Tomasz Tunguz on Raising Theory, Portfolio Concentration and AI Investing

Harry Stebbings welcomes venture investor Tomasz Tunguz, founder of Theory Ventures, to discuss fund formation, portfolio construction and AI. This conversation was published in April 2023. Both speakers are investors; Tunguz's fund plans, portfolio examples and forecasts reflect his financial interests and the information available then.

Tunguz says he left Redpoint after roughly 15 years because he wanted to experience building a venture firm himself and experiment with a different structure. He values thesis-driven investing: spending months studying a market so he can be a more useful partner to founders, even though founders know their businesses more deeply.

He also favors concentration. His view is that venture returns follow a power law and that a fund can benefit from putting more capital behind a small number of strong outcomes. Theory was designed around deep research and concentrated positions, a strategy involving both conviction and risk.

Tunguz says raising Theory required about 150 limited-partner meetings in a difficult environment and produced roughly $230 million. He approached the process like software sales, building a pipeline around an estimated close rate. He recalls knowing 40 to 50 of the prospective LPs beforehand and says roughly half the capital came from new relationships.

He approached large institutional anchors early, hoping their commitments and participation in the limited-partner advisory council would reassure others. Those investors also offered references and advice. He says the largest LP represented no more than about 12% of the fund, reflecting his chosen approach to diversification.

The fundraising materials included a track record, pitch deck, biography, selected writing and metrics. Tunguz sent concise materials before meetings to help LPs decide whether the fund matched their mandates, such as stage, geography or willingness to back a solo general partner. He used document engagement to understand where prospective investors focused.

Reasons for declining included key-person risk in a solo-manager fund and uncertainty about portfolio allocations. He describes the denominator effect: falling public-market values can increase the private share of a portfolio before private valuations adjust. His description is a 2023 account of the fundraising market.

Theory had one closing, but Tunguz rejects the idea that a single close is inherently superior. A close is a way to organize commitments and establish a cadence; multiple closes can be useful too. Stebbings notes that waiting too long can allow a verbal commitment to go stale.

To maintain momentum, Tunguz updated prospective LPs when he received additional commitments. He says credible evidence of progress can make a fundraising outcome feel more likely and encourage introductions. His description is a sales practice, not proof that verbal commitments are binding.

In hindsight he would have spoken earlier with expert fundraisers and better appreciated that some geographies require longer lead times. Long-standing relationships worked well. He recalls meeting a meaningful share of LPs in person only after they committed, although he did not calculate a precise comparison of remote and in-person conversion.

He says LPs asked detailed questions about the fund's business model: numbers of investments by stage, failure rates, ownership and expected multiples. For fund managers, he recommends making these assumptions explicit rather than relying only on access or a narrative.

Tunguz used historical data and Monte Carlo simulations to design a portfolio of roughly 12 to 15 companies. He envisages 40% to 50% or more of the fund in its top three holdings. These are planned allocations and model-based judgments, not evidence that concentration will produce higher returns.

He expects initial Series A checks of roughly $8 million to $12 million, with flexibility to co-lead larger rounds. Concentration allows time and capital to be allocated to the highest-conviction companies, but the fund could not lead every very large AI round on its own.

To decide where to increase a position, he combines market research, customer understanding, benchmarking, entry prices and possible exit outcomes. He argues that an exit-multiple exercise is useful as a discipline, especially when recent peak valuations might make unrealistic assumptions feel normal.

He also uses a conditional-probability approach inspired by Superforecasting: list milestones a company must reach, estimate the likelihood of each and consider the resulting range of outcomes. He gives examples such as hiring a research team and raising later rounds. The figures are illustrative assumptions, not guaranteed probabilities.

Asked whether deeper market knowledge justifies a higher price, he agrees up to a point: more knowledge may narrow uncertainty and improve confidence. He still wants meaningful ownership in a concentrated fund and says a position can sometimes be built over several rounds.

Later ownership is more expensive as valuations rise. Tunguz discusses starting with a meaningful stake and adding in subsequent rounds, while reserving capital for all portfolio companies and running diligence to decide where to concentrate further. This is Theory's proposed strategy, not an assurance of allocation or returns.

He recounts a difficult financing period at Snowflake when the company was spending heavily and existing investors supported a round. Tunguz says that episode taught him the value of backing a company through both good and difficult times when the underlying thesis is sound. He is uncertain about the exact round, and his account should not be treated as a verified financing chronology.

Stebbings asks how a thesis-driven investor avoids confirmation bias. Tunguz says repeated conversations with buyers and evidence of actual demand are essential. He gives his then-current interest in blockchain-related marketing infrastructure as an example: if he cannot find customers willing to spend, he should abandon or set aside the thesis.

Market timing also returns to customer demand. Tunguz looks for a pipeline and tests whether one buyer's need can be generalized across a broader population. A compelling view of the future is not sufficient if customers are not ready to buy.

On AI architecture, he predicts both integrated proprietary systems and open models will coexist. He imagines an interface that routes requests across models, developer platforms that make models easier to use and enterprise providers delivering complete solutions. These were scenarios in 2023, not established market structure.

He argues that training foundation models is capital intensive, while applications can serve a much wider range of needs. As an analogy he compares estimated market values of major cloud infrastructure businesses with a group of public cloud applications, saying aggregate values were similar but concentrated in far fewer infrastructure companies. The estimates depend on his definitions and should not be treated as audited, directly comparable totals.

Stebbings raises enterprise concerns about sensitive data. Tunguz describes an architecture in which a model runs near data held in the customer's environment and returns a result without moving the underlying data. He expects some highly regulated users to prefer on-premises deployment, but acknowledges security, compliance and legal questions remain.

He sees enterprise readiness—including data security, compliance and legal protections—as a potential business opportunity. Generated code and personal information can raise ownership and privacy questions. These are issues to investigate rather than legal conclusions.

In a new market, Tunguz expects customers initially to prefer bundled solutions that work end to end because they do not yet know how to evaluate every layer. As needs become clearer, they may replace individual components with specialist products. This is his market hypothesis, not a universal adoption pattern.

He thinks enterprise buyers need reliable results and compliance more than a full technical explanation of how a model works. He uses an example about people struggling to draw a functioning bicycle to illustrate the difference between familiarity and understanding. His suggested accuracy and productivity levels are hypothetical thresholds, not measured performance guarantees.

Asked about AI-generated code, Tunguz predicts a much larger share in ten years, partly because many programs contain common patterns. He cites economic forecasts and suggests productivity could increase growth. His figures, comparisons and 2023 macroeconomic interpretation are speculative and do not establish future employment or GDP outcomes.

Stebbings raises wealth concentration. Tunguz says private markets drive innovation while government expresses social values through law, acknowledging tension between network effects and distribution. This is his political and economic view rather than a proposed comprehensive policy.

On regulation, Tunguz argues that compliance costs can favor incumbents and that rules can have unexpected effects. He recommends incremental adjustment as problems become visible, using aviation as an analogy. His examples and policy conclusions are his opinions, not a complete legal or historical analysis.

He says his thinking on startups versus incumbents evolved. Microsoft appeared to have a distribution advantage and was moving quickly with OpenAI, but he believes good execution can still let startups win against large firms, citing Notion and Snowflake. Both speakers have professional interests in startup and venture outcomes.

Tunguz names his former employer Google as a company he believed was challenged by conversational AI in 2023. He interprets its response through the innovator's dilemma and suggests query cost may have made the economics look unattractive. These claims about strategy and future competition are his assessment, not inside confirmation.

He describes his changing understanding of language models, comparing learning from written material with learning through practice and discussing emergent behavior. His analogy is a simplified personal interpretation, not a technical demonstration that deployed models learn from every interaction or behave like humans.

Stebbings asks how publishers can sustain their businesses when AI systems synthesize their material. Tunguz says ownership, attribution and revenue sharing are unresolved and notes existing disputes around search snippets. He expects arrangements may emerge, but the interview does not settle copyright or fair-use questions.

Asked whether the macroeconomic situation would improve by the end of 2023, he predicts it would worsen and cites interest rates, money supply and geopolitical risk. This is a historical forecast, not a report of the subsequent outcome. He also praises Adobe's then-current generative AI work and sees enterprise readiness as an opportunity.

In quick-fire questions, he names Founders Fund and Goodwater as firms he would invest in, explaining that the latter's consumer focus would complement his B2B work. These are personal preferences, and no fund performance is established by the interview.

He cites Datadog and Twilio among investments he missed and says startups can create markets that are hard to recognize early. He also mentions investments in Omni and MotherDuck and says he did not yet see an obvious recent mistake. These comments involve his own investment relationships and are not realized-return figures.

Tunguz wants LPs to explain their goals and portfolio construction more openly so managers can understand whether a fund fits. Stebbings agrees that the LP market can be opaque and hopes for greater transparency.

Asked in April 2023 about the 2024 US election, Tunguz predicts Ron DeSantis rather than Donald Trump would win and speculates about legal and party dynamics. He then offers views on Republicans and entitlement reform. These are his political opinions and pre-election forecast, not statements of later events.

He names Guy Podjarny and former Zendesk CFO Alan Black as people whose operating advice he values. He describes Black's experience in difficult markets as useful to founders; these are professional relationships rather than independent endorsements.

Tunguz identifies Looker as his strongest cash-generating investment and describes a 2012 thesis that cloud data infrastructure would create room for a new business-intelligence product. He credits founder Lloyd Tabb for trusting him early. His account does not provide a full audited return calculation, and the interview closes with thanks.