Home › Blog › Founding an AGI Startup, Part 2
Founding an AGI Startup · Part 2 of 11 📜

Legal Foundations — Entity Structure, IP, and Jurisdiction

Part 1 laid out three lanes instead of picking one — this article, and every one after it, goes three-lanes-deep instead of forking. Here's how incorporation, IP ownership, and cap table structure actually differ depending on whether you're building a scale-first lab, an applied product, or a research bet.

FL
FrontierAGI Team
Startups Legal Simulation
What this series is. Part 2 of a standalone founder-playbook simulation. Per reader direction, this series now runs all three strategic lanes (frontier scale-first, applied/agentic layer, narrow research bet) in parallel through every step, rather than choosing one — so each article below is organized as three explicit lane sections. This is educational simulation content, not legal advice; consult an actual startup attorney before incorporating anything.

1. 🧭 Why Legal Structure Comes Before Anything Else

Every later step in this series — who owns equity, how you raise money, whether a research breakthrough belongs to the company or to an individual researcher who leaves — traces back to decisions made in the first weeks of incorporation. Fixing a bad entity structure later is expensive, slow, and sometimes impossible without giving up leverage in a fundraise. This article covers the four decisions that matter most: what kind of entity to form, who owns the IP, where to incorporate, and how the cap table starts out — for all three lanes, since the right answer genuinely differs by lane in ways that surprise most first-time founders.

Delaware C-Corp
The default entity for any lane expecting US venture capital — the starting point, not the whole decision
PBC
Public Benefit Corporation — now the standard for-profit structure for AGI-mission labs, used by Anthropic and xAI
Trust + Board
Anthropic's added layer — a Long-Term Benefit Trust with disinterested trustees holding voting stock
1 Bad Clause
The realistic number of IP-assignment mistakes needed to blow up a future acquisition or fundraise

2. 🏛️ Entity Structure

🔴 Lane 1: Frontier Scale-First

A standard Delaware C-Corp is the baseline, but the mission-driven framing that helps attract long-horizon capital and top research talent increasingly points toward a Public Benefit Corporation (PBC) — a for-profit structure with a legal mandate to pursue a stated mission alongside shareholder returns. This has become the standard structure for AGI-labeled labs specifically: both Anthropic and xAI operate as PBCs. Anthropic went further, layering a Long-Term Benefit Trust on top — five disinterested trustees holding voting stock that can influence board composition, explicitly designed to keep mission prioritized over pure shareholder return even as outside investment scales. The tradeoff: a PBC's expanded fiduciary duty (considering stakeholders beyond shareholders) can complicate a straightforward acquisition or maximize-return exit later, and it requires real governance discipline to be more than words on an incorporation filing.

OpenAI's experience is the cautionary tale worth knowing here: its original nonprofit-plus-capped-profit structure made sense when it looked like there might be one dominant AGI effort, but became a genuine legal and governance liability once the field turned out to have many well-funded competitors — OpenAI has since been restructuring away from that original model. Lesson: don't build your entity structure around an assumption about the competitive landscape that might not hold in three years.

🔵 Lane 2: Applied / Agentic Layer

A plain Delaware C-Corp is almost always the right answer, full stop. Standard venture investors expect it, it's well understood by every lawyer and accountant you'll work with, and there's no mission-framing benefit to a PBC that outweighs the added governance complexity for a product company competing on execution speed, not a stated AGI-safety mission. Unless your specific product has a genuine public-benefit angle investors and customers would actually value (e.g., an AI-safety-adjacent auditing tool), keep this simple — a plain-vanilla C-Corp is one less thing to explain in every future fundraising conversation.

⚪ Lane 3: Narrow Research Bet

This is where the PBC-plus-trust model matters most, because the entire pitch to investors is "trust us to spend years on unproven research" — a structure that visibly, legally constrains you to the stated mission is itself a credibility signal to patient, thesis-driven capital, not just a compliance formality. SSI's own structure follows this logic; a research-bet company with a plain C-Corp and no mission-lock mechanism has a harder time explaining why investors should tolerate years of zero revenue without the safety net of "this is legally what we said we'd do."

3. 🔐 IP Ownership — The Mistake That Costs You Later

🔴 Lane 1: Frontier Scale-First

Every researcher and engineer needs an ironclad IP assignment agreement from day one — signed before they touch any code or research, not retroactively. A frontier lab's entire value is the model weights and training methodology; a single ambiguous IP clause with an early researcher who later becomes a competitor's key hire (a very real pattern in this industry, per the founder-alumni thread covered elsewhere in this series) can create years of legal exposure. Data licensing agreements for training data also need airtight provenance documentation from the start — retroactively proving you had rights to use training data is far harder than establishing it upfront.

🔵 Lane 2: Applied / Agentic Layer

Your actual IP is usually the product, the workflow integration, and any proprietary fine-tuning data or prompt/orchestration logic you develop — not a foundation model. Standard invention-assignment agreements for all employees and contractors are still essential (a surprising number of early-stage disputes come from a contractor who built a "quick prototype" without a signed agreement), but the bigger IP risk in this lane is usually the reverse: over-relying on a frontier lab's API under terms that could change, rather than under-protecting your own output.

⚪ Lane 3: Narrow Research Bet

IP assignment here needs to account for academic-style researcher expectations — many top researchers you'll want to hire come from academia or open-research labs where publication norms and IP ownership work differently than standard corporate employment. Being explicit early about what can be published, what stays proprietary, and how that balance might shift as the research matures avoids a painful renegotiation later when a genuine breakthrough happens and the stakes suddenly get much higher.

4. 🌍 Jurisdiction — Where You Incorporate and Why

🇺🇸 Delaware (Default for All Three Lanes)
Well-established corporate case law, investor familiarity, and predictable governance rules make Delaware the default for any lane expecting US venture or growth capital — this is true regardless of which lane you're in, and deviating from it needs a specific, strong reason.
🇪🇺 EU / Sovereign Alternative
Relevant mainly for Lane 1 founders explicitly pursuing a sovereign-compute or non-US-dependent narrative (echoing Mistral's European positioning) — a real strategic choice, not just a tax or convenience decision, since it shapes your entire investor base and regulatory environment.
🏝️ Offshore Holding Structures
Occasionally relevant for later-stage tax optimization, but a genuine trap for an early-stage company in any lane — adds complexity and investor skepticism at exactly the stage when you need fundraising to be simple and legible.

4b. 🇮🇳 Building From India — A Separate, Real Path

Everything above assumes a Delaware-first default, which is correct for a founder targeting primarily US venture capital. But a growing share of applied-layer and even some research-bet founders are building from India first, and the legal path there has matured enough by 2026 that it deserves its own treatment rather than a footnote.

🔵 Lane 2: Applied / Agentic Layer — The Realistic India-First Path

A Private Limited Company under the Companies Act, 2013 is the standard entity — minimum two directors, two shareholders, a registered office address, with MCA (Ministry of Corporate Affairs) approval typically taking 7-10 working days in normal conditions. This is the direct Indian equivalent of the Delaware C-Corp default in Section 2, and it's the right starting point if your investor base, customers, and team are primarily India-based.

DPIIT recognition (Department for Promotion of Industry and Internal Trade, under the Startup India scheme) is worth pursuing early and is not automatic — it requires meeting eligibility criteria (entity age under 10 years, or 20 years for Deep Tech classification; annual turnover under ₹200 crore, or ₹300 crore for Deep Tech) but unlocks real, concrete benefits: an 80% patent filing fee reduction, a three-year income tax holiday under Section 80-IAC, angel tax exemption, self-certification under nine labour and three environmental laws, access to the Startup India Seed Fund Scheme, and preferential access to government e-marketplace (GeM) procurement. For an AI company specifically, the "Deep Tech" classification's extended 20-year age window and higher turnover ceiling is worth confirming eligibility for explicitly, since it changes your compliance runway substantially.

An AI-specific compliance layer sits on top of the standard registration: DPDP Act (Digital Personal Data Protection Act) compliance for any personal-data handling, plus standard IP protection (trademark and provisional patent filings) — sector-specific add-ons to an otherwise standard registration flow, not a fundamentally different process.

🔴 / ⚪ Lanes 1 & 3 — The Flip Structure Question

If you're building a compute-intensive Lane 1 company or a research-bet Lane 3 company and need access to US-scale venture and growth capital, you'll almost certainly face the flip structure decision: incorporating a Delaware C-Corp as the parent, with your existing (or newly formed) Indian entity converted into a wholly-owned subsidiary. This is a genuinely consequential decision, not a formality — it permanently changes the tax profile of every founder and investor on the cap table, your compliance obligations across two jurisdictions simultaneously, your ESOP structure and its US tax treatment (ISOs, 83(b) elections, QSBS eligibility), and your eventual exit mechanics. It's also a proven, well-trodden path — Freshworks, Chargebee, and BrowserStack are among the many Indian-founded companies that flipped to a Delaware parent specifically to access US VC financing, standard stock option frameworks, and US enterprise procurement processes.

The counter-trend worth knowing about: a "reverse flip" — moving the parent entity back to India — has become increasingly common among more mature companies (Razorpay, Pine Labs, and others have pursued this) as Indian public markets matured, angel tax was abolished, and the domestic VC ecosystem deepened enough that US domicile stopped being a prerequisite for a strong exit. GIFT City IFSC (India's International Financial Services Centre, in Ahmedabad) has emerged specifically as a structure competing with a full Delaware flip: it offers a 10-year tax holiday, exemptions on capital gains, securities transaction tax, and stamp duty under Section 80LA, plus FEMA non-resident status — global capital access without full US domicile. For a Lane 1 or Lane 3 founder in 2026, GIFT City is now a real third option worth evaluating alongside "flip to Delaware" and "stay India-only," not just a niche alternative.

The flip-vs-stay decision isn't really about where to file paperwork — it's about which capital pool, which talent pool, and which exit market you're building toward for the next decade.

5. 📊 Cap Table Basics — Founders, Early Employees, and Advisors

🔴 Lane 1: Frontier Scale-First

Expect founder equity to dilute fast and early — a $1B+ capital requirement means multiple large funding rounds close together, and founders in this lane typically retain meaningfully less ownership by Series B than their Lane 2 counterparts, in exchange for capital access most companies never get at all. Standard four-year vesting with a one-year cliff still applies, but option pools for research talent tend to run larger than typical Lane 2 pools, since competing for the same researchers as well-funded incumbent labs requires genuinely competitive equity offers.

🔵 Lane 2: Applied / Agentic Layer

Standard startup cap table norms apply cleanly here: founders typically retain 70-85% post-seed, standard four-year vesting with a one-year cliff, a 10-15% employee option pool. This is the lane where "standard startup playbook" advice actually transfers directly, because the funding path (seed, Series A, Series B) matches the standard venture pattern much more closely than the other two lanes.

⚪ Lane 3: Narrow Research Bet

Founder credibility substitutes for revenue traction in these fundraising conversations, which shows up directly in the cap table: a founder with strong enough personal credibility (a departed frontier-lab executive, for instance) can often retain more favorable early-round terms than the capital amount alone would suggest, precisely because investors are underwriting the person and thesis, not a product metric. This lane also sees larger research-staff option pools relative to headcount than either other lane, reflecting how much of the company's value is concentrated in a small number of senior researchers.

6. 🏛️ Case Studies: Real Structures, Real Lanes

Anthropic — PBC + Long-Term Benefit Trust
Lane 1 Structure
Public Benefit Corporation5 disinterested trusteesVoting stock held by Trust
Structured as a PBC from the outset, layered with a Long-Term Benefit Trust holding voting stock — a governance model explicitly designed to be resistant to the kind of structural transformation pressure OpenAI later faced. The lesson: building the mission-lock into the legal structure from day one, rather than retrofitting it after outside investment scales, avoided a whole category of governance dispute that OpenAI had to work through publicly and expensively.
OpenAI — The Cautionary Restructuring Tale
Lane 1/3 Structure
Nonprofit + capped-profit → restructured
OpenAI's original nonprofit-plus-capped-profit structure was designed around an assumption — that there might be one dominant AGI effort — that stopped holding as soon as a genuinely competitive multi-lab field emerged. The company has since undertaken a significant structural evolution as a result. The lesson directly relevant to this article: don't encode a competitive-landscape assumption into your legal structure that a few years of market evolution could invalidate — build in flexibility deliberately, or accept that restructuring later will be public, slow, and costly.

7. 📋 Side-by-Side: Legal Foundations by Lane

Factor🔴 Lane 1: Scale-First🔵 Lane 2: Applied Layer⚪ Lane 3: Research Bet
Entity typePBC (+ Trust layer common)Plain Delaware C-CorpPBC (+ Trust layer common)
IP priorityModel weights, training methodology, data provenanceProduct, workflow integration, orchestration logicResearch output, publication norms
Jurisdiction defaultDelaware, or sovereign alternative if EU/regional strategyDelawareDelaware
Founder dilution by Series BFast, often steepStandard venture paceOften slower — credibility substitutes for capital-driven dilution
Option pool sizeLarge, competitive with incumbent labsStandard 10-15%Large, concentrated in senior researchers
India-first pathFlip to Delaware almost always needed for compute-scale capitalPrivate Limited Company + DPIIT recognition often sufficientFlip to Delaware, or GIFT City IFSC as an emerging middle path

8. ⚠️ Risk Flags

✍️
Unsigned or Late IP Assignments
The single most common early-stage legal defect across all three lanes — a contractor or early collaborator who did real work before signing an assignment agreement creates an ownership cloud that surfaces, predictably, during due diligence for your next round.
🏛️
Over-Engineering Structure Too Early
A Lane 2 founder adopting a PBC-plus-trust structure because it "sounds more serious" adds real governance overhead with no corresponding benefit — match the structure to the actual lane, not to how the biggest labs in the news look.
🔓
Mission-Lock Without Governance Discipline
A PBC structure is only as good as the board's actual practice — OpenAI's experience shows that legal structure alone doesn't resolve mission-vs-capital tension if the underlying incentives and competitive pressure are strong enough.
🌐
Jurisdiction Decisions Made for the Wrong Reason
Choosing a non-Delaware jurisdiction for tax reasons alone, without weighing the investor-familiarity cost, is a common early mistake that creates friction in every subsequent fundraising conversation.
The legal structure you choose in week one is a structural bet about what kind of company you're going to become — get it wrong and every later step in this series gets harder, not just this one.

9. 🧪 Filing Checklist (All Three Lanes)

1
Incorporate in Delaware as a C-Corp or PBC depending on your lane's mission-framing needs, before any funding conversation, not after a term sheet arrives.
2
Get signed IP assignment agreements from every founder, employee, and contractor before they do any substantive work — no exceptions, no "we'll paper it later."
3
Set up standard four-year vesting with a one-year cliff for all founders and early employees, sized appropriately for your lane's dilution pattern (Section 5).
4
Document data/training provenance from day one if you're anywhere near Lane 1 — retroactive documentation is far harder and a real diligence risk in any future fundraise or acquisition.
5
Decide your governance philosophy explicitly — mission-locked (PBC + Trust) versus standard shareholder-primacy — before your first outside investor, not after, since it's much harder to add mission-lock provisions once outside capital already has a seat at the table.

9b. 🤔 Do You Actually Need Any of This on Day One?

Honestly: usually not yet. Everything in this article — PBC structures, Trust layers, flip decisions, GIFT City — is about when you take outside institutional money and start hiring beyond a founder or two, not about whether you're allowed to start building. Nothing above is a legal precondition for writing code, talking to customers, or even taking your first payment. What actually forces the issue is one of three specific triggers: (1) a co-founder joining, where a handshake equity split without a formal cap table and vesting agreement is a lawsuit waiting to happen; (2) an institutional investor writing a check, since no serious VC or angel fund will wire money to an unincorporated founder's personal bank account; or (3) hiring your first employee, where payroll, benefits, and liability protection start to matter in ways they don't for a solo founder. Before any of those three triggers, a lean starting point — a single-owner entity, or even operating unincorporated while validating the idea (Section 3 of Part 1) — is not just acceptable, it's the more common real-world path than founders assume from reading startup media.

🇮🇳 Zerodha — Simple Structure, No Flip, No VC
Founded in 2010 with ₹2 lakh of founder savings, Zerodha never raised external venture capital and never needed a Delaware flip, a PBC structure, or a GIFT City arrangement — a straightforward Indian private entity was sufficient for its entire growth into one of India's largest stock brokerages, precisely because it never took the institutional-capital step that makes the heavier structures in this article necessary.
🇮🇳 Zoho — Decades Without a Cap Table Fundraise
Zoho, founded in 1996, grew into a major global SaaS company entirely on customer revenue, with zero external VC funding across its history. No flip structure, no investor-driven PBC conversion, no Series A term sheet to negotiate — the entire apparatus of Sections 2-7 above simply never became necessary, because the trigger event (outside institutional capital) never happened.
🇺🇸 Mailchimp — Bootstrapped for Nearly Two Decades
Mailchimp operated as a straightforward, non-VC-backed company for close to twenty years before its acquisition — no PBC, no Trust layer, no complex multi-round cap table. It used a simple entity structure the entire time and only faced acquisition-specific legal complexity once, at the actual exit, rather than repeatedly across years of institutional fundraising rounds.

The pattern across all three: the heavier legal structures in this article are a function of the capital-raising path you choose, not a function of being a "real" company. A Lane 2 founder who successfully bootstraps to revenue, in the Zoho/Zerodha/Mailchimp pattern, can legitimately skip most of Sections 2 and 5 for years. Where this gets risky is a founder who assumes they can stay this simple while also planning to raise from institutional VCs later — investors will require you to fix an underlying entity structure, IP assignments, and cap table cleanliness retroactively before they'll invest, and doing that cleanup after early collaborators, informal advisors, or a co-founder dispute have already muddied the picture is far more expensive and slower than setting it up properly before any of that happens. The honest rule: match the legal investment to your actual capital-raising intent, decided honestly now, not to what you're doing this week.

10. 🧭 What's Next in the Series

Part 3 covers The Founding Team — co-founder selection, early hires, and equity splits, again across all three lanes: who you actually need on day one differs sharply depending on whether you're racing to a product MVP, building a research org, or trying to attract top ML talent away from incumbent labs.