20VC x SaaStr with MongoDB’s Dev Ittycheria: OpenAI Reaccelerates to a $70B Run Rate, a CRO Jumps to a Direct Competitor, and Agents Start Picking Your Vendors

extract

With Harry Stebbings, Jason Lemkin, Rory O’Driscoll, and special guest Dev Ittycheria, back in the CEO seat at MongoDB Dev Ittycheria joined us one week into his second run as CEO of MongoDB. CJ Desai stepped down on September 28, effective immediately, and the board named Dev interim President and CEO. Desai left to run Meta’s new enterprise platform, reporting to Mark Zuckerberg. Dev ran MongoDB from 2014 to 2025 and took it from about $35 million to more than $2.3 billion in annual revenue. The loudest story of the week was about an executive walking out the door, and we had a guest who had handled exactly that seven days earlier. Here’s what the four of us got into. #1. OpenAI Nears a $70B Run Rate, Up More Than 70% Since July OpenAI’s annualized revenue is approaching $70 billion, up more than 70% since the start of the third quarter, with enterprise sales more than doubling since July. Jason’s read: the speed is the story. At the start of the year, agentic coding only really worked on Anthropic. By summer OpenAI looked like it was on the ropes. It’s October, and portfolio companies are swapping out models because Sol and Luna are better and cheaper. Dev’s read: after 12-plus years selling to developers, he has learned there is very little loyalty. Developers switch tools fast or run several at once. The stickier revenue is in the enterprise, where you sign contracts, train thousands of employees, and build workflows. Inside MongoDB, his conversations with his CIO are about token costs, token budgets, and who gets the data. Rory’s read: OpenAI grew 18% quarter over quarter from Q1 to Q2, under 100% annualized, while Anthropic was still growing 10x. By his numbers, Q2 was the first quarter Anthropic’s GAAP revenue was larger, roughly $11 billion against $6.8 billion. A 60% to 70% quarter resets that story. It also brings relief, given OpenAI’s forward purchases of compute, data centers, and land in Virginia. Harry’s question: almost every team he talks to has moved a meaningful share of work from Anthropic to Codex, so what happens if Anthropic’s Q3 isn’t a blowout. Rory noted the IPO timing means that number will be in the prospectus before the stock trades. #2. OpenAI Seeks $30B at $1.4 Trillion While Anthropic Preps a November IPO Near $2 Trillion OpenAI is in talks to raise at least $30 billion at about a $1.4 trillion valuation, before the new money, as a bridge to a later IPO. Anthropic is targeting a listing as early as mid-November, with a roadshow the week of November 9 and a valuation discussed at $1.8 trillion to $2 trillion. Rory’s read: this would be the first private round above a trillion dollars. He remembers the first round above a billion in the mid-1990s, three orders of magnitude ago. Harry’s hypothetical: you run directs at CalPERS and can buy OpenAI private at $1.4 trillion or Anthropic public at $2 trillion. Rory’s answer: lean to liquidity. When this much is unknown, you either own the liquid stock or accept that you’re holding for three or four years. #3. Factory Cuts Ties With Adviser Chris Degnan, Who Is Cognition’s CRO Two Hours Later On September 30, Factory CEO Matan Grinberg posted that he had fired Chris Degnan as a board advisor, alleging he shared confidential information with Cognition. Two hours later Degnan announced he had joined Cognition as chief revenue officer. Degnan, a former Snowflake CRO who had advised Factory for about a year, says he resigned and never shared anything. Cognition also rejected the claims. Dev’s read (he disclosed he is an angel investor in Factory and that MongoDB partners with Cognition): it depends what kind of adviser. Someone who vets a hire in Europe is a light relationship, and moving on is fine. Someone inside the tent on product plans, board plans, and financials is different. A CRO has also recruited people on the vision, and jumping to a direct competitor burns that reputation with all of them. Jason’s read: 95% of CROs would say this is how they found their last job. Degnan was an adviser and board observer, not an employee or a director. Many CROs treat competition as a game and are friends with their counterparts at rivals. More time between the two roles would have been better. Rory’s read: a full-time CRO can do the same thing and knows everything up to the minute he gives notice. What failed here was an implied understanding that nobody made explicit. Telling Grinberg up front about Cognition’s approach would have fixed it. Harry’s read: Rory is wrong. An employee sees a silo. A board observer sees strategy across every function. The argument widened from there. Jason said half the frontier-lab talent rotates between labs and that loyalty in the AI era has changed for good. He used to see a “rule of two,” where a departing exec could take two people before the relationship broke. Now he sees eight to ten leave in the first week. Dev disagreed. Recruiting, developing, and holding people accountable are the same problems he had when he started his first company in 1998. Having “dealt with this last week,” he asks a departing executive one question: did it have to be a direct competitor. #4. Vinod Khosla Calls His Own Portfolio Company “a Struggling Second Tier Competitor” Khosla Ventures is an investor in both Cognition and Factory. KV’s Keith Rabois led Factory’s $110 million Series C in April and sits on its board. Vinod Khosla sided with Degnan, accused Grinberg of lying, and wrote that Factory was a “struggling second tier competitor.” Rabois publicly defended Factory. Dev’s read: he was flummoxed. The tweet handed every firm that competes with Khosla Ventures a line to use in deals: is this the partner you want when things go bad. Rory’s read: it was a mistake Vinod will probably regret. Holding two directly competitive investments only works if you say you love your children equally and say nothing in public about either. “It was not venture’s finest hour.” He also recalled a rule about praising in public and chastising in private. Jason’s read: he offered the Pollyanna theory that it was meant as tough love for a struggling founder. Nobody on the call bought it. #5. Reflection Ships Beam, a US Open-Weight Model Claiming 3-4x Less Inference Compute Reflection AI unveiled Beam, its first open-weight model, at 501 billion parameters. The company says it matches Z.ai’s GLM-5.2 on reasoning benchmarks while using three to four times less inference compute. The claims have not been independently verified, and the weights are due this month. Dev’s read (he disclosed Sequoia is an investor): enterprises get to have their cake and eat it. A near-frontier US model at a fraction of the cost removes the hesitation regulated industries have about Chinese open models. A model within six months of the frontier covers about 90% of use cases. He expects Jevons paradox to follow, with lower costs driving more deployment, and he doesn’t see the market as zero-sum. Jason’s read: at Dreamforce, nobody he spoke with wanted to run a Chinese-sourced model. Those who did felt forced by cost. Internal demand for tokens is roughly ten times what these companies can serve. He is skeptical of published evals, which need “three asterisks and four daggers.” He checked OpenRouter mid-show and Beam wasn’t there yet. If parity is real, US models could carry half of open-source tokens in 12 months. Rory’s read: two labs at $70 billion each is $140 billion, more than Microsoft collected across Windows and Office for most of its history. At that size the CFO tells the CIO to cut the token bill. The fix is routing: Jev for the cheap 20%, something like Reflection for non-frontier work, Anthropic for the hard problems. He wouldn’t guess 50% share. Even 10% to 20% would matter, because frontier models take 20% to 30% of OpenRouter tokens and 90% of the dollars. Harry’s read: as a Fireworks investor, his fear was that enterprises avoid Chinese models and go straight to the frontier labs. A credible US open model de-risks the inference providers. #6. ElevenLabs Doubles to $22B in a $300M Employee Tender ElevenLabs closed a $300 million employee tender offer led by Wellington and T. Rowe Price at a $22 billion valuation, double the $11 billion from its $500 million Series D in February. Jason’s investment committee pitch: put 10% of the fund in, in one round. “I don’t need to see it at 50 or 100.” Voice is the workload where model quality is not interchangeable. When a flower shop’s agent picks up the phone, it has to answer in seconds and get the order right. Mid-size companies are paying hundreds of thousands of dollars a year and telling reference callers it’s a good deal. “This isn’t Jev-level risk.” Harry’s challenge: Jason said earlier this year that cost would drive substitution in voice models. Jason’s answer: ElevenLabs covered the low end and the high end with multiple models. He also noted the fund is short-staffed, since “one of our most esteemed partners is going back to run a public company.” #7. Salesforce Buys Three-Year-Old Listen Labs for a Reported $2B Salesforce signed a definitive agreement on September 29 to acquire Listen Labs, a three-year-old startup whose AI agents recruit, interview, and analyze research participants. The reported price is about $2 billion. Listen Labs walked away from a signed $125 million Series C term sheet at a $1.5 billion valuation, and its revenue was estimated at about $30 million annualized. Harry’s math: about $850 million back to investors on $96 million invested. Sequoia led the seed and the A for roughly 25x. Ribbit led the B and made 4x in eight months. Rory’s read: market research is a large, fragmented, pre-technical category, and LLMs do it better. The old survey asks if you liked your stay, hears that you never checked in, and then asks if the bed was comfy. Jason’s read: he doesn’t see how tens of millions in revenue moves the needle for a company approaching $50 billion, and it may be forgotten in two years. The founders’ side is different. If you can sell for $2 billion in the first five years and you aren’t building something generational, take it. He had a company get an offer near this price this year, and every VC at the table said don’t sell. Dev’s read: expect more of these, because many AI app companies are a feature built on someone else’s platform. The durable ones run a data loop, where usage creates data nobody else has and the data improves the product. BladeLogic had offers at every stage, including a meeting in John Chambers’s office at Cisco, and sold to BMC for about $900 million in 2008. When Harry needled that Intercom’s founders must be annoyed, Dev said: “Comparison is the thief of joy.” #8. Agents Now Trigger More Than Half of Vercel’s Deployments, and They Choose the Vendors Vercel says agents went from under 3% of deployments to more than 50%. Harry’s number on the pod was $600 million in ARR with agents driving half of new business. The last public figure was a $340 million run rate in February. Jason’s read: Guillermo Rauch didn’t plan for agents to become the majority of his business. Vercel had the right infrastructure before they arrived. With 20-plus agents in production and 8 to 10 hours a day building, Jason’s experience is that agents have opinions and are hard to argue with. - Resend: it was his first agentic purchase. His agent recommended it after SendGrid setup kept breaking. By his numbers, Resend’s MCP calls went from 106,000 in April to 3 million in September. - Render: he asked an agent live on the show what to use for hosting. Render came first and was labeled “my pick,” then Railway and Fly.io. Vercel came fifth as “not suitable for this.” He barely knew Render. “Render just got a pretty good lead.” - His advice to founders: find 10 people with agents in production and ask them every two weeks what their agents recommend in your category. Dev’s read: for 25 years every marketer optimized for page one of Google. Now agents have to find you and have current, accurate information about you. It is harder for startups, because the corpus of data favors the incumbent. Documentation and APIs have to make you a first-class experience for an agent. Harry’s question: is the category, including his own Peec and Sequoia’s Profound, just next-generation Semrush. Rory’s read: it has more depth. Showing up for agents is the natural next product after showing up in ChatGPT. Someone at MongoDB will spend this afternoon checking that MongoDB comes up first when Jason’s agent needs a database. #9. Two Former Groq Engineers Sue Over Nvidia’s $20B License-and-Hire Deal Two former Groq engineers filed suit in Delaware Chancery Court on October 2. The complaint describes a $20 billion transaction: $17 billion for a non-exclusive license and $3 billion in Nvidia RSUs for the 150 to 200 engineers who moved over. Both plaintiffs left Groq before the deal but still held shares. Rory’s read: the non-negotiable principle of Delaware law is that everyone holding the same security is treated the same. These structures let the buyer reallocate consideration, and nobody allocates to former employees. He expects a settlement, because the class is small. Nvidia took the risk knowingly, in his telling: the lawyers offered a Sunday-night close instead of nine months of review, Jensen asked if there was risk, and said he could live with it. Dev’s read: investors learned to negotiate for odd exits up front. Employees never thought to ask. Founders should expect the question: what happens to my equity if the tech and part of the team get bought and the company doesn’t. Jason’s read: some of these are mergers in substance. If one lawsuit wins, the structure dies, since it already carries double taxation. #10. Meta’s Muse Tops the App Store and OpenAI Answers With Dots Muse reached the top of Apple’s US free iPhone chart within 10 days of its September 8 launch, overtaking ChatGPT. OpenAI unveiled Dots, its always-on agent, on September 29 and put it behind some of its most expensive subscriptions. Harry’s read: Muse was a phenomenal launch. The Dots demo was tech-centric and out of touch. Rory’s read: he had been cynical about Meta’s AI effort and wanted to give credit. Zuckerberg and Alexandr Wang shipped a killer product and resourced it with compute, and shipping cures everything. Jason’s read: Dots is a 24-hour persistent agent tied to Codex. Give developers 30 to 90 days before judging it. Nerdy products make money. Dev’s read: doing one thing well is hard, doing two at once is very hard, and OpenAI struggles with that across its product surface. #11. Oura Pulls a $2.2B IPO That Was Four Times Oversubscribed Oura postponed its $2.2 billion Nasdaq IPO days before pricing. It had planned to sell 50 million shares at $40 to $44, valuing it at up to $15.62 billion, and orders were about four times the shares on offer. Existing holders were selling 36.5 million of the shares, including Forerunner’s entire 9.3% stake, worth roughly $1.2 billion. Revenue for the first nine months of the fiscal year was $1.21 billion, up 74%. Jason’s read: this is terrible news for venture. If a company at $1.2 billion growing 74% with real recurring revenue can’t get out, the “liquidity is back” story has a hole in it. Dev’s read: Occam’s razor says price. The bankers likely told the board one number, the book came in lower, and a holder selling its entire position didn’t help. He stressed this was speculation. Rory’s read (he disclosed a small passive position): it’s the classic banker dynamic. They pitch $50 to win the mandate and the book comes in at $40. Harry’s read: hit the bid. The history of consumer hardware in public markets is Jawbone and Peloton. Dev closed with BladeLogic. It priced in July 2007 as the debt markets started to wobble, and got out. Then: “I don’t mean to be rude, but I got to go sell some software.” Quotable Moments Dev Ittycheria “I was just totally flummoxed by that tweet. It basically handed every other firm a weapon when they’re competing on deals: is this the partner you want when things go bad?” “I’ve never seen a CRO flip from one company to a direct competitor. The reputational damage you can do long term is going to come back and bite you.” “If you’re missing from an AI-generated answer, it’s like being invisible on Google.” Jason Lemkin “In the AI age, the definition of loyalty has changed. I think it’s changed permanently.” “When agents pick you, it’s a force of nature. My agents have opinions, and they’re really hard to argue with.” “Just because you say it’s not a duck doesn’t mean it’s not a duck. Some of these are ducks.” Rory O’Driscoll “You were all in for team A, and you didn’t go to another sport. You went to team B in the same damn league.” “I don’t care that it’s easier to just use Anthropic, dude. We need to shave $3 million off this token bill.” “Jensen said, ‘Is there some risk?’ They said, ‘Yeah.’ He said, ‘I can live with it.’ Done.” Harry Stebbings “When we hear the word loyalty, I just think of Vinod Khosla.” “ABBA’s answer to Justin Bieber comes along and three years later sells for a comparable number, and you’re like, oh bugger, there was an easier route, wasn’t there.” “I would have hit the bid. Look at the consumer hardware companies that have gone before, from your Jawbones to your Pelotons. Hit the bid at 13 and take it out already.” This post is part of the ongoing 20VC x SaaStr collaboration with Harry Stebbings and Rory O’Driscoll.