Ad Narrator | The a16z Show Summary

1. Key Themes

Idea selection is empirical, not analytical

Dharmesh argues that founders over-invest in spreadsheet rigor when the quality of an idea can't be known in advance. The winning move is to shorten the learning loop and make lessons cheap. He says: "it's almost impossible to know a priori what the quality, quote, unquote, of an idea is, or its kind of probability of success, or what the outcome is going to look like. And the only real way to test an idea is to actually, or know if an idea is good, is to actually test it." 00:07:04 He adds that the eventual winner usually only loosely resembles the original: "the thing that you end up with that succeeds will oftentimes, or more often than not, resemble the original idea you started with... sometimes doesn't resemble that." 00:07:24 His operating principle: "the people that win are the ones that have the shortest iteration time and the cheapest learning time." 00:08:55

Narrow the aperture, because AI makes distribution the harder problem

Building is easier, so noise is higher. Dharmesh says: "The one thing that's still hard, and I would argue maybe even harder now than it was, is distribution, right? Because it's so much easier to build software now, many more people will be doing it." 00:08:03 His prescription is a narrow focus with a concrete 30-day plan: "the mistake number two that entrepreneurs make is that they feel like they need to start with this massive TAM... tell me what you're going to do in the first 30 days." 00:08:23 HubSpot itself started as "LegalSpot," an ERP for law firms, purely "as a way to kind of test the market early on." 00:12:09

Stay off the Yellow Brick Road: the "yet" test and the product-vs-company test

Dharmesh gives founders two filters for avoiding collision with frontier labs. First: "if you can just append the word 'yet' to it and it still holds water, that's a problem. It's like, 'Oh, ChatGPT can't do that yet.'" 00:24:37 Second: "measuring the product versus the company. It's like, 'Oh, ChatGPT can't do that.' It's like, 'Okay, but can OpenAI, the company, do that?'" 00:24:54 He recommends picking domains where the labs face a structural constraint (regulatory, for example) or an "incentive constraint" — they could do it, but the odds are low. Joe Schmidt's framing of the "fleet of tanks" on the Yellow Brick Road anchors the discussion.

Frontier labs need resellers: AI as infrastructure and the commoditization of models

Because the labs have sunk "hundreds of billions of dollars" into pre-training, their business is monetizing tokens, and they can go direct or through intermediaries. Dharmesh says: "think of AI as infrastructure, in the same way we thought of AWS and in cloud compute as infrastructure 15 years ago." 00:27:36 Model fungibility reduces supplier risk: "things like OpenRouter exists because it's easy to sort of swap out models... most SaaS companies are doing some form of use case decomposition and model routing to figure out what's the right model for this task." 00:29:15 Result: "it is completely possible now that you're not going to get margined out or whatever, because you have options." 00:29:33

The interface is a means to an end: ChatGPT and Claude Cowork are the new browser/OS

Responding to the fear that connectors to ChatGPT erode HubSpot, Dharmesh argues that business software sits closer to Uber Eats than to a pure-UI product: "the actual thing that's valuable is the work that's happening, the interface is a means to an end." 00:00:11 He draws the browser analogy: "ChatGPT and Claude Co-work or whatever, they're effectively the new OS... They are the new browsers, like this is the way people are consuming value." 00:34:07 HubSpot shipped a ChatGPT connector "over a year ago" and has one for Anthropic. His advice: "don't overthink the interface, figure out what's the best possible way to achieve the outcomes your customers are looking for." 00:34:50

Use deterministic queries where they work; don't throw inference at everything

Dharmesh: "anything that can be solved with a structured query deterministically, solve with a structure query deterministically. Don't use inference to solve something that a query can solve." 00:35:17 His example: revenue from Ireland "should be a database query, that should not be a, you know, solved through inference and throw a 300 IQ model at it." 00:35:54

Rearchitecting for AI: primitives, "system of record" plus "system of context"

HubSpot's durability came from reusable primitives: marketing automation became the base for sales automation. Now the question is "what is the framework we need for the next 20 years of HubSpot, now knowing that we have this AI world?" 00:00:43 His answer pairs a structured store with unstructured context: "some persistent mechanism for storing kind of known facts and data, some sort of system of record, if you will, alongside this kind of system, we think of as a system of context, which is this unstructured data." 00:38:45 The likely structure is "some sort of context graph." 00:39:38

Agentic coding lets incumbents replay history with "alternate-universe" rewrites

Dharmesh describes using evals plus cheap code generation to re-run architectural decisions: "What would happen if we said we're going to spend $200,000 on tokens to go just redo that... let's jump to that branch of the universe where we had made a different choice, and then see what happens." 00:41:04 Incumbents have an edge because "we have what people used to call unit tests, and now show up as evals" to verify results. 00:41:21

Incumbent innovation: "startup within a startup" (HubSpot Next)

HubSpot Next is an innovation lab where teams have "none of the rules, you didn't have to build it on our stack, you didn't have to use the brand." 00:43:16 Winning projects get absorbed, as agent.ai's builder became HubSpot's Agent Builder: "half of the idea was right, but actually belonged and needed the full HubSpot team behind it. So we did effectively an acquisition." 00:46:25 Dharmesh's thesis on incumbents: "it's not the absence of something... the issue is the baggage, and the weight that the organization has. And if they can find a way to kind of shed that, then you get all the advantages of being larger." 00:47:57

The solopreneur market is exploding

Dharmesh cites Census-based numbers: "in 2025, we had the most number of new company formation the United States has ever seen, I think it was like 5.6 million, closing on 6 million... This year... by August, we were already over 4 million." 00:50:26 Further: "the percentage of those new company formations that are expected to remain one-person companies is the highest it's ever been... over 70%." 00:50:42 YouSpot, his one-person CRM, saw unusual willingness to pay: "willingness of people to put a credit card down on a consumption-based model... in that regard, the most successful." 00:50:08

2. Contrarian Perspectives

Don't "do one thing really well": do the whole customer problem

Conventional startup advice says to pick a single wedge. Dharmesh says HubSpot deliberately did the opposite: "even when HubSpot started, we had a very broad base of marketing applications, and which goes against every piece of advice you get, and that I've given, it's like do one thing really, really well. We did all the things... The reason we chose to do it all is the only way for us to solve the actual problem customers had, which is they're intimidated by this fragmentation." 00:18:07 The nuance: narrow the market and entry point, but don't narrow the solution below the customer's whole problem.

Making the CRM free, not freemium, to attack the 60% who weren't using one

HubSpot's second product emerged from a customer survey: "20% were using Salesforce.com, another 20% were using some other CRM... but 60% were not using a CRM at all." 00:15:18 Rather than rip-and-replace, they chose greenfield by removing the two objections: "Let's make the easiest CRM that's ever been conceived... And we're going to take the 'it's too expensive' off the table because we're going to make it free." 00:16:53 The market-making insight was hiding inside their own customer base, not in an external TAM spreadsheet.

The best AI opportunities exist because the labs' ambition makes them unlikely to enter your market

Counterintuitively, Dharmesh argues the labs' incentive structure protects incumbents in some categories: "he could wake up tomorrow and say, 'You know, that CRM industry, that looks awfully tempting'... Are they smart enough? Yes. Do they have ambition? Yes. But it's that very ambition that makes it highly unlikely that the team at OpenAI or Sam or anyone... they're unlikely to do that." 00:25:49 The sheer scale of ambition makes a $3 billion revenue category an unattractive distraction.

Being "a little further off the road" beats chasing today's gap

Most founders look for gaps in the labs' current product ("Claude Cowork doesn't do X"). Dharmesh calls those the ideas "that make me nervous" 00:24:19 and advises: "maybe the road itself is getting wider, fine. So maybe stay off the road, but maybe also stay, you know, a little bit outside." 00:24:54 Don't solve horizontal problems like memory: "Don't solve the horizontal use case or whatever, but say, 'Hey, within this class of customer, here's a class of problems that they have that would be benefited by thinking about this notion of having a second brain.'" 00:44:59

AI is software, and the SaaS-pocalypse thesis is backwards

Dharmesh pushes back on "SaaS is dead": "AI, by the way, for the record... it is software. It's not an alien, it's not some new energy source, it's software. Really, really smart software, but software." 00:52:45 He argues AI expands the set of problems software can solve, just as the internet made marketing a software-shaped problem: "now there is a massive set of these new problems... so the opportunity set is way larger." 00:54:36 While conceding "some software companies will die," he concludes: "in aggregate, the opportunity is bigger than we've ever..." 00:54:57

3. Companies Identified

HubSpot

CRM and marketing platform, now ~20 years old and public, serving SMBs. Mentioned as the case study throughout. Dharmesh states "you know, $3-plus billion of revenue that HubSpot has" 00:26:03. It was the first CRM company to ship a ChatGPT connector: "HubSpot launched our, we were the first CRM company to do so, our connector to ChatGPT over a year ago, and we have a similar one for Anthropic." 00:31:30

YouSpot (HubSpot Next)

A one-person CRM for solopreneurs, built inside HubSpot Next, exactly 30 days old at recording. Joe describes it: "they've launched YouSpot, a CRM for one-person businesses." 00:01:21 Dharmesh: "I monetized really early... the most successful" project in terms of consumption-based willingness to pay 00:50:08. It serves as a test bed for the "second brain"/context graph idea.

HubSpot Next

HubSpot's internal innovation lab. Dharmesh: "it's effectively an innovation lab, so it's a startup within a startup." 00:42:50 Projects are free from the stack, brand, and rules, and can be absorbed later.

agent.ai

A marketplace for AI agents launched by Dharmesh, with an Agent Builder component later folded into HubSpot. Dharmesh: "agent.ai was a marketplace for AI agents back in the day... half of the idea was right, but actually belonged and needed the full HubSpot team behind it." 00:45:41

OpenAI

Frontier lab; DevDay was held the day of recording. Dharmesh calls plug-in marketplaces inside ChatGPT sensible: "They launched, you know, a marketplace for people to plug in and, you know, plug-in architecture and all those things, which makes complete sense to me." 00:34:25 He also notes he is "an investor in OpenAI" 00:25:49. Cited as the paradigm lab that is unlikely to attack HubSpot's category.

Anthropic / Claude Cowork

Frontier lab and product; HubSpot has a connector for Anthropic. Dharmesh calls "ChatGPT and Claude Co-work" "effectively the new OS." 00:34:07

OpenRouter

Model routing layer used as evidence of model fungibility. Dharmesh: "things like OpenRouter exists because it's easy to sort of swap out models." 00:29:15

Typeface (the "Jev" model)

Dharmesh references a new non-transformer model released "by the Typeface folks": "we have the Jev model, released by the Typeface folks. And what that really takes it back to is like, 'Oh, well, it is really just software.'" 00:04:46 (The name is garbled in the transcript, so this is flagged only loosely.)

Uber Eats / DoorDash

Used as an analogy for businesses where the value is the network and logistics rather than the app. Dharmesh: "the app is trivial, what's actually valuable is the network that they have, the partnerships with the restaurants." 00:31:57

Twitter (early API ecosystem)

Cautionary tale of platform behavior. Dharmesh: "literally 1,000 flowers bloomed back then... But then they sort of closed it down... everything from developer-unfriendly to almost developer-hostile." 00:20:27

Salesforce

Incumbent CRM referenced in HubSpot's early customer survey: "20% were using Salesforce.com." 00:15:18

Intuit

The only SMB software multibillion-dollar business cited at the time HubSpot started. Dharmesh: "no software company has ever, other than Intuit, has ever made a multibillion-dollar business in SMB." 00:51:27

Oracle

Reference point for HubSpot's original idea: "ERP software, like an Oracle for small business." 00:11:37

MIT $50K competition

Where HubSpot's original business plan was entered. Dharmesh: "our original idea that we took to the MIT $50K competition was like ERP software." 00:11:37

GitHub

Referenced as the system of record for code in the rewrite-the-past idea. Dharmesh: "we have a system of record for code, which most people, which is GitHub." 00:40:11

Workday

Joe used it as an example in an essay on needing to restart to be AI-native: "I actually used Workday as an example on that front." 00:42:07

Google / Mozilla / Internet Explorer

Used in the browser-as-interface analogy: "Google did not own the entire application stack... lots and lots of software companies that got built." 00:33:41

JP Morgan

Example of a Fortune 500 the labs might sell tokens to directly using "an army of FDEs." 00:27:17

AWS

Analogy for AI as infrastructure. Dharmesh: "in the same way we thought of AWS and in cloud compute as infrastructure 15 years ago." 00:27:36

Dunkin'

Dharmesh's lightning-round pick over Starbucks, though he doesn't drink coffee. 00:55:27

a16z

Hosting firm; Joe's Yellow Brick Road essay was the genesis of the conversation. Joe: "the idea that there's plenty of room to build software outside the path the frontier AI labs are racing down." 00:01:21

4. People Identified

Dharmesh Shah

Co-founder and CTO of HubSpot; 30+ years building commercial software; no direct reports ("I've never had direct reports, so like I'm an IC"). Joe calls him "a generational entrepreneur, a wonderfully kind human, and a generational thinker." 00:57:41 His stated best time at HubSpot is now: "I have all the benefits of scale... but I have like none of the downsides." 00:57:04

Joe Schmidt

a16z partner and host; author of the "Yellow Brick Road" essay and a Workday/AI-native essay. His framing: "I wrote that essay a few months ago, really in response to this nihilistic nothing matters feeling that was going on in Silicon Valley." 00:01:59

Sam Altman

CEO of OpenAI. Dharmesh's "never compete with Sam Altman" rule 00:28:20 and his thought experiment about Altman targeting CRM 00:26:02. He also says he knows Sam and is an investor in OpenAI.

Brian Halligan (implied: HubSpot's co-founder, unnamed)

Dharmesh refers to "my co-founder" as the marketing and sales-side half of the founding team: "he on the kind of marketing and sales side, me on the kind of product development and engineering side." 00:10:14 Not named in the transcript.

5. Operating Insights

Measure time to joy for developers

Treat your ecosystem like a product. Dharmesh: "in the same way we think about time to value for the core customers, we have to think about time to joy for a developer. How fast can someone get to, 'Oh, I got something working with the HubSpot APIs,' to the, 'Oh, I made my first dollar'?" 00:21:59 The metric to optimize: the interval from first API call to first dollar earned in your ecosystem.

Give partners a candid "white space" map

Rather than promising never to compete, publish where you will and won't go. Dharmesh: "We can't promise you we're not going to do this, but here's what we're doing. If we were looking for white space, if we were you, here's the kinds of things we would do that we think... we're highly, highly unlikely to kind of traverse into that territory." 00:21:30 Trust compounding across partners is what built the ecosystem.

Monetize early to learn pricing, not to cover costs

On YouSpot, Dharmesh deliberately charged from the start: "not to kind of cover COGS or something like that, it's like that's one of the lessons I wanted to learn, is I wanted to learn like the pricing lessons." 00:49:39 Early monetization is an experiment design choice for learning willingness to pay.

Use evals as a switching-cost-free decision engine

HubSpot runs "our battery of evals across new models" to quantify tradeoffs: "we lose about three percentage of quality points, but we gain this much in terms of latency or economic." 00:41:38 Evals turn model selection and even architectural rewrites into measurable decisions.

Structure the absorb-or-spin-out path for internal startups up front

HubSpot Next projects can be folded in "like an acqui-hire" at "the low, low price of it's included" 00:47:08. Pre-defining the options (absorb the product, absorb the idea/injection points, or leave standalone) lets the team move without bureaucracy while preserving the option to use the HubSpot brand and distribution later.

6. Overlooked Insights

Incumbents' real moat is a 20-year codebase plus evals, usable as training data for the next product

Mentioned in passing during the "baggage vs. assets" discussion, this reframes legacy software. Dharmesh lists the incumbent's assets: "you have distribution, you have customer relationships, you have evals, you have, you know, 20 years of a codebase that you can use as data to kind of inform the next product that you build." 00:48:09 Combined with the earlier idea of replaying architectural decisions for "$200,000 on tokens" 00:41:04, this implies large incumbents may be uniquely positioned to rewrite themselves AI-natively, since greenfield startups lack both the test suites and the historical code. Investors who assume incumbents are doomed by legacy architecture may be missing that the legacy is now an asset for automated refactoring.

"Always on the P&L": missing-software line items as a market-sizing heuristic

Dharmesh's HubSpot origin story contains a repeatable screening tool: "every line item on the P&L has a $100-plus billion software company name attached to that line item, except for marketing." 00:53:30 The thesis was that marketing was becoming software-shaped because of the internet. He then applies the same logic to AI: "before there was a class of problems that software wasn't really equipped to solve. Now there is a massive set of these new problems." 00:54:20 The actionable insight is to scan a business's P&L for categories historically done by human labor and "arts-craftsy" work, and ask which just became software-shaped. These are the likely homes of the next large software companies.