Overview of 20VC: Anthropic, Compute Wars, IPOs, and the Public-Market SaaS Reset
Harry Stebbings and guests Rory O’Driscoll and Jason Lemkin unpacked a huge week in tech: Anthropic tightening control over secondaries, its compute deal with Elon Musk’s xAI/SpaceX ecosystem, a massive $200B compute commitment to Google, the Cerebras IPO, Ramp’s $40B valuation, and a wave of confusing public-market reactions across SaaS. The episode’s throughline was clear: AI demand is exploding, compute is becoming the new strategic bottleneck, and many legacy software businesses are being forced to prove they still matter in an agentic world.
Anthropic, Compute, and the AI Supply Chain
Anthropic tightening the cap table
- Anthropic said all secondary sales and SPVs now need board approval.
- The discussion framed this as:
- a normal move for a company trying to regain cap-table control ahead of an IPO,
- a way to stop messy, unauthorized economic transfers,
- and a signal that AI “greed” in secondary markets had gotten out of hand.
- The hosts noted that this kind of restriction is not unusual in modern startup charter docs.
Compute deal with Elon’s ecosystem
- Anthropic’s deal to use Elon Musk’s compute capacity was interpreted as a smart, pragmatic market reset.
- The key idea: companies are reallocating scarce compute toward the models that can generate the best returns fastest.
- In their view, this shows:
- AI infrastructure is consolidating,
- weaker players may become net sellers of capacity,
- and Anthropic is acting like the most aggressive buyer in the market.
The $200B Google commitment
- Anthropic’s reported $200B commitment to Google over five years was described as a major sign of the circular AI economy:
- Google supplies TPU capacity,
- Google also has Gemini,
- and Anthropic is both a customer and a quasi-competitor.
- The hosts emphasized that this underscores how dependent hyperscalers are becoming on a small number of frontier model companies.
Tokens, Parallel Agents, and What AI Does to Software Economics
The token consumption debate
- Goldman’s estimate that agentic workflows could drive 24x token growth by 2030 was debated.
- Jason Lemkin argued that figure may be too low because parallel agents could multiply usage far beyond current assumptions.
- Rory added that token economics are hard to forecast because:
- chip performance keeps improving,
- inference gets more efficient,
- but usage also expands as outputs get better.
Parallel agents change the game
- A major theme was that software is moving from sequential to parallel work.
- Example: an AI can generate 10 versions of a feature, compare them, and surface the best ones for approval.
- That means:
- more tokens consumed,
- more throughput,
- and potentially much faster product iteration.
But not everyone needs more tokens
- The hosts also discussed a counterpoint from strong CTOs:
- many teams are producing too much code,
- much of it never reaches production,
- and “token maxing” can be wasteful.
- Their split view:
- top engineers may become dramatically more productive with AI,
- mediocre engineers may simply consume huge amounts of tokens for limited value.
Software categories at risk
- They drew a distinction between:
- new software being eaten by models
- and old software decaying because agents make it obsolete
- Examples of software that may decay faster:
- marketing automation tools like HubSpot/Marketo,
- basic workflow products that agents can now handle directly.
- Core verticals like legal, accounting, customer support, and financial modeling were seen as more defensible, but still under pressure.
Public Markets: SaaS Winners, Losers, and Valuation Reality
Monday vs HubSpot
- Monday.com was praised for raising guidance and showing at least some resilience.
- HubSpot was punished for lowering guidance despite a decent quarter.
- Main takeaway:
- if you are not accelerating, you need to at least raise guidance,
- because markets still fear terminal decay in the AI era.
Cloudflare, AppLovin, and price sensitivity
- Even strong quarters did not always translate into stock gains.
- The episode emphasized that valuation matters as much as fundamentals:
- high-multiple stocks are fragile,
- and even good results can disappoint if the story is messy.
- Their broader point:
- public markets are repricing old-school SaaS based on AI risk and growth durability.
ZoomInfo as a cautionary tale
- ZoomInfo was held up as a brutal example of AI and adjacent tools stealing growth.
- Clay was credited with commoditizing data-provider workflows by letting customers compare multiple sources and automate data enrichment.
- The result: what was once a differentiated data business now looks increasingly replaceable.
IPOs and Big Financings
Cerebras IPO
- The Cerebras IPO was described as one of the most anticipated of the year.
- Key points:
- heavily oversubscribed,
- range raised materially,
- likely to pop on debut.
- The long-term view was more cautious:
- the company is betting on inference demand and speed,
- but the real test will be whether it can sustain a durable business versus Nvidia and other chip platforms.
- They also highlighted Cerebras’ journey as a major venture win and a strong founder story.
Ramp’s $40B valuation
- Ramp’s new valuation was discussed as a sign of both strength and market exuberance.
- Why investors like Ramp:
- it started with cards,
- but expanded into a much broader spend-management and automation platform,
- including AI agents for procurement and spend optimization.
- Concerns:
- the valuation implies years of very strong growth,
- and fintech businesses can still be heavily exposed to multiple compression if growth slows.
Other names
- Parker’s Chapter 7 filing was noted as a contrast to Ramp’s success.
- Gusto crossing $1B in revenue was mentioned as another sign that payroll/fintech infrastructure can scale very far.
- Lime’s IPO prep was called a positive sign that the micromobility business has become more durable than many expected.
Musk vs. Altman and the Legal Overhang
- The hosts briefly covered the Musk vs. Altman trial.
- Their view:
- a lot of noise will come out in discovery,
- but the judge will decide the core legal issues,
- and their gut is that OpenAI likely keeps the key deal structure intact.
Founder Intensity, Mental Health, and What It Takes to Win
The “sacrifice” debate
- Harry referenced the idea that success often requires sacrificing mental health and balance.
- The guests agreed that:
- real success usually requires serious sacrifice,
- founders often give up time, relationships, and stability,
- and the job is emotionally rewiring over time.
But don’t burn yourself out
- They also argued that being completely on tilt makes people worse decision-makers.
- The more mature view:
- intensity is necessary,
- but founders still need enough health and clarity to function well.
- Their practical takeaway:
- if you’re overloaded and spiraling, step back before you make bad decisions.
“Take the offer” principle
- Jason emphasized that founders should often take a strong acquisition offer if they’re not fully committed to the next level of intensity.
- The subtext:
- only unusually intense founders can survive the long, punishing path to outsized outcomes.
Conference and Content Trend: Podcasts Beat Firesides
- Jason argued that podcasts have largely replaced traditional conference firesides.
- Why:
- people can hear the same ideas in a better format,
- and workshops or live demos are more valuable than repeated speaking panels.
- His view for SaaStr:
- fewer “speaker” sessions,
- more live demos of agents and real product workflows,
- more practical, technical content.
Main Takeaways
- Anthropic is acting like a company preparing for a much larger future: tighter cap-table control, massive compute commitments, and aggressive capacity acquisition.
- AI infrastructure is consolidating around a small set of winners, with compute becoming the new strategic moat.
- Parallel agents could massively expand token usage, but the market may still be overestimating how much software actually needs.
- Legacy SaaS is being re-rated brutally; the market wants growth, guidance raises, and evidence of AI defensibility.
- Cerebras and Ramp both show that investors will pay up for credible AI and fintech infrastructure stories, but only if they believe the growth runway is real.
- The episode’s broader thesis: in the agentic era, software either accelerates, adapts, or decays fast.