Sameer Dholakia on AI Investing, Bessemer's Growth Strategy and Company Culture

Sameer Dholakia says that in his 31 years in software, the few moments that felt like technological magic were followed by very large market cycles. Harry Stebbings introduces the Bessemer Venture Partners partner, a former SendGrid CEO, and asks how he views the current AI wave.

Dholakia recalls graduating in Silicon Valley in 1995 as the internet made it possible to publish globally and conduct commerce online. Later, virtualization at VMware made it possible to run multiple virtual machines and move live workloads between physical machines. He came to see that flexibility as a foundation for cloud computing.

During a break after his time at Twilio, he saw an early demonstration of Jasper using GPT-3 for marketing and recognized the same feeling. Dholakia believes the AI cycle could be larger than the internet and cloud cycles combined. He compares intelligence to oil as a broad input, arguing that every kind of organization can use more of it.

Stebbings asks about Bessemer's Anthropic investment. Dholakia recalls a Series D at approximately a $15 billion pre-money valuation, led by Menlo with participation from Lightspeed and Bessemer. He credits an introduction through partners Byron Deeter and Matt Murphy for bringing the opportunity to his attention.

His framework was an analogy with cloud infrastructure: enormous compute and capital requirements could lead to a small number of major model providers, with many applications built on top. He looked at which companies had access to the resources needed to pursue scaling and which cloud platforms had already chosen partners.

At the time, he viewed Anthropic's relationship with AWS as important for both compute and distribution. The outcome was not obvious: several other labs were trying to build frontier models. Dholakia says the investment was contentious inside Bessemer and, according to Deeter, received unusually low support in an internal vote.

He values a firm culture that allowed the partners with conviction to proceed despite disagreement. Looking back, his main regret is not having persuaded the firm to invest a larger amount. At the time, he says, the risks in the business were substantial even though he believed the cloud relationship was pivotal.

Stebbings asks whether open models will undermine the economics of frontier labs. Both investors also have exposure to Fireworks. Dholakia argues that the market can support both open and closed approaches: customers from consumers to governments will need different kinds of intelligence, and he believes the overall market will be very large.

He says he previously underestimated how large a frontier business could become and does not think investors need to predict the precise open-versus-closed token mix. Bessemer's investments in Anthropic and Fireworks give it exposure to both. This is his portfolio view, not proof that both business models will succeed.

The speakers consider whether US enterprises will use open models that originated in China. Dholakia argues that hosting can change how a model is delivered, but acknowledges that some chief information officers remain concerned about its provenance. He would also welcome a strong US-based open-weight model.

Dholakia believes an open-source provider can build a substantial business, citing historical examples of commercial support around open software. Many enterprises, he argues, prefer a provider with expertise to taking full responsibility themselves. He does not commit to a particular valuation for any such company.

Asked for a bear case against frontier labs, Dholakia says he remains strongly optimistic about demand for intelligence that can solve previously difficult problems. He disagrees with an extreme proposed split in which almost all traffic would be open and almost all spending closed, but again says the exact ratio is not central to his thesis.

Stebbings recalls passing on an opportunity that later looked cheap. Dholakia says he made a similar mistake with OpenAI after ChatGPT appeared. Rapid growth in the following period stretched his imagination about the possible size of the market and influenced how he evaluated Anthropic.

For Dholakia, the lesson is that investors will make important mistakes but may have another opportunity to act. He does not treat that as a reason to ignore risk; he describes reassessing the scale of a market after observing new evidence.

Does entry price still matter if companies can become extraordinarily valuable? Dholakia says it always matters. His central question is whether a company can compound at scale for long enough to produce a venture return from the price paid.

He gives examples of investments made at very high multiples of current recurring revenue. Rather than relying only on the current figure, he evaluates what he believes the company can reach over the next twelve months and whether the implied forward multiple is reasonable.

Dholakia is wary of paying for projections much further out because rapid growth makes distant outcomes difficult to predict. Depending on the business, he says a multiple of roughly 10 to 15 times his estimate for the next twelve months can be acceptable. That is his investment framework, not a general valuation rule.

Stebbings asks whether customer mix and revenue quality change the analysis. Dholakia says they matter greatly. Drawing on his enterprise software background, he prefers strategically important products with strong customer retention and expansion; he would pay less for a business with more volatile consumer or prosumer retention.

On gross margins, Dholakia says he looks beyond the next twelve months because a growth investment may be held for five to seven years. An investor must nevertheless defend the mechanism behind any projected improvement rather than assume that a spreadsheet will make margins attractive.

He argues that scale and market position can improve margins and says some of his AI investments have seen substantial changes over one to two years. Stebbings then asks how he thought about Legora's opportunity in legal services and the price paid in its financing.

Dholakia says his analysis included global legal spending rather than only the US. He believes lower-cost services could also reach legal needs that currently go unmet. His estimate of market size and his view of Legora's potential are investment assumptions; he and Stebbings disclose that they have invested in companies under discussion.

Stebbings asks whether competition has changed because AI companies have raised so much capital. Dholakia distinguishes the number of competitors from the intensity with which they can compete. Larger war chests can support faster scaling and create stronger competitive pressure than he experienced as an operator.

Dholakia calls capital a moat: funding can help a company scale, attract talent, build brand recognition and win customers, creating a reinforcing cycle. He compares that with the developer adoption that made SendGrid and Twilio familiar building blocks in their respective markets.

But he qualifies the idea that investors can simply crown a winner. Capital may be necessary in a compute-intensive market, he says, but it is not sufficient. A company also needs strong leadership and a product customers love enough to recommend.

The conversation turns to older software growth benchmarks, often described as triple-triple-double-double. Dholakia says that pace may struggle to attract some venture capital when alternative opportunities are growing much faster. He nevertheless believes investors can make good returns in less explosive businesses if the entry price is appropriately different.

Stebbings asks whether capital-efficient application companies can still resemble successful growth investments of the past. Dholakia thinks such companies will exist, but expects near-term venture dollars to favor the fastest-growing opportunities. For a portfolio, he says, the issue is how to mix those different risk and return profiles.

Stebbings asks whether seed investing now requires a billion-dollar fund, pointing to the reported size of Bessemer's seed and early-stage vehicle. Dholakia says high entry prices and the number of investments required to find winners make the arithmetic challenging for a small fund.

He qualifies that view by asking where a fund invests. Hardware, robotics and model training can be particularly capital-intensive; other areas may require less. He recalls much smaller seed checks and valuations earlier in his career, while declining to give a precise current market-wide figure.

Dholakia sees venture as an increasingly global and professionalized asset class, spanning more sectors, geographies and stages than it did in the 1990s. He expects firms to need either the breadth of a large platform or the differentiated expertise of a specialist. In his opinion, firms between those positions face a difficult market.

He compares that structure with other industries, where a few large companies coexist with distinctive specialists. Dholakia acknowledges there may be other paths to success; his argument is an interpretation of how industries change as they grow, not a demonstrated rule for every venture firm.

Stebbings raises the idea that an early investment can function as an option to invest more later. Dholakia says the financial logic is straightforward, though founders may understandably dislike being described as an option. In his view, additional capital still depends on the business demonstrating the right to keep scaling.

He says Bessemer's early-stage fund includes reserves, but the firm can also make larger follow-on investments from growth capital when a company appears to be taking off. He cites an announced investment in EliseAI as an example of backing a company the firm believes can build a large business.

Stebbings asks whether following on in one portfolio company but not another creates a negative signal. Dholakia says he does not focus on the optics: building a business can take time, and a company that has not yet reached the next stage may simply need more work. He believes founders generally understand their own progress.

On constructing a $4 billion growth portfolio, Dholakia says the principles still involve concentrated bets. He describes the possibility of investing hundreds of millions in a company across multiple rounds when conviction is high. He distinguishes late-growth portfolio construction from the firm's early-growth practice.

He says a late-growth position of roughly 5% to 10% of a fund can feel appropriate, with flexibility for exceptional cases. He recalls that a previous growth fund had a much larger concentration in Anthropic. These are descriptions of the firm's approach, not recommendations for individual investors.

Stebbings questions whether early and late growth need separate categories. Dholakia concedes that the dividing line is partly arbitrary but says the information available, risk profile and check size differ significantly between a company with tens of millions in revenue and a much later-stage business.

Stebbings argues that the multiple on invested capital ultimately matters more than the label. Dholakia agrees that strong returns can be found at both stages, though he describes a high bar for growth investments. Their hypothetical multiples depend on future performance and are not assured outcomes.

Dholakia mentions Fieldguide, an AI company serving auditors, as an early investment he wishes the firm had backed more heavily. He says better model capabilities made its product more compelling. Stebbings questions whether a vertical application can become as large as an infrastructure company, while acknowledging that entry price can make the investment return attractive.

Dholakia says the same growth team works across the firm's early- and late-growth opportunities. He expresses confidence in the portfolio and team they have built, a view that reflects his position inside the firm rather than an independent assessment of future performance.

Stebbings turns to culture and asks about Dholakia's belief that teammates and friends outperform mere colleagues. Dholakia traces the idea to his years at SendGrid. He says aligning many people around a common goal is one of the hardest parts of building a company.

In his experience, people who care about one another are more willing to work for the team and less willing to let teammates down. To create that environment, he says, an organization must define how it works and consistently apply its values, including to high performers.

At SendGrid, Dholakia used four values he called the four H's: happy, hungry, humble and honest. Stebbings challenges whether such words are sufficiently distinctive. Dholakia responds that the meaningful part is how a particular organization defines and enforces each value.

He explains humility as a preference for 'we' over 'I': in interviews, he listened to how candidates described an achievement and the role of their teams. He does not claim that every company must choose the same value, but says a team needs a shared way of working and should be clear when it is not a fit.

His second practice was spending time together outside day-to-day work. He recalls bringing the entire SendGrid company together for an annual kickoff: one day aligned people on vision and goals, and another encouraged colleagues from different offices to spend time together.

Dholakia says those relationships matter when a business encounters problems. Colleagues who lack trust may blame one another or leave; teammates who know and care about each other are more likely to work across functions to solve the issue. He presents the offsite as one way he tried to build that trust.

Stebbings agrees with the principle but argues that values should be specific enough to be debated. He describes his own team's emphasis on hard work and direct discussion. Dholakia says a culture can differ from SendGrid's as long as people understand and choose the way the team works.

Dholakia clarifies that 'happy' at SendGrid did not simply mean being cheerful. He wanted an optimistic disposition: people who would approach hard problems constructively rather than bring the team down. The practical definition, he says, mattered more than the word alone.

Returning to venture, Stebbings asks when investors should sell shares as companies remain private for longer. Dholakia says limited partners ultimately need cash returned to them. Although he thinks the AI wave is still early, he considers partial sales along the way a prudent way to lock in some gains and return capital.

He says late-stage founders and investors increasingly understand that selling a portion of a large position does not necessarily mean losing faith in the company. Many founders also sell a small amount to provide financial security for their families. Stebbings agrees that this can enable a longer-term perspective.

Dholakia expects companies to stay private longer because private capital markets have grown. At the same time, he predicts a reopening of the IPO market in later years as public investors seek direct exposure to AI companies. He frames this as a forecast based on his view of supply and demand.

Drawing on his experience taking SendGrid public, Dholakia says preparing for public scrutiny can make a company more operationally disciplined. Delivering predictably quarter after quarter is difficult, but he believes the process can build a stronger organization.

Stebbings argues that staying private can also let companies make large product investments without the same quarterly pressure. Dholakia acknowledges the tradeoff: private ownership offers flexibility, while a public listing can provide acquisition currency and, in some markets, credibility with customers.

They discuss whether private companies already have enough liquidity and acquisition capacity to weaken the case for an IPO. Dholakia does not offer a definitive answer, but says some companies still view public status as meaningful to customers assessing their scale and reliability.

Stebbings asks whether individual investors should have greater access to venture returns, while raising the risk that they might be left with less attractive funds. Dholakia supports broader access in principle but emphasizes that venture is high risk and illiquid; he believes investors should understand those tradeoffs.

Asked for the investment case for Fireworks, Dholakia praises its technical leadership and its go-to-market team. He says the combination of technology and commercial execution underpinned Bessemer's investment. Both speakers are investors in the company, so their enthusiasm comes with a direct financial interest.

The product thesis, as Dholakia describes it, is that enterprises will want to own at least some of their intelligence rather than rent all of it. He expects a mix of owned and rented capabilities, and sees demand for tools that let organizations exercise more control.

On sovereign models, Dholakia mentions Bessemer's involvement with Sarvam in India and Mistral. He argues that local languages and data can make models tailored to a region valuable. In other cases, governments or regulated industries may prefer systems developed under their own jurisdiction.

Dholakia does not argue that every buyer will require a sovereign model. In a large market, he believes a subset of customers and use cases can support substantial businesses. The strength of that business case remains a question for each product and market.

The speakers compare required outcomes across early- and growth-stage funds. Dholakia says early-stage portfolios need some very large multiples to offset investments that do not work, while later-stage businesses may be more developed and require a different distribution of outcomes.

Asked whether very large funds can still return five times capital, Dholakia says he believes they can in the current environment. He points to unusually rapid value creation, while acknowledging that producing such returns is difficult at any stage. This is his expectation, not evidence that a particular fund will achieve it.

Stebbings asks him to choose between early and late growth. Dholakia likes the opportunities available at the later stage but also values early ownership obtained with smaller checks. He says a broad platform can combine early-stage, early-growth and late-growth practices.

In a quick-fire round, Dholakia praises several Bessemer investors. He ultimately distinguishes finding unconventional companies from persuading founders to choose an investor, naming David Cowan for the former and Byron Deeter for the latter. He recounts seeing Deeter's repeated presence at company IPOs during SendGrid's own listing.

Asked which missed company the firm regrets, he names Stripe, noting Bessemer's history investing in developer-facing API companies. He also praises an investor at Lightspeed for deep technical thinking and a willingness to make unusually large bets.

Dholakia says he has revised his view of the potential size of the open-model token economy over the previous year. He links that change to investment in infrastructure for generative media and open-model inference, including Fireworks.

Looking forward, he is excited about robotics and embodied AI. He compares the state of robotics with generative AI several years earlier and expects rapid progress, while recognizing that the field has more work ahead. He mentions possible applications in construction, mining and agriculture alongside autonomous transport.

Stebbings questions the valuation paid for Waymo relative to ride-hailing companies. Dholakia argues that the eventual opportunity could extend beyond ride-hailing into personal vehicles, trucking and logistics. He explicitly notes that personal-vehicle licensing is not Waymo's near-term focus; his trillion-dollar valuation is a speculative projection.

His second area of excitement is using AI in therapeutics. Dholakia hopes automated laboratories and models that can analyze large datasets will accelerate biological discovery and drug development. He repeats a prediction that a decade of progress could otherwise have taken much longer; the outcome and timing remain unproven.