20VC: Bessemer's Sameer Dholakia on $1BN seed funds and AI

Summary

• Dholakia underwrites AI as a larger opportunity than the internet and cloud combined. “Oil powers all the machines. Intelligence powers all the humans.” Bessemer’s Anthropic investment at $15 billion pre-money followed a cloud-style thesis: capital intensity would concentrate frontier intelligence among a few providers, with hundreds of applications above them.
• Open models and frontier labs can both become enormous businesses without investors correctly forecasting their market-share split. Dholakia rejects the claim that open models will capture 99% of tokens while closed models capture 99% of dollars. He admits he underestimated the open-token economy 12 months ago; Fireworks and fal now express that revised conviction.
• Entry price matters, but Dholakia tests high current ARR multiples against what a business could become over the next 12 months. He considers 10–15x on an NTM basis reasonable for extraordinary businesses; Legora’s move from $20 million to $200 million illustrates his argument. Revenue quality still affects the multiple, while weak margins require a defensible improvement mechanism over his five-to-seven-year holding period.
• Capital can reinforce market leadership, but it cannot substitute for an exceptional product. Dholakia sees funding driving faster expansion, stronger branding, better recruiting and more customer wins. Legal AI illustrates the opportunity: he frames worldwide legal services as roughly $1 trillion, with lower prices potentially unlocking previously unserved demand.
• Venture’s strongest positions are giant global platforms and specialists with a clear edge. Bessemer raised $1.75 billion for seed and early stage plus $4 billion for growth; Dholakia struggles to see how small seed funds obtain enough “shots on goal” in capital-intensive AI. He calls the undifferentiated middle a difficult place to make money, while allowing that cheaper, slower-growing businesses can still deliver venture returns.
• Dholakia believes large growth funds can still produce 5x-plus returns because outcome sizes have expanded. Bessemer may invest $200–400 million per company across rounds, and its previous growth fund put approximately 20% into Anthropic. Both speakers favor late-stage opportunities offering potential 10–20x outcomes with more developed businesses; those remain underwriting ambitions.
• Longer private-company lives do not eliminate the need for liquidity or the possibility of an IPO reopening. Dholakia supports partial secondary sales and predicts an IPO unlock “later next year, and in 2028,” driven by demand for pure-play AI exposure. Stebbings challenges the structural appeal of listing when private companies can already finance acquisitions, support employee liquidity and make substantial long-term investments.
• Autonomy and therapeutics are Dholakia’s next major expressions of the intelligence thesis. Bessemer invested nearly $200 million in Waymo; he thinks personal-car subscriptions, trucking and logistics could extend it into “a trillion-dollar business. Easy.” In therapeutics, he believes AI and automated wet labs could compress half a century of discovery into the next decade.

Deep dive

1. Anthropic required a reset in the imagination of scale

• After 31 years in software, Dholakia associates rare moments of technological “magic” with trillion-dollar cycles. Jasper’s marketing application on GPT-3 in winter 2021 supplied the latest recognition. He thinks this cycle exceeds the internet and cloud combined because intelligence serves consumers, businesses and governments: “Oil powers all the machines. Intelligence powers all the humans.”

• His Anthropic underwriting was “a cloud redux”: compute intensity implies capital intensity, concentrating frontier providers while allowing hundreds of applications above them. Microsoft had OpenAI, Google was building Gemini, and AWS had made a substantial Anthropic investment. AWS mattered twice—training compute and early distribution—making its relationship pivotal to the investment case.

• Bessemer joined Anthropic’s Series D at what Dholakia recalls as $15 billion pre-money and $18 billion post-money, alongside Menlo and Lightspeed. According to Byron Deeter, it was the lowest-voted investment during his tenure at the firm. Partner empowerment allowed conviction to prevail: “If you guys really believe, go do it.” Dholakia’s regret: “I just wish we had written a bigger check.”

• Stebbings says he passed on an OpenAI opportunity at $30 billion, reasoning that even a $100 billion outcome would be only about a 3x before dilution. Dholakia says he also passed on an OpenAI opportunity after the ChatGPT moment, but gives no valuation for it. He recalls that OpenAI had grown from about $40 million to well north of $1 billion—perhaps $1.5–2 billion—by the time Bessemer looked at Anthropic’s first round. “Our imaginations were capped.” Anthropic offered a chance to correct an important mistake rather than treating the missed investment as final.

2. Open intelligence does not require a closed-model collapse

• Dholakia rejects forecasting a precise open-versus-closed split in a market this large: “This is ‘and,’ not ‘or.’” He also rejects the suggested 99% open-token/99% closed-dollar allocation. Asked whether he sees a frontier-lab bear case, he answers, “I really don’t,” arguing that intelligence can address vast numbers of unsolved problems and solve others at previously unimaginable speed.

• On Chinese open models, Dholakia argues that once hosted by Fireworks, a model is “no longer a Chinese model.” Stebbings pushes back: some CIOs reject Chinese provenance regardless of where downloaded code runs. Dholakia acknowledges encountering exactly that objection and sees a gap for a strong US open-weight model. The exchange leaves buyer perception as a commercial constraint.

• A leading open-model provider can generate venture returns through enterprise delivery and support, Dholakia argues, citing Red Hat. Whether it can reach $500 billion: “I don’t know.” His own changed view concerns the open-token economy’s size: “I just didn’t see it 12 months ago.” That reassessment supports larger commitments to Fireworks and fal, the latter spanning generative media and infrastructure.

3. Forward growth can justify the multiple, but revenue quality still matters

• Dholakia’s examples look extreme on current ARR: Abridge at $5 million ARR and roughly $700–800 million valuation; Anthropic at approximately $100 million ARR and $15 billion; Legora at $20 million ARR and $1.8 billion. These required difficult investment-committee conversations. His question was whether the companies could grow into a reasonable multiple over the next 12 months.

• The horizon is specifically next-twelve-months, rather than the end of the following calendar year, which could mean underwriting 18–24 months ahead. His operating experience makes that distance uncomfortable. Depending on the business, 10–15x NTM is reasonable when growth is unprecedented. Legora subsequently going “20 to 200” makes him comfortable with the $1.8 billion entry.

• Stebbings asks whether customer mix and creative run-rate calculations change that comfort. “Quality of revenue matters deeply,” Dholakia replies. He favors sticky, strategic enterprise products with strong gross and net dollar retention; consumer or prosumer businesses deserve substantially lower multiples because “churn, it can kill you.”

• Margins receive a different time horizon: a growth investment may last five to seven years. Dholakia is not principally worried about the next 12 months, but demands an explanation for any forecast reaching 60–70% gross margins. Scale economies and market position can support improvement; spreadsheets cannot establish it. He reports “factors improvement inside of 12 to 24 months” in AI investments.

4. Legal AI exposes both the market opportunity and the limits of vertical upside

• Stebbings tests Legora’s outcome math: a $300 billion US legal market, a hypothetical 20% software share yielding $60 billion, and competition including Harvey might support $5–10 billion in company revenue. Can an investor entering near $2 billion still make 10–20x? Dholakia says that framework is “not terribly off,” but the geography is too narrow.

• Dholakia uses roughly $1 trillion in worldwide legal-services spending and says the US represents only 40% of the market. Lower prices could also unlock needs currently left unserved. A business reaching $5–10 billion in revenue need not capture much of that spending to become enormous; its eventual valuation would determine the investment return.

• Competitive breadth has not fundamentally changed, he argues; the capital behind competitors has. Funding accelerates expansion, builds brand, attracts talent and reinforces customer wins. “Markets pick winners,” creating a self-reinforcing cycle. Dholakia explicitly endorses king-making, but calls capital “necessary but not sufficient”: scarce compute requires a war chest, while leadership and a product customers recommend still determine whether it works.

• Fieldguide reveals a boundary. Dholakia thinks it was below $10 million ARR when he invested, probably below a $200 million valuation; model improvements around January–February accelerated adoption of its auditing product. He wishes he had invested more. Stebbings questions whether even a $10–20 billion outcome is “generational.” If that requires $100 billion, Dholakia concedes Fieldguide is “probably not that.”

5. Venture’s middle faces pressure from both scale and specialization

• Bessemer’s $5.75 billion raise comprises $1.75 billion for seed and early stage and $4 billion for growth. Dholakia struggles to see how small seed funds obtain enough “shots on goal,” particularly in compute-intensive AI. Historical seed checks were $3–5 million at $20–40 million valuations; he supplies no specific current figures, only that they are “factors higher,” while allowing less capital-intensive niches.

• Venture has expanded across geography, sectors and stages, from pre-seed through enormous pre-IPO checks. Dholakia sees professionalization favoring either a giant global platform or an extreme specialist with differentiated picking and support. “If you’re in between, good luck.” His automotive analogy: mass-market giants at one end, Porsche or Ferrari at the other, with a difficult middle.

• Traditional triple-triple-double-double growth struggles to attract capital when alternatives are growing 10x at scale. Nevertheless, Dholakia says these businesses can still generate excellent returns at dramatically different entry multiples. Capital-efficient application companies resembling the older 1Password or Qualtrics growth-investment pattern may emerge over time; he simply does not expect them to attract the near-term concentration of venture dollars.

• Stebbings challenges retail access through adverse selection: oversubscribed firms already serve institutions such as hospitals and pension funds, potentially leaving retail investors weaker managers. Dholakia favors access but says investors should choose strong platforms or genuine specialists and understand the risk. For those uncomfortable with that, he explicitly raises public-market compounding without venture’s liquidity sacrifice as an alternative.

6. Early options become large positions only after businesses earn them

• Defending Menlo’s Venky’s description of early investments as options, Dholakia distinguishes the portfolio logic from entrepreneurs feeling diminished. Follow-on capital depends on progress: “You have to earn the right to keep scaling.” Bessemer’s $1.75 billion early fund includes reserves, while the growth fund can invest far beyond pro rata. He dismisses signaling anxiety when a business has not yet demonstrated the necessary acceleration.

• In the $4 billion growth vehicle, Bessemer may deploy $200–400 million into one company over several rounds. Dholakia describes 5–10% as a comfortable late-growth allocation, with flexibility for extraordinary conviction. Its previous growth fund, Century II, placed approximately 20% into Anthropic. The larger fund preserves concentrated underwriting.

• Stebbings mocks dividing growth into early and late categories. Dholakia concedes the cutoff is arbitrary but distinguishes $20–40 million-revenue companies, perhaps worth $500 million–$1 billion, from entering Fireworks at $15 billion: information, risk and checks differ. Stebbings insists the decisive metric is invested-capital multiple, observing that infrastructure can have a larger addressable market than a vertical application.

• Dholakia wants 5x-plus funds. Early portfolios need some 50x outcomes to offset lower hit rates; growth investments can require less upside because more business risk has been resolved. He still seeks 10–20x opportunities and says, “We want nothing less than a five,” while acknowledging some investments may return less. Forced to choose one vehicle personally, he favors late-stage growth.

7. Culture becomes concrete through hiring and behavior under stress

• “Teammates and friends beat colleagues and coworkers every day and twice on Sunday,” Dholakia argues. Coordinating large groups is difficult; people who care about one another work harder for the team. SendGrid’s four Hs—happy, hungry, humble and honest—were operating standards. A high performer could not receive an exemption, and everyone needed permission to call out violations.

• Stebbings objects that values should permit disagreement: nobody advocates unhappy, unambitious liars. Dholakia makes the definitions concrete. Humility meant “we not I,” tested by counting pronouns in interview answers about accomplishments. Happiness meant optimism under pressure rather than cheerfulness. Someone consistently bringing the team down, or preferring individual credit, could be a poor fit despite otherwise strong performance.

• SendGrid flew the whole company to Mexico each January: one day aligning on direction, another spending time with people from other offices. The mechanism mattered when things broke. Colleagues might blame sales, blame marketing or leave; friends who had kayaked or mountain-biked together were more likely to work through the problem. Stebbings separately endorses blunt IC debate—“feelings are expensive”—while Dholakia emphasizes “intellectual honesty.”

8. Longer private lives still require DPI and leave room for IPOs

• “Our LPs are not in the business of handing us the money for perpetuity,” Dholakia says. Although he thinks the AI wave remains early, prudent investors should take some chips off the table and return capital. Late-stage partial sales are increasingly treated as ordinary portfolio management. He likewise encourages founders to establish family financial security, potentially making them more willing to swing for larger outcomes.

• Longer private-company lives are a “permanent structural shift” in his view, supported by a larger venture asset class. Crossovers and LPs must respond because substantial appreciation happens before listing. Public investors waiting for an IPO may miss much of the earlier value creation.

• Nevertheless, Dholakia predicts an IPO unlock “later next year, and in 2028.” His mechanism is supply and demand: successful listings demonstrate appetite for pure-play AI exposure, public investors want dozens or hundreds of companies rather than a handful, willingness to pay improves, and more issuers follow. This is his forecast for reopening the window, not an established timetable.

• Having taken SendGrid public, Dholakia believes being “IPO worthy” forces predictable execution and operational discipline. Stebbings counters that large private companies already face high expectations, while Stripe’s private status permits strategic investment, acquisitions and liquidity. On diminishing IPO attractiveness, Dholakia concedes, “Maybe. I don’t know.” His portfolio companies still want to list, sometimes because customers view public status as evidence of credibility.

9. Fireworks combines enterprise ownership with a stronger commercial engine

• Asked for Fireworks’ $250 billion bull case, Dholakia begins with people rather than a detailed valuation model: its technical leadership and team from Meta, now joined by president George Hu. Having worked alongside Hu at Twilio, he considers him exceptionally strong at strategy, analysis and building go-to-market engines. Hu joined a month or two before Bessemer’s investment: “That is the pitch right there.”

• The enterprise proposition is owning intelligence rather than exclusively renting it. Dholakia expects “a healthy, healthy amount of both,” but finds it difficult to imagine businesses wanting to own none of their intelligence. Stebbings sees a possible 20x return from their approximately $15 billion entry; Dholakia agrees.

• Sovereign models need only win particular buyers and use cases. Bessemer backs Sarvam in India and is “invested in or investing in Mistral right now.” Indic languages and local multimodal content underpin the Indian case; government, military and regulated use cases support sovereign preferences elsewhere. Dholakia’s argument rests on segments of a huge market choosing local models, rather than every buyer doing so.

10. Judgment includes contrarian selection and the ability to close

• Dholakia initially names Byron Deeter Bessemer’s best picker, then revises his answer when asked separately about closing founders. David Cowan gets the selection distinction for apparently strange investments such as Rocket Lab; Deeter gets closing for his experience, intelligence and ability to build relationships.

• Stripe is the miss that particularly hurts because Bessemer backed developer-API companies including Twilio, SendGrid and Auth0. His dream partnership addition is Lightspeed’s Ravi, whom he praises for deep technical and market thinking—and willingness to take risks that look “borderline insane,” exemplified by a $1 billion Anthropic check in the $60 billion round.

11. Autonomy and therapeutics extend the thesis beyond software

• Bessemer invested nearly $200 million in Waymo. Stebbings challenges its roughly $115 billion valuation against Uber’s $120–125 billion market capitalization. Dholakia’s objection: “You’re looking at the ride-hailing market as the only market.” His case requires thinking beyond today’s principal business.

• Having bought his 16-year-old son a car, Dholakia would readily pay a monthly subscription for a Waymo driver. He sees potential across other automakers, followed by trucking and logistics. Personal-car licensing is not Waymo’s near-term focus, he acknowledges. Combining those possibilities with expanded ride-hailing use cases, he thinks it becomes “a trillion-dollar business. Easy.”

• Embodied robotics is “probably” around 2021 relative to generative AI, with a couple or few years still to go, Dholakia thinks; progress could arrive faster than most expect. He is interested in autonomy across construction, mining and agriculture, but cannot recall offhand whether Bessemer owns a robotics-data company.

• Therapeutics offers what he considers AI’s greatest potential impact: “drug the undruggable.” Automatable wet labs paired with systems reasoning across datasets beyond a single human’s capacity could accelerate biological discovery. Echoing Dario’s forecast, Dholakia believes the next decade could deliver what otherwise would have taken half a century: “I really believe that.”