SaaStr Takes on This Week’s 20VC: Why OpenAI Giving the Govt 5% Might Make Sense, the Death of Block Risk, and Why Frontier Models Are the Cheapest

A few days after every 20VC x SaaStr pod, I’m gong to write up the stuff going foward I went deep on, and wanted to go … even deeper on.

We covered a lot this week with Harry Stebbings and Rory O’Driscoll, from Washington lifting the Fable 5 ban to Sam Altman floating a 5% government stake to China owning the top open models. Our weekly full recap the other day runs through all of it across the three of us. This is the narrower version about the SaaStr AI perspective.

1. Small ownership stakes can, sometimes, buy shockingly large alignment

Everyone’s first reaction to Sam floating 5% of every AI lab to the U.S. government was that it’s insane. My reaction was the opposite, and it comes straight from watching my own portfolio. When a founder sells 5% to a giant strategic partner, the way Klaviyo gave Shopify a slug so Shopify didn’t crush them, that stake often drags the big partner into your corner far more than the math says it should. Five percent is immaterial to a $100B+ partner’s balance sheet. And yet it consistently changes how they show up for you. Not always in my experience, but often enough.

So if bringing the federal government on to your cap table for <=5% makes you the good guy and buys you real alignment heading into a decade of regulation, you take the dilution. The only nuance: Sam anchored at 5%. If AI really is the labor threat these labs keep saying it is, the political appetite may not stop at 5.

Learning → A 5% stake can buy 50% alignment with the #1 partner in your space, at least, sometimes

2. In the Age of AI, seed investors really pay about 4x the headline price

I used to tell myself my real entry price as a seed investor was double the headline, because of dilution over the coming years. In 2026 I’ve moved to four times.

A round at a $60M valuation is really a round priced in a sense at $240M once you count every round still to come, and there are often a lot of rounds still to come now. Founders are often running 12, 16, 20 rounds, each at 5% or 6% dilution. It’s probably the play today in most cases to build an AI decacorn. It’s just much more dilution than in the past.

Learning → Your real entry price in seed investing is effectively 4x the headline for many AI leaders.* Underwrite the dilution along with the round.*

3. The decline of VC “block risk” changed founder behavior more than any valuation trend

Founders have in many cases stopped worrying about returning their last high-priced round. Investors have learned to take the 1x and move on. No drama, no blocking, no threats. The exception is non-standard money, but the standard players don’t block exits.

That single change removed the fear that terrified my generation of founders. We were scared of getting blocked by the last money in, of those 2x-or-you-can’t-sell terms. That fear is mostly gone, and it freed up an enormous amount of risk-taking.

Learning → *Blocking risk from your VCs in M&A is muchly though not totally dead. Don’t let anxiety around it necessarily stop you from raising the capital you need. *

4. When your core prints cash, you’ve earned the right to play, not the obligation to have answers

Meta doesn’t know if much of its $70B in AI spend pays off. It might end up like VR for them. Doesn’t matter. The core, WhatsApp, Instagram, Facebook, throws off so much cash that there’s no fatal-error risk in continuing to invest. When that’s true, you go all-in without having the answers yet. You don’t sit it out.

I give the same advice to every founder whose base business is healthy: stay in the game. A strong core doesn’t obligate you to be right. It gives you permission to swing.

Learning → A cash-rich core doesn’t owe anyone answers. It’s earned the right to swing before it has them.

5. Love your early adopters like your life depends on it, because it does

Nvidia’s compute-now-pay-later structure is arguably aggressive accounting, but the instinct behind it is correct: capture the customer early and lock in the relationship before anyone else can. Showering startups with infinite love in their first 24 months is the highest-ROI thing an incumbent can do. If there’s any lock-in at all, any switching cost, the payoff compounds for years.

I’m constantly shocked at how many companies get this wrong. They optimize the enterprise logo and treat the tiny early customer with big potential as just … a tiny customer. That tiny early customer is the one who compounds. Under-invest here and you pay for it later, watching a competitor own the relationship you could have had for the cost of some attention.

Learning → Your smallest early customer compounds harder than your biggest logo. Love them accordingly.

6. The frontier model can be the cheaper model

I spent about 10 hours in Replit on a cheaper model stack trying to crack an algorithm I’ll admit I’m not smart enough to fully understand. I got nowhere. I passed the same problem to Fable and Opus and solved it in 20 minutes. The cheaper model cost me a full day and $500. The frontier model was faster and effectively free inside my existing plan.

Price in your own time and “expensive” and “cheap” flip. On genuinely hard problems, the frontier model is the low-cost option, because the alternative is a day of your life on a mediocre answer that doesn’t work. Reach for the best model first when the problem is hard and unbounded. Save the cheap models for the commoditized answers where you already know the shape of the solution.

Learning → Price in your own time and the frontier model is often the cheapest thing you’ll buy. Use it first on anything hard.

7. The cheapest token isn’t the cheapest answer

AI customer support is standardizing around 50 cents per resolution, down from a dollar, and everyone is racing to open source to hit that number. Fine. But I’m seeing genuine plateauing in resolution quality as a result. Getting from 40% resolution of not-that-hard problems to 95% resolution of genuinely complex ones may require going back to premium models.

Optimize purely for cost per call and you cap your resolution rate, then hand the hard tickets back to humans anyway. The customer doesn’t buy tokens, they buy resolved problems. Optimize for the outcome. What makes this market good is that the ROI is legible: customers already say “I’ll happily pay 50 cents to resolve what used to cost me three dollars an email.” When they can see the return that clearly, you can afford to spend up for quality.

Learning → Customers buy resolved problems, not tokens. Optimize the outcome, not the invoice.

8. Restrict access and you don’t stop a competitor, you fund one

We just spent two weeks in China and Hong Kong, and one thing didn’t hit me until I was on the other side of the Great Firewall: you cannot access Claude or OpenAI at all. It isn’t throttled, it’s unavailable, and the VPN workarounds are partially blocked too.

So of course the second-largest economy in the world, full of excellent engineers who’ve been building software for decades, went and built world-class models of its own. The top six models on OpenRouter are Chinese. If we don’t like what China is building, we should understand that we helped create it by locking them out. Jensen was right. Cut off access and you don’t slow the competitor down, you guarantee they build the thing themselves.

Learning → Cut off a market and you don’t slow the competition down. You fund it.

9. AI services businesses will be gated by talent depth, not demand

Microsoft and Amazon are both throwing thousands of people at embedding engineers inside enterprise clients. On a spreadsheet it makes sense. In practice I think it struggles, and the reason is depth. We work with some of the best forward-deployed engineers at the leading vendors, and even at the hottest companies the bench is one or two people deep.

One story from our own vendors makes the point. A public-company vendor