SaaStr 的 Jason Lemkin 谈 B2B+AI 创始人十个经营决策 换模型降本 72% 与 1 亿美元收购门槛

The decisions behind this week’s headlines: how to price for falling model costs, how to report growth when your percentages are shrinking, and why you need to test whether you’re expanding the TAM or just taking share.

On Thursday’s 20VC x SaaStr, Harry, Rory and I covered OpenAI cutting its valuation to $1.25T, Anthropic’s slowing quarterly growth, DoorDash buying Instacart, and a round of IC picks across Cognition, Baseten and Serval. That recap covered what all three of us said. This post is my own read: the 10 decisions I think B2B + AI founders should be making based on this week.

1. Our Costs on SaaStr Connect Fell 72% When We Changed Models

SaaStr Connect is our AI matching and recruiting app. I moved it onto Haiku 5.5 and a cheaper model mix, and total costs dropped 72% for the same workflow. It got faster and the output got better. Since early 2025, despite all the falling token prices and eval charts, nothing we ran agentically had gotten cheaper per use. This was the first time I saw it in our own bill.

The decision for founders is how often you re-run your model mix. If you set it once at launch and haven’t touched it since, you’re probably overpaying by a wide margin, and your competitors who re-benchmarked are underpricing you with healthier margins.

My learning → Re-benchmark your production workflows every time a new small model ships, using your real tasks, not public evals. A 70% drop in inference costs changes your gross margin enough to justify the work.

2. Report Net New ARR in Dollars Alongside Your Growth Rate

Anthropic’s GAAP revenue went from roughly $5B to $11B to $17B over the first three quarters of the year. Rory laid out how the same facts give two opposite stories. One: quarter-over-quarter growth was cut from 100% to 50% in a single quarter. Two: they added roughly the same dollars each quarter and are still on pace for about 5x year over year. Both are true. The market will pick whichever story fits its mood that week.

Your board and your next investors will do the same thing to you. As you scale, percentage growth falls automatically, even when you’re adding more dollars than ever.

My learning → Put absolute net new ARR on the first slide, next to the percentage. If net new dollars are flat or rising, that’s the story. If net new dollars fall, you’ll hear about it, so address it before someone else brings it up.

3. A 30% Discount From a Competitor Can Halve Your Growth Rate

My explanation for Anthropic’s deceleration is pricing pressure. Demand hasn’t dropped. Anthropic is still around 0.1% of Fortune 500 spend. But OpenAI is selling roughly equivalent products at 30% off, and open-weight models are pulling prices down too. Once a competitor offers 30% off, you need a huge amount of momentum to overcome it.

If the leader in the category has hit a ceiling on its premium pricing, so has every B2B + AI company priced as a markup on that capability.

My learning → If most of your value is model capability, plan for your price to fall 30 to 40% within a year. Shift pricing toward the outcome you deliver and the workflow you own, because those don’t deflate the way tokens do.

4. I Turned Off Canva After 12 Years Because Our Agents Don’t Log Into It

We run 21 agents at SaaStr. They use Higgsfield, Opus and other tools that can be driven programmatically. Canva wasn’t one of them, so after 12 years I turned it off. That isn’t surprising for a company like ours. What surprised me is that the broader market is pulling back too. Canva’s growth has reportedly slowed sharply. Squarespace just cut 20% of its staff.

The question for every B2B app is whether an agent can use your product without a human logging in. DoorDash saw this coming and has already shipped an MCP integration for external agents.

My learning → List the five most common jobs customers do in your product and check whether an agent can do each one through an API or MCP today. Any job that still requires a human clicking through the UI is a job an agent will route to someone else.

5. Persistent Agents Cost More Than $20 Per User Per Month to Serve

I underestimated the cost of persistent consumer agents. Replit and Lovable have teams tuning VMs to spin up and down, and they’ve gotten cost-to-serve down to about $4 to $5 per user per month. I assumed always-on agents would be close to that. They’re not. They can run north of $20 a month per user, and power users cost far more. Elon had to cap free usage of his agent product. Amjad says the era of free agent subsidies is ending.

Rory’s counterpoint is fair. Every digital product launches before its economics work, and the bet is that costs fall faster than the money runs out. That’s fine if you know which side of the bet you’re on.

My learning → Measure cost-to-serve separately for your top 10% of users before you set a price. Blended averages hide power users who cost you 5 to 10x what they pay, and those are the users who post about you most.

6. DoorDash’s Third Competitor Is Nobody

DoorDash said agents have placed only a few hundred orders against about 3 million deliveries a day. I think that’s a self-serving comment, much like Brian Chesky and Parker Conrad saying their categories didn’t need AI 18 months ago. DoorDash’s real competitors are Uber Eats and the restaurant itself, since most restaurants offer pickup or their own delivery to avoid fees. An agent will find the restaurant’s $2 delivery option and save you $14 per order. When an agent can go straight to the source, a marketplace that takes 20%+ ends up competing with no middleman at all.

Harry thinks people love browsing food too much for this to matter. Maybe. But “few hundred orders” is the starting point for agentic ordering, not the ceiling.

My learning → If you run a marketplace or take a fee on transactions, find the direct alternative an agent would route a customer to, and price the value you add against it. If you can’t explain why your fee beats going direct, assume an agent will skip you.

7. Test Whether You Expand the TAM by Comparing Deal Size to the Incumbent’s

Harry picked Serval at $3B. I picked Cognition at $80B. My one question about Serval, and about every company built to unseat an incumbent, is whether it expands the TAM or just competes for a share of ServiceNow’s existing TAM. I think 2027 will show that many vertical AI companies are nifty startups competing for slices of Salesforce, ServiceNow and Atlassian against five to eight times as many competitors, while the incumbent is still very much alive.

Rory proposed the right test: look at ACV like-for-like. If a mid-market company spends $200K a year on ServiceNow, will it pay $300K to $400K for Serval because Serval resolves the tickets ServiceNow can only track? That’s the same bar I use for my own fund. I won’t invest in an agentic B2B company unless its deal sizes are at least 2x what pre-AI competitors get.

My learning → Pull your last 10 deals and compare each one to what that customer paid the incumbent. If you’re at 1x or below, you’re taking share, and you’ll be fighting on price within two years. If you’re at 2x or more, put that ratio in your next deck.

8. Most Acquirers Won’t Commit Their Sales Team Below $100M in Revenue

If you’re confused about why acquirers take your meetings, get excited and then go quiet, here’s why. Almost every large acquirer will tell you they want at least $100M in revenue before they’ll put their go-to-market machine and platform behind a deal. Below that, they either buy a small team cheaply or wait. Fin’s agentic business had to hit $100M before it mattered to a $50B company. Listen Labs, at about $20M ARR for around $2B, was the exception. Acquirers will tell you deals like that are their worst-performing category, and they only do them under competitive pressure.

Salesforce is buying a company every couple of weeks, and every one of them is an SVP’s best idea. To be one of those ideas, you need to be big enough to move their numbers.

My learning → If acquisition is your likely outcome, plan to $100M in revenue, not to the first inbound from corp dev. Between $20M and $100M, the conversations are mostly about price discovery.

9. “ARR Grew 165%” Without a Base Number Signals Weakness

Cohere announced a big round, with Canada putting in up to $1.14 billion, and said ARR grew 165% year over year without disclosing the base. As Harry said, not disclosing revenue in 2026 is hardly a flex. In a market where Anthropic files audited quarterly numbers and Kalshi talks openly about its run rate, a growth rate with no base number reads as a small base.

My learning → When you announce a round, include one real number: ARR, revenue or net new. A percentage on its own invites the reader to assume the base is small.

10. ARTEX Shows That Any AI Security Tool Can Be Used Against You

ARTEX started as a well-regarded open-source pen-testing tool from China. It was a couple of months old, free, and could be plugged into DeepSeek or another LLM to attack your own systems, which is exactly what you’re supposed to do. Then it was turned outward. Microsoft says state-linked Chinese hackers used it to compromise environments linked to ConnectWise, and WorkOS reported a breach this week too. People could always do this. Now they can do it with LLMs, at scale, against anyone.

Every MCP endpoint and agent integration you ship this year is a new attack surface that tools like this will probe automatically.

My learning → Pen-test agent-facing endpoints in the same quarter you ship them, and run the same open-source tools attackers are using. Assume they’ll find anything you haven’t found yourself.

Jason’s Takes is the SaaStr AI companion to our weekly 20VC x SaaStr recap with Harry Stebbings and Rory O’Driscoll.