Jason’s Takes on This Week’s 20VC: The Token Governor, the Net-New-Logo Test, and Why Renewal Stopped Being Safe

Going a level deeper on the points Harry, Rory and I hit this week, and what they actually mean if you’re building a B2B company right now.

Here were the top SaaStr learnings I wanted to share from this weeks’ deep dive with the 3 of us.

1. If You Can Hire the Expertise Legally, Never Steal It

The Apple lawsuit against OpenAI is going to end a few careers, and none of it was necessary. In California, non-competes don’t hold and the knowledge in someone’s head travels with them. You can hire the domain expert and use what they know. Anthropic is the proof: those founders walked out with nothing but their brains and built a company worth more than $50 billion. The person who carries the actual files for their new employer isn’t being loyal. They’re the first one thrown overboard when discovery starts.

Just don’t do this. Hire the expertise, tell them to always leave the laptop behind. Always.

My learning → The talent is legal to hire. The theft is what gets you sued. Never confuse the two.

2. Tokens Are a Real Line Item Now, and Left Alone They Go 60x

ClickHouse’s AI spend is up 60x since February. That’s not a rounding error, that’s a new cost center that appeared inside a year. And the incentive structure is brutal if unmanaged: any individual can press the magic button, look like a hero if the KPI is tokens used, and spin up a huge bill that accrues to the company, not to them. Sometimes that’s twenty dollars of tokens saving five hundred of labor and you’re thrilled (as with our AI Agents at SaaStr AI). Without a governor, you eventually spend six hundred to save five hundred, and nobody notices until the quarter closes.

My learning → Put a spend governor on AI before your CFO puts one on you. This line item does not self-regulate.

3. Stop Evaluating AI on Price Per Token

The number every vendor markets, cost per input and output token, is close to useless. A model with cheap base tokens can burn expensive reasoning tokens in volumes you can’t predict, and lose badly on the only thing that matters, which is what it costs to actually finish the task. Different models win different jobs. You end up managing a portfolio, not picking a winner, and the infrastructure around the models matters as much as the models.

My learning → Cost per completed task is the metric. If a vendor only quotes you cost per token, they’re hiding the bill.

4. But Also, You Are Not Consuming Enough Tokens

I burned a full month of Claude Design credits redesigning one page. What I actually want is to hand it my entire hundred-page site and wake up to better versions of all of them, then keep the two or three that win. That’s orders of magnitude more consumption and a far better outcome. The old habit is to choose between A, B and C. The new move is to build A, B, C, and their prime versions, ship them all to staging, and let the results decide. Every strong builder on your team could consume 100x what they do today and produce better work.

My learning → The ceiling on token spend is your imagination, not your budget. Run more variants than feels reasonable.

5. Net New Logos Is the Survival Metric

Everyone obsesses over retention, and in the AI era it’s the wrong number to lead with. The thing that predicts whether a public B2B company makes it is net new logo growth. Above 15% a year and you’ll figure the rest out. The moment that cracks, you can raise prices for a while, but you’re on a timer, because you can’t raise your way out of a funnel that stopped bringing in new customers.

My learning → Watch net new logos, not just NRR. The funnel tells you the future. Retention only tells you the past.

6. If Buyers Never Start With You, They Never Graduate to You

Claude Design won’t take Figma’s million-dollar accounts this year. It takes the single-seat deals at the very bottom, the ones nobody tracks. But that’s exactly where the danger is. The next generation is starting agentically and never developing the habit of your product. They don’t graduate up to you later, because they never began with you, and the loss shows up years out as a funnel that quietly stopped refilling. This is the risk to Salesforce and every workflow tool, and it’s far quieter than the vibe-coding story everyone tells.

My learning → Losing the bottom of your funnel doesn’t hurt this quarter. It hollows out the next five years. Defend where people start.

7. Know Your Real TAM, Not the Fantasy One

Rory did the math that stuck with me. There are only 1.8 million developers in the US and about $250 billion in total developer wages. If most of the frontier labs’ enterprise revenue is coding, they may already be at a fifth of that entire market, with no layoffs to show for it. Even hypergrowth eventually meets a physical ceiling, because you cannot spend more than you take in, even at 100% gross margins. The point is discipline: size off the real denominator, not the 30-million-developers fantasy number everyone repeats.

My learning → The fastest-growing companies in history can still hit the size of the till. Size your market off real data, then plan for the ceiling.

8. Model AI as About 10% of Revenue in Your COGS

The calmer, bigger number is the one to plan around. Software is a $1.4 trillion market and it’s all going agentic. Companies will tolerate roughly a 10% gross-margin spend on AI, the same way they swallowed a 7% Amazon tax a decade ago. Salesforce can comfortably pay 10% of revenue for tokens. It will never pay 40%, because it only runs 22% operating margins. Outside of coding, agentic token costs land closer to 10% than 50%, which is livable if you plan for it.

My learning → Budget tokens as a permanent ~10% tax on revenue. If your model only works at 2%, your model doesn’t work.

9. Never Take Debt Instead of an Equity Round on a Slow-Growth Business

Constellation just bought TouchBistro, a $70 million ARR company, for $70 million. One times revenue. The mechanism that killed the equity was venture debt from Francisco Partners that converted to senior when they missed and wiped out common. A big slug of debt on top of a slow-growth business is a trap: it puts your equity at risk and hands control to lenders whose only job is getting their principal back. I used to like debt. I hate it now.

My learning → Debt in place of equity on a slow-growth company is the sucker bet. You’d better be the hottest thing on earth, or it kills you.

10. Renewal Is No Longer Safety

What surprised me most this week is how fast switching costs are collapsing. Marketo is going to zero: no net new logos, everyone hates it, and they keep raising prices while deprecating the API. What used to protect it was a year-long migration. Salesforce did an LLM-powered lift and it took one day to leave. When switching costs go from a year to a day, sticky revenue stops being sticky, and terminal decay stops being slow. Rollup math that assumes a long, gentle decline is going to be wrong.

My learning → Your renewal rate is only as safe as your switching cost, and AI just cut switching costs by orders of magnitude. Earn the renewal every year.

If You Only Keep One, Keep Number Five

Every other lesson here, the token governor, the COGS model, the funnel defense, is downstream of whether you’re still adding new customers fast enough to matter. Net new logos above 15% forgives a lot of mistakes. Net new logos stalling exposes all of them, and in the AI era it exposes them faster than it used to, because the tools that erode you at the bottom and the migrations that free your customers at the top are both accelerating. Growth was always the answer. It’s just less forgiving now if you don’t have it.