By Magic AILearning Objectives
- Understand what Magic is building and why 100 million token context windows matter for coding
- Evaluate the technical achievement of LTM-2-mini and its implications for software engineering
- Assess the risks of a pre-revenue company valued at $1.5 billion
What Is Magic?
Magic (magic.dev) is building AI models specifically designed for software engineering — with a singular technical breakthrough: 100 million+ token context windows. That is enough to read approximately 10 million lines of code or 750 novels simultaneously. For comparison, the next largest context window belongs to Google Gemini at 2 million tokens — 50 times smaller.
The vision is an "AI colleague" (not copilot) that understands your entire codebase at once — every file, every dependency, every commit — and can write code that is architecturally consistent with the whole project. The company has raised $466 million+ from Eric Schmidt, Atlassian, Sequoia, and CapitalG.
⚠️Warning
Magic has not publicly launched a product as of March 2026. The company operates in stealth/pre-launch mode with a waitlist. The 100 million token context is a demonstrated research capability (LTM-2-mini), not a shipping product. This page covers what is known about the technology and company for educational purposes.
The 100 Million Token Context
Magic's core innovation is the LTM (Long-Term Memory) architecture:
| Model | Context Window | Equivalent |
|---|---|---|
| LTM-1 | 5 million tokens | ~500,000 lines of code |
| LTM-2-mini | 100 million tokens | ~10 million lines of code (~750 novels) |
| LTM-2 (full) | In training | Expected to exceed LTM-2-mini |
The sequence-dimension algorithm behind LTM-2-mini is approximately 1,000 times cheaper than standard attention mechanisms at 100 million token context lengths. This makes impossibly long contexts computationally feasible.
💡Key Concept
Why Context Length Matters for Code: Most AI coding tools see only the file you are editing (a few thousand tokens). Some see a handful of related files (tens of thousands of tokens). Magic's vision is to see your entire codebase — every file, package, configuration, and test — simultaneously. This would enable architecturally consistent code generation, cross-file refactoring, and whole-project understanding that no existing tool can match.
Infrastructure
Magic is building dedicated AI supercomputers on Google Cloud:
- Magic-G4 — NVIDIA H100 GPU cluster (current)
- Magic-G5 — NVIDIA GB200 NVL72 GPU cluster (in development)
- Google Cloud selected as preferred cloud provider, using Google's AI Hypercomputer architecture
Google Cloud Partnership
Google Cloud and Magic announced a strategic partnership providing Magic with dedicated supercomputer infrastructure for training its LTM models. This gives Magic access to the latest NVIDIA hardware at scale without building its own data centers.
Company Details
| Detail | Info |
|---|---|
| Company | Magic AI (magic.dev) |
| Founded | 2022 |
| CEO | Eric Steinberger (co-founder) |
| Co-Founder | Sebastian De Ro |
| Headquarters | San Francisco, California (offices also in New York and Seattle) |
| Employees | ~97 |
| Valuation | $1.5 billion (mid-2024) |
| Total Raised | ~$466-515 million across 4 rounds |
| Latest Funding | $320 million Series C (August 2024; Eric Schmidt, Atlassian, Sequoia, CapitalG, Jane Street) |
| Earlier Investors | Nat Friedman; Daniel Gross; Elad Gil |
| Revenue | $0 (pre-revenue as of latest reports) |
| Website | magic.dev |
Magic vs. Competitors
| Company | Context Window | Product Status | Revenue |
|---|---|---|---|
| Magic | 100 million tokens (research) | Pre-launch / waitlist | $0 |
| Cursor | Up to ~100,000 tokens | Shipped; $500 million+ ARR | $500 million+ |
| Devin (Cognition) | Standard (sandboxed environment) | Shipped; $73 million ARR | $73 million+ |
| Poolside | 1 million+ tokens (Malibu) | Enterprise-only | ~$50 million |
| Augment Code | 200,000 tokens | Shipped; generally available | ~$20 million |
Magic's differentiation: No other company has demonstrated anything close to 100 million token context for code. If the technology translates to a shipping product, it would represent a category-defining capability. The risk: it has not shipped yet, and well-funded competitors are advancing rapidly.
Strengths
- Category-defining context window — 100 million tokens is 50 times larger than the next best (Gemini at 2 million); would enable true whole-codebase understanding
- 1,000 times cheaper attention — novel sequence-dimension algorithm makes long contexts computationally feasible
- Strong backers — Eric Schmidt, Atlassian, Sequoia, CapitalG, and Jane Street provide funding and strategic connections
- Google Cloud partnership — dedicated supercomputer infrastructure for training
- Clear vision — "AI colleague, not copilot" represents a bold bet on fully autonomous software engineering
Limitations and Considerations
- No shipped product — waitlist only as of March 2026; no generally available tool, demo, or API
- Zero revenue — $466 million+ raised with no reported revenue; high burn rate assumed
- No public benchmarks — no SWE-bench, HumanEval, or any standard evaluation published
- Tiny team — approximately 97 employees building one of the most ambitious AI products
- LTM-2-mini is a research demo — the blog post describing it states it is "still unreleased"; translating research capability to production product is a significant gap
- Execution risk — competitors (Cursor, Devin, Claude Code) are shipping real products while Magic remains in stealth
Key Takeaways
- Magic is building AI coding models with 100 million+ token context windows — enough to read 10 million lines of code simultaneously, using a novel algorithm 1,000 times cheaper than standard attention
- Backed by $466 million+ from Eric Schmidt, Atlassian, and Sequoia at a $1.5 billion valuation; Google Cloud partnership provides dedicated training infrastructure
- Has not publicly launched as of March 2026 — no product, no revenue, no benchmarks. This is a company to watch, not a tool to use today
- If the technology ships as demonstrated, it would enable whole-codebase AI understanding that no competitor can match — but execution risk is significant
