@SemiAnalysis_ declared pretraining dead for non-frontier labs. Enterprises and startups suffer from 'Pretrainitis,' pursuing vanity pretraining for career advancement.
Key facts
- Pretraining ROI negative for non-frontier organizations.
- @SemiAnalysis_ coined term 'Pretrainitis'.
- Frontier labs include OpenAI, Google DeepMind, Anthropic.
- Prompt engineering offers higher ROI according to analysis.
- SemiAnalysis is a respected AI infrastructure analyst firm.
According to @SemiAnalysis_, the economic calculus for pretraining large language models has shifted decisively. The analyst firm argues that only frontier labs—OpenAI, Google DeepMind, Anthropic—can justify the capital and compute expenditure required for pretraining runs that now cost tens of millions of dollars.
"Pretraining fundamentally does not make sense anymore for anyone other than frontier labs," the thread states. It identifies a widespread pattern across enterprises and startups: teams pushing to pretrain their own models as a way to demonstrate impact internally and secure promotions, a phenomenon the firm dubs 'Pretrainitis.'
The diagnosis is blunt: "There are a lot of people at enterprises & startups who have 'Pretrainitis' to show 'impact' and get promotions, fundamentally, it doesn't make sense." Instead, @SemiAnalysis_ recommends a different allocation of resources. "There is probably higher ROI in partnering with a frontier lab to do prompt engineering, although it isn't as 'sexy' as pretraining."
This view aligns with a broader industry trend. In 2025, multiple enterprise AI projects pivoted from pretraining to fine-tuning existing frontier models, driven by the escalating cost of training runs. Meta's Llama 3.1 405B, for instance, required an estimated $60M+ in compute alone. For a non-frontier organization, replicating such capability is economically irrational when API access to frontier models costs pennies per query.
The 'Pretrainitis' critique extends to startup funding dynamics. VCs have grown skeptical of pretraining-heavy pitches, preferring startups that build applications on top of existing models. The signal: if your startup's core moat is a pretrained model not built by a frontier lab, your runway may be shorter than your training epoch.
What to Watch
Watch for enterprise AI budget allocations in Q2 2026 earnings calls—specifically whether companies disclose reducing internal pretraining spend in favor of API-based fine-tuning or prompt engineering services. A single Fortune 500 CFO calling out 'Pretrainitis' by name would validate the thesis.
Key Takeaways
- @SemiAnalysis_ declares pretraining dead for non-frontier labs, citing 'Pretrainitis' as vanity-driven waste.
- Prompt engineering offers higher ROI.
[Updated 12 Jun via gn_dc_power]
OpenAI is reportedly in talks to lease a 10-gigawatt data center in Ohio, with backing from Nvidia, according to The Information. The facility would require an estimated $500 billion investment and could be built on federal land. This massive infrastructure push underscores that even as pretraining becomes uneconomical for non-frontier labs, the leading frontier labs are doubling down on compute scale to maintain their edge.
