In August 2023, Microsoft's $13 billion commitment to OpenAI and the subsequent $650 million acqui-hire of most of Inflection AI's team by Microsoft in March 2024 put a spotlight on just how differently AI ventures are structured. Inflection had raised $1.3 billion at a $4 billion valuation in June 2023, yet within nine months the majority of its core team — including CEO Mustafa Suleyman — departed under terms that regulators in the UK and EU scrutinized as a de facto acquisition engineered to avoid merger review. The term sheet provisions that governed investor rights, team retention, and IP ownership were at the center of that scrutiny.
A conventional venture term sheet concerns itself with three core economic questions: how much equity does the investor receive, what rights protect that equity (liquidation preference, anti-dilution), and what governance rights come alongside it (board seats, protective provisions). For a SaaS company the primary asset is recurring revenue; for a marketplace it is network effects; for a consumer app it is user data. The term sheet is designed to protect those assets.
For an AI venture the primary asset is often a combination of trained model weights, proprietary datasets, compute infrastructure access, and the researchers who built them. None of these map cleanly onto standard IP assignment clauses, and their valuation is radically harder — model weights have no GAAP book value, and a team of five researchers can represent $500 million in competitive moat or nothing at all depending on what they build next.
This asymmetry creates specific pressure points in AI term sheets that standard templates do not address. Founders who sign unmodified NVCA model documents often discover — at Series B or during an acqui-hire conversation — that they have inadvertently signed away rights they assumed they retained.
Before going deeper into AI-specific provisions, the foundational terms must be precise. Founders frequently confuse pre-money and post-money mechanics, and that confusion has measurable dollar consequences.
Coatue Management and other Stability AI investors filed suit in early 2024 alleging that CEO Emad Mostaque had misrepresented the company's financial condition and GPU contract obligations. Central to the dispute were information rights — investors claimed they were not receiving the financial disclosures the term sheet required. The case illustrates how AI-specific liabilities (compute contracts worth hundreds of millions) can make standard information rights provisions inadequate if they don't specifically require disclosure of infrastructure commitments.
A term sheet is typically two to five pages and non-binding except for exclusivity (no-shop) and confidentiality clauses. It is organized into three sections: economic terms, control terms, and other terms. Understanding what belongs where is essential before evaluating AI-specific modifications.
| Section | Key Provisions | AI-Specific Pressure Point |
|---|---|---|
| Economic Terms | Valuation, investment amount, option pool, liquidation preference, anti-dilution | Option pool shuffle erodes founder ownership before compute and researcher equity grants |
| Control Terms | Board composition, protective provisions, voting rights, drag-along | Safety covenants written as protective provisions give investors veto over model releases |
| Other Terms | No-shop, information rights, pro-rata, ROFR, co-sale | Pro-rata fights in oversubscribed AI rounds; compute contract disclosure in info rights |
The single most expensive line in most AI term sheets is not the liquidation preference — it is the option pool expansion required before closing. A $50M round at a $200M pre-money valuation sounds like 20% dilution. If investors require expanding the option pool from 10% to 20% before closing, the effective pre-money drops to ~$160M, and founders and early investors absorb the full dilution of that pool expansion. This dynamic was documented in detail in Brad Feld and Jason Mendelson's analysis of the "option pool shuffle" — and it hits AI companies particularly hard because researcher comp packages demand large pools.
Term sheets are not take-it-or-leave-it documents, even when VCs present them that way. Every provision is negotiable, and the negotiating leverage of an AI founder in 2024–2025 is historically high given competition among investors. The key discipline is identifying which terms are economically material versus which are standard boilerplate that investors expect to keep. Fighting over boilerplate signals inexperience; failing to fight over material terms signals naivety.
In AI specifically, the most material provisions — in rough order of financial consequence — are: (1) liquidation preference structure, (2) option pool size and timing, (3) anti-dilution mechanism (broad-based weighted average vs. full ratchet), (4) pro-rata rights, and (5) IP assignment and model weight ownership. The AI-specific additions — safety covenants, compute disclosure, key-scientist provisions — matter for governance but are rarely the primary value driver.
You are evaluating a Series A term sheet for your AI infrastructure startup. The advisor can walk you through any provision, explain AI-specific implications, and help you identify which terms to prioritize in negotiation. Ask about specific clauses, the option pool shuffle, liquidation preferences, or AI-specific additions you've read about.
In 2023, Anthropic raised two major rounds: a $450 million Series C in May and then secured commitments totaling $4 billion from Google and $1.25 billion from Amazon — part of a broader $7.3 billion raised over the year at valuations peaking around $18 billion. Anthropic's annualized revenue at the time of the Google commitment was estimated at under $100 million. By conventional SaaS multiples of 10–15x revenue, the company was worth $1–1.5 billion, not $18 billion. The gap — roughly 12–18x — represents the AI premium: the valuation that investors assign to competitive positioning, safety research moat, and potential market capture in a winner-take-most dynamic.
But Google and Amazon did not simply write equity checks. Both deals were structured as combinations of equity investment and cloud compute commitments — Amazon's $4 billion included a requirement that Anthropic use AWS infrastructure, and Google's investment was bundled with Google Cloud credits. This structure fundamentally changes the term sheet: compute commitments are both an asset (free infrastructure) and a constraint (vendor lock-in), and the term sheet must account for both.
For revenue-generating AI companies, investors use a blended approach combining forward revenue multiples (projecting 24–36 months out), comparable transactions, and qualitative moat assessments. For pre-revenue or early-revenue frontier AI labs, the methodology shifts entirely to option value — the probability-weighted value of capturing a share of an enormous future market.
The standard formulation: if the total addressable market for general AI capabilities is $1 trillion (a commonly cited figure from analysts at Goldman Sachs and Morgan Stanley in 2023–2024), and the company has a plausible 5% probability of capturing 10% of that market, the risk-adjusted value is $5 billion. At a discount rate appropriate for venture-stage risk, this can support a $1–2 billion valuation today. Apply a "strategic scarcity premium" for the small number of credible frontier labs, and the math begins to justify $10–20 billion.
This is not precise finance — it is structured storytelling with numbers. The term sheet provisions that protect investors against this speculation are doing real work.
Anti-dilution protection is the term sheet mechanism that compensates investors if a future round prices lower than their entry (a "down round"). There are three mechanisms, in order of investor friendliness:
Stability AI's valuation fell dramatically from its $1 billion peak (October 2022 funding at that valuation) as the company struggled financially through 2023–2024. While specific down round terms were not publicly disclosed, the company's fundraising difficulties illustrate precisely the scenario anti-dilution provisions address: an AI company raised at a premium valuation based on narrative, then failed to convert that narrative into revenue growth fast enough to sustain the valuation in subsequent rounds. Founders at companies with full-ratchet anti-dilution would face near-total equity wipeout in such a scenario.
One of the most consequential — and underappreciated — valuation-related provisions in AI term sheets is the compute commitment structure that hyperscalers use when investing. When Microsoft invested in OpenAI, when Google invested in Anthropic, and when Amazon invested in Anthropic, the deals bundled equity investment with cloud compute credits or preferential pricing. This creates a hybrid structure that traditional term sheet analysis does not capture.
From the AI company's perspective: $4 billion in investment that includes $2 billion in AWS credits at market rates is economically equivalent to $2 billion in cash plus a multi-year infrastructure contract. The term sheet must specify: Are compute credits counted as part of the investment for valuation purposes? Do they dilute existing shareholders as much as cash? What happens to the credit obligation if the company changes cloud providers or is acquired?
| Deal | Stated Amount | Structure | Key Term Sheet Issue |
|---|---|---|---|
| Microsoft → OpenAI (2023) | $10B (part of cumulative $13B) | Cash + Azure compute credits | Compute credits structured as revenue-sharing arrangement; complex liquidation waterfall |
| Google → Anthropic (2023) | $300M–$400M equity tranche | Equity + Google Cloud credits | Cloud commitment creates vendor dependency; exit rights if cloud SLA not met |
| Amazon → Anthropic (2023–2024) | $4B committed | Equity + AWS as primary cloud | AWS exclusivity provisions; Trainium/Inferentia chip requirements |
| Inflection AI (June 2023) | $1.3B | Pure equity | Standard preferred; acqui-hire nine months later exposed key-person provision gaps |
For AI companies raising at high valuations, the liquidation preference stack becomes a critical factor in any M&A outcome. Consider an AI company that has raised three rounds: seed at $10M at $30M pre-money, Series A at $30M at $90M pre-money, Series B at $100M at $300M pre-money. The liquidation preference stack — assuming 1x non-participating preferred throughout — is $140M. This means the first $140M of any acquisition goes to preferred shareholders before common (founders, employees) see anything.
If the company is acquired for $200M — a seemingly positive outcome — founders receive only $60M split among all common holders, while investors received 2.3x their Series B investment and only 1x their earlier rounds. At acquisition prices below $140M, founders and employees receive nothing. This is not a theoretical concern: the acqui-hire market for AI teams frequently produces transaction values in the $100–300M range (talent value, not enterprise value), and the liquidation preference stack determines who captures that value.
The most founder-friendly modification to liquidation preferences in AI deals is a "carve-out" for key employee retention in acqui-hire scenarios — a provision that allocates a specified percentage (typically 5–15%) of any acquisition proceeds directly to the engineering and research team regardless of preference stack order. This provision is increasingly standard in frontier AI term sheets because investors recognize that acquiring parties are primarily buying talent, and talent that receives no acquisition proceeds has incentive to leave immediately post-acquisition. The Google/DeepMind acquisition in 2014 included retention provisions of this type.
The advisor can help you model valuation scenarios, evaluate anti-dilution mechanisms, and think through the economic consequences of compute-for-equity structures. Bring hypothetical deal parameters or questions about Anthropic, OpenAI, or Stability AI's funding structures.
On November 17, 2023, OpenAI's six-person board fired CEO Sam Altman, citing a loss of confidence in his candor. Within five days Altman was reinstated, three board members resigned, and the governance structure was reformed. The episode is the most scrutinized AI governance event in the industry's history — and it occurred precisely because OpenAI's structure deviates fundamentally from standard VC-backed governance.
OpenAI is a capped-profit LLC controlled by a nonprofit board, not a standard Delaware C-Corp with investor-controlled preferred shares. Microsoft, despite committing $13 billion, had no board seat and no standard protective provisions. The board that fired Altman was selected for mission representation, not capital representation. This structure — which OpenAI chose deliberately to prevent capital from controlling AI development decisions — meant that when the board acted, no investor had a legal mechanism to stop them.
In a standard VC-backed company, board composition is negotiated explicitly in the term sheet. The typical progression: at seed stage, founders hold majority control (2-1 or 3-1 founder to investor seats). At Series A, the lead investor takes one seat, creating a 2-1 or 2-2 structure with an independent tiebreaker. At Series B, investors often seek equal representation (2-2-1 or 3-3-1 with independent chair). By Series C, investors frequently have board majority.
For AI companies, this trajectory carries specific risks. An AI company's decisions about which capabilities to develop, what safety standards to apply, and when to release models are both technical and ethical decisions where investor-majority boards may prioritize commercial velocity over risk management. This tension is not hypothetical — it is the explicit reason OpenAI chose its non-standard structure, and it is why Anthropic was structured as a public benefit corporation.
| Stage | Typical Board Structure | AI-Specific Concern | Founder Protective Approach |
|---|---|---|---|
| Seed | 3 founders, 0–1 investor observers | Few; founders control all decisions | Avoid giving full board seats at seed |
| Series A | 2 founders, 1 investor, 1 independent | First external check on model release decisions | Negotiate independent seat appointment rights |
| Series B | 2 founders, 2 investors, 1 independent | Investor parity; safety covenants get teeth | Ensure founder protective provisions survive |
| Series C+ | Negotiated; often investor-majority | Commercial pressure may override safety judgment | Dual-class shares; mission-lock provisions |
Protective provisions are a list of actions that require preferred shareholder approval (by vote or consent) regardless of board composition. They are the primary mechanism through which investors exercise governance rights without needing a board majority. Standard protective provisions include: issuing new shares, amending the certificate of incorporation, selling the company, declaring dividends, and taking on debt above a threshold.
AI-specific protective provisions added since 2022 have included requirements for: investor consent before releasing models above a specified capability threshold; investor consent before entering data licensing agreements covering training data; and investor consent before making regulatory submissions regarding AI safety classification. These provisions give investors genuine power over core technical decisions.
In 2023, multiple AI companies including Cohere and Mistral reported that investors were negotiating protective provisions requiring board notification (and in some cases consent) before entering enterprise agreements with defense or intelligence agencies. This reflects a pattern documented by Bloomberg Law in mid-2023: investors concerned about regulatory risk and reputational exposure are using protective provisions to establish an approval layer over commercial decisions that would be entirely unremarkable in a SaaS company.
A drag-along provision allows a specified majority of shareholders (often a combination of preferred majority and common majority) to force all other shareholders to approve a sale of the company. This is designed to prevent a small minority from blocking an otherwise-approved transaction. In AI companies, drag-along provisions interact in complex ways with mission-preservation goals.
Consider: an AI safety-focused company receives an acquisition offer from a company with a significantly different approach to AI development. The board wants to accept; a group of early employees holding common stock wants to block the sale on mission grounds. A well-drafted drag-along provision would override the employee objections. Poorly drafted drag-along provisions — those that don't specify required thresholds clearly — have been litigated in Delaware courts in non-AI contexts and the same vulnerability exists in AI deals.
Several prominent AI companies have adopted or considered dual-class share structures — where founders hold high-vote shares (typically 10:1 or 20:1 voting ratio) that preserve control regardless of equity dilution. Google (2004 IPO), Facebook/Meta (2012 IPO), and Snap (2017 IPO) used this mechanism. In the AI context, Elon Musk's xAI and several private AI infrastructure companies have adopted dual-class structures in their early-stage term sheets.
The investor negotiating tension is significant: sophisticated institutional investors often prohibit or penalize dual-class structures because they eliminate the governance rights that make equity investment valuable. The Institutional Shareholder Services and Glass Lewis both recommend against dual-class structures at IPO. Yet for AI founders who believe mission-critical decisions must not be subject to commercial investor override, dual-class structures may be the only mechanism that reliably preserves founder control through late-stage funding and eventual public market listing.
The OpenAI board crisis did not destroy the company — Altman was reinstated within five days and Microsoft emerged with an observer seat. But it revealed a structural fragility: a governance architecture designed to prevent commercial capture of AI decision-making can also prevent the efficient resolution of leadership conflicts. For founders designing governance: the goal is not to maximize control or maximize investor protection, but to create a governance structure that can make fast, legitimate decisions under pressure. That requires explicit term sheet provisions about board decision thresholds, emergency governance procedures, and what happens when the board is deadlocked.
Work through governance design scenarios for your AI company. The advisor can help you evaluate board composition options, draft protective provision language, analyze drag-along risks, and think through how OpenAI's governance lessons apply to your specific situation.
Mistral AI was founded in April 2023 by former DeepMind and Meta researchers. By June 2023 — just weeks after founding — the company had raised a €105 million ($113M) seed round at a €240 million pre-money valuation, with investors including Lightspeed, Index Ventures, and Xavier Niel. The round was oversubscribed with multiple competing term sheets. In December 2023, Mistral raised an additional €385 million Series A at a €2 billion valuation — approximately 8x the seed valuation in six months, with no shipped product and minimal revenue.
The negotiating dynamics in both rounds were unusual: Mistral's founders — having deep credibility from Llama and Gemini work — were able to run a competitive auction process that generated multiple competing bids, drove valuation significantly above initial offers, and resulted in founder-favorable governance terms including no safety-covenant protective provisions in the standard form. This outcome was not accidental — it was the product of deliberate process management.
The fundamental principle of venture fundraising negotiation is that leverage comes from alternatives. A founder negotiating with a single interested investor has almost no leverage. A founder with three competing term sheets has significant leverage on every economic and governance provision. Building that competitive dynamic is the primary goal of a fundraising process, and it is substantially more achievable for AI companies in 2024 than for almost any other category of startup.
The mechanics of a competitive process: (1) identify and approach all plausible investors simultaneously, not sequentially; (2) create artificial time pressure by setting a decision deadline that aligns with multiple investors' internal processes; (3) use anchoring — share the highest offer (or a credible number) with other interested investors to pull their bids upward; (4) leverage strategic investor interest (hyperscalers) to pressure pure financial investors, and vice versa.
Founders often spend negotiating capital on visible but economically minor issues (valuation) while ceding invisible but material provisions (option pool size, liquidation preference structure, anti-dilution mechanism). The effective negotiating approach priorities provisions by their expected financial impact across likely exit scenarios.
| Provision | Fight Hard? | Expected Financial Impact | AI-Specific Note |
|---|---|---|---|
| Option Pool Size / Timing | Yes — highest priority | Direct equity dilution, often 5–15% of founder ownership | AI researcher comp requires large pools; fight pre vs. post-money expansion |
| Liquidation Preference Structure | Yes — 1x non-participating is the line | Can determine whether founders receive anything in an acqui-hire | Acqui-hire is the modal AI exit; preference structure determines outcome |
| Anti-Dilution Mechanism | Yes — broad-based weighted average only | Potentially total equity wipeout in down round | High AI valuations make down round risk real; full ratchet unacceptable |
| Board Composition | Negotiate carefully, not aggressively | Moderate — governance control has indirect economic value | Independent seat appointment is the key leverage point |
| Pro-Rata Rights | Grant standard, fight super pro-rata | Moderate — determines future round dynamics | Super pro-rata in AI rounds can lock out better investors in later rounds |
| Safety Covenants | Negotiate scope carefully | Low economic impact, high operational impact | Define "material release" precisely; broad language creates operational risk |
| Information Rights | Grant standard; add compute disclosure | Low direct economic impact | Ensure compute contract disclosure is included; investor suit risk if omitted |
| Valuation | Important but often overweighted | High visibility, moderate actual impact if other terms are right | Inflated AI valuations create down-round risk — be careful what you win |
The no-shop clause — binding even in an otherwise non-binding term sheet — prevents founders from soliciting competing offers for a specified period (typically 30–60 days). In AI markets where deal velocity is high and investor interest is intense, the no-shop period is a significant concession. Once signed, founders lose the competitive leverage that generated the term sheet in the first place.
Negotiating strategy on no-shop: (1) shorten the period to 30 days maximum; (2) add a right to continue discussions with investors already in conversation before signing; (3) add a "fiduciary out" allowing termination if a substantially better offer arrives; (4) add a reciprocal exclusivity — the investor agrees not to invest in directly competing companies during the no-shop period. The fourth provision is rarely granted but signals negotiating sophistication.
According to reporting by The Information and Bloomberg in 2023, multiple AI companies including Character AI, Cohere, and Adept AI received simultaneous term sheets from competing investors including Andreessen Horowitz, Sequoia, Khosla Ventures, and Google Ventures. In several cases, the competitive dynamic drove valuations 40–60% above initial offers. Character AI's August 2023 round ($150M at $1B valuation from a16z) followed a competitive process in which Google reportedly offered a lower valuation. The tactical lesson: creating genuine competition among investors is worth more than any single negotiating tactic on individual provisions.
The period between term sheet signing and closing — typically 30–90 days for a venture round — is particularly fraught for AI companies because of factors that don't apply to most other sectors. Regulatory review risk: several AI deals in 2023–2024 attracted antitrust scrutiny (Microsoft/OpenAI, Google/Anthropic) that delayed or complicated closing. Key-person departure risk: AI research teams are highly mobile, and a key researcher departure between term sheet and close can trigger investor renegotiation or withdrawal. Model capability announcements: a competitor's major model release between term sheet and close can reset market valuation expectations and trigger investor attempts to renegotiate.
Term sheet provisions addressing closing risk: (1) a "bring-down" certification from founders at closing confirming no material adverse change; (2) a material adverse change (MAC) clause that specifies what events give investors the right to renegotiate or walk away; (3) an explicit definition of "material adverse change" that excludes market and industry conditions (standard in M&A but less common in early-stage term sheets). AI-specific MAC definitions should address team departures, regulatory actions, and compute access disruption.
The term sheet is the beginning of an investor relationship, not the end of negotiation. The definitive documents — stock purchase agreement, investor rights agreement, voting agreement, right of first refusal agreement — are where the term sheet provisions are converted into binding legal language. This conversion introduces ambiguity that founders often allow investors' lawyers to resolve in investors' favor.
The three most common post-signing errors: (1) accepting investor-drafted definitive documents without parallel founder counsel review; (2) allowing vague term sheet language to be resolved toward investor interpretation in definitive docs (e.g., "reasonable best efforts" vs. "best efforts" in safety covenants); (3) failing to negotiate the sequence of closings in tranched rounds — an investor who commits $20M in two $10M tranches holds leverage over the second tranche that can be used to renegotiate terms after you've spent the first tranche.
The best documented advice on AI venture term sheet negotiation comes from founders who have done multiple rounds: the most expensive mistake is optimizing for valuation at the expense of structure. A $500M valuation with 2x participating preferred and a large option pool requirement is economically worse for founders than a $400M valuation with 1x non-participating preferred and a reasonable pool. Investors know this arithmetic; many founders, especially first-time founders in their first institutional round, do not. The purpose of this module is to ensure you are not in that category.
The advisor plays the role of a strategic negotiation coach. Describe your deal situation — multiple term sheets, a specific investor's demands, a preference structure dispute — and work through the negotiating strategy, priority stack, and likely outcomes. You can also model acquisition scenarios to test whether specific term combinations work in your favor.