20VC: Why OpenAI and Anthropic Won't Win the App Layer | Why Teams Will Get Bigger Not Smaller in a World of AI | Why AI Removes Incumbents Advantage of Bundling | China vs America: Who Wins the AI War with Arvind Jain, Co-Founder @ Glean

Summary

Arvind Jain makes a broad case that frontier and open-source models are increasingly interchangeable for most enterprise use cases, so application companies should treat model labs as enabling infrastructure rather than existential competitors. He says the real enterprise risk is operational dependence: if AI agents run the work, they also accumulate institutional learning, so enterprises must retain control of the agent and its learning. Jain argues open-source adoption is accelerating mainly because of cost and sovereignty, while fears about model providers training on enterprise data have largely eased through contracts. He also says AI’s ROI is uneven: clear in support and some coding tasks, but less obvious in end-to-end product shipping. On organizational structure, he predicts teams will get larger, not smaller, because AI raises per-person productivity while also raising customer expectations and the bar for output.

Notes

• Most enterprise AI use cases are already partially commoditized across models, including open source; Jain says non-frontier AI companies should treat model providers as an asset, not a threat.
• He says bundling advantages fade in a consumption-based world because customers pay per unit of work, not per suite relationship.
• He frames enterprise fear of frontier-model dependence as operational, not just technological: the agent learns the work over time.
• Open-source demand is being pulled by cost pressure and, for some firms, the requirement to keep inference inside private datacenters.
• He describes Glean as a horizontal enterprise AI platform that started with workplace search and now acts as a superset of ChatGPT, Claude, and Gemini connected to company context.
• He says Anthropic-style vertical packs are shallow and mostly expand the market rather than replace tools like Figma.
• Glean’s competitive response to Claude is to explain why context is hard to build and why MCP connections are not enough on their own.
• Being first helped Glean, but Jain says first-mover status is not enough to preserve leadership.
• Enterprise use cases are often solvable by many models, so Glean uses model selection as a cost-control layer.
• Jain says GLM 5.2 is the first open-source model his team feels comfortable using for most workloads.
• His concern about Chinese open-source models is framed as paranoia, possible backdoors, and competitive/reputational risk.
• He predicts open source will handle the majority of enterprise workloads within three years.
• Microsoft is one of Glean’s most important competitors because Copilot bundled into the suite often blocks evaluation.
• He says pricing is often the real reason enterprises choose Microsoft over best-of-breed alternatives.
• AI returns are uneven today: support is measurable, coding is faster, but shipping products is not necessarily faster.
• At Glean, nearly all initial code is AI-generated, but human code reviews remain mandatory.
• He warns that AI-generated code may become difficult to maintain, understand, and manage over time.
• The company accepts slower review cycles because AI still speeds up writing enough to justify the tradeoff.
• A naive MCP-first rollout forces the model to brute-force context gathering, which makes AI slow and expensive.
• He rejects the goal of replacing people with AI, arguing the technology is not ready for that framing.
• Jain says AI should make teams larger and more ambitious, not smaller, because rivals can use the same tools too.
• Glean expects to grow from a little over 1,000 employees to about 5,000 in five years.
• He expects inference costs to eventually fall by orders of magnitude, even though per-token prices rose recently.
• If AI gets much cheaper, Jain thinks many loss-making businesses that underpin the economy become vulnerable.
• He expects per-person productivity to rise, but also says demands will rise fast enough that teams may still need to stay large.
• Glean did not impose formal token-budgeting early on; it let employees explore AI use cases organically.
• Internal and customer AI usage follows a power law, with heavy users spending orders of magnitude more than casual users.
• Basic AI use is widespread for search, question answering, summarization, and synthesis, but advanced workflows are still rare.
• He dislikes token-maximization dashboards and prefers social proof through demos and town-hall showcases.
• Recruiting had gotten easier versus the SaaS peak because major tech employers were not expanding headcount.
• AI/ML talent is still fiercely contested, and even startups now pay much more than before.
• Seed-stage compensation can be so high that a small round may not fund a viable team.
• He prefers raising as much capital as possible upfront and sees large rounds as recruiting signals.
• A large valuation with limited revenue was, in his view, mainly useful as a signal to prospective employees.
• Investor brand matters for hiring because top candidates respond to the signaling power of well-known firms.
• He says his own discipline may be too conservative for a land-grab environment, even though he still believes in pricing power and marketing ROI.
• The market is a land grab, and missing now may make later entry dramatically harder.
• Composite roles should become common across product and go-to-market functions.
• AI should broaden what each employee can do, even if specialization declines.
• Analyst, BI, and sourcer roles are most exposed to being absorbed into self-serve AI workflows.
• He sees demand for sovereign models, but says it is less intense than a year ago.
• Model training is capital intensive, which makes classic open-source dynamics harder to apply.
• He says U.S. open-source model development needs active investment and tailwinds, not just regulation.
• There is too much capital in startups, and it can produce unhealthy comp structures and weak execution.
• He describes founder life as stressful and mission-driven rather than glamorous.
• He says founder success does not reduce the need for relentless effort and irrational ambition.