Key Takeaways

  • Bessemer raised $5.75 billion on the thesis that artificial intelligence represents a larger supercycle than the internet and cloud combined.
  • Backward ARR multiples make early rounds look impossible, such as paying 150x ARR for Anthropic or pricing Legora at $1.8 billion on $20 million in ARR.
  • Sameer Dholakia underwrites fast-growing AI startups at 10x to 15x next-twelve-months (NTM) revenue instead of trailing figures.
  • Bessemer strictly caps projection horizons at 12 months, refusing to pay forward 18 or 24 months of revenue.
  • Low gross margins caused by heavy compute are acceptable during a 5-to-7-year hold period because model efficiencies improve over time.

The Disagreement

Harry Stebbings challenges whether traditional entry discipline has been abandoned in the rush to fund AI. Software venture capital built its reputation on strict rules: never overpay on current run rate, and never compromise on 80 percent gross margins. When investors pay 150x current ARR for Anthropic, or price Legora at $1.8 billion when it only generates $20 million in revenue, classic venture math breaks down. Stebbings asks whether entry prices and software gross margins still carry weight when hardware and inferencing bills swallow revenue.

Dholakia acknowledges the historical rule. “Entry price always matters Every venture capitalist has taught me that,” he concedes. Yet he insists that evaluating hypergrowth companies on trailing ARR leads to bad decisions: “If I look at some of my the investments of which I'm most proud they were at current ARR multiples that would make most venture capitalists throw up.”

Instead of looking back, Dholakia prices deals strictly on next-twelve-months numbers. “Something that's in the 10 to 15 times NTM, I don't think is crazy, especially if there are these businesses that are growing at rates that I have never been seen in the history of business.” If a company at $20 million ARR reaches $150 million within a year, an entry price that looks insane on current revenue becomes a standard 12x forward multiple. He takes the same position on compute costs: “Am I worried about their gross margins in the next 12 months? I am not. I have got a hold period of 5 to 7 years at a growth stage check.” In his view, scale and model efficiency will restore margins before exit.

Who's Right (and When They're Wrong)

Dholakia is right when an AI company possesses clear distribution and structural retention that sustains extreme growth. If a startup multiplies top-line revenue tenfold within twelve months, anchoring on trailing figures blinds you to the math. High growth compresses an eye-watering entry multiple into a defensible valuation almost immediately.

Yet Stebbings's caution proves correct when startups mistake transient compute experiments for durable software contracts. Traditional software commanded high multiples because enterprise subscriptions retained revenue at 80 percent margins for five years. If an AI startup surges from $5 million to $30 million on pilot project budgets that churn after six months, forward revenue never materializes.

Dholakia protects against this failure mode by drawing a hard line on the calendar: “I'll go NTM next 12 months. And yeah, depending on the business, next 12 months only. Because you know, depending on where you are in the year and that could be more like 18 or 24 months and I won't pay that far forward.”

His discipline on the 12-month boundary is essential. Underwriting 24 months of AI revenue assumes market stability that does not exist. Similarly, relying on gross margin recovery over a 5-to-7-year hold period works only if model compute expenses fall faster than competitor price cuts. If basic models commoditize enterprise pricing, gross margins remain crushed, destroying the terminal value at exit.

What to Do With This

Recalculate your startup's pitch deck using contracted pipeline and customer expansion to show your exact NTM revenue rather than trailing ARR. If an investor pushes back on your price, defend your valuation at 10x to 15x forward revenue alongside a 12-month unit economics model showing how compute costs fall as query volume scales.