Magic has moved deeper into the race to automate software development, raising $320 million while also outlining a major compute partnership with Google Cloud. The San Francisco-based startup is building AI models intended to help engineers write, review, debug and plan code changes.
The new financing brings Magic’s total raised to $465 million and places it among a better-funded group of AI coding startups that includes Codeium, Cognition, Poolside, Anysphere and Augment.
A larger bet on AI coding tools
Magic said in a blog post on Thursday that the round included contributions from ex-Google CEO Eric Schmidt, Alphabet’s CapitalG, Atlassian, Elad Gil, Jane Street, Nat Friedman and Daniel Gross, Sequoia and others. Schmidt is also backing Augment, another company in the AI coding category.
The round appears to have exceeded what had previously been reported. In July, Reuters reported that Magic was seeking to raise over $200 million at a $1.5 billion valuation. The company’s current valuation could not be ascertained, while Magic was valued at $500 million in February.
Magic’s products are not yet for sale, but the company is aiming at a large and increasingly crowded market. Its tools are designed to act like an automated pair programmer, helping with software engineering tasks while trying to understand the broader context of a project.
That context matters because code work is rarely isolated to a single file or short prompt. A useful assistant must often understand how a change affects surrounding code, project conventions and future maintenance. Magic’s pitch is that larger context windows can make its models more useful for that kind of work.
Google Cloud and Nvidia hardware are central to the plan
Alongside the funding, Magic announced a partnership with Google Cloud to build two “supercomputers” on Google Cloud Platform. The Magic-G4 will use Nvidia H100 GPUs, while the Magic G5 will use Nvidia’s next-gen Blackwell chips scheduled to come online next year.
GPUs are widely used for generative AI because they can perform many computations in parallel. For a company training and serving code generation models, access to large GPU clusters can shape both model development and product performance.
Magic says it wants to scale the Magic G5 cluster to “tens of thousands” of GPUs over time. Together, the clusters are expected to reach 160 exaflops, with one exaflop equal to one quintillion computer operations per second.
The company framed the Google Cloud partnership as a way to move faster on infrastructure. Magic co-founder and CEO Eric Steinberger said Google Cloud gives the company “the fastest timeline to scale” while Nvidia’s Blackwell system is expected to improve inference and training efficiency for its models.
Long context is Magic’s main technical claim
Magic was co-founded in 2022 by Eric Steinberger and Sebastian De Ro. Steinberger previously worked at Meta as an AI researcher after beginning a computer science Bachelor’s program at Cambridge and dropping out after a year. De Ro came from German business process management firm FireStart, where he became CTO.
The two met through ClimateScience.org, an environmental volunteer organization Steinberger co-created. Steinberger has said in a previous interview that his interest in AI began early, including a high school project in which he and friends wired school computers for machine-learning algorithm training.
Magic’s core technical focus is its model architecture, which it calls “Long-term Memory Network,” or “LTM.” The company says this architecture gives its models ultra-long context windows.
A context window is the information a model can take into account before producing an answer. For coding, that information can include code and related project material. Larger context windows can allow more of a codebase to fit into the model’s working view.
Magic claims its latest model, LTM-2-mini, has a 100 million-token context window. The company says that equals around 10 million lines of code, or 750 novels. TechCrunch described it as the largest context window of any commercial model, with Google’s Gemini flagship models next at 2 million tokens.
Magic says LTM-2-mini used that long context to implement a password strength meter for an open source project and to create a calculator using a custom UI framework pretty much autonomously. The company is now training a larger version of the model.
The opportunity comes with unresolved concerns
Magic is pursuing a market that could be worth $27.17 billion by 2032, according to an estimate by Polaris Research. That potential helps explain why investors are willing to fund a company with a small team, around two dozen people, and no revenue to speak of.
Developer interest in AI coding assistants is already visible. GitHub’s latest poll found that the vast majority of respondents had adopted AI tools in some form. Microsoft also reported in April that Copilot had over 1.8 million paying users and more than 50,000 business customers.
Still, the category has open questions. AI-powered assistive coding tools face security, copyright and reliability concerns. For software teams, those issues matter because code assistants can influence how production systems are written, reviewed and maintained.
Magic’s ambitions also go beyond routine development support. On its website, the company describes a path to AGI, meaning AI that can solve problems more reliably than humans can alone.
To support that work, Magic recently hired Ben Chess, a former lead on OpenAI’s supercomputing team. It also plans to expand its cybersecurity, engineering, research and system engineering teams.
What the funding signals
The $320 million round gives Magic more room to compete in an expensive field where talent, compute and model training all matter. Its Google Cloud partnership also shows how closely AI coding companies are tied to the infrastructure needed to train and run large models.
The central question is whether Magic’s long-context approach can translate into practical tools that developers trust. The company has money, major investors and a large compute plan. It still has to turn those advantages into products engineers can actually use.