Why Sequoia Heritage quietly funds private LLMs.

What Sequoia Heritage actually is

The confusion begins with the name. Sequoia Heritage is not a venture fund. It is the multi-family office arm of Sequoia Capital — the vehicle through which Sequoia manages capital for its own partners, select LPs, and a limited number of ultra-high-net-worth families who came into the relationship through Sequoia's direct investment history. Assets under management sit above $12 billion as of the end of 2023, according to two people with direct knowledge of the firm's reporting. The investor base is small, stable, and holds capital on timelines measured in decades, not quarters.

Heritage operates with meaningful autonomy from Sequoia's venture operation. It runs its own investment committee. Its portfolio construction logic differs substantially from the venture fund's: the aim is capital preservation and compounding across asset classes — public equities, private credit, real assets, and increasingly, direct positions in private technology infrastructure. The LLM commitments fall into that last category. They are not venture bets on model companies competing with OpenAI. They are infrastructure positions: compute orchestration, fine-tuning tooling, private deployment platforms, and what one person close to Heritage described as "the picks-and-shovels layer for families who want a model that is genuinely theirs."

That framing matters. Heritage is not funding foundation model labs. Its six 2023 commitments targeted the stack that sits between an open-weights base model and a production deployment inside a family office or multi-family office client. The thesis is vertical: Heritage's own clients — and prospective clients — are the end market. The fund is building, on behalf of its capital base, the infrastructure that capital base will eventually use.

The long-duration capital advantage

Venture capital has a structural problem with AI infrastructure. The best infrastructure bets — compute networks, private deployment platforms, model-serving layers — require patient capital and do not fit neatly into a ten-year fund with a two-year investment window. The return horizon on a private-LLM deployment platform, if it compounds across a client base of family offices over fifteen years, exceeds what most venture fund structures can hold. Sequoia Heritage's capital does not have that constraint. Its flagship vehicle has an indefinite horizon, with liquidity windows managed on a rolling basis rather than a fixed fund clock. That gives Heritage the ability to make commitments that venture cannot price.

Reinholt's 2021 memo — portions of which were described to us by two people who read it — made the case in blunt terms. Long-duration capital, she argued, should be the natural funder of AI infrastructure that would not yield to normal venture timelines. The analogy she used was telecommunications infrastructure in the 1990s: the companies that built the cable and the switching hardware were not venture returns in the classic sense, but they compounded at rates that dwarfed public-market alternatives for patient capital that held through the construction period. She argued that private LLM infrastructure, specifically the deployment and fine-tuning layer, would follow the same curve — slow to monetize, then durable once client relationships locked in.

The argument gained traction inside Heritage in late 2022, after the firm's technology allocation committee — chaired by senior partner Marco Feltrinelli — ran a comparative analysis of AI infrastructure return profiles against Heritage's existing private-credit book. The analysis concluded that deployment-layer AI infrastructure, bought at the right valuation and held for twelve to fifteen years, offered a risk-adjusted return superior to mid-market private credit at prevailing spreads. The six 2023 commitments followed from that conclusion.

The multi-asset allocation logic

Heritage's AI infrastructure commitments do not sit in a dedicated technology sleeve. They are classified as private equity within the firm's multi-asset framework — alongside direct co-investments in industrial companies, healthcare services, and financial infrastructure. That classification is deliberate. It signals that Heritage views private-LLM infrastructure as a business-quality investment, not a speculative technology bet. The underwriting criteria applied to the six companies reflect that posture: the team required evidence of recurring revenue, existing institutional client relationships, and a credible path to profitability within seven years. None of the six were pre-revenue at the time of commitment.

The six positions range in size from $18 million to $74 million, based on two people with knowledge of the portfolio. The largest — the $74 million commitment — went to a platform that provides private model deployment infrastructure for institutional asset managers and family offices. The smallest went to a fine-tuning tooling company whose primary client base, at the time of Heritage's commitment, was three large European multi-family offices and a sovereign wealth fund operating a private AI program. Heritage structured four of the six as equity with a preferred return; two are structured as hybrid instruments combining equity participation with a revenue-sharing component that kicks in at defined revenue thresholds.

Feltrinelli's team ran the deals. The sourcing came through two channels: Heritage's own LP network — specifically the subset of LPs who are also operating family offices running or considering private AI programs — and referrals from Sequoia Capital's venture investment team, which sees early-stage companies in the space before they are large enough for Heritage's ticket size. That referral channel is selective. Sequoia Capital's venture partners do not systematically funnel deals to Heritage; the relationship is collegial, not contractual. But the proximity gives Heritage visibility into company trajectories earlier than most multi-family office competitors, several of whom have launched comparable programs in 2023 and 2024.

Client retention: the real thesis

The investment logic is inseparable from Heritage's own client strategy. The families that Heritage manages capital for are the same families that are, or will be, the end customers of private-LLM deployment infrastructure. Heritage's portfolio companies sell to exactly the client segment that Heritage serves. That alignment is not coincidental — it is the core of the investment thesis, and it creates a dynamic that Heritage's leadership describes, in internal discussions, as "the flywheel."

The flywheel works as follows. Heritage makes a commitment to a private-LLM deployment platform. Heritage introduces that platform to two or three of its own family-office clients as a preferred vendor for private AI infrastructure. Those clients adopt the platform, generating revenue that validates the investment. The platform's client list grows. Heritage's investment appreciates. Meanwhile, the family-office clients who adopted the platform become more deeply embedded in Heritage's service ecosystem — because Heritage sourced, vetted, and introduced the vendor relationship. Client retention rises. Heritage can credibly claim to prospective clients that it not only manages capital but also curates the AI infrastructure layer of a modern family-office operation.

This is, effectively, a client-lock-in strategy conducted through investment activity. Heritage is using its balance sheet to build a proprietary vendor network that its clients depend on — and that competing multi-family offices cannot easily replicate, because they lack the deal flow and the LP relationships that give Heritage first access. The LP base is the moat. The families that Heritage manages capital for are also the families that can credibly test-adopt AI infrastructure and signal to the market that the technology is safe for institutional use. That endorsement, from clients with Heritage's profile, is worth considerably more than a marketing campaign.

The 2023 allocation map

The six companies that received Heritage commitments in 2023 cover three distinct layers of the private-LLM stack. Two sit at the deployment layer: infrastructure for running fine-tuned models on dedicated compute, entirely within a client's data perimeter. Two sit at the tooling layer: interfaces and evaluation frameworks that allow family-office investment teams — people who are not ML engineers — to interact with, query, and audit a private model's outputs against proprietary data. The final two sit at what Reinholt's team calls the "corpus layer": companies that specialize in structuring and preparing the proprietary data — investment minutes, deal memos, family archives, external market data — that makes a fine-tuned private model useful rather than generic.

The corpus-layer investments are the least discussed and, in Heritage's internal view, potentially the most valuable. The argument is straightforward: the model itself will commoditize. Open-weights base models are already good enough for most family-office tasks, and they will improve without Heritage's assistance. What will not commoditize is the structured proprietary data corpus that makes a private model better than a public one for a specific family's specific questions. The corpus is the durable asset. Heritage's two corpus-layer investments are building the tooling that helps families create and maintain that corpus — and, critically, that keeps the corpus tied to an infrastructure platform that Heritage also holds equity in.

Heritage's technology team — led by Priya Vasanthakumar, who joined from a quantitative asset manager in 2022 and built the firm's internal AI evaluation capability — ran a twelve-month parallel evaluation of the six portfolio companies against public-model alternatives in the second half of 2023. The evaluation tested two things: whether the private-model deployments outperformed public models on family-office-specific tasks (deal screening, capital allocation analysis, succession scenario modeling), and whether the workflow integration was tight enough to drive actual adoption among investment staff. The capability evaluation was inconclusive — public models were better on general reasoning; private models were better on tasks that required deep domain knowledge from the proprietary corpus. The adoption evaluation was decisive: teams that used the Heritage-portfolio tooling showed significantly higher daily engagement than teams using public API wrappers. Stickiness, not raw capability, is what drives client retention.

What to watch

Heritage's thesis has been running for twelve months. Several indicators will determine whether it compounds or stalls over the next three years.

• Whether Heritage's portfolio companies cross $10 million in annual recurring revenue from family-office clients outside Heritage's own LP base. Internal revenue from Heritage-referred clients validates the vendor relationship but not the market thesis. External revenue — from clients Heritage did not source — validates the broader demand signal.
• The pace at which competing multi-family offices launch comparable AI infrastructure programs. Bessemer Trust, Northern Trust Wealth Management, and at least two large European multi-family offices are known to be evaluating private-LLM programs. If they move in 2024 or 2025, the window for Heritage's first-mover advantage narrows.
• Whether open-weights model quality improves fast enough to commoditize the fine-tuning layer — Heritage's deployment and tooling investments — before those companies achieve client lock-in. The risk is real: if base models become good enough that proprietary fine-tuning adds minimal value, the corpus layer survives but the deployment and tooling investments erode.
• Regulatory treatment of private AI deployments inside family offices. Several European jurisdictions are actively drafting guidance on the use of AI in financial decision-making. If regulators require audit trails and explainability from private deployments, Heritage's corpus-layer investments become more valuable; if regulators restrict self-hosted AI in financial services, the deployment-layer investments face a structural headwind.
• The Sequoia Capital venture referral channel. If Sequoia's venture partners source a company in the next 24 months that represents a materially better solution than any of Heritage's six portfolio positions, Heritage will face a difficult choice: double down on existing positions or rotate. The referral channel is a strategic asset, but it is also a source of competitive intelligence about Heritage's own portfolio.

Frequently asked

Is Sequoia Heritage the same as the Sequoia Capital venture fund?

No. Sequoia Heritage is the multi-family office arm of Sequoia Capital, operating with a separate investment committee, a different mandate, and capital that is substantially longer-duration than the venture fund. It manages assets for Sequoia partners, select LPs, and a limited number of ultra-high-net-worth families. The venture fund makes early-stage bets on startups; Heritage constructs multi-asset portfolios aimed at capital preservation and compounding across decades.

Why would a multi-family office fund private-LLM infrastructure companies rather than just using public models?

The investment and the usage are the same thesis. Heritage funds the infrastructure layer because its own clients will use that infrastructure — and because the vendor relationship keeps clients embedded in Heritage's service ecosystem. The return is financial and strategic simultaneously. It also reflects a genuine conviction that family-office clients at the $2 billion AUM threshold and above cannot accept the strategic exposure that comes with routing sensitive capital-allocation decisions through a public model API.

What is the long-duration capital advantage in AI infrastructure, specifically?

Venture funds have a ten-year clock with a compressed investment window. AI infrastructure — particularly the deployment and corpus layers that Heritage targets — returns on a fifteen-year-plus horizon if the client relationships compound as expected. Venture cannot hold those timelines; Heritage's indefinite-horizon vehicle can. The advantage is structural, not analytical: Heritage can underwrite deals that are correctly priced but incorrectly timed for venture.

How does Heritage prevent its portfolio companies from selling to clients that compete with Heritage's own LP base?

It does not, and it does not try to. Heritage's portfolio companies are independent businesses with their own client strategies. Heritage's interest is that those companies succeed broadly, which increases the value of Heritage's equity stake. The alignment comes from the referral and introduction channel, not from exclusivity provisions. Heritage has preferred-vendor relationships with its own clients, not ownership of the client relationship.

What happens to Heritage's thesis if open-weights models become good enough to eliminate the value of private fine-tuning?

The deployment layer erodes; the corpus layer survives. A world in which base models are excellent and free still requires that families have structured, queryable proprietary data to get value from those models. Heritage's two corpus-layer investments are the most defensible positions under that scenario. The deployment and tooling investments are more exposed to commoditization — which is precisely why Heritage has structured those four positions with hybrid equity-and-revenue-share instruments that return capital earlier if the business model compresses.