Overview of 20VC
In this episode, Harry Stebbings, Rory O’Driscoll, and Jason Lemkin dig into the biggest AI and tech-market stories of the week: China’s rapidly improving open-weight models, whether the U.S. should restrict access to them, whether OpenRouter should sell, why the inference layer is getting so valuable, and what Stripe’s reported move on PayPal could mean for payments and private-tech M&A. The throughline is simple: AI demand is still overwhelmingly concentrated in infrastructure, pricing power matters, and the next 12–24 months of growth at OpenAI and Anthropic may determine a huge amount of the market’s direction.
Chinese Open-Weight Models: Kimi, Qwen, and the U.S. Response
What the hosts think the model announcements mean
- China’s recent model releases, especially Kimi and Alibaba’s Qwen, are notable but not shocking.
- The panel’s view: Chinese model teams have been steadily closing the gap for years; this is more acceleration than a true step-change.
- One key point: evals are not the same as real-world performance. The hosts want field usage, not just benchmark tweets.
Why the market is reacting
- Demand for cheaper equivalent models is exploding.
- OpenRouter data reportedly already shows a large share of traffic going through China-created or China-origin models.
- The hosts argue this is no longer niche; cheaper models could become mainstream across consumers and enterprises.
Should the U.S. ban Chinese models?
- The group is skeptical of a blanket ban.
- They agree there are real security, IP, and data-exfiltration risks, especially when models are connected to tools and enterprise data.
- But they also think a total prohibition would be overkill and would likely hurt U.S. buyers who just want lower-cost inference.
- Their likely prediction: some restrictions and tighter scrutiny, not a full shutdown.
Policy and politics
- The episode notes the backlash around OpenAI policy adviser Dean Ball and comments framing Chinese models as “AI communism.”
- David Sacks and Emil Michael are cited as prominent anti-ban voices.
- The hosts repeatedly return to the same tension: security concerns are real, but so is the market demand for cheaper intelligence.
OpenAI, Anthropic, and the Economics of the Frontier
The key question
- The hosts repeatedly say the most important variable is OpenAI and Anthropic’s growth rate in 2026 and 2027.
- If they can keep growing fast with improving gross margins, the current economics hold together.
- If growth slows materially, the whole market could reprice.
Why this matters
- Frontier model companies are now massive businesses with huge capex, training, and inference costs.
- Their economics depend on:
- sustaining very fast growth,
- keeping margins improving,
- and avoiding a race-to-the-bottom on price.
Main takeaway
- The hosts think this is the real “million-dollar question” behind almost every AI valuation debate right now.
OpenRouter and the Routing Layer: Great Time to Sell?
Why they think OpenRouter is attractive M&A
- The hosts say it may be a great time for OpenRouter to sell because:
- routing is becoming a standard part of the stack,
- more platforms are building their own version,
- and the market is moving fast enough that strategic buyers may pay for distribution and plumbing.
Who might want it
- Potential buyers include:
- hyperscalers,
- enterprise infrastructure platforms,
- and companies trying to reduce model dependency.
- The strategic value is less about pure standalone NPV and more about control of the plumbing.
Founder perspective
- The discussion stresses the difference between:
- a VC’s return math, and
- a founder’s life decision.
- Their view: if an offer is “nosebleed” in absolute terms, founders should only pass if they truly believe the company can be 10x bigger.
Fireworks, Inference, and Why Infrastructure Is Winning
Fireworks as a case study
- Fireworks’ growth is used to show how big inference has become.
- The hosts see inference as a huge business because:
- open-weight model usage is rising,
- enterprises want model choice,
- and many workloads are still routed through inference providers.
Why margins may improve
- Even if these companies buy compute from near-clouds at first, demand is now so strong that they can raise prices and improve gross margin.
- The hosts think some of these infrastructure businesses are shifting from “maybe low-margin” to very attractive.
The big risk
- Inference companies may eventually need to verticalize into data centers to control destiny.
- That would mean a lot more capex and a lot more complexity.
- So yes, the business is strong — but commoditization risk is still real.
Custom Models, Labeling, and the Rise of Domain-Specific AI
What the episode says about “your own model”
- The hosts strongly agree that every company eventually wants its own model or reasoning layer.
- Even if it’s not a literal foundation model, companies want:
- their own workflow logic,
- their own labels,
- their own domain-specific tuning.
Why data labeling matters
- Jason says building and labeling his own recruiting app made the output dramatically better.
- The key lesson: generic LLMs are good, but domain expertise creates a step-function improvement.
- That supports the thesis that data-labeling and model-tuning companies can still be very valuable.
Stripe Buying PayPal: Why It Makes Sense
Strategic logic
- The hosts think a Stripe-led acquisition of PayPal is plausible and potentially smart.
- Stripe can use PayPal’s scale, consumer assets, and legacy distribution to expand its footprint.
- PayPal is viewed as a company with big assets but major operational drag.
Why the deal is tricky
- Stripe is much faster-growing than PayPal, so the acquisition could dilute Stripe’s growth rate in the short term.
- Integration will be messy:
- different codebases,
- different operating cultures,
- and a big turnaround challenge.
Deal mechanics
- The board’s rejection is interpreted as part of the normal negotiation dance.
- The hosts think the eventual outcome is likely a higher price and a completed transaction, especially if advisors and bankers do their usual back-and-forth.
Bigger picture
- The conversation highlights how private markets, liquidity windows, and strategic M&A can create unusual opportunities when an iconic asset is temporarily depressed.
Broader Market Themes
Late-stage growth has been a great place to make money
- The hosts say the last 2–3 years have been especially favorable to growth-stage investing.
- When a hot company is still growing quickly, even large checks can re-rate fast.
Tranche rounds and transaction-heavy markets
- They discuss tranche pricing as a symptom of a more transactional market.
- It creates headline optics and bragging rights, but also more complexity and resentment.
- Their take: it’s not ideal, but it’s not the worst thing a founder can do if the capital is valuable.
Infrastructure is still where the money is
- The recurring refrain: the biggest dollars in AI are still in infrastructure, not apps.
- App-layer revenue is real, but still tiny relative to model training, inference, and compute.
Additional Notes and Side Topics
- Databricks’ huge raise is another example of infrastructure commanding massive valuations.
- Valor Atomics / nuclear came up as a side note on private-company progress in energy.
- TSMC, ASML, and DRAM were discussed to show how AI demand changes supply-chain behavior:
- some parts of the stack behave like long-term partnerships,
- others behave like brutal commodities.
- C-Square’s IPO was used as an example of a less compelling, slower-growth asset not getting the same AI premium.
Main Takeaways
- China’s open-weight models are a real competitive force, even if the benchmarks aren’t the whole story.
- The U.S. is more likely to impose targeted restrictions than a full ban.
- Inference and routing layers are becoming increasingly valuable as AI adoption broadens.
- OpenRouter and similar businesses may be at an M&A inflection point.
- The future value of the AI market still hinges on OpenAI and Anthropic’s growth and margins in 2026–2027.
- Stripe buying PayPal would be a bold strategic move, but one that fits the current private-market and payments landscape.
Notable Lines from the Discussion
- “The only thing that matters is the OpenAI and Anthropic growth rate in 26 and 27.”
- “I think it’s a great time for OpenRouter to sell.”
- “The quest for equivalent models at a cheaper price is just going to keep going up.”
- “All the good investments sure seem to be in the infrastructure.”