Most B2B companies claiming to be “AI-powered” aren’t seeing growth accelerate. Not because their technology doesn’t work, but because they’re playing the wrong game entirely.
The Framework: Three Paths to AI Revenue
Path 1: Attach to Compute Infrastructure
Path 2: Replace Human Headcount
Path 3: Massively Displace an Incumbent
Everything else is noise in AI and B2B. Let’s break down why each works, how to know if you’re actually doing it, and where most companies go wrong.
Path 1: Attach to Compute Infrastructure — The DataDog Playbook
DataDog’s 23% stock surge after earnings wasn’t luck. It was the inevitable result of a simple principle: sell to the people building AI, and as they grow, you grow automatically.
“The AI leaders, the hyperscalers — they’re starting to buy like classic B2B companies,” Jason explains. “They’re recycling the same people in procurement. So if you’re attached to AI budget and you’re a DataDog era, you’re actually going to have a great 2026.”
Rory adds critical context: “DataDog is a core piece of compute infrastructure and these hyperscalers are the most compute-intensive companies that have ever been known. If you’re selling compute stuff, you should be having a great quarter. If you’re selling routers, switches, interconnects, whatever it takes to stand up Stargate — you’re going to be golden.”
What This Actually Means
The hyperscalers (OpenAI, Anthropic, Google, Microsoft, Meta) are deploying unprecedented amounts of compute. They’re not just buying different things — they’re buying MORE of everything:
- More observability (DataDog)
- More networking equipment (Broadcom crushing it)
- More security infrastructure
- More data pipeline tools
- More developer tooling for ML workflows
Here’s the key insight: These companies buy like traditional B2B enterprises, just at 10x the scale and 10x the speed.
OpenAI isn’t some mystical different creature when it comes to procurement. They need observability. They need security. They need data tools. They need all the same infrastructure any rapidly-scaling company needs — they just need more of it because they’re deploying more compute than any companies in history.
How to Know If You’re Actually Playing This Game
Ask yourself these questions:
- Is my revenue directly correlated to compute deployment? If OpenAI spins up 10x more GPUs, do they automatically need 10x more of my product?
- Am I selling into AI infrastructure budgets? Not “AI product budgets” — infrastructure budgets. The money being spent to build and deploy models, not the money being spent to use them.
- Do my customers include the hyperscalers themselves? If you’re not selling to OpenAI, Anthropic, Google AI, Microsoft AI, Meta AI — you’re probably not truly attached to compute.
- Does my product become MORE essential as models get bigger? As context windows expand and training runs get more expensive, does the need for my product increase?
If you can’t answer yes to at least 2-3 of these, you’re not actually playing the compute infrastructure game.
The 2026 Opportunity
Jason’s prediction: “If you have attached to AI budget and you’re a DataDog era, you’re actually going to have a great 2026.”
Why? Because the infrastructure layer is just getting started. The capex spend on AI is still accelerating. Data center buildouts are still in early innings. The hyperscalers are still ramping their compute deployments.
Every single piece of software that sits in that stack — from chip to model to deployment — is going to see explosive growth. But you need to actually be IN that stack, not adjacent to it.
The Cautionary Tale: Being Compute-Adjacent Isn’t Enough
Here’s where companies get confused. They think, “We help AI companies be more productive, so we’re attached to AI.”
No. That’s not the same thing.
Compute-attached means your revenue scales automatically with compute deployment. It means you’re essential infrastructure, not a productivity enhancer. It means when Stargate comes online with $100B of compute, your revenue goes up whether you do anything or not.
If you’re a project management tool that AI companies use, you’re not compute-attached. You’re just selling to hot companies. That’s fine, but don’t confuse it with the DataDog playbook.
Path 2: Replace Human Headcount — When AI Actually Joins Your Team
“Where are you replacing humans for real?” Jason asks. “Where are you going to go in and reduce the headcount that vendor needs by half?”
This is the path most companies claim to be on but aren’t actually delivering. The difference between real human replacement and “AI-enhanced productivity” is stark.
The Replit Example: What Real Human Replacement Looks Like
Jason runs through his experience deploying 20+ AI agents at SaaStr:
“We have Reply for Replit (we call it ‘Reply’). We have RD for Artisan which is an SDR. We have Quali for Qualified. They all have these little desks where there’s no human at it anymore.”
But here’s the critical distinction he makes about Replit’s V3 agent: “It is the first one that jumped the line to not being one of the parts of our team, but literally being part of our team just like a human being.”
What makes it different?
It’s sufficiently autonomous, knowledgeable and powerful to complete material high-value tasks on its own with some daily discussions just like on our team.
Not “it helps us be more efficient.” Not “it automates some tasks.” It completes high-value work autonomously with the same level of oversight you’d give a human team member.
The False Replacement Problem
Jason makes a crucial point: “Most of the time, ‘replacing humans’ is a story.”
Here’s how to know if you’re actually replacing humans versus just telling yourself you are:
Real Human Replacement:
- The human role literally doesn’t exist anymore
- Headcount actually goes down or doesn’t scale with growth
- The agent handles end-to-end workflows, not just tasks
- The company’s org chart has fewer boxes
- The cost saving is measurable and significant (30%+ reduction)
Fake Human Replacement (productivity story):
- Humans use the tool to be “more efficient”
- Headcount still scales with growth, just slightly slower
- The tool handles pieces of workflows, requiring human assembly
- Everyone still has the same job titles
- The “savings” are theoretical or marginal (10-20%)
The Gamma Evolution: From Tool to Team Member
Jason walks through how Gamma could evolve from Path 3 (displacement) to Path 2 (replacement):
“When Gamma can do this — go into my Google Calendar, create prospectuses and sales collateral for all 20 sales calls this week, pull all the data on them last year from Salesforce and HubSpot and Marketo, put them all apart, review them once, and then distribute them to the team — Gamma’s part of our team.”
That’s not “AI-enhanced PowerPoint creation.” That’s replacing a marketing operations person entirely.
“It’s not that far away,” Jason notes. “Not just make me PowerPoint. When Gamma is part of your marketing team rather than a marketing tool, there’s a lot of revenue expansion if they can pull it off.”
The 2025-2026 Transition
Jason identifies the key shift happening right now:
“The co-pilot was the 2024 story. It didn’t work. It was a ripoff. Spend $30 more a month on Office like no one wanted the ripoff. This year OpenAI and Claude finally actually got good. That’s why Lovable, Replit, Gamma exploded. Next year is AI is part of your team.”
2024: Co-pilots (didn’t work)
Tools that made you more efficient but required constant human oversight and direction.
2025: Agents that work (breakthrough year)
Applications where AI finally got good enough to deliver real value autonomously.
2026: AI team members (what’s coming)
AI that’s sufficiently autonomous to be actual members of your team, not just tools your team uses.
That transition from tool to team member is where the revenue explosion happens. Because when AI is a team member, companies will pay team-member prices ($50K-$150K+ per year), not tool prices ($10-$100 per month).
How to Build for Path 2
If you’re building a company to replace human headcount:
- Start with the highest-volume, lowest-judgment work. Customer support, SDR outreach, data entry. These are the easiest to replace completely.
- Design for autonomy from day one. Don’t build a tool that helps humans do work faster. Build an agent that does the work completely.
- Price like you’re replacing a human. If you’re replacing a $75K/year employee, charge $30K-$50K/year, not $50/month.
- Market the headcount reduction explicitly. Don’t shy away from it. CFOs want to reduce headcount. Lean into it.
- Build for the day-to-day management experience. If your AI team member needs more oversight than a human, you’re not there yet.
The Uncomfortable Truth
Jason’s right that most companies avoid saying this explicitly, but the economics are undeniable: “If you are not removing humans from the equation, you are going to be heavily discounted.”
The market is dividing into:
- Companies that actually replace humans (valued highly)
- Companies that just make humans more efficient (valued as marginal improvements)
You have to choose which game you’re playing. And if you choose Path 2, you need to actually deliver on human replacement, not just talk about it.
Path 3: Massively Displace an Incumbent — The Hardest Path
“You have a third option,” Jason concedes. “Use AI to massively displace an incumbent and steal all the revenue. In fact, that’s the history of B2B software mostly.”
This is the path that sounds most familiar to traditional B2B investors. It’s what Salesforce did to Siebel. What HubSpot did to Marketo (before Marketo became HubSpot). What Slack did to email for internal communication.
But Jason is deeply skeptical this path works well in the AI era: “I just don’t know how many of our public leaders are in a place to steal their own revenue.”
Why Incumbent Displacement is Harder Now
In the traditional SaaS era, you had ~3 years before incumbents would really compete:
“You’d be like, well, Salesforce or HubSpot, whoever, it’ll take them like a year and a half to decide if it’s worth cloning. Then they’ll launch something and it’ll be okay because they have smart engineers, but it won’t actually do anything. The first six months it’ll be so feature poor. Then after two years they’ll decide, should I put 100 people on this?”
That timeline has collapsed from 3 years to 30 days.
Jason shares a brutal example: “I can think of one investment I’ve made that has had five clones in the first 30 days, including one from a cloud leader.”
The AI-enabled cloning speed changes everything. Canva is now “borderline competitive” with Gamma — and Canva wasn’t even a competitor when Gamma’s CEO was on the show just months earlier.
The Crowded CRM Problem
Jason’s skepticism crystallizes around the most obvious displacement opportunity: CRM.
“There’s like 400 AI CRM startups out there all saying they’re going to eat HubSpot’s and Salesforce’s lunch. That’s not exciting to me as an investment. That’s much less exciting than truly replacing 90% of your GTM team.”
Why isn’t this exciting?
- Everyone sees the opportunity. When 400 startups are all attacking the same incumbent, the odds of any one winning are low.
- Incumbents can add AI too. Salesforce and HubSpot aren’t standing still. They’re adding AI features. Maybe not as well, but they have distribution and existing relationships.
- The switching costs are still real. Yes, your AI CRM might be better. But is it 10x better? Enough to justify migration pain? For most companies, probably not.
- You’re competing for the same budget. Unlike Path 1 (new compute budget) or Path 2 (headcount budget), you’re fighting over existing CRM budget. That’s a zero-sum game.
When Incumbent Displacement Can Work
Jason isn’t saying Path 3 is impossible. He’s saying it’s the hardest path and requires specific conditions:
You need at least 3 of these 4:
- Massive technological discontinuity. Not just “better with AI” but fundamentally different because of AI. The product should do things that literally couldn’t exist before.
- Vertical specialization. Don’t try to displace Salesforce horizontally. Displace Salesforce for patent attorneys specifically. Rory’s example: “Solve Intelligence does AI for patent law. The more patents that go through their algorithms, the better they are at writing, editing, predicting.”
- 10x better, not 2x better. The switching costs and incumbent advantages mean you need to be dramatically superior, not marginally better. “Our AI CRM is cleaner” isn’t enough. “Our AI CRM eliminates 80% of data entry” might be.
- Weak incumbent position. You’re not displacing Salesforce in enterprise. But maybe you can displace some legacy vertical software that never went cloud-native and is now run by private equity.
The Education Example: Duolingo’s Dilemma
The Duolingo discussion illuminates why incumbent displacement is so hard for public companies.
Duolingo displaced Berlitz and traditional language schools. Hooray. But now what?
Jason: “Where are you going to disrupt humans? This is your job. You already disrupted those humans. Unfortunately, they’re gone. Where’s the next level of human disruption?”
Rory sees the opportunity: “The next generation of companies that are going to be teaching a foreign language will be LLM-based. We’ve seen companies doing more professional, more interactive teaching using LLMs. There’s a whole ton to be done in terms of one-on-one instruction.”
But here’s the problem: Duolingo can’t easily displace itself. They have a business model built around light learning at consumer prices ($10-20/month). The AI opportunity is intensive learning at tutoring prices ($50-200/month). That requires cannibalizing their existing revenue model.
This is why Path 3 is especially hard for incumbents. They’re not in a position to displace themselves. Which creates the opening for startups, but also means you’re starting from zero on distribution against companies with millions of users.
So Should You Play Path 3?
Jason’s answer: Only if you can’t play Path 1 or Path 2.
“We can fund those deals as investors, but I think the first two categories are much easier — attached to the compute or replace humans — rather than just steal.”
The math is simple:
- Path 1: New budget being created, you’re capturing net new dollars
- Path 2: Headcount budget being redirected, you’re capturing existing dollars going to humans
- Path 3: Software budget being redirected, you’re competing with existing software providers who have relationships, data, and switching costs
All three can work. But Path 3 requires the most capital, the longest timeline, and has the highest failure rate.
The Captain Obvious Test
Jason drops a truth bomb midway through the discussion: “We’re in the Captain Obvious era of investing. You either got something or you ain’t getting funded anyway.”
What does this mean for these three paths?
If you’re on Path 1 (compute-attached):
Your growth should be automatic. If the hyperscalers are deploying 3x more compute and your revenue isn’t growing 2-3x, something is broken. This should be the most “Captain Obvious” path of all.
If you’re on Path 2 (replacing humans):
Your customers should be measurably reducing headcount or headcount growth. If they’re not, you’re not actually replacing humans — you’re just a productivity tool.
If you’re on Path 3 (displacing incumbents):
You should be winning deals against the incumbent consistently, not just in small accounts. If you’re only winning when there’s no incumbent, you’re not displacing anyone.
The test is simple: Is your success obvious? Or are you explaining why your metrics are actually good despite not looking impressive?
What Doesn’t Count: The “AI-Enhanced” Trap
Throughout the discussion, the panel keeps coming back to what doesn’t work: using AI to make your existing product better.
Jason is blunt: “Using AI to make your product better? That’s table stakes now, not a competitive advantage.”
This is the Duolingo problem. They’re using AI to improve their language learning. Great. Every single competitor can do that too. There’s no moat in “we use AI to make our product better.”
Rory agrees: “Every single portfolio company at scale should be using AI by this point to make your product better. This is not 2023. You don’t get any kudos for sprinkling AI dust on your product.”
The 2023 vs. 2025 Divide
In 2023: Adding AI features was novel and impressive. Companies could raise on “we’re adding AI to X.”
In 2025: Adding AI features is expected. It’s like saying “we have a mobile app” in 2023. Cool, so does everyone else.
The market has moved from “are you using AI?” to “how are you using AI to capture revenue in one of the three ways that actually work?”
Examples of What Doesn’t Count
Let’s be explicit about what isn’t one of the three paths:
Not Path 1: “We sell to AI companies”
Unless you’re infrastructure they can’t build AI without, this is just selling to hot companies. When the AI boom ends or slows, your growth ends.
Not Path 2: “We make teams more efficient with AI”
Unless headcount actually goes down or dramatically slows its growth, you’re not replacing humans. You’re selling productivity software. That’s fine, but it’s not Path 2.
Not Path 3: “We’re better than the incumbent because we have AI”
Unless you’re 10x better and winning head-to-head deals consistently, you’re not displacing anyone. You’re just another alternative.
The Brutal Market Reality Check
Harry shares a data point that shows how much the market has shifted:
A classic enterprise SaaS company grew from $400K to $3M ARR (10x growth). Five years ago, this would have generated five term sheets from top firms. Today: 120 meetings, one term sheet at 12x revenue.
Why? Because it’s not playing any of the three games.
It’s not compute-attached. It’s not replacing humans. And it’s not dramatically displacing an incumbent with AI. It’s just a good SaaS company growing well.
That’s not good enough anymore.
Jason: “It’s the most binary fundraising environment in our lifetimes. You’re either Captain Obvious getting funded, or you’re not seeing it anywhere.”
The Action Items: Which Path Should You Choose?
If you’re building a company or investing in one, here’s how to think about which path to pursue:
Choose Path 1 (Compute-Attached) If:
- You can build infrastructure that scales with compute deployment
- You have or can build relationships with hyperscalers
- Your product becomes MORE essential as models get bigger/better
- You’re comfortable with infra sales cycles and pricing
Best for: Infrastructure engineers who understand ML workflows, companies with existing relationships in cloud/DevOps
Choose Path 2 (Replace Humans) If:
- You can identify high-volume, low-judgment work that’s expensive
- You can build true autonomy, not just automation
- You’re comfortable pricing like you’re replacing a human ($30K-$100K/year)
- You can stomach the optics of explicitly reducing headcount
Best for: Founders who have done the job they’re replacing, deep understanding of workflow completion, strong AI/agent expertise
Choose Path 3 (Displace Incumbent) If:
- You’ve identified a weak incumbent in a large market
- You can be 10x better, not 2x better
- You have patient capital for a longer fight
- You’re willing to specialize (vertical, use case, etc.) rather than go head-to-head horizontally
Best for: Domain experts in industries with weak incumbents, founders with deep incumbent knowledge, situations with massive technological discontinuity
The Priority Order
If you can do Path 1, do Path 1. It’s the easiest and most capital-efficient.
If you can’t do Path 1, do Path 2. The market for replacing humans is massive and underpenetrated.
Only do Path 3 if you can’t do the first two or if you have specific advantages (domain expertise, weak incumbent, vertical specialization) that make it compelling.
The 2026 Prediction
Jason’s final point ties it all together: “If you’re in software today and this isn’t the most exciting time of your lifetime, you’re doing it wrong. This is the first time software has gotten better since the three of us met.”
But that excitement only matters if you’re playing one of the three games that actually capture revenue:
- Attach to compute (DataDog’s 23% surge shows this works)
- Replace humans (the 2026 opportunity as agents mature)
- Displace incumbents (hardest path but still viable with the right conditions)
Everything else — all the “AI-enhanced” products, all the “using AI to be better” positioning, all the sprinkling of AI dust on existing products — that’s not going to generate venture-scale returns.
The market is telling us this clearly. DataDog up 23%. Duolingo down 25%. Same week. Both “using AI.” One playing a winning game, one not.
Choose your path wisely. Because in 2025/2026, there are only three that actually work.
