By Anna Tong and Krystal Hu
(Reuters) - Magic, a U.S. startup developing artificial-intelligence models to write software, is in talks to raise over $200 million in a funding round valuing it at $1.5 billion, just several months after its last capital raise, three sources told Reuters.
Investors including Jane Street are expected to participate in the round, which could triple Magic's valuation from its last round, despite earning no revenue and having no product for sale, according to sources who requested anonymity to discuss private matters.
The startup, which has about 20 employees, was last valued at $500 million after the previous February fundraising, according to PitchBook data. It has raised $140 million since it was founded in 2022, from backers including Nat Friedman and Daniel Gross’ NFDG Ventures, as well as Alphabet (NASDAQ:GOOGL)'s CapitalG.
Magic declined to comment. Jane Street did not respond to a request for comment.
Magic's new round is the latest fundraising in one of the most promising applications of generative AI technologies. Software developers represent a significant cost for tech companies, and tools that can generate code or assist developers in more efficient programming are highly attractive.
The early success of Microsoft’s GitHub has further stoked investors’ enthusiasm. In April, coding-assistant startup Augment raised $252 million, while Cognition, developer of the AI-powered coding assistant Devin, secured $175 million in a round led by Founders Fund at a $2 billion valuation.
Investors say they see GitHub Copilot's success as proof of how large the market is. GitHub reported 40% higher year-over-year revenue in the latest quarter, driven by Copilot, which has 1.3 million paid subscribers.
“The success of Microsoft (NASDAQ:MSFT) has validated the commercial market for AI code assistants, leading everyone to believe there is clear market demand and a customer willingness to pay for the right product. The opportunity is enormous, with likely multiple winners in this category,” said Brian Dudley, partner at Adams Street Partners.
While commercially available products like GitHub Copilot or OpenAI's ChatGPT can suggest how to complete lines of code, the next frontier for coding assistants is to design and write entire software applications without human help.
Magic and a few other startups are trying to achieve this by training their own large language models for coding-specific tasks. This effort is cost-intensive, however, for purchasing data, chips and electricity.
Magic plans to spend its funding on improving its own models that support long-context windows, which refer to AI systems that can process more data in one query, the sources added.
Magic's ability to understand and process a large amount of context at once is due to an innovative design that goes beyond the traditional "transformer model" commonly used in large language models such as OpenAI's GPT models, according to one of the sources.
Poolside AI, a Paris-based startup that is taking a similar approach to Magic, is in talks to raise $450 million at a $2-billion valuation, TechCrunch reported in June. Poolside also has no commercially-available products.