costs an extra forward pass per step, adding roughly 30% to the compute required for the run [3]. This is a direct trade-off: you pay more in compute time to avoid buying more expensive high-memory GPUs.
Finally, there is idle time. In many legacy cloud environments, you pay for the GPUs from the moment they are provisioned, even if your data pipeline is still loading or your environment is being set up. Lyceum addresses this with per-second billing and no subscription or base fee: you pay for the seconds you actually hold the GPUs, not for rounded-up hours.
Calculate your full training run cost across providers. Try the GPU Pricing Calculator →
Sovereignty as a Strategic Cost Advantage
For European startups and enterprises, the cost of a training run isn't just the cloud bill; it's the long-term risk of data dependency. Relying on non-European infrastructure introduces regulatory overhead and potential litigation risks under GDPR that are rarely factored into a simple calculator. By using a sovereign European GPU cloud like Lyceum, you eliminate the 'compliance tax' associated with moving sensitive data across borders.
Beyond compliance, sovereign infrastructure provides predictable pricing. We have seen US-based providers fluctuate rates based on domestic demand, leaving European teams with unexpected cost spikes. Our clusters in European data centres in Spain, Paris and the Nordics offer high-performance compute with a radically transparent pricing model, allowing CTOs to forecast their R&D spend with precision. When you own the data and the orchestration layer, you aren't just running a model; you are building a strategic asset that is protected from external geopolitical shifts.
Sources
[1] Epoch AI, "How much does it cost to train frontier AI models?", 3 June 2024; [2] Grattafiori et al., "The Llama 3 Herd of Models", arXiv:2407.21783, 31 July 2024; [3] Chen et al., "Training Deep Nets with Sublinear Memory Cost", arXiv:1604.06174, 21 April 2016; [4] NVIDIA, "NVIDIA DGX B200", read 3 August 2026; [5] Amazon Web Services, "Amazon EC2 On-Demand Pricing", read 3 August 2026
Frequently Asked Questions
What is a realistic MFU to use in my calculator?
For most well-optimized transformer training runs on H100 or B200 clusters, assume an MFU of 38% to 43%, the band Meta reported for Llama 3. If you are using experimental frameworks or unoptimized data pipelines, drop this estimate to 25%.
Does activation checkpointing increase the cost?
Yes. Activation checkpointing saves memory but requires recomputing the forward pass during the backward pass. That extra forward pass typically adds around 30% to total compute time and cost.
Should I use H100 or B200 for fine-tuning?
For fine-tuning models under 70B parameters, H100s are often more cost-effective due to their lower hourly rates and high availability. B200s are better suited for massive pre-training runs where memory bandwidth is the primary bottleneck.
How does interconnect speed impact training cost?
Slow interconnects (like standard Ethernet) create bottlenecks during gradient synchronization, which can drop your MFU below 20%. High-speed interconnects like InfiniBand or NVLink are essential for maintaining cost-efficiency in multi-node clusters.
What are the benefits of a sovereign European cloud for AI training?
Sovereign clouds like Lyceum offer GDPR-compliant processing in European data centres and predictable per-second pricing. Lyceum's S3-compatible storage is free of ingress and egress charges.
Lyceum Technology