Observed Signal · Sep 8, 2026 · Technical Release · Source: Trending Topics (DACH/CEE Innovation & Tech) · Impact: 4/5 · Sentiment: Positive
Magic AI Claims Frontier-Level Pretraining for Under $1M
Magic, an AI startup co-founded by Austrians Eric Steinberger and Sebastian De Ro, claims a major breakthrough in pretraining efficiency. Its new recipe reportedly matches DeepSeek V4 Pro's base model quality using 50x less compute, costing about $500,000 on Nvidia GB200 systems, and is over ten times more compute-efficient than leading open-weight models. Scaling to roughly $4 million, Magic says it outperforms all public base models on perplexity evaluations, comparing against DeepSeek V4 Pro, Kimi K2, and Nvidia's Nemotron 3 Ultra, while excluding closed models from Anthropic, Google, and OpenAI. The company has raised over $460 million, with a $320 million round valuing it at $1.5 billion, and partners with Google Cloud for tens of thousands of GB200 chips. Magic has not yet released a model, and all claims are self-reported.
Magic's claim of frontier-level pretraining at a fraction of the cost could disrupt the AI model training landscape, making it relevant for AI infrastructure and potentially affecting the broader tech ecosystem.
Wichtigste Kernpunkte & Evidenz
- Magic claims its pretraining recipe matches DeepSeek V4 Pro's quality with 50x less compute, costing ~$0.5M on GB200, and is 10x more efficient than leading open models.
- Scaling the recipe to ~$4M, Magic says it beats all public base models on perplexity evaluations.
- Magic has raised over $460M, with a $320M round valuing it at $1.5B, and partners with Google Cloud for GB200 compute.
- Magic has not yet released a model, and its claims have not been independently verified.
- Magic's model achieves a 72% pass rate on math problems in a short RL run, compared to GPT-6 Astra at 100% and Kimi K3 at 75%.
Verknüpfte Unternehmen
10 verknüpfte UnternehmenAtlassian
Enterprise-Software für Zusammenarbeit, Workflows und Team-Produktivität.
“Atlassian backed a $320 million round for Magic....”
Magic
Frontier code-model developer for autonomous software engineering and research.
“Magic is an AI startup founded by Eric Steinberger and Sebastian De Ro, claiming frontier-level pretraining efficiency....”
Fireworks AI
B2B-Plattform für hocheffizientes Modelltraining, intelligentes Routing und extrem latenzarme Inferenz.
“Magic brought in the inference provider Fireworks for an independent check of baseline figures....”
Sequoia Capital
Eine führende Venture-Capital-Gesellschaft, die Technologieunternehmen über alle Wachstumsphasen hinweg mit institutionellem Kapital und strategischer Expertise begleitet.
“Sequoia backed a $320 million round for Magic....”
Alphabet
Digitales Plattform-Konglomerat mit Fokus auf Werbung, Medien und Software.
“Alphabet participated in a $23 million funding round for Magic....”
CapitalG
Alphabet growth fund investing in transformational technology companies.
“CapitalG backed a $320 million round for Magic....”
NVIDIA
Ein führendes Unternehmen für Accelerated Computing, das KI-Software, Cloud-Infrastruktur und Gaming-Technologien bereitstellt.
“Magic's compute costs refer to Nvidia's GB200 systems, and Nvidia's Nemotron 3 Ultra is used as a comparison model....”
DeepSeek
DeepSeek ist ein führender LLM-Entwickler, der hocheffiziente KI-Modelle über eine performante API und Consumer-Chat-Schnittstellen bereitstellt.
“Magic compares its pretraining recipe to DeepSeek V4 Pro and V4 Flash....”
Moonshot AI
Chinese AI company delivering Kimi models, agents, enterprise software and APIs.
“Moonshot's Kimi K2 and Kimi K3 are used as comparison models....”
Google Cloud
Enterprise cloud, data and AI services for businesses.
“Magic announced a partnership with Google Cloud to build compute capacity with tens of thousands of GB200 chips....”
Ontology Mapping & Concepts
Track Real-Time Market Signals & Shifts
Set up custom watchlists to receive automated, evidence-grounded executive digests whenever material signals or shifts occur across your tracked landscape.