20VC x SaaStr: When the Agents Pick the Models, OpenAI Comes Back to Life, and Thoma Bravo Just Wiped Out $5.1B on Medallia

With Harry Stebbings, Jason Lemkin, and Rory O’Driscoll

🎤 Breaking!! Rory is joining LIVE at SaaStr AI 2026, May 12-14 in the SF Bay Area for an AMA on Investing In the Age Age and more, along with Amelia and me. Come see the 20VC x SaaStr conversation & meet the team live in person. Get your tickethere.

OpenAI missed Q4 numbers (apparently). Anthropic just took $45 billion from Google and Amazon to keep up with compute demand. Nvidia hit $5 trillion. Google hit $4 trillion. Thoma Bravo handed Medallia to its creditors in a $5.1 billion equity wipeout. China blocked Meta’s $2 billion acquisition of Manus. And Gary Tan just published the closest thing the seed market has to a revenue accounting standard.

But the story underneath all of these stories is the same one. The agents are going to pick the models, the vendors, and the workflows. Not humans.

That’s what changes everything. It changes who wins the foundation model wars. It changes which B2B companies have terminal value. It changes whether Canva’s IPO becomes a generational outcome or a sensibly-priced PE deal in disguise. And it explains why the entire PE-for-B2B playbook that powered Thoma Bravo, Vista, and Francisco Partners for a decade may be quietly breaking.

Here’s the breakdown:

Top Takeaways

1. The Agents Pick the Models, and That’s Why OpenAI Just Got Back in the Game

Last year’s narrative was simple. Anthropic had something special in coding. Claude was the model humans loved. OpenAI’s models slipped, market share moved, and Wall Street is just now waking up to numbers that were knowable three months ago.

Jason’s view: that whole story is yesterday’s war. Going forward, the agent picks the model. The agent picks the vendor. And humans become bystanders to those decisions.

“I see no competitive advantage to Claude for most workflows,” Jason said. “Open AI, whether it’s Codex 5.5, was just state-of-the-art of the models. It’s so good for my workflows that I think the advantage that humans get out of Claude and Claude Code… I’m not sure our agents are going to get the same advantages.”

His own agents, the AI VP of Marketing and AI VP of Customer Success at SaaStr, prefer the OpenAI API. So that’s where the money goes. “Just like you got to back your team of humans in the old days like 2024, today I have to back my team of agents. If they pick OpenAI, I’m I got to be on the team.”

This reframes the entire Anthropic vs. OpenAI debate. Claude won 2025 because humans loved it for vibe coding. The 2026 battle is whether agents prefer the same models humans did. Early signal says no.

2. The Three Buckets of B2B Software in the Agent Era

Rory laid out a framework that’s worth pinning to the wall. Every B2B software company now lives in one of three buckets:

**Bucket 1: Eroding terminal value.**Melting iceberg. Low stock price. If you’re levered, you’re dead. Medallia ended up here.**Bucket 2: System of record.**They’ll keep you forever, but no real agentic activity sits on top. Calculable terminal value, fair price, no premium. Workday lives here. Atlassian is being interrogated about whether it lives here.**Bucket 3: Agent-leveraged.**The agents are using you. You get increasing returns from AI leveraging your platform. ServiceNow is being graded daily on whether it can credibly claim this status.

Most of B2B is currently being repriced from bucket 3 (where the market gave SaaS pixie dust credit until 2025) into bucket 2. Some is sliding into bucket 1. The single most important question for any B2B company right now is: do agents want to use this?

3. Canva Will Have a Very Successful IPO. Agents Still Won’t Use It.

This is the cleanest illustration of the new framework. Canva 2.0 is great. The agentic suite is real. The product is shipping. The IPO will price.

It will not get pixie dust credit.

“Would an AI agent use it?” Jason asked. “No. An AI agent is not going to go in and move assets around, vibe them. It’s just going to create the assets.”

Same logic applies to Jira, Confluence, and most of the project management category. Agents don’t need project management tools. They have no use for them. The market figured this out before the narrative did. That’s why Atlassian and Monday have struggled while Twilio and Cloudflare (which agents do still use) have outperformed.

The lesson for IPO-stage companies: scale plus profitability gets you a real-company multiple. The 30x revenue premium of 2021 is not coming back. SaaS pixie dust credit expired in 2025. As Rory put it, “good SaaS companies that aren’t AI-first are going to trade at fair value, which means if you’ve created value, you’ll get value. But what they won’t get is that stupid 30 times revenue premium.”

4. The Compute Equals Revenue Thesis Just Got Its First Crack

Sam Altman’s framing has been: compute equals revenue. Build the data centers, the demand will come.

OpenAI just printed a quarter that suggests this isn’t always true. They had the compute. They didn’t have the model. Demand softened. Anthropic ran the opposite play, built better models, ran out of compute, and is now scrambling to catch up via the $45 billion Google and Amazon deals.

Rory’s read: “It’s a stupid statement by Sam. It implies causation. It’s just correlation. No compute equals no revenue. But compute and a shitty model also equals no revenue. See Grok for details.”

The real version of the thesis is harder. You need enough compute to meet demand AND a good enough model to generate demand. Both at the same time. Both two years out. The capital intensity makes this a four-to-one ratio: every $1 of run-rate revenue requires roughly $4 of capex. If you’re forecasting 5x growth this year and 4x next year on a $10B base, someone between you and your hyperscaler partners has to find $300 billion of capex in the next 24 months.

Get it wrong on the high side, you’re capacity-stranded. Get it wrong on the low side, you can’t serve the demand. There is no version of this where the bet is small.

5. Google Wins No Matter Who Wins

The under-appreciated story in the Anthropic raise is what it does for Google. Google now wins three different ways: if Gemini wins, if Anthropic wins (because Anthropic is now deeply tied to GCP and TPUs), and if neither wins (because Google has the spare capacity to redirect from internal workloads to whichever foundation model needs it).

The chip economics matter here. GPU spend is roughly 50-55% of total capex on any new buildout. Nvidia’s gross margins are 70%, which means roughly $14 billion of profit per gigawatt of data center built. Google and Amazon are bundling their custom chips, their capital, and their equity to shift that gross margin away from Nvidia and into themselves.

Asked which to buy at maximum upside, Nvidia at $5T or Google at $4T: Jason and Rory both came down on Nvidia for raw upside, Google for risk-adjusted return. The one thing that breaks Google: ChatGPT eroding search. As long as that doesn’t happen, Google has multiple ways to win.

6. Thoma Bravo Just Handed $5.1B of Medallia to Creditors

This is the story most people will sleep on. It shouldn’t be.

Thoma Bravo bought Medallia in 2021 for roughly $6 billion, with about $5 billion of equity and $2 billion in debt. Not over-levered by historical PE standards. They just handed the keys to creditors. The full equity is gone. It’s the second total wipeout in the category after Pluralsight.

The brutal insight: it wasn’t the leverage that killed it. It was the price. As Jason put it, “You can’t service 2 billion plus of debt on a 1 billion low growth company with a pre-AI story that has to transform to AI.”

Customer survey software is exactly the kind of category CIOs cut first. Vendor consolidation is taking 30-50% of new AI dollars. Medallia’s net retention had reportedly dropped to 21%. There’s no AI story, and writing one would require a full rewrite to compete with much better AI-first produc