Overview of 20VC: Open Models vs Frontier Models: Who Actually Wins? with Clay Bavor
Harry Stebbings interviews Clay Bavor, co-founder of Sierra and former long-time Google leader, about the future of AI models, enterprise adoption, token economics, and how Sierra is building and operating an AI-native company. The core thesis: the future is not “open vs frontier,” but a hybrid world where frontier intelligence remains essential for high-stakes, complex work, while open-weight and distilled models win on cost and task-specific efficiency. Clay also shares how Sierra uses forward-deployed engineers, internal AI agents, and a highly structured operating cadence to ship faster and stay close to customers.
Frontier Models vs Open Models
Clay’s main view
- The market has underappreciated demand for frontier-level intelligence.
- Open models are improving quickly and will cover more workloads, but they do not eliminate the need for frontier models.
- The likely future is mix-and-match:
- Frontier models for high-complexity, high-stakes, or highly creative work.
- Open-weight / fine-tuned models for cheaper, repeatable tasks.
Why frontier models still matter
- In domains like:
- coding
- science
- materials discovery
- law
- There may be effectively unbounded demand for more intelligence.
- Clay argues that as intelligence gets cheaper, people simply find more things to do with it.
Open model growth, especially in China
- He attributes part of the strength of Chinese open models to scale distillation from U.S. frontier models.
- If you can’t build frontier models yourself, distilling them is a rational next-best strategy.
Token Economics and Compute
Tokens are becoming a real budget line
- Clay expects companies to eventually allocate budgets as:
- salary
- equity
- token spend
- His rough expectation: token spend could grow to around 20% of developer salary in many environments.
- He contrasts that with the current low single-digit estimates some executives cite.
Why token usage is rising
- Reasoning models “think out loud” more, increasing token consumption.
- More capable models encourage more agentic workflows, which use more inference.
Compute remains the real constraint
- Whether models are open or frontier, GPU capacity and power still set the floor.
- Running locally on laptops or phones may help some consumer use cases, but it won’t solve frontier-scale compute needs.
- Clay sees a future with:
- more efficient hardware
- some local inference
- but frontier workloads still running in large data centers
Sierra’s Enterprise AI Strategy
Enterprise is still a team sport
- Sierra serves highly complex, regulated organizations.
- Clay says enterprise AI adoption is not just a model problem; it’s also:
- application design
- workflow integration
- change management
- trust-building
Forward-deployed engineers are central
- Sierra borrowed the forward-deployed model from Palantir.
- Engineers work closely with customers to:
- understand business context
- build the first deployment quickly
- shorten time to value
- Examples mentioned:
- Next going live in about six weeks
- Cigna going live in under two months
Sierra’s customer scope is expanding
- They started in customer support, but the platform is moving into broader lifecycle use cases:
- sales
- inbound conversion
- outbound engagement
- product discovery
- personalized recommendations
- Clay points to customers like Rocket and Next as examples of this expansion.
How Sierra Runs Internally
AI-native internal tools
Sierra has built a set of internal systems to help the company operate faster:
- MCP gateway: a central server that connects major internal systems
- Pinecone: Sierra’s internal agent, used to navigate company knowledge and workflows
- Sierra Brain: a strategy/thinking layer grounded in company docs, board letters, and operating reviews
These tools let employees reason over:
- Slack
- documents
- reviews
- board letters
- internal knowledge
Internal productivity gains
- Engineers using Claude Code, Codex, and internal tools report being 3x to 20x more productive.
- Clay sees the whole company becoming more agentic over time, not just engineering.
Token budgeting inside the company
- Sierra is not fully at per-employee token budgets yet.
- But Clay expects that to become normal.
- He views higher token usage as a sign that teams are leaning into AI productively.
Operating Cadence and Board Process
Faster board rhythm
- Sierra runs board meetings every six weeks, not quarterly.
- The format is a tick-tock:
- one 3-hour meeting
- one 1.5-hour meeting
Memo-based boards
- No slide decks.
- Brett and Clay write 6–10 page memos instead.
- This forces clarity, honesty, and better discussion.
Milestone-based fundraising
- Sierra thinks about fundraising as:
- how much capital is needed to reach the next major milestone
- not maximizing valuation at all costs
- Clay says they have often taken lower prices than they could have.
Hiring, Culture, and Values
AI-native hiring
- Sierra has redesigned engineering interviews.
- Candidates now:
- choose their coding setup
- use their own tools
- build something real with AI
- The company tests for:
- architecture
- systems design
- product thinking
- culture fit
- “smart, nice, intense”
The three values
Clay highlights Sierra’s values as:
- Craftsmanship
- Do things with excellence.
- Details matter because they compound into company quality.
- Intensity
- The company has to move fast to win in a giant market.
- Founders must set the pace.
- Family
- Work matters, but not at the expense of being a whole person.
- Clay and Brett both have large families, and they want Sierra to reflect that balance.
In-person culture
- Clay is strongly pro-in-person, especially for a young company.
- He believes:
- culture is easier to build face-to-face
- mentorship happens more naturally in person
- apprenticeship matters for younger employees
Leadership Lessons from Google
What Clay took from Google
- Invest deeply in the tech stack when needed.
- Work with great people and learn from them.
- Think at multiple zoom levels:
- strategy
- product detail
- execution
What he learned from Sundar Pichai
- Extraordinary ability to zoom from macro to micro.
- Strong product focus paired with human decency.
- A mission-driven culture can produce enormous invention.
Notable Insights and Quotes
- “We have not yet appreciated the unbounded demand for frontier levels of intelligence.”
- “Work expands to the room that you give it.”
- “If you can’t build frontier models yourself, the next best approach is to distill them.”
- “Here’s your salary, here’s your token budget, have at it.”
- “Agents all the way down.”
Key Takeaways
- The future of AI is likely hybrid, not winner-take-all.
- Frontier models will remain crucial for the hardest problems.
- Open and distilled models will dominate many cost-sensitive workflows.
- Token spend is becoming a real operational input, like headcount.
- Enterprise AI winners will need:
- deep customer intimacy
- fast implementation
- forward-deployed teams
- strong internal AI tooling
- Sierra’s advantage comes from combining:
- product rigor
- operational speed
- enterprise trust
- AI-native company design