HubSpot’s Dharmesh Shah: AI Won’t Kill SaaS — It Just Made the Problem Set Bigger

Add preferred source
HubSpot co-founder and CTO Dharmesh Shah argues the “SaaS apocalypse” thesis gets the economics backwards: AI is not a new energy source but very smart software, and it hands builders a far larger set of solvable problems rather than shrinking the pie. In an a16z interview he lays out a simple test for defensible startup ideas — append “yet” to your differentiator, and if the claim still holds it’s a red flag — and separates the labs’ *structural* constraints (regulation) from their *incentive* constraints (a $3B+ horizontal CRM is off their yellow brick road because their ambition points at monetizing inference tokens). He details how HubSpot itself is being re-architected into two stores — a system of record for structured facts and a system of context, likely a graph, for unstructured data — with evals as the guardrail, plus an internal startup arm, HubSpot Next, whose current project is Uspot, a CRM for solopreneurs. That market is the fastest-growing customer segment: 2025 saw roughly 5.6 million new U.S. company formations, and the Census predicts over 70% will never hire an employee within two years. For investors, the read is that AI-native software’s risk is not a single-supplier margin squeeze but distribution — and incumbents with data, distribution and evals are re-platforming, not being displaced.

Key Elements

HubSpot made its name selling marketing software to small businesses. Twenty years later its co-founder, Dharmesh Shah, is making a significantly larger wager: that the same structural logic that built a $3 billion-a-year company applies to AI — and that most founders trying to get rich off OpenAI’s blind spots are fishing in a pond that will be drained within a year.

His rule is almost insultingly simple, and he laid it out on the a16z podcast. If you can append the word “yet” to your differentiator and the claim still holds — “ChatGPT can’t do that yet” — that is not an opportunity. It is a warning. “If you can just append the word ‘yet’ to it and it still holds water, that’s a problem,” Shah said. “Measure the product versus the company. ChatGPT can’t do that — okay, but can OpenAI the company do that? Maybe it’s a new product line.”

The distinction matters because it splits the reasons a frontier lab stays out of a market into two very different buckets. A structural constraint is something a lab legally cannot cross. An incentive constraint is something it could do tomorrow but almost certainly won’t, because its ambition points somewhere else. Shah’s example: Sam Altman could wake up and decide to take HubSpot’s $3 billion-plus in CRM revenue, and OpenAI has the talent and the ambition to try. But “it’s that very ambition that makes it highly unlikely.” The labs are racing to monetize enormous capital spending by selling inference tokens, not chasing domain-specific sales software. Stay a little further off that road than strictly necessary, he advises, because the road is widening and its expansion is fairly predictable.

That framing is the counterweight to the “SaaSpocalypse” narrative that has dominated the past two years — the fear that agents and chat interfaces hollow out application software the way the browser once threatened desktop apps. Shah’s answer is to reframe AI entirely.

The hottest AI take of 2026: AI is just software

“AI is not some new energy said. It sounds almost deflating. It is actually the bullish case

His historical parallel is HubSpot’s own founding pitch, which he describes as the thing that made him laugh all the way to the bank. “Every line item on the P&L has a 100 plus billion dollar software company attached to it except marketing — because marketing was historically arts-and-craftsy,” he told investors then. When HubSpot was raising money, venture capitalists told him no marketing software company had reached billions, and that HubSpot had two strikes against it: small business, and marketing. The internet turned marketing into a problem software could solve.

“Before, there was a class of problems software wasn’t equipped to solve. Now there’s a massive set of new problems — so the opportunity set is way larger,” he said of the AI shift. Some software companies will die, just as some failed to make the cloud transition. But in aggregate, “if you’re anywhere near software, have the agency, and are entrepreneurial, there’s never been a better time to build a software company.”

He is also arguing that the supply-chain risk that worried builders in 2023 has largely dissipated. Then, the labs looked like they held the only magic elixir and a single critical ingredient. Today models are trending toward infrastructure: open-source alternatives, multiple commercial vendors, and routing layers like OpenRouter that let a company swap models per task. Shah’s prediction is that this fungibility holds — meaning an AI-native software company need not fear being margin-compressed by a single supplier, an important point for anyone valuing these businesses on long-run gross margins.

What HubSpot actually learned — and it wasn’t in the plans

The CRM platform, Shah is careful to point out, was never in the original deck. He and his co-founder met at business school on the East Coast and picked small business for a reason that has aged well: it combines the enterprise virtue of being able to charge money with the consumer virtue of scale and no revenue concentration. Their first idea, entered into the MIT $50K competition, was ERP for small business — “an Oracle for small business.” They then cut the aperture down to ERP for law firms, an imaginary “LegalSpot,” and learned in customer conversations that lawyers cared far more about finding clients than about tracking time.

That detour produced the marketing software, and a deliberate choice that ran against the standard startup playbook. Where most advice says build one thing well, HubSpot shipped a broad suite — CMS, SEO, landing pages, forms — because the customer’s real problem was the fragmentation. Its first product was deliberately a bundle.

The CRM decision is the more instructive one. Five or six years into the business, a couple of years before the IPO, HubSpot surveyed its paying customers: 20% used Salesforce, 20% used another CRM, and 60% used no CRM at all. Shah’s read was that these were companies paying HubSpot to generate leads, so where were the leads going? A CRM felt overwhelming and expensive. HubSpot built “the easiest CRM ever conceived” and made it free — not freemium, free — subsidized by the marketing business. “Create customer value before you try to extract it,” he said. There was also a defensive motive: the incumbent CRM players were likely to move into marketing, so HubSpot needed the adjacent product regardless.

Both lessons — that the platform was an accumulation of wedges, not a master plan, and that a customer’s fragmentation is your product opportunity — are the ones he now applies to AI.

The interface is a means to an end

The question every SaaS investor is asking in 2026: if customers reach HubSpot’s marketing engine through ChatGPT or Claude, doesn’t the app — and its pricing power — evaporate?

Shah reframes it as a spectrum. At one end sits DoorDash or Uber Eats: the app itself is trivial, and the value is the restaurant network and the logistics behind it. At the other end are products where the interface is the product. Most business applications sit at the DoorDash end, because marketers and salespeople are trying to get something done, not to admire a UI. “The app is trivial. What’s actually valuable is the network they have… the actual thing that’s valuable is the work that’s happening. The interface is a means to an end,” he said.

His analogy is that ChatGPT and Claude are “the new browsers.” When software moved from client-server to the web, Google did not end up owning the application stack, and a generation of companies built on top of the browser. HubSpot was, by Shah’s account, the first CRM company to ship a ChatGPT connector — more than a year ago now — followed by an Anthropic connector. He expects the labs to keep enriching their own surfaces, citing OpenAI’s Dev Day announcements of a marketplace and plug-in architecture, and says that makes sense to him. The advice to B2B builders: don’t overthink the interface; find the lowest-friction, lowest-latency path to the customer’s outcome.

He adds a technical rule with a sharp edge worth internalizing: “Anything that can be solved with a structured query deterministically, solve with a structured query deterministically. Don’t use inference to solve something that a query can solve.” If the question is how much revenue came from Ireland, you should not trawl eight million Slack messages and emails and throw a 300-IQ model at inferring the answer. That should be a database query. It is a good proxy for where AI budgets will actually flow, and where they will get wasted.

Two stores, and the evals that hold it together

Shah’s account of how HubSpot is re-architecting itself is the most concrete part of the conversation, and it is a useful template for judging every software incumbent facing the agent era.

HubSpot’s original architecture was a set of core primitives — an operating framework on which marketing automation was built — that could be recomposed. When the company entered sales and CRM, its marketing-automation primitive was reusable, so the sales product launched with more workflow capability on day one than rivals had accumulated over years. He calls this “compounding loops all the way down.”

For the next 20 years, he argues, the framework needs two stores side by side.

The system of record holds known facts — relational or otherwise — and Shah will “happily debate” anyone who claims it is obsolete. The system of context holds the unstructured data that always existed inside customer organizations but could not be used meaningfully at any price a decade ago; he expects it stored in some kind of graph, and he uses the terms “second brain” and “context graph.” Combine the two and you feed inference and agents that actually know the business.

What he says incumbents underappreciate is the eval advantage. Because GitHub is a system of record for code, and because HubSpot maintains evals (formerly unit tests) that measure whether a change still produces the right outcome, the company can branch to an alternate universe — “what if we had made a different primitive choice” — spend a defined token budget replaying that history, and check the result against the eval suite. The same machinery lets HubSpot run its eval battery against each new model release and trade quality points for latency or cost.

That is the less-discussed half of the AI moat debate: not the model, not the interface, but the ability to measure whether an agent is doing its job. It is also why the “model vendors will eat the app layer” thesis has a hole in it for large, workflow-heavy businesses.

A startup inside the startup — with the liabilities removed

Shah runs an internal innovation arm called HubSpot Next, structured as a startup within a startup. The model dates back to the CRM build, when the team was cordoned off: no requirement to use HubSpot’s stack, brand or rules — just solve the problem. Projects target a new product category, a new market, or a new technology. The point is degrees of freedom; experiments run at the pace he wants only if the organization’s baggage is set aside.

His framing of incumbent disadvantage is precise. The problem is not that startups have something large companies lack; it is the weight of the organization. Shed the liabilities and you keep the assets — distribution, customer relationships, evals, and 20 years of code as data to inform the next product.

The current project is Uspot, a CRM for solopreneurs, which was exactly 30 days old at recording and which Shah says he works on every night. He describes the absorption mechanism as “a little bit like an aqua-hire”: if a project works, its pieces get pulled into the mothership. The prior example is Agent.ai, a marketplace for AI agents plus an agent builder for non-technical people. Shah says half the idea was right — and the agent-builder half belonged inside HubSpot with the full team behind it, so it was effectively acquired internally and now lives as a HubSpot product.

There is an ecosystem dimension here too. Shah has been a consumer of developer platforms most of his career and describes the common failure mode: a platform courts developers, then “stomps over” them by closing APIs or becoming developer-hostile. His example is early Twitter, where an open ecosystem let “a thousand flowers bloom” before it closed down and became “almost developer hostile.” HubSpot’s approach, by his account, is to be the platform he wished had existed — an open book about where the white space is and which territories it is “highly highly unlikely” to enter. He calls the company “warm-hearted red-blooded capitalists”: everyone is there to make money, and trust comes from transparency about where.

The operating metric is borrowed straight from product. Just as HubSpot thinks about time-to-value for customers, it thinks about time-to-joy for developers: how fast someone gets something working on the HubSpot APIs, and how fast they make their first dollar in the ecosystem. Building an ecosystem, he said, “is a lot like building a product. You have to think of it that way.”

The solopreneur math behind the bullish case

The most investable number in the conversation is not about AI at all. It is about company formation.

Shah cites U.S. Census data: 2025 saw the most new company formations in American history, roughly 5.6 million, approaching 6 million. By August 2026 the count had already passed 4 million, putting the year on track for another record — so the 2025 spike was not a one-time pandemic-era echo. More striking, the share of new companies the Census predicts will not hire an employee within two years is now over 70%, the highest on record and rising.

Shah says Uspot’s early signal surprised him: willingness to put a credit card down on a consumption-based model was near-instant, making it the most successful launch on that measure of any HubSpot Next project. He deliberately monetized early to learn pricing and value lessons, not to cover cost of goods.

Whether one-person businesses become a durable software category is the open question. The historical knock on small-business software was that acquisition costs ate the revenue. Shah’s rejoinder is that Uspot is cheap to build and cheap to learn from, and that HubSpot itself is the existence proof that small business can support multi-billion-dollar software.

Iteration speed is the strategy — and distribution is the bottleneck

His advice on picking ideas is deliberately anti-analytical. Founders treat idea selection as rigorous analysis — “an 18-column spreadsheet” fed by Census data on 18,000 law firms — but it is nearly impossible to know an idea’s quality in advance. The only real test is to ship it, and even successful companies rarely resemble their original idea; the gene pool survives, the shape changes.

He names distribution as the thing that has gotten harder, precisely because building is easier. More people will build, so cutting through the noise is the binding constraint. His prescriptions: narrow the aperture early, and don’t open with a $100 billion TAM slide — answer what you will do in the first 30 days and how you get to the first dollar before the hundredth. Winners have the shortest iteration time and the cheapest learning time, and AI compresses the feedback loop itself: lessons that once took six months to two years can now be learned in a fraction of that.

Two tensions are left open. Shah’s own advice to “do one thing really well” is one he consciously violated at HubSpot, and his justification — that fragmentation was the customer’s problem — is a claim about a specific market at a specific moment, not a general rule. And the “incentive constraint” test depends on reading the labs’ ambitions correctly; if a lab decides a vertical is strategic, the moat he describes is a prediction, not a guarantee.

For investors, the implication is a reframing of where AI risk actually sits. The existential fear — that a foundation model vendor absorbs the application layer — is probably misplaced for large workflow businesses; the real pressure is competitive, from a much larger cohort of builders and a distribution channel that keeps getting noisier. The incumbents worth watching are the ones re-platforming around context and measurement, not the ones merely bolting a chatbot onto an old UI. The concrete markers to track: Uspot’s trajectory past its first 30 days, whether the context-graph primitive ships into HubSpot’s core platform, whether the one-person-company share keeps climbing past 70%, and — the tell for the whole thesis — whether model pricing keeps behaving like infrastructure rather than like an elixir. Shah’s own summary of the moment is that the best day he has had at HubSpot was about a week before recording. “Things are starting to fall into place,” he said — twenty years in, with no direct reports, and by his account the best of both scale and startup.

Full content available at:Where Startups Can Win Beyond OpenAI and Anthropic | HubSpot’s Dharmesh Shah

• State Street CEO Ron O’Hanley: AI Investment Backed by Real Orders, Hard to Call It a Bubble at This Stage
• ByteDance Tightens External AI Tool Access, Bans L4-Level Code Repositories and Feishu Documents
• Hayes’s AI Chain FLOP Opens Developer Testnet, Token Not Yet Issued
• AI Stocks Take a Beating, but Morgan Stanley Says ‘Selloff Is Overdone’: Data Centers Are Delayed, Not Canceled
• Blackstone, Goldman Say AI Boom Needs Mega-Fundraisings as Spending Tops $730 Billion