Overview of 20VC with Harry Stebbings
This episode is a fast-moving weekly tech roundup with Rory O’Driscoll and Jason Lemkin, centered on the biggest AI, software, and market-structure stories of the week. The discussion is blunt and opinionated: it argues that AI spending is entering a more mature phase, that frontier model vendors may face real competition from open source, that prediction markets are increasingly mainstream, and that software roll-ups like Bending Spoons may be one of the most interesting public-market plays of 2026.
Coinbase, AI Spend, and the “Show Me the Money” Era
The conversation opens with Coinbase’s disclosure that it cut AI spend by about 50% while increasing usage, largely by shifting toward open-source models and better token routing.
Main takeaways
- Rory’s view: the post was valuable because it was concrete, fact-based, and showed how a real company reduced AI costs without sacrificing usage.
- Jason’s view: he’s tired of “performative AI” posts from CEOs who are not running AI-native companies; he wants to see revenue lift, not just impressive token graphs.
- The key point is that many companies have ramped AI spend aggressively, but:
- they are not seeing proportional productivity gains,
- and CFOs are now demanding ROI discipline.
- For software companies, Jason argues, AI should either accelerate growth or it risks being irrelevant.
Bigger implication
The Coinbase example is less about Coinbase specifically and more about a broader market reset:
- AI spend is moving from “experiment everywhere” to cost-justified deployment.
- Frontier model vendors may see growth pressure if customers realize they can do more with cheaper open-source alternatives.
Anthropic vs Open Source and the Distillation Fight
A major segment of the episode focuses on Anthropic’s claims that Chinese open-source model makers are distilling U.S. foundation models by sending large volumes of prompts to extract behavior and training data.
Core discussion points
- The hosts acknowledge the irony:
- foundation model companies were themselves trained on other people’s IP,
- yet now they are arguing that competitors are stealing theirs.
- Still, the practical question is whether this is:
- a terms-of-service violation,
- a copyright/trade secret issue,
- or a national security issue that could trigger government intervention.
What Anthropic may be trying to do
The discussion suggests Anthropic may be pushing for a policy regime where:
- U.S. companies are discouraged or banned from using Chinese models,
- especially if those models were distilled from U.S. frontier systems,
- and open-source competitors are framed as both IP thieves and security risks.
Risk to the market
- The hosts see this as potentially regulatory capture dressed up as national security.
- Their bottom line:
- If distillation happened, punish it.
- But they are skeptical that this should justify broad bans on models, especially if there is no real security threat.
Microsoft’s Weak AI Narrative and Azure Deceleration
Microsoft comes up as another example of a giant tech company whose AI story is looking less compelling.
Why the stock is under pressure
- Microsoft is having one of its worst stretches in years, with the market reacting to:
- Azure growth deceleration,
- lack of a strong standalone frontier model,
- and the idea that much of its AI story is really just selling inference to OpenAI.
- The hosts argue Microsoft’s core franchises—knowledge work and developer tools—are being challenged by:
- ChatGPT/Copilot-style tools for knowledge workers,
- Claude Code and similar tools for developers.
Key framing
- In AI, the market wants acceleration, not just “good enough” growth.
- A decelerating Azure number is seen as a canary in the coal mine.
- The broader message: even large incumbents need a clear, defensible AI product of their own.
Kalshi, Prediction Markets, and the Casino Economy
The episode also covers Kalshi’s reported move toward a $40 billion valuation, up sharply from its prior round.
Why investors care
- Kalshi is being viewed as part of a broader shift toward:
- legalized betting,
- prediction markets,
- and “casino-like” consumer behavior in finance and sports.
- The hosts see two huge TAM drivers:
- sports betting,
- and financial/perpetual markets that function like betting on price action.
Takeaways
- Political prediction markets are interesting, but probably too small to justify massive scale on their own.
- The real upside is in:
- sports,
- and financial speculation.
- The discussion frames this as a major behavioral trend: people love to bet, especially on sports and money.
Bending Spoons and the New Age of SaaS Roll-Ups
One of the strongest segments is about Bending Spoons, which is about to go public and is seen as the anti-AI, but very compelling, software story.
What Bending Spoons does
- It buys aging consumer/software products and optimizes them:
- raises prices,
- cuts costs,
- improves operations,
- and extracts more value from sticky customer bases.
- The transcript references assets like AOL and Evernote as examples of the kind of products they own.
Why the model works
- These are products with:
- large existing user bases,
- decaying or stalled organic growth,
- and a lot of room for operational improvement.
- The hosts think there may be dozens or hundreds of similar targets.
The debate on valuation
- Bending Spoons may come public at roughly $20 billion, implying a rich revenue multiple.
- Jason and Rory think that, while the valuation looks high, the company may deserve a premium if it can keep executing.
- They also think this could be a template for:
- B2B SaaS roll-ups, not just consumer software.
B2B roll-up opportunity
They name examples like:
- Marketo
- PagerDuty
- Asana
- SEMrush
The thesis:
- many legacy software companies have:
- sticky customers,
- broken cultures,
- poor product velocity,
- and no AI strategy.
- A disciplined acquirer could buy them cheaply, install motivated operators, and re-accelerate growth.
Chamath, 80/20, and AI Software Factories
The podcast briefly discusses Chamath Palihapitiya’s new AI startup, 8090.
General reaction
- The hosts are skeptical of founders who are split across too many projects.
- Still, they give Chamath credit for actually “being in the arena” and trying to build something in a crowded market.
Broader principle
- They are wary of wealthy VC-style CEOs who are not fully committed.
- Their point is that AI software building requires intense focus, not side-project energy.
The Series A Market and Brutal Honesty
A side discussion focuses on Harry’s tweet about turning down a founder whose revenue was projected to go from $1.5M to $5M, which he argued was not strong enough for a Series A in today’s market.
Consensus
- The underlying statement was broadly correct:
- the bar is much higher now,
- and opportunity cost of capital matters more than it used to.
- But the panel also notes this can be hard to say without sounding dismissive of a founder’s work.
The practical advice
- Founders need to understand that:
- what used to be top-quartile growth may no longer be enough,
- investors are hunting for standout outcomes,
- and raising money requires an honest read on market expectations.
Claude Tag, Slack, and the Enterprise Battlefield
The show closes with a discussion of Claude Tag in Slack: Anthropic’s idea of embedding Claude as a fully present agent in a Slack channel.
Why it matters
- This could be a major enterprise wedge if it works well:
- it can observe how teams actually work,
- capture context outside core software systems,
- and potentially automate workflows across tools like Salesforce and HubSpot.
- That said, the hosts are skeptical it will immediately replace existing tools like Slack’s own bot.
Strategic significance
- If AI agents become truly autonomous in Slack, then traditional SaaS applications could become “dumb databases.”
- This is one of the most important enterprise software battlegrounds:
- whoever owns the context layer may own the workflow layer.
Final Takeaways
The episode’s big themes
- AI spending is entering a discipline phase: companies want measurable ROI, not just impressive usage.
- Open source is a real threat to frontier model economics, not just a side story.
- Regulation may become a weapon in the AI model wars, especially around Chinese models and distillation.
- Legacy software roll-ups may be one of the most underappreciated opportunities in public markets.
- AI winners will be companies that can show lift—in revenue, productivity, or both.
The hosts’ shared message
The era of hype-only AI narratives is ending. The market wants proof:
- proof of savings,
- proof of revenue lift,
- proof of durable advantage,
- and proof that AI is actually changing the business.