The Twenty Minute VCWhy OpenAI and Anthropic Won't Win the App Layer | Glean Founder
CHAPTERS
- 0:00 – 2:26
Arvind Jain’s mindset: paranoia, disruption, and why AI changes the “double down” playbook
Arvind explains that he’s motivated more by fear of losing than thrill of winning, and that building a new company always feels like starting from scratch. In AI especially, he argues daily disruption makes “keep building what worked” an insufficient strategy.
- •Founder psychology: paranoia as a durable driver even after prior success
- •New company ≠ compounding advantage; lessons help but environment resets
- •AI era demands constant adaptation vs simply doubling down on winners
- 2:26 – 3:13
What Glean is: enterprise search → enterprise AI coworker connected to company context
Arvind gives a compact overview of Glean’s evolution from “Google for work” enterprise search to a broader AI platform. He frames Glean as a unified experience across frontier models (ChatGPT/Claude/Gemini) grounded in enterprise context and systems.
- •Started with enterprise search across hundreds/thousands of internal systems
- •Evolved into an AI platform as models improved
- •Positioning: a “superset” interface over multiple frontier models
- •Core differentiator: connected enterprise context about how work happens
- 3:13 – 6:21
Why enterprises fear frontier model providers: operational dependence and loss of institutional learning
Responding to Alex Karp’s claim about enterprise skepticism, Arvind argues enterprises are genuinely “terrified” of model providers—not just for data/IP, but because agents will accumulate institutional learning over time. If enterprises don’t control agent execution and the learning loop, they risk outsourcing operations to external labs.
- •Enterprise fear goes beyond data leakage: it’s dependence on external operators
- •Institutional learning often isn’t documented; it compounds through repetition
- •Agents will become the repository of optimized workflows and tacit knowledge
- •Key question: how enterprises keep ownership/control of agent learning
- 6:21 – 8:19
The shift to open source: inflection point driven primarily by cost (not data paranoia)
Arvind says enterprises have long wanted multi-model control, but open source only recently became good enough to be credible near frontier capability. The immediate catalyst is cost blowouts and CFO shock, while fears about model providers training on enterprise data have eased with contracts.
- •Open source desire has existed for years; capability is now catching up
- •AI budgets routinely get blown in weeks/months; cost pressure accelerates adoption
- •Data-training fear has declined with contractual guarantees
- •Open source momentum is mostly cost-led today
- 8:19 – 12:14
Will Anthropic/OpenAI “eat the app layer”? Why vertical packs are shallow and context is hard
Harry presses on whether Anthropic could cannibalize Glean the way model companies push into adjacent apps. Arvind argues most lab “vertical packs” are shallow and often expand the market rather than fully replacing incumbents, though Glean competes daily against lab products connecting to enterprise systems.
- •Skepticism that lab vertical packs truly replace best-of-breed tools
- •AI enables non-experts to do some tasks; experts still rely on core tools
- •Competitive reality: customers compare Claude + MCP vs Glean
- •Differentiation: building reliable enterprise context is complex and non-trivial
- 12:14 – 15:01
Model layer commoditization and routing: 90%+ of enterprise use cases can run on many models
Arvind claims that for enterprise workloads, most use cases can be handled by multiple models—including open source—driving commoditization. He positions Glean’s value in model selection/routing for quality and cost control, with a new emerging question: whether enterprises will accept Chinese open-source models.
- •Enterprise AI: majority of tasks don’t require a single “best” frontier model
- •Product strategy: automatically pick the right model per task for cost/quality
- •Open source is newly “close enough” to frontier for many workloads
- •The real adoption friction may be China-origin models, not open vs closed
- 15:01 – 17:32
Have we mispriced frontier models? Pricing pressure, open-source deflation, and labs becoming platforms
The discussion turns to whether frontier model businesses are overvalued if open source covers most enterprise needs. Arvind expects intense pricing pressure (plus open-source being ~10x cheaper) and notes labs are evolving beyond “model companies” into platform/ecosystem players—especially Anthropic via MCP ecosystems.
- •Competition among labs already creates pricing pressure
- •Open source introduces a step-change lower cost curve
- •Rumors of major price cuts as labs respond to competition
- •Anthropic’s ecosystem positioning: platform + developer automation layer
- 17:32 – 20:48
Microsoft bundling vs best-of-breed—and how consumption pricing might break bundles
Arvind calls Microsoft one of Glean’s most formidable competitors and admits bundling works, especially when “free” is the killer. He argues the transition to consumption-based AI changes the dynamics by weakening traditional bundle advantages, though Harry challenges this due to enterprise vendor-management realities.
- •Copilot bundling is a major enterprise adoption wedge
- •Best-of-breed still wins where quality/context matters
- •Consumption-based pricing can reduce bundling leverage over time
- •Enterprise procurement friction still favors large incumbents
- 20:48 – 25:56
Why “AI isn’t working” in enterprise: ROI is uneven, and coding gains don’t equal shipping velocity
Arvind nuances the ‘AI isn’t working’ claim: some areas like customer support show measurable productivity improvements, while software delivery is harder to quantify. Even if coding is faster and most code is AI-written, overall product shipping speed may not rise because coding is only one part of delivery and review/maintenance remain bottlenecks.
- •Clear ROI pockets exist (e.g., support ticket throughput)
- •Coding is the biggest spend area, but shipping speed often doesn’t improve
- •At Glean, near-100% of initial code is AI-generated with strict human review
- •Maintenance and comprehension of AI-generated code are real constraints
- 25:56 – 27:34
The throughput and token-waste problem: brute-force context assembly makes agents slow and expensive
Arvind argues many enterprise rollouts connect models to systems in a rudimentary way and let agents brute-force retrieval and workflow steps. This burns tokens assembling context, slows execution, and uses AI where it’s not needed; the remedy is investing in structured context and surrounding systems so AI can act efficiently.
- •Naive enterprise deployments rely on brute-force MCP-style connectivity
- •Token spend is dominated by context gathering rather than task completion
- •Speed and cost degrade when AI must discover raw materials each time
- •Fix: invest ‘around’ AI—context pipelines, structure, orchestration
- 27:34 – 32:52
Stop trying to ‘replace yourself’: competitive bar rises, team sizes won’t necessarily shrink
Arvind rejects “replace yourself with AI” as the goal, arguing AI can cover many subtasks but not the last-mile intangibles required for top performance. He predicts per-person productivity will rise, but so will competitive demands; companies that simply shrink may be outbuilt by competitors using AI to do more, not less.
- •AI can handle many components of roles but not full job replacement (yet)
- •Competition effect: peers also get AI—humans + AI can outperform AI-only
- •Arvind expects Glean to grow headcount (1,000 → 5,000/10,000)
- •Thesis: productivity up, but expectations and required output rise faster
- 32:52 – 37:16
The $1M/month agent case study: replacing 15 engineers still may not be cost-efficient
Arvind shares a concrete example: an engineering triage/on-call agent that handled ~95% of issues but cost about $1M per month to run—more than the team it partially replaced. The exchange underscores his belief that AI pricing is currently ‘absurdly expensive’ and will need to fall dramatically, especially with open-source alternatives.
- •Triage agent automated ~95% of production-issue triage workload
- •Operating cost reached ~$1M/month, challenging human-vs-token economics
- •Open source can deliver similar work at a fraction of cost
- •Arvind’s bet: inference costs must drop by orders of magnitude
- 37:16 – 39:58
Token budgeting realities: no planning, power-law usage, and lightweight adoption programs
Discussing internal AI spend, Arvind says many companies (including Glean initially) didn’t proactively budget—preferring exploration. He observes a strong power-law in token consumption: everyone uses basic Q&A, but only a small minority uses advanced workflows heavily; Glean encourages sharing wins in town halls rather than gamifying token usage.
- •Early phase: ‘let people figure it out’ rather than strict budgeting
- •Power-law adoption: a few users spend massively; many spend little
- •Baseline usage is broad: Q&A/info seeking is universal
- •Adoption tactic: showcase wins (town halls) vs rewarding token maxing
- 39:58 – 45:33
Talent, fundraising, and land grab dynamics: raising bigger seeds, discipline vs speed, and market signaling
The conversation shifts to recruiting in a world of inflated AI compensation and how that changes early-stage fundraising needs. Arvind argues larger seed rounds can be rational, explains why an expensive Series C valuation was strategic signaling for talent, and wrestles with tension between disciplined spending and the reality of an enterprise-AI land grab.
- •Recruiting: overall easier than SaaS peak, but elite AI/ML talent is pricier than ever
- •Startups paying top-of-market comp pressures seed sizing upward
- •High valuations can be used as talent/market validation signals
- •Arvind feels tension: discipline vs ‘land grab’ urgency in enterprise AI
- 45:33 – 54:31
Which jobs change first, sovereign AI, and China’s open-source lead (plus politics and regulation)
Arvind predicts composite “generalist” roles will become more common, while certain analyst and recruiting-sourcer roles shrink or merge as AI enables self-serve answers and automation. The discussion then turns to sovereign models and the geopolitical reality that China leads in open models, with debate over whether the US will respond via innovation, investment, or regulation.
- •Future roles: composite builders (PM+design+eng), broader GTM generalists
- •Roles at risk: dashboard-focused analysts, some BI intermediaries, recruiting sourcers
- •Sovereign model appetite exists but execution is hard; China is the main non-US model producer
- •US response: push for domestic open models (e.g., NVIDIA support) vs regulatory barriers
- 54:31 – 1:00:08
Quick-fire: studying CS, Google as ‘legacy’ AI winner, too much capital, and the reality of being a CEO
In the closing quick-fire, Arvind encourages CS students not to panic and credits Google with strong AI adoption. He argues excess startup capital can create unsustainable behaviors, downplays ‘lack of exits’ as uniquely bad, and emphasizes that CEO life is stressful and requires mission orientation and relentless drive.
- •Advice: CS is still worth studying; don’t overreact to hype cycles
- •Legacy AI adopter pick: Google (with caveats)
- •Critique: too much startup capital can incentivize unsustainable comp/spend
- •Founder reality: high stress, little glamour, requires mission-first mindset
