20VC: Open Models vs Frontier Models: Who Actually Wins? | The $100,000 Token Budget Every Engineer Will Need | Why Forward-Deployed Engineers Are the Future of Enterprise AI with Clay Bavor, Co-Founder of Sierra

Summary

Clay Bavor says Sierra deliberately avoided building its own foundation models and instead fine-tunes on top of open-weights systems, while still believing demand for frontier-level intelligence remains effectively unbounded in high-stakes domains like coding, science, materials, and law. He argues that reasoning models are driving token usage sharply higher because they “think out loud,” making token spend a real operating line item that may eventually be budgeted per employee. Bavor walks through Sierra’s internal AI stack, including an MCP gateway that connects Claude, Codex, and Sierra’s own Pinecone agent to company systems, plus Sierra Brain, which grounds agents in company context and leadership docs. On the enterprise side, he says Sierra’s forward-deployed model helps customers go live quickly, with examples like Next in six weeks and Cigna in 58 days, and claims the company now serves 40% of the Fortune 50. He also explains how AI is changing hiring, board cadence, and the bottleneck in software development, which he thinks is moving from writing code to reviewing code to deciding what should be built.

Notes

• Sierra chose not to do its own frontier pretraining.
• The company viewed full pretraining as too capital intensive for most startups.
• Sierra now uses proprietary fine-tuned models on top of open-weights models.
• Bavor frames model strategy as slipstreaming behind labs and hyperscalers.
• He says frontier intelligence demand is still underappreciated.
• He links part of that dynamic to Chinese scale distillation.
• If a company cannot build frontier models itself, distillation may be the next best option.
• High-complexity domains still justify frontier models.
• Bavor names coding, science, materials science, and legal as examples.
• For simpler consumer tasks, open-model capability is already sufficient.
• Reasoning models are a major driver of rising token consumption.
• Bavor highlights OpenAI's late-2024 o-model as an important inflection point.
• He describes test-time compute as an increasingly important performance lever.
• Inference capacity and power are the practical bottlenecks for frontier use.
• Training still requires massive compute, from petaflops to exaflops.
• Local inference on phones remains constrained by thermals.
• Some model families are already dramatically cheaper than the GPT-4-era baseline.
• Sierra says AI-assisted engineers can be 3x to 20x more productive.
• The company believes the bottleneck is shifting away from code writing.
• Bavor expects review to become the immediate constraint.
• He thinks the next bottleneck after review is product judgment.
• Top engineers at Sierra are already spending over $100,000 a year in tokens.
• Bavor expects token budgets to become a formal planning item.
• He says token spend may ultimately be a sizable fraction of engineering cost.
• Sierra built a single MCP gateway for internal systems access.
• That gateway can plug into Claude, Codex, and Pinecone.
• Access is permissions-aware across internal data.
• Employees can access their own Slack messages through the system.
• Pinecone is Sierra's purpose-built internal agent harness.
• The company has harnesses around engineering, the core platform, and agent studio.
• Sierra also maintains a shared internal skills library.
• Sierra Brain is a company context layer for strategic reasoning.
• It is grounded in board letters and operating reviews.
• Bavor says the system helps him reason about company direction.
• Sierra serves 40% of the Fortune 50.
• Half of Sierra's customers are over $1 billion in revenue.
• Thirty percent of customers are over $10 billion in revenue.
• These are unusually large and regulated enterprise accounts.
• Sierra built a forward-deployed team in early 2024.
• The rationale was that nobody had deployed AI agents in this customer-facing way before.
• The model embeds Sierra deeply in customer business mechanics.
• Next went live in six weeks from kickoff.
• Cigna went live in 58 days.
• Sierra says its platform is extensible and transparent.
• Bavor sees customer interaction becoming agent-mediated.
• He expects conversation to become the interface.
• Intensity is required because the market is competitive and fast-moving.
• Founders must set the tone for intensity.
• Founder mode should be applied selectively.
• Sierra uses a six-week board cadence.
• Board meetings alternate between three hours and ninety minutes.
• The company uses board memos instead of decks.
• Those memos are usually six to ten pages long.
• Writing is used as a forcing function for clearer thinking.
• Bavor says an early miss was hiring too slowly.
• Demand was visible but recruiting did not scale fast enough.
• Sierra thinks about fundraising as milestone-to-milestone capital planning.
• Across rounds, Sierra says it accepted lower prices than it could have gotten.
• Craftsmanship means excellence across people, process, product, and culture.
• During a major retail launch, leadership personally reviewed every agent conversation.
• Sierra uses that scrutiny to ensure customer outcomes are correct.
• Bavor says young companies need in-person work to build culture and norms.
• He argues apprenticeship and mentorship are stronger in person.
• Bavor says young graduates have an AI-learning advantage.
• Sierra's engineering interviews now require candidates to build with AI tools.
• The interview tests architecture, systems design, product thinking, culture, and values.
• Bavor sees cybersecurity as a strong bet because offensive AI has improved sharply.
• He wonders whether the same model vendors will become the defenders.
• Brett and Clay use independent thinking before converging.
• Values work starts with separate drafts and then comparison.
• Family is a core value for both founders.
• Bavor says family matters more than Sierra.
• He emphasizes rituals and good habits in parenting.
• He describes family dinner and maker mornings as recurring practices.
• He recommends making children's interests part of his own life.
• Bavor frames a great marriage as a partnership with shared goals.
• He credits his parents with supporting his early computer obsession.
• The Wright Brothers is his recommended entrepreneurship book.