Topics Pioneers Insight Method Author
Back

Where Startups Can Win Beyond OpenAI and Anthropic | HubSpot’s Dharmesh Shah

·The a16z Show

Rendering the long image…

Topics Commercial Pitfalls of AI-Native Applications AI Inference Gateways and Intelligent Routing Autonomous Agents Reshaping Enterprise SaaS Legacy Software Buyouts and AI Overhauls One-Person Companies via End-to-End AI
Pioneers Dharmesh Shah

Summary

  • Shah’s startup filter is whether an opening survives the word “yet.” A missing ChatGPT feature is vulnerable as the labs’ “yellow brick road” widens; founders must assess OpenAI and Anthropic as companies, including products they could launch. His proposed shelter is structural or incentive constraints: pursuing HubSpot’s more than $3 billion revenue base is technically feasible, but extremely unlikely given the labs’ broader ambitions, in his view.
  • Model interchangeability and multiple options reduce the risk of building an AI application business. Shah says that in 2023 he feared the labs were so far ahead he did not know whether others could meet the capital needs; “Fortunately, I was wrong about that.” He stops short of calling models commodities, but OpenRouter’s easy model switching, task-specific routing, open-source code and commercial products reduce dependence on one supplier.
  • AI can own the interface without automatically capturing the underlying application’s value. Shah compares business software to delivery platforms, where fulfillment and relationships matter beyond the app: “The interface is just a means to an end.” His priorities are customer outcomes, cost, latency and friction—including using database queries wherever inference is unnecessary.
  • Faster coding increases the importance of distribution, customer focus and inexpensive learning. More software means more competition for attention, while spreadsheets still cannot establish an idea’s quality in advance. Shah’s rule is that “whoever has the shortest iteration time and the lowest learning cost wins”; a $100 billion ambition must translate into a concrete first 30 days.
  • Existing software, structured records and evaluations remain useful assets in an AI rebuild. Shah envisions a “second brain” combining unstructured context with persistent facts, supported by reusable software primitives. His hypothetical $200,000 token-funded rewrite illustrates how incumbents might test alternative architectures against accumulated evaluations.
  • HubSpot is using internal startups to explore new markets and absorb useful components. Its 30-day-old CRM for solo entrepreneurs tests customer-specific context and agents with fewer organizational constraints. The precedent is agent.ai: Shah concluded “half of the idea was right,” and HubSpot incorporated its agent builder while recognizing the horizontal marketplace’s exposure to OpenAI.
  • Solo businesses give Shah an expanding customer market and an early willingness-to-pay signal. He cites a record roughly 5.6 million new U.S. companies in 2025, approaching 6 million, with more than 70% of new companies not expected to hire an employee within two years. Early credit-card adoption for his solo CRM has exceeded his previous HubSpot projects in that respect; his broader argument is that AI expands software’s addressable problems, making “the opportunities … greater now than ever.”

Deep dive

1. Faster building makes customer learning the scarce advantage

  • For Shah, AI continues software’s progression toward higher abstraction, from machine language through C and C++ to tools that shorten the distance between an idea and implementation. The underlying business problems remain. After more than 30 years building commercial software, he calls this “the most interesting development experience I’ve ever had.”

  • Idea selection remains empirical: no “18-column table” can establish quality in advance, and successful businesses may retain only the “gene pool” of their original concepts. AI makes shipping, testing with customers and rejecting weak premises cheaper. Lessons that previously took six months, a year or two can arrive much faster.

  • Schmidt asks how faster development changes the value of starting narrowly. Shah argues distribution becomes harder as software supply expands, making focus more valuable for reaching customers and learning. A $100 billion ambition still needs a credible first 30 days: “First, income, and then we’ll reach 100 billion.” His rule: “whoever has the shortest iteration time and the lowest learning cost wins.”

2. HubSpot discovered its platform through unfinished customer problems

  • The original SMB thesis combined enterprise willingness to pay with consumer-scale customer counts. Enterprise software meant long sales cycles and customers steering the roadmap; consumer businesses had bimodal outcomes and often needed advertising subsidies. SMBs offered paid problem-solving without revenue concentration. The first concept, entered in the MIT $50K competition, was an “Oracle for small businesses,” narrowed to ERP for law firms under LegalSpot.

  • Customer conversations redirected that plan. Law firms cared less about time tracking than websites and marketing: “Can you help me find more clients?” Shah says choosing between SMB ERP and SMB marketing rested on limited evidence—almost a coin flip, though not literally one. Becoming a CRM platform appeared in none of the original presentations.

  • Five or six years after founding, roughly two years before the IPO, a customer survey found 20% using Salesforce, 20% another CRM and 60% none. These were paying lead-generation customers, exposing an unfinished workflow. Complexity and expected implementation expense explained non-adoption. HubSpot built the simplest CRM it could for that 60%, subsidized by marketing revenue: “Not freemium, but free.”

  • The move also anticipated CRM vendors entering marketing. Shah’s qualification to “do one thing very, very well” is to solve the customer’s whole problem. HubSpot bundled content management, SEO, landing pages and forms because fragmentation itself frightened SMBs. Schmidt’s observation is that the apparently deliberate greenfield strategy emerged from discovering customers lacked those systems; winning CRM subsequently required becoming a platform.

3. Developer trust belongs inside the product strategy

  • To Shah, Twitter illustrates platform risk: an open API ecosystem supported clients and companies, then became unfriendly or hostile to developers. HubSpot tried to be “the platform we dreamed of when we were doing things ourselves.” Its approach includes candid communication about what it is building, where partners might find openings and where HubSpot is unlikely to go, without promising never to compete.

  • An ecosystem needs product design, with developers treated as customers. Shah’s metric is “time to satisfaction”: shorten the path from getting something working with the HubSpot API to “I made my first dollar working in the HubSpot ecosystem.” The developer experience includes both technical success and reaching revenue.

4. The labs leave openings where their incentives diverge

  • Schmidt’s “yellow brick road” looks inviting until “a whole tank unit” of frontier labs races down it. Shah adds two tests: can “ChatGPT doesn’t do that” become “ChatGPT doesn’t do that yet,” and are you evaluating today’s product or the company behind it? OpenAI could launch new product lines, widening the road. All else equal, he prefers opportunities farther away.

  • His proposed boundary is structural constraints, including regulation, or “incentive constraints.” He discloses being both HubSpot’s co-founder and an OpenAI investor. OpenAI could pursue HubSpot’s more than $3 billion of revenue. Precisely the labs’ broader ambitions, Shah argues, make a focused CRM assault extremely unlikely.

  • The mechanism starts with hundreds of billions invested in pre-training: labs need to sell inference. They can reach consumers and Fortune 500 customers such as JPMorgan directly, while supplying SaaS companies with industry knowledge and distribution. “It’s not an ‘either-or’ choice.” Shah’s infrastructure analogy is AWS and cloud computing 15 years earlier.

  • Shah says that in 2023 he feared the labs were so far ahead he did not know whether others could meet the capital needs: “Fortunately, I was wrong about that.” He adds, “I wouldn’t say they’ve become a commodity.” But open-source code, many commercial products, OpenRouter and use-case-based model routing reduce dependency on one critical supplier. Conditional on staying off the labs’ road, he thinks building an AI business is less scary now.

5. The work matters more than the screen displaying it

  • On whether ChatGPT access reduces HubSpot’s long-term value, Shah acknowledges interface value varies. Most business apps, he argues, are closer to Uber Eats or DoorDash: the app matters, but logistics, restaurant relationships and fulfillment carry much of the value. He explicitly avoids claiming business apps share those network effects. The analogy is that “the value lies in the work that is done.”

  • HubSpot launched its ChatGPT connector more than a year before the conversation and also built one for Anthropic; Shah says HubSpot was the first CRM company to do so. He sees these assistants as “a new OS” or new browsers; browser owners did not capture every web application’s economics. AI differs because it performs cognitive work, he concedes. Still, “The interface is just a means to an end”; optimize outcomes, cost, latency and friction.

  • His limit on model use is categorical: “Do not use inference where the problem can be solved with a query.” Asking how many customers or how much revenue comes from Ireland should trigger a database query. Searching eight million Slack messages and emails, even if the answer is somewhere inside, wastes inference on a task deterministic software can solve.

6. Context expands the architecture while evaluations make changes testable

  • Reusable primitives underpin Shah’s architectural case. HubSpot’s marketing automation—send an email, wait three days, respond to a click—became the foundation for sales automation. Recombining existing technology gave a new product capabilities it otherwise would have needed years to build. The question now is what primitives should support HubSpot’s next 20 years.

  • His answer retains structured records for known facts alongside a system of context for unstructured information. Such data existed before; using it meaningfully at a reasonable price was the constraint. Shah imagines a “second brain,” perhaps organized as a graph, combining context and structured records so agents can infer, act and produce outcomes.

  • Agentic programming also makes past architectural decisions testable. Shah imagines spending $200,000 on tokens to rebuild differently, or trying Rust on a small portion, then checking behavior against existing tests and evaluations. Incumbents have those measurement assets. Evaluations also compare new models—for example, whether roughly 3% less quality is worth equivalent latency or economic gains. The rewrite is a hypothetical experiment.

7. HubSpot Next gives experiments freedom to borrow incumbent assets

  • HubSpot Next isolates experiments from core-company obligations. The earlier CRM team could ignore HubSpot’s stack, branding and ordinary rules; current projects similarly explore categories, markets or technologies. Schmidt calls incumbent constraints “baggage”: distribution, customer relationships, evaluations and 20 years of code remain assets once teams escape organizational weight. They can draw on those assets when useful.

  • Shah’s current project, YouSpot, is a 30-day-old CRM for solo entrepreneurs, where a “second brain” can be tested faster than inside the main platform. Shah says the concept works. His warning is to avoid generic memory gaps that ChatGPT or Claude will improve; instead choose a customer category, its data entry points and the agents or work integrated context enables.

  • The precedent is agent.ai: a horizontal marketplace plus a builder mixing deterministic and non-deterministic agents. Shah considers the marketplace too horizontal; “half of the idea was right.” HubSpot absorbed the builder. Experiments can become full HubSpot offerings, remain separate or contribute only an underlying idea; the core team chooses what is useful.

8. Solo businesses expand the market for software

  • Shah monetized YouSpot early to study pricing and willingness to pay. Credit-card adoption under pay-as-you-go has been stronger than across his previous HubSpot projects in that respect. Schmidt adds that vertical software could keep annual or monthly costs low and make up for that with payment revenue and interchange fees; AI creates additional ways to deliver value to small customers.

  • At the September 30 recording, Shah cites a record roughly 5.6 million new U.S. companies in 2025, approaching 6 million, and over 4 million through August of that year. Citing Census data, he says more than 70% are not expected to hire an employee within two years. He predicts another record year and sees a growing market for solo-entrepreneur software.

  • Shah’s counter to a “SaaS apocalypse” reprises HubSpot’s founding pitch: marketing budgets existed, but the internet made marketing a problem software could solve. AI similarly expands the set of addressable problems. He expects some software companies to disappear, as others did during previous platform shifts. His aggregate conclusion remains categorical: “the opportunities are greater now than ever.”