Magic AI

1.1 Identity and product

Magic AI is identified by Tracxn as a San Francisco company founded in 2022 by Eric Steinberger and Sebastian De Ro, operating in AI-based code completion and code generation. Independent coverage describes a broader ambition than inline completion: Magic's tools were designed to help engineers write, review, debug, and plan changes, while its LTM architecture was intended to ingest unusually large codebases. The Decoder reported the company's LTM-2-mini claim of a 100-million-token context window, equivalent to about 10 million lines of code or 750 novels. These are company or third-party reports of product capability, not independent proof of production reliability, customer deployment, or realized revenue.[CO001, CO002, CO003, CO004]

1.2 Founders and leadership

The public record supports a founder-led operating model centered on Eric Steinberger, who TechCrunch described as co-founder and CEO, and Sebastian De Ro, his co-founder. TechCrunch linked Steinberger's background to Cambridge computer science study and Meta AI research, and described De Ro's prior CTO role at FireStart. The same report said Magic had a small team of around two dozen people in August 2024, while later profiles give materially different staffing signals: approximately 50-100, 109 as of an unspecified August 26, and 19 for a legal-entity row dated December 31, 2024. These figures are not comparable enough to establish current headcount. The evidence also does not establish a board, independent directors, formal governance controls, or a complete executive roster, leaving founder dependence as a material diligence issue.[CO001, CO005, CO006, CO007, CO008]

1.3 Financing and operating scale

Financing is the clearest externally visible scale signal. TechCrunch reported a $320 million round announced in August 2024 and total funding of $465 million, while The Decoder reported $320 million from investors including Eric Schmidt, Jane Street, and Sequoia and $465 million total. Tracxn's later profile reports $466 million across four rounds, whereas AI Market Watch reports $515 million and lists a $5 million seed, $23 million Series A, $117 million Series B, and $320 million Series C. The resulting $465 million, $466 million, and $515 million totals are not reconciled. Tracxn also shows a historical peak valuation of $1.5 billion in December 2023 and says the latest valuation was 7% lower as of December 31, 2024; the actual latest valuation is masked. Revenue, ARR, customers, burn, runway, and financing terms remain unestablished.[CO009, CO011, CO012, CO013]

1.4 Milestones and unresolved diligence

The milestone record shows rapid movement from a 2022 founding to a major 2024 financing and infrastructure build-out. TechCrunch reported a Google Cloud partnership for Magic-G4 and Magic-G5 systems using Nvidia H100 and planned Blackwell chips, while The Decoder described the 100-million-token LTM-2-mini announcement and the HashHop benchmark. The New Stack provides an important adverse check: in early 2026 it stated there was no public evidence of LTM-2-mini being used outside Magic, despite the earlier financing and long-context claims. That does not prove the technology failed, but it limits public evidence of commercial adoption. A board roster, named customers, recurring revenue, production usage, security posture, training-data provenance, and current valuation should be requested before relying on the financing headline as evidence of operating maturity. The retained Polaris, BLS, and Artificial Analysis pages provide category, labor, and benchmark context, but no Magic-specific operating metrics; they should not be treated as company proof.[CO014, CO015, CO016, CO017]

1.5 Exhibits

2.1 Market boundary and adoption job

The relevant market is AI code tools used inside software delivery: code generation, completion, debugging, testing, review, and increasingly agentic planning and execution. Polaris describes the category as software that writes, checks, and optimizes code and automates activities, while its use-case list includes generation, completion, bug detection, test creation, DevOps automation, and API development. Magic's narrower wedge is whole-codebase engineering: TechCrunch reports that its tools are designed to help engineers write, review, debug, and plan changes, with LTM-2-mini intended to keep very large codebases, documentation, and libraries in context. Excluded from this boundary are general-purpose foundation-model spend, conventional IDE licenses without AI functionality, outsourced software-development labor, and infrastructure GPU spend. The status quo is therefore a mix of human engineering time, conventional IDEs, search and retrieval, tests, and manual review rather than a single substitute product. The market is attractive because automation can compress repetitive work, but buyer value is realized only when generated changes are reliable enough to merge and deploy.[CM001, CM002, CM003, CM004]

2.2 Sizing lenses and what remains unproven

The strongest quantitative lens in the assigned evidence is a third-party estimate for global AI code tools: Polaris reports USD 7.34 billion in 2025, USD 9.23 billion in 2026, and USD 58.96 billion in 2034, implying a 26.05% CAGR from 2026 to 2034. Those figures are category revenue estimates, not Magic revenue, ARR, or a SAM. They include tools and services across offerings, technologies, applications, industries, and regions, so the estimate is a broad TAM-like ceiling for the defined category. A second lens is technical demand: Magic and external coverage position 100-million-token context as a way to reason over entire codebases, but The New Stack notes that competing long-context architectures and benchmark claims remain contested and reports no public evidence of LTM-2-mini use outside Magic as of early 2026. A constrained SAM and SOM cannot be isolated from the pool because Magic has no public pricing, customer count, production deployment, conversion, or usage data in these sources. The correct valuation implication is thus category exposure with substantial execution and commercialization uncertainty, not a precise share-of-market calculation.[CM005, CM006, CM007, CM008, CM009]

2.3 Buyer segmentation, drivers, and constraints

The primary economic buyer is likely an engineering organization that owns delivery throughput, developer productivity, and software quality; users are developers, reviewers, and technical leads; and the payer is typically an enterprise software or engineering budget. Polaris identifies enterprise adoption of agentic assistants, governance, observability, and security as market trends, and describes SMEs as an opportunity because cloud subscriptions reduce the need for large engineering teams. Magic's long-context proposition is most relevant to teams with large, interdependent repositories where retrieval and context assembly are bottlenecks. Adoption triggers include faster implementation, debugging, testing, and review, but the sources do not establish a Magic-specific ROI or procurement path. Constraints are material: Polaris flags security, accuracy, licensing, and intellectual-property risk; TechCrunch reports the wider coding-agent market is moving quickly and that early reviews found Devin made mistakes; and The New Stack cautions that benchmark margins can reflect the harness as much as the model. GPU economics and dependence on cloud infrastructure also matter: TechCrunch describes Magic's Google Cloud supercomputer partnership and plans to scale to tens of thousands of GPUs. These constraints make trust, integration, security controls, and measurable production outcomes the gating factors for expansion.[CM010, CM011, CM012, CM013, CM014, CM015]

2.4 Exhibits

3.1 Competitive landscape and buyer alternatives

Magic sits in a crowded coding-assistance market with several distinct ways to solve the same job. Cursor is a dedicated AI coding environment with substantial reported investor demand; Amazon Q Developer is embedded across IDEs, the AWS console, CLI, chat applications, and GitHub; and Claude Code provides workflows for exploring, debugging, refactoring, testing, and pull requests. These are direct or adjacent product alternatives because a buyer can adopt them for repository understanding and software changes without waiting for Magic. Internal build, conventional IDE tooling, retrieval-augmented workflows, and human review remain status-quo substitutes. The evidence does not establish Magic's current distribution, production customer base, or realized pricing, so its competitive position should be treated as a product thesis rather than a demonstrated commercial advantage.[CP001, CP002, CP003, CP004]

3.2 Capability, packaging, and trust comparison

Magic's technical proposition is unusually ambitious: the company describes a 100-million-token LTM-2-mini context window and a HashHop evaluation intended to test retrieval across long contexts. That capacity can support whole-codebase workflows, but the evidence explicitly warns that context size alone does not guarantee correct reasoning, tool execution, or test-passing changes. Competitors expose more observable workflow surfaces: Amazon Q describes autonomous implementation, testing, documentation, review, refactoring, upgrades, vulnerability scanning, and private-repository connections, while Claude Code documents repeatable engineering workflows. Artificial Analysis provides benchmark categories for coding-agent performance, token usage, cost, execution time, safety refusals, and reward hacking, but the captured page does not provide a Magic score. Accordingly, unsupported matrix cells remain unknown rather than being scored as wins or losses.[CP005, CP006, CP007, CP008, CP009]

3.3 Moat durability and displacement risk

Magic's potential moat is the interaction of long-context architecture, training efficiency, and agentic reinforcement learning rather than context length alone. The company says its pretraining recipe matches a named open model with substantially less compute, but the reporting also states that the figures are company-supplied, have not been independently verified, and that Magic had not shipped a model at the time of publication. The competitive field is moving quickly: Subquadratic reports a 12-million-token model and a coding agent, while The New Stack notes that long-context approaches have repeatedly traded one property for another and cautions that Magic's earlier 100-million-token claim had no public evidence of use outside Magic as of early 2026. A durable advantage therefore requires public production outcomes, reliable code-change quality, procurement controls, and distribution—not just a larger context headline.[CP010, CP011, CP012, CP013, CP014]

3.4 Exhibits

4.1 Revenue model and monetization visibility

Magic's public materials describe a code-focused model and product workflow, but the assigned evidence does not disclose a price list, contract structure, realized revenue, ARR, or customer-level conversion. The careers page shows that the company is building both core systems and product surfaces, including APIs, backend services, and frontend workflows, which is consistent with a product-led or enterprise software motion but does not prove a particular revenue model. A third-party funding profile describes AI-powered code completion and generation and reports four funding rounds, yet that profile is a directory-style source rather than audited financial disclosure. The appropriate underwriting stance is therefore to separate capability from monetization - the product may support usage-based inference, subscriptions, enterprise contracts, or a mixture, but the current public record does not identify the unit actually sold or the realized price per unit. List prices from competing platforms can frame infrastructure economics, not Magic's revenue.[CI001, CI002, CI003, CI004]

4.2 Cost structure and margin drivers

The cost base is likely shaped by model training, inference, data acquisition, engineering, and specialized compute, but only some of those drivers are directly evidenced. Magic's recruiting copy explicitly calls out internet-scale data acquisition, long-horizon post-training loops, production reliability, and trade-offs between quality, performance, and cost. The Decoder reports that Magic was building supercomputers with Google Cloud and Nvidia and says the company raised $320 million in 2024, while Nvidia describes the GB200 NVL72 as a 72-GPU liquid-cooled rack designed for large-model training and inference. These sources establish technical and infrastructure context, not Magic's actual utilization, depreciation, cloud bill, or gross margin. Google Cloud's published token prices illustrate how model-serving costs can vary by input, output, caching, and endpoint, reinforcing that realized margin requires Magic's workload mix and commercial terms. No public source in this pool provides cash burn, gross margin, or working-capital data.[CI005, CI006, CI007, CI008, CI009]

4.3 Traction, financing dependency, and capital adequacy

The evidence supports a large historical financing headline but not cash available today. The assigned funding directory reports $466 million raised over four rounds and a $1.5 billion valuation tied to a $321 million Series C; The Decoder independently reports $465 million total funding and a $320 million 2024 raise. Those figures are close but not identical, so the chapter preserves the discrepancy rather than treating either as an audited total. The New Stack adds an adverse diligence signal - it says Magic announced a 100-million-token model in 2024, raised over $500 million on that strength, and had no public evidence of LTM-2-mini use outside Magic as of early 2026. That observation does not prove a failed commercialization effort, but it weakens confidence that technical differentiation has translated into repeatable customer revenue. Capital adequacy remains unresolved because there is no current cash balance, burn rate, runway, debt schedule, or next-round trigger in the retained pool.[CI010, CI011, CI012, CI013]

4.4 Financial verdict and diligence blockers

Magic's financial profile is best characterized as well-funded historically but opaque operationally. The company has credible evidence of a high-cost frontier-model build - its own hiring materials describe distributed pre-training pipelines and production systems, while external reporting describes long-context performance claims, supercomputer infrastructure, and a large 2024 financing. That combination can create meaningful technical option value, but it also creates financing dependency if inference, training, and data costs scale faster than paid usage. A financial underwriting decision should not convert the $1.5 billion reported valuation into a current valuation conclusion, or convert total funding into cash on hand. The immediate diligence requests are current revenue and ARR by product, realized price and gross margin by workload, customer concentration and renewal, monthly net burn, cash and liquid investments, debt or committed cloud obligations, and the financing milestone that would trigger another round. Until those are supplied, the financial verdict remains research-more with high sensitivity to commercialization evidence.[CI014, CI015, CI016]

4.5 Exhibits

5.1 Product and customer workflow

Magic is described as building AI-driven tools for software engineers to write, review, debug, and plan code changes. The workflow is closer to an automated pair programmer than a single autocomplete feature: the intended system ingests a project’s code, documentation, and libraries, then uses that broader context to generate or modify code. Public reporting also says the tools were not yet for sale in August 2024, so the commercially delivered SKU, deployment model, support process, and realized customer benefit are not established by this evidence set. The product thesis is strategically coherent for large repositories, but the key diligence question is whether long context improves verified task completion enough to offset inference cost and operational complexity.[CE001, CE002, CE003, CE004]

5.2 Model architecture and evaluation

The central technical differentiation is the Long-term Memory Network, or LTM, with LTM-2-mini reported to support a 100 million-token context window—roughly 10 million lines of code or 750 novels. The Decoder reports that Magic created HashHop to test retrieval over random, non-compressible hashes and multi-step jumps through context, rather than relying only on semantic “needle” retrieval. The same report attributes a roughly 1,000-times efficiency comparison against the attention mechanism of Llama 3.1 405B to Magic, while noting that these are company claims. SWE-bench is a useful external task definition, but the retained SWE-bench repository documents the benchmark and reproducible Docker harness, not a Magic score. Accordingly, the evidence supports a differentiated evaluation approach, not a verified production-quality or independent benchmark lead.[CE005, CE006, CE007, CE008]

5.3 Operating dependencies, maturity, and trust

Magic's disclosed scaling plan depends on specialized compute: TechCrunch reports a Google Cloud partnership for Magic-G4 with H100 GPUs and Magic G5 with next-generation Blackwell chips, while The Decoder describes a larger LTM-2 model and supercomputer work with Google Cloud and Nvidia. This creates a credible path to training and serving capability, but also concentrates execution on capital access, cloud capacity, accelerator supply, and software efficiency. The available sources do not establish customer-facing security architecture, data retention defaults, tenant isolation, privacy controls, copyright provenance, reliability SLAs, incident history, or certification status. OpenAI's Codex description illustrates the level of isolated execution, test evidence, and manual review expected in this category; it is a competitor reference, not evidence that Magic provides those controls. Product maturity should therefore be treated as promising but unverified outside the company's reported demonstrations.[CE009, CE010, CE011, CE012]

5.4 Exhibits

6.1 Buyer, user, and use-case fit

Magic presents an autonomous software engineer rather than a conventional code-completion utility. The available company-profile material describes natural-language planning, writing, review, debugging, and management of large code changes, while also framing the product as an “AI colleague” for complex engineering work. That positioning implies a primary user set of software engineers and engineering managers, with an economic buyer likely in a technology organization that can value throughput, migration speed, or reduced technical debt. The evidence does not identify a paying customer, a contract structure, a deployment boundary, or whether a human engineer must approve every production change. Accordingly, buyer fit is plausible from the workflow described, but payer, procurement owner, and realized business value remain unverified.[CU001, CU002, CU003]

6.2 Adoption proof and evidence quality

Magic's public customer proof is materially thinner than the category benchmarks in the assigned pool. AIHub describes the company and its models but does not name a customer deployment or outcome, and Startuply explicitly says that commercial details and named customers are not yet public. By contrast, competitor-owned pages provide long lists of named organizations and quantified workflow outcomes, while a Cognition engineering post describes a concrete internal deployment across web, Slack, Linear, CLI, and API interfaces. These comparators do not prove Magic's adoption; they establish the level of evidence a buyer or investor could expect from a mature coding-agent go-to-market motion. The key conclusion is therefore not that Magic has no customers, but that public evidence cannot distinguish pilots, evaluations, internal use, and production accounts.[CU004, CU005, CU006, CU007]

6.3 Retention, expansion, and concentration

No assigned source reports Magic's net revenue retention, gross retention, churn, renewal rate, contract length, cohort behavior, customer count, or concentration. The absence of those disclosures prevents a durability judgment and also blocks a reliable land-and-expand analysis. A plausible expansion path would move from a bounded engineering workflow such as repository understanding or code migration toward broader agent access and integrations, but that is an analytical hypothesis rather than observed Magic behavior. Competitor evidence suggests that expansion can be measured through organization-wide rollout, repeated workflows, or quantified engineering output, yet those metrics remain competitor-reported and cannot be transferred to Magic. Commercial diligence should therefore prioritize cohort-level usage and renewal evidence rather than additional logo collection.[CU008, CU009, CU010]

6.4 Investment implications and diligence priorities

The customer case is investable only as an open proof question, not as demonstrated traction. Magic's long-context and autonomous-engineering positioning may create a differentiated wedge for large codebases, but customer value depends on reliability, security review, workflow integration, and repeat use. The assigned benchmark material shows that named customers often disclose the use case and an outcome, such as migration acceleration, incident triage, or organization-wide adoption; Magic's retained evidence does not reach that standard. The next decision should be driven by verifiable customer artifacts: a segmented account list, production-versus-pilot status, usage and renewal cohorts, realized savings or output lift, procurement blockers, and the share of revenue or usage represented by the largest accounts. Until those are available, customer confidence should remain low.[CU011, CU012, CU013]

6.5 Exhibits

7.1 Risk ranking and thesis-break conditions

Magic's risk profile is dominated by commercialization and infrastructure execution. TechCrunch reported in August 2024 that Magic's tools were “not yet for sale,” that the company had around two dozen people and “no revenue to speak of”; the assigned company and independent profiles also describe a research-first strategy and very large compute ambitions. Those facts support a high execution and financing-risk assessment, but they do not establish cash burn, runway, or insolvency. The practical thesis break is a failure to convert the long-context research advantage into a sellable product with repeatable production use before the compute and hiring burden outruns financing capacity. A second thesis break is an inability to document customer-facing security, privacy, authorization, and reliability controls suitable for enterprise code access. The monitoring plan is to obtain monthly cash runway, launch and deployment milestones, production customer references, incident metrics, and an enterprise control matrix; absent that evidence, residual exposure remains high rather than “managed.”[CR001, CR002, CR003, CR004, CR005]

7.2 Legal, regulatory, and data-governance exposure

The assigned evidence does not show a Magic-specific lawsuit, enforcement action, license failure, or infringement finding, so the report should not convert general AI litigation or regulation into a company allegation. It does show why training-data provenance, copyright allocation, privacy controls, and deployment governance are decision-critical. The EU Artificial Intelligence Act is a binding harmonized framework whose application depends on system role and use case; NIST's framework is voluntary and intended to improve trustworthiness risk management. A court docket involving Stability AI records a complaint and later motion practice, but it is a competitor precedent rather than evidence against Magic. Magic's public materials describe frontier pretraining, reinforcement learning, ultra-long context, and code generation without providing the requested provenance, evaluation, or indemnity package. The residual legal exposure is therefore unquantified and should be treated as material until counsel reviews datasets, licenses, model cards, customer terms, and jurisdiction-specific obligations.[CR006, CR007, CR008, CR009, CR010]

7.3 Operational, security, partner, and financial dependencies

Magic's operating model creates linked dependencies: large-scale compute, cloud and accelerator availability, model quality at long context, secure handling of customer code, and enough capital to reach product-market proof. TechCrunch reported a Google Cloud partnership involving H100 and planned Blackwell systems and an ambition to scale to “tens of thousands” of GPUs; that is a disclosed plan, not evidence that the infrastructure was delivered or economically sustainable. The New Stack's 2026 comparison shows that long-context claims are contestable and that benchmark results can carry methodological caveats, while the MCP security document demonstrates how authorization proxies can create confused-deputy attacks if per-client consent and redirect controls are missing. Magic has not publicly supplied enough architecture, uptime, benchmark-reproduction, security, or unit-cost data in the assigned pool to rate these controls mature. Diligence should therefore test failure modes with production telemetry, red-team reports, cost per task, cloud commitments, and customer reference checks.[CR011, CR012, CR013, CR014, CR015, CR016]

7.4 Monitoring indicators and diligence triggers

The risk register is intentionally evidence-led. Positive evidence would include a shipping product, named production deployments, repeatable task-level quality against relevant baselines, documented security controls, and disclosed unit economics that reconcile compute commitments with gross-margin potential. Negative evidence would include delayed launches, inability to reproduce long-context results, outages or data incidents, customer restrictions on code use, adverse legal findings, or financing that increases dilution without improving commercial proof. The absence of a recorded incident in the assigned sources is not evidence that no incident occurred. Investors should require a current litigation and regulatory search, dataset and license schedule, SOC 2 or equivalent control evidence where applicable, cloud contract and GPU capacity terms, cash-burn and runway schedule, and customer-level conversion and retention data before reducing the high residual-risk assessment.[CR003, CR004, CR009, CR012, CR014, CR016]

7.5 Exhibits

8.1 Investment call and evidence boundary

The evidence supports a differentiated technical thesis but not a fully underwritten valuation. Magic states that its LTM-2-mini model was trained for 100 million-token contexts and that it has raised $515 million, including a recent $320 million investment; these are company claims rather than independently audited financial results. The same source says Magic had 23 people and was still hiring, which underscores an unusually concentrated execution model relative to the compute ambition. The assigned evidence set does not provide current revenue, ARR, gross margin, retention, customer count, price realization, or cap-table preferences. That missing denominator makes a revenue-multiple or cash-flow valuation speculative. I therefore recommend research-more, with high risk and an expensive-or-unknown entry stance until management supplies current financing terms and operating metrics. The positive case is that a useful long-context coding system could create strategic value for developers and infrastructure partners; the anti-thesis is that technical demonstrations may not translate into production demand before compute and model competition compress economics.[CV001, CV002, CV003, CV004, CV005, CV006]

8.2 Thesis, anti-thesis, and comparables

Magic's strongest valuation support is capability scarcity: its post says the company built a custom training and inference stack, reported a 100 million-token model, and planned large Google Cloud supercomputers using H100 and GB200 systems. Those statements indicate ambition and a potentially valuable technical option, but they do not prove commercial scalability. The competitive record is a warning. The New Stack notes that Magic had raised more than $500 million on its context-window strength and that, as of early 2026, there was no public evidence of LTM-2-mini being used outside Magic. This is an adverse commercialization signal, not proof that Magic has no customers. Meanwhile, Sacra estimates that Cursor reached $4 billion of annualized revenue in May 2026, while TechCrunch reported Cognition at a $25 billion pre-money valuation and $492 million annualized revenue run-rate. These are category comparables, not direct multiples for Magic: they have different products, scale, disclosure quality, and financing terms.[CV007, CV008, CV009, CV010, CV011, CV012]

8.3 Scenarios, entry discipline, and diligence

A scenario framework is more defensible than a point estimate. In the bull case, independent production usage validates the 100 million-token architecture, Magic converts strategic cloud and developer relationships into recurring revenue, and the company raises on terms that preserve meaningful common-equity upside. In the base case, technical progress continues but commercialization remains under-disclosed, so an investor tracks the company and waits for revenue, retention, and unit-economics evidence. In the bear case, competing models make long context less scarce, compute commitments create financing pressure, and a down round or preference stack absorbs much of the value implied by the prior financing. The entry rule should therefore require a verified current price, fully diluted capitalization, liquidation preferences, debt or cloud commitments, and evidence of paid production workloads. Exit readiness is not demonstrated by the assigned sources: a financing, acquisition, or IPO path remains contingent on durable usage and efficient inference. The highest-value diligence asks are current ARR and revenue recognition, cohort retention, named paid deployments, gross margin and inference cost per task, financing instruments and preferences, and independent benchmark or customer validation.[CV013, CV014, CV015, CV016]

8.4 Exhibits