Vivek Vaidya has sold two companies — one to Microsoft, one to Salesforce — and now runs a venture studio that starts about three new companies a year and expects 10 to 15 out of a fund.
Most advice to technical founders is about building the product. His is the reverse: get comfortable selling something that does not exist yet, because the enterprise procurement cycle gives you enough time to go and build it.
"Building technology is easy in the grand scheme of things. Building companies is hard, right?"
Vaidya was the first engineer his co-founder hired at Rapt, then started Crux with him, then ran Salesforce Marketing Cloud's engineering after the acquisition — a jump from 80 or 90 people to almost a thousand overnight.
I listened to the full interview so you can skip it. 44 minutes of audio, 16 minutes of reading.
Here are the 12 takeaways that matter.
👤 Guest: Vivek Vaidya, founding general partner of the venture studio super{set}, previously chief technology officer of Salesforce Marketing Cloud and a co-founder of Rapt, which Microsoft bought, and Crux, which Salesforce bought
🎙️ Host: Nick Moran, of New Stack Ventures, who has run The Full Ratchet for more than 500 episodes
📰 Published: 31 August 2026 on the show's own feed
🟣 Apple Podcasts | 🔗 Episode page | ⏱️ 44 min | ✅ Time saved: 28 min
Key Takeaways
The hardest commercial skill is selling a product you have not finished
He says the enterprise contracting cycle buys you the time to build it
A customer naming a problem is not enough; ask what they already tried to fix it
The signal is a workaround held together with tape and glue, because maintaining it hurts too
A finance chief will not buy a bill that changes every month, whatever the technology costs you
His example of the answer they want: $10,000 a month, or $60,000 a year
He expects inference prices to keep falling, on the evidence of cloud storage
His team ran a leave-the-cloud analysis every few months for three and a half years and never got close
The AI acquisition multiples price adoption, and most AI products will not have Cursor's
Data only becomes a moat if you built the contract and the product to earn it
He wrote data rights into Crux's customer contracts, and says it moved the acquisition economics
His answer to the will-a-frontier-lab-build-this question is about staffing, not capability
Your A team against Google's C team, and he says he will take those odds
When anything can be built, the moat is execution, not technology
Never send personally identifiable information to a frontier model, and check the retention settings
1. Two ways super{set} starts
Vaidya described super{set} as a venture studio with two ways of forming a company, six or seven years into a model he says was still rare when they began.
The first model starts with a founder and no idea. Someone arrives saying they want to build a company and has domain expertise; the studio works with them to shape the idea, test whether the problem is real, and build around it
The founder is the chief executive and runs the company. The studio's job at that point is to judge whether the idea is good, whether the team is right and whether that person is the right chief executive
The second model starts with the studio's own thesis. super{set} hatches the company itself, then brings people in from outside once it is real
Sometimes a chief executive is never installed: at Catch, his co-founder Tom Chavez is still chief executive, and at Kana, launched recently, Chavez is chief executive and Vaidya is chief technology officer
At Habu they brought in Matt Kilmartin after the two of them had held the roles for a couple of years
The studio funds the first stage, so an early chief executive is not needed to raise money. That is what lets the second model run without one
The differentiator he claims is the equity split. Founders working with super{set} get founder-level equity and autonomy, and he described the studio as helpers and guides rather than a firm that dictates how a company is built
Deep operational involvement caps the volume. About three new companies a year on top of the existing backlog, and 10 to 15 over the life of a fund
2. Be clear why you are selling
Moran asked what Vaidya has learned about selling companies that founders should hear.
The first distinction is who initiated it. "Be very clear about why you're selling or why you want to get acquired, right?" followed by: "Because it's two different things: being bought versus being sold, right?"
Crux was sold from a position of strength: things were going well, and the company had become strategic enough that the buyer could not afford not to acquire it
The other case is wanting out after six or seven years, which he said carries no shame as long as the founder is honest about it
The second thing to settle is what happens after the deal. Whether the founder stays, what happens to the team, what roles people get — worked out in their own head first and then discussed with the acquirer during the negotiation
His reasoning for that is expectation management. A founder who has not thought it through will carry expectations into the deal that are not met afterwards
"And as I like to say, expectations reduce joy."
On timing a sale he declined to give a rule and described the arithmetic instead. The founder owes investors a return, so the calculation is the certain offer today against the expected value of a much better outcome in three years, discounted by the chance of getting there
Team fatigue belongs in that calculation too. The founder may not be tired; the team may be, and the founder cannot do it alone
3. Tech is easy, selling is not
Asked how his approach to company building has changed across being a chief technology officer, a founder and now a studio partner, Vaidya gave the view he says he has held for years.
"Building technology is easy in the grand scheme of things. Building companies is hard, right?" The rest is hiring, getting results through people, the market cooperating, the right product, and the systems and processes to do all of it
What has changed is how much faster a first product can be built, which moves the constraint to sales. His term for the skill: "you have to embrace this notion of living in the declarative present tense"
In practice that means selling tomorrow's technology today — telling a customer you have solved their problem once you have worked out how to solve it, not once you have built it
The cover is procurement. By the time an enterprise gets through negotiation and contracts, there is enough time to build the thing
"So selling ahead, getting comfortable in selling what's not on the back of the truck is become even more important now than it was five years ago"
4. Do not lead with the tech
Moran asked how a technical chief executive with little commercial experience builds that muscle.
Vaidya said he had the same problem and named the failure exactly. "Oh, look what cool features we have, and look, we've used this cool technology to build our product. Nobody cares."
The customer does not care which model is underneath — he listed Anthropic, OpenAI and Kimi as interchangeable from the buyer's point of view
The sequence he prescribes is problem, value, then technology. "First, get alignment on the problem that you're solving and why it matters to the customer and what value it will provide before leaning in with the technology."
Technology has one job in the sale, and it comes later. When the customer asks whether they could do this themselves, or how this compares with a competitor, that is when the technical differentiators and moats come out
5. Look for the duct-tape fix
Moran quoted Vaidya's own line — that solving the right problem matters more than solving a problem well — and asked how to know which problem to solve first.
The opening question is the standard one. In customer interviews, whether for design partners or early customers, super{set} asks what keeps you up at night, or what problems the customer would make disappear with a magic wand
The answer to that question is not the signal. An executive will name several things that are simply on their mind, which he said is not enough
The follow-up is what does the work: what have you already done about it? "And you're looking for problems that keep them up at night, for which they have gum tape and glue solutions already"
The workaround itself becomes a second pain point, because someone has to maintain it
The reason that combination matters is that it proves two things at once. "That's very important because now they're telling you two things: one, it's important, and two, I've actually tried to do something about it, and it hasn't worked, right?"
6. CFOs want a fixed number
Moran asked how to engineer the cost structure of an AI-native product, where the underlying costs are usage-based.
Vaidya said the industry has not solved this and then gave the constraint that decides it. "CFOs hate unpredictability." followed by: "They want predictable cost."
"If I'm selling something to them, they want to know that every month it's going to cost them $10,000 a month, or every year it's going to cost them $60,000 a year."
The failure mode is the honest answer: if month one costs $5,000 and month two costs $20,000, "CFOs hate that, right?"
So the founder's job is to present a near-fixed price on top of usage-priced inputs. That means engineering the system to make the margin work rather than passing the variability through
The first lever is that customers do not consume evenly. One customer may hold 10 terabytes of data and another six gigabytes, and a wide spectrum of customers is something a pricing model can exploit
The second is buying compute opportunistically. At Crux they ran batch data processing on spot instances — spare cloud capacity sold cheaply and reclaimable at short notice — which he said extracted more from the cloud at a positive margin
The third is not defaulting to a single model provider. "You don't always have to use Anthropic, but use it. Use it if it's giving you the quality and the value you you're after, but you don't always have to."
What that requires is instrumentation: tracking cost, and having the switches to route work to a different provider depending on the situation
7. Inference prices will fall
Asked whether falling inference costs will expand margins or be eaten by rising demand, Vaidya said he does not know and told the cloud story instead.
Crux went to Amazon Web Services in early 2010, when it was the only real option, and stayed. For the first three years he had to defend that choice at nearly every board meeting against building a data center
The company kept waiting for the price rise that never came. "And Amazon just kept lowering prices."
"Storage was the perfect example. S3 store cost just kept going down and down and down and down, right?"
They checked the alternative repeatedly and it was never close. For three and a half years his head of infrastructure ran a migrate-off-the-cloud analysis every three or four months, and the numbers never came near making a move worthwhile
He expects inference to follow the same path. "I think the same thing is going to happen with inference as well. Costs are going to come down. People are going to build more."
His mechanism is volume: the hyperscalers, the intermediate inference providers and the frontier labs will see enough usage to keep cutting prices and hold their revenue anyway
8. Multiples price adoption
Moran said the market has not reached consensus on how to value AI businesses. Vaidya agreed and said he has no real view, then explained the one pattern he can see.
The high multiples so far are being paid for adoption, and developer tools are the clean case. "Take Cursor for example. Great acquisition, very happy for the Cursor folks. But the reason for that high multiple is adoption."
As the number of developers rises, demand for a product like Cursor rises with it, which he called obvious
That logic does not transfer to most AI products. His counter-example was agentic finance: the population is not expanding, because not everybody is about to become a hedge fund analyst
One recent deal he cannot explain. "OpenRouter is another great example. Stripe acquired them. Great, very good outcome for OpenRouter, but not completely clear why they paid that much for OpenRouter."
His forecast is the usual shape. "There's going to be hype, and then things are going to kind of come down to some accepted normal in the next three to five years."
His warning to anyone holding these positions was blunt. "But just know that the music is going to stop, and you don't want to be left without a chair, when the music stops, right?"
9. Good data is the real moat
Moran put it to him that investors have argued data is a moat for years and it did not hold in earlier machine-learning cycles.
Vaidya pushed back on the premise using his own company. At Crux they architected both the systems and the customer contracts to give them access to anonymized data they could use to build new products and improve existing ones, and he said that played a decent role in the economics of the acquisition
The hard part is the incentive, not the engineering. The product has to have features that make the customer see the value in the exchange, which he said sounds easier than it is
What has changed is what the data is for. Enterprises have been told for years that first-party data is key, and used it to improve the business; now it has to improve the operations
"If you have crappy data, doesn't matter what kind of AI you put on top of it won't work, right?"
He hears the same objection in nearly every enterprise conversation, and gives customers three options. Enterprises tell him their data is not ready or not high quality. He said that either means they are not interested in solving it, or they are afraid of what an assessment would show, or they try it, find out and fix it
The connection to cost is direct. Fine-tuning small or open-weight models on your own data to get a proprietary model only works if the data is good
"The data does become your moat, if you will." He added that he does not know how many companies are doing it well right now
10. Your A team vs Google's C
Asked how he decides which companies a frontier lab might build itself, Vaidya said the question is old and the answer has not changed.
The labs employ some of the smartest people alive, and so did Google and Oracle before them. Startups have been asked why the incumbent could not just do this for decades
"The answer to the can question is almost always yes. The question is, will they? Should they? Does it make sense for them?" And after that: whether they will put their best people on it
"Your A team is competing with the C team of Google, and I'll take those odds any day."
His second argument is domain pain. A founder who has stood in the customer's problem for years has expertise the labs do not, and the people who want to build companies are not the ones taking jobs at the labs
The counter-example answers itself. "Like, why do the Frontier Labs use Slack, for example?" If building anything were the whole story, they would not buy software — the answer, he said, is that maintaining and enhancing it is where the difficulty lies
11. Execution is the only moat
Moran raised Whisper Flow, a dictation tool he uses and likes, and said he struggles to see the defensibility: others have built the same thing, and it is native to what the chat products already do.
Vaidya said defensibility is being measured on the wrong axis. Assume anything is easy to build, and then: "The moat really is execution. Are you maniacally focused on execution, right?"
Moran's own experience was the evidence: he was onboarded free, graduated to paid, and the time to a first useful result was fast
What a large lab will not give a customer is attention, and he used himself as the example. "I pay 200 bucks a month for Claude. It gives me a crappy answer, which is wrong. I can't do anything about it."
The second thing a lab will not build is the boring surface coverage. Moran noted the tool works across everything he uses; Vaidya said "Anthropic is not going to take the time to build that seamless experience across all of the surfaces where you work"
12. Never send PII to a model
Moran asked how Vaidya feels about founders sending data to closed frontier labs, and when to start protecting it.
Start with ownership. "So if you're a founder building in B 2b, you don't own the data. Data is the customer's data, right?"
Then accept that frontier models are unavoidable today. He said nobody can currently take the risk of building enterprise software only on open-weight models, and that the industry may get there but is not there yet
The controls he named are specific. "Make sure you have all the zero data retention policies and boolean flags and everything else." — and the hard rule: "You should not be sending your PII or a customer's PII out to the frontier models."
There are tools that prevent it, and he said to use them
He does not think the labs are singling anyone's data out for training, on the grounds that it would be too much of a risk for them, and said he does not worry beyond the steps above
On open source he is a supporter with a warning about the switching cost. Swapping models is not a one-line change: it needs real evaluation infrastructure and a decision about which problems go to which model and which inference provider
"I'm starting to use Kimi K3 instead of Opus five"
On five- and ten-person startups claiming they will run their own inference servers: "Bullshit. I call bullshit on that."
"So you're going to be using Baseten or Nebius or Fireworks or any of these providers that are out there." His point is that the provider's ecosystem becomes something you have to understand too
Bonus Insights
The studio's own track record is the credential behind the advice. Habu, from the first batch, was acquired by the data clean-room company LiveRamp, which has since been acquired by Publicis; Catch, a privacy company from the same batch, is past its Series B
Vaidya and Tom Chavez have worked together for 26 years, since Chavez hired him as the first engineer at Rapt
The guest he would most like the show to book is Demis Hassabis, to talk about the DeepMind build-out and the Nobel Prize. He had just finished The Infinity Machine, a book about Hassabis and DeepMind, and recommended it
Asked for a secret weapon, he said he has none and named two habits anyway. He makes his own tea and reads a book for the first half hour of every morning
He has worked with a personal trainer for two and a half years, having spent his younger years wondering why anyone would need one: "And now, having worked with a trainer for the last two and a half years, I I'm like, what an idiot you were back then!"
He told listeners to reach him on LinkedIn or by email, and offered more book recommendations
Vaidya's bottom line is that in a world where the product is cheap to build, the durable advantages are commercial rather than technical: selling before the thing exists, choosing a problem the customer has already tried and failed to fix, pricing it so a finance chief can sign it, and out-executing companies that will never put their best people on your market.
Products, Companies & Tools Mentioned
super{set} (His venture studio: two formation models, about three new companies a year, 10 to 15 per fund, with founder-level equity for the founders it backs)
Salesforce and Microsoft (The two acquirers in his own history — Rapt went to Microsoft, Crux to Salesforce, after which he ran Marketing Cloud engineering)
LiveRamp and Publicis (Habu, from super{set}'s first batch, was acquired by the data clean-room company LiveRamp, which Publicis has since bought)
Amazon Web Services (The cloud Crux adopted in early 2010 and never left; the repeated price cuts are the basis of his forecast for inference costs)
Cursor (His clean example of an AI acquisition multiple justified by adoption, because the developer population keeps growing)
OpenRouter and Stripe (The acquisition he says he cannot explain the price of)
Anthropic, OpenAI and Kimi (The model providers he says a customer does not care about, and that a founder should not be locked into)
Baseten, Nebius and Fireworks AI (The inference providers he says a small startup will use instead of running its own servers)
Slack (His counter-example: the frontier labs could build it and buy it instead)
Whisper Flow (The dictation tool Moran raised as a defensibility test; Vaidya's answer was that its moat is onboarding and cross-platform coverage, not the technology)
Books & Resources Mentioned
The Infinity Machine (The book on Demis Hassabis and DeepMind that Vaidya had just finished and recommended; he also named Hassabis as the guest he would most like the show to book)
If this was worth your time, send it to someone closer to the industry than you are.
Get the latest market chatter as it happens: