⁠Why OpenAI and Anthropic Won't Win the App Layer | Glean Founder

Summary

• Open source just hit its enterprise inflection. Arvind Jain says 90%+ of enterprise use cases can now be fully handled by many models including open source, and GLM 5.2 — arriving within 3 months of frontier capability, “literally a month back” — is the first open model Glean’s own team trusts with “majority of our workloads.” His call: majority of enterprise workloads run on open source within 3 years “for sure,” and the only gate is comfort with Chinese models — “it’s not open source versus closed source.”
• The frontier model business may be mispriced. Standalone, it’s “probably not as lucrative as everybody believes”: fierce competition even in a three-way lab race plus open source at “an order of magnitude” cheaper, and Jain has “heard rumors that OpenAI was going to drastically reduce their model prices.” Meanwhile every model bizarrely raised per-token prices in the last 6-9 months; Stebbings’ jab — “they needed to prove that they were good businesses before they went public” — and his warning that if AI gets much cheaper, these loss-making labs “which prop up our entire global economy” are very threatened.
• The labs’ app-layer push is shallow. Anthropic’s vertical packs (Figma, legal, finance) are “quite shallow” — net-new usage that expands the market, not workload displacement — while enterprises are “terrified” of operational dependence: institutional learning accumulates inside the agent doing the work, so enterprises need control of the agent and its compounding learnings. Jain’s advice to founders: treat the labs as “a huge asset, not a competition.”
• Jain argues teams should get bigger — the episode’s sharpest disagreement. Glean is 1,000+ people and Jain wants 5,000 in five years: with symmetric AI access, the competitor that keeps headcount ships a 10x better product and “they’re going to beat you.” On the labor-versus-tokens framing, Jain says technology costs should fall: “we’ve not put technology cost and labor cost in the same sentence ever before… this is not how technology works” — inference costs will fall by orders of magnitude.
• AI ROI is a throughput problem, not a model problem. Nearly 100% of Glean’s code is AI-written, yet across companies “the actual shipping speed of products has not increased.” His triage agent resolves 95% of production issues for a 15-person on-call team — at $1M/month, a cost he questioned against the humans. Advanced use cases sit with only ~5% of employees; the fix is investing in context rather than letting models “brute force their way” through raw MCP connections.
• Consumption pricing can break bundling. Microsoft is a significant competitor today (“hard to compete with free”), but “once you move towards consumption, there’s no inherent bundling advantage” — enterprises pay per unit of work wherever users choose to do it, weakening the Copilot lock-in argument Stebbings pushed.
• China has the leading open models; the US is playing catch-up. On OpenRouter the top six models by usage are Chinese, with Anthropic the first US entry at seventh; Jain says China is the only country producing models outside the US, while Stebbings mentions perhaps a little activity in France. Jain’s explanation: model training needs upfront capital that skunkworks open source can’t fund. The US needs to build its own open models, with Nvidia among the motivated funders.

Deep dive

Enterprises Face Operational Dependence

• Stebbings opens with Alex K. from Palantir’s CNBC claim that the largest enterprises are more skeptical than ever of frontier model providers. Jain’s confirmation: they’re “terrified” — worried their core IP, data, and “way of doing things” end up hostage. If the majority of work is done by an agent fully powered by a frontier lab, “you’ve now transferred a lot of your operations to these technology providers… This is more than technology dependence.”
• The mechanism worth keeping: work starts as a documented 10-step process, then gets optimized through tacit, undocumented learning — and “all of that institutional learning is actually going to accumulate in that agent that is doing that work.” If you don’t run the agent yourself or own its learnings, “you’re basically fully dependent on these AI companies to get your work done.” The compounding learnings, he insists, “belong to the enterprises.”

Open Source Reaches Inflection

• The headline stat: 90% or greater of enterprise use cases can now be fully handled by many different models, including open source. Jain says open source is now within 3 months of frontier capability: “GLM 5.2 is the very first time where our own team feels comfortable that now we can run majority of our workloads on that model” — something that happened “literally a month back, or not even.”
• The driver is cost, not data fear: companies set annual AI budgets and “run past that within a month or two” — CFOs have noticed — while the old fear of labs training on enterprise data “is no longer there” given the right contracts. Open source prices are “an order of magnitude” cheaper.
• His prediction, stated categorically: “majority of enterprise workloads will actually be on open source models in 3 years for sure.” The only residual question: “are they okay with the Chinese model or not. That’s the only question here.” The objections are paranoia (“some back door that we don’t even understand”) and competitive optics — resolved when bold early movers normalize it.

Models Commoditize, Labs Expand

• On the Anthropic-will-eat-you question every Glean investor told Stebbings to ask: the vertical packs for Figma, legal, and finance are “quite shallow in my opinion” — Jain knows of nobody moving workloads off Figma. “It’s actually net new… expanding the market” — designers still use Figma while non-designers design with Claude.
• But he doesn’t pretend the labs aren’t competitors: Claude’s largest use case is question answering — Glean’s home turf — so “they probably started to compete with us before others.” Customers ask why Claude-plus-MCP isn’t enough, forcing Glean to explain “what context really is and why it is actually complicated to build.”
• His posture for founders losing sleep: “absolutely don’t worry about that… they should see the model companies as a huge asset, not a competition” — the labs enabled a product Glean could never have built alone. First-mover credit (first enterprise AI company, first to bring RAG into the enterprise) “is a huge asset, but neither is it a requirement nor is it a savior.”
• The model business standalone is “probably not as lucrative as everybody believes” — but the labs are no longer just model companies. Anthropic’s MCP-server, skills, and automation ecosystem means “you should consider them an application-level company, not just a model company.”

Consumption Pricing Can Break Bundling

• Microsoft is “one of our most significant competitors” and “the bundling strategy actually works” — in prospecting, “we’re a Microsoft customer and we already have Copilot” comes up more often than any lab-product objection. Best-of-breed survives, but companies crushed by Microsoft historically name pricing as the killer: “it’s hard to compete with free.”
• The structural counterweight: “once you move towards consumption, there’s no inherent bundling advantage” — a business can deploy six tools, let users choose, and pay only for the unit of work wherever it happens. Stebbings’ pushback — approving 15 vendors instead of one creates a vendor-management problem that compliance-heavy enterprises won’t tolerate — gets a concession (“that is true”) but not a retreat.

AI ROI Depends on Throughput

• Value pockets exist where productivity is measurable: customer support agents resolving 12 cases a day instead of 10. But the majority of AI spend is coding, and “the actual shipping speed of products has not increased even though coding speed increased significantly” — writing code is only a small part of shipping.
• At Glean, “almost 100%” of code is AI-written, yet human reviews are enforced — a proposal to eliminate code reviews (as “many companies are doing”) was rejected, because a million lines of AI code “becomes incredibly hard to actually maintain and understand.” Stebbings notes the tension: a stringent review process almost removes the point of fast generation; Jain accepts the cost — the writer does the first review, and they’re still net faster.
• The specimen that carries the cost argument: Glean’s engineering triage agent handles 95% of production issues automatically for what was a 15-person on-call team — at $1M a month, a cost Jain questioned against the cost of humans. Stebbings: “Are you buying Cristiano Ronaldo?”
• The diagnosis: most enterprises “just throw AI in,” connecting it rudimentarily via MCP and letting models “brute force their way” into assembling context — slow, and “most of the tokens are being burnt just trying to assemble the right context.” The fix: “you have to invest around it” so AI works faster at lower cost.

Jain Argues for Larger Teams

• Glean is over 1,000 people; asked about five years out, Jain says “hopefully 5,000.” Against every shrinking-CEO consensus Stebbings cites, his Coke-vs-Pepsi logic: both competitors have identical AI access, so the one that keeps people chooses “to elevate and build a 10x better product… they’re going to beat you.” Exhibit A: the model companies themselves are hiring aggressively.
• Stebbings’ counter-thesis in full: more people slow everything down; cut headcount, lift technology spend from 8-12% to 16-20%, and the best people follow the best tools — Mark B.’s $300M on Anthropic is only 3.7% of developer salaries, so dev tools are “dramatically underpriced.” Stebbings calls current pricing “absurdly expensive”; Jain says it depends on what the technology is doing and cites the $1M/month triage agent. Jain rejects the idea that technology costs should be treated like labor costs: “we’ve not put technology cost and labor cost in the same sentence ever before.” The “I would rather have fewer humans and more tokens” framing is Stebbings’ — Jain bets inference costs fall by orders of magnitude.
• The anomaly both flag: in the last 6-9 months “every model actually increased their per-token price” when everyone expected continued declines. Stebbings’ explanation — “they needed to prove that they were good businesses before they went public” — and his conclusion: if AI gets much cheaper, “these already loss-making businesses which prop up our entire global economy are very threatened.” Jain, laughing: “my take remains the same.”
• His synthesis of the whole debate: “per person productivity is going to shoot up. But so will the demands… You have to produce a 10x better product in the future to make the same amount of revenue.”

AI Adoption Is Uneven

• On token budgeting with his CFO: “we did probably what most companies did, which is we didn’t do anything.” The result is a power law — some employees spend $10,000-15,000 a month in tokens, others $20. Everyone does basic Q&A (“the number one use case for AI in the world today”), but “advanced use cases are limited to like 5% of the employee base.” He limited token-maxxing dashboards — rewarding consumption was the wrong idea — and showcases new AI agents at every town hall instead.
• The EA exchange is the adoption debate in miniature: Stebbings argues a role gets cannibalized once AI does 90% of it; Jain — “I don’t think you’re going to take a 90% solution” — because your competitors have the same AI tools plus a human on top.
• The roles that emerge: composites — one person acting as engineer, PM, and designer; a seller who can also demo and talk about use cases. Stebbings catches the contradiction (composite roles are smaller teams); Jain’s reconciliation: “you have to do 10 times the work to get the same amount of revenue… we just are forced to do more.” Going away: data-analyst roles “which are not business thinkers,” and recruiting sourcers, consumed into full-cycle roles.

Talent Costs Reshape Seed Rounds

• Recruiting had actually been getting easier than at the SaaS peak — big tech stopped growing, Meta likely below its 2021-22 headcount — until AI/ML pay scales “completely changed,” startups included. At $300-500k per great dev, Stebbings’ arithmetic stands: “the $2 million seed round just doesn’t go anywhere” — four hires need $6M. Jain’s advice: raise as much as you can from the get-go.
• His biggest change of mind in 12 months, offered as self-doubt rather than conviction: “my style has been a little bit too disciplined to be the right strategy anymore… we may lose the land grab.” And the land grab is real: “every single company in the world wants a product like ours today, and either we get in today or it’s going to be 10 times harder.” Yet he still holds that “a business is always built on discipline” — Uber is the counterexample that makes him wonder if he’s wrong.
• Glean’s Series C — north of $1B valuation with barely any business (“definitely under $2M or $3M, maybe $5M; I don’t remember exactly”) — “felt the most expensive,” done as “a statement to prospective employees.” And employees absolutely care who your investors are: “investor reputation directly impacts your reputation.”

China Has Open Models

• Sovereign-model desire was stronger a year ago — nations “figured out that’s not going to be the way” and settled for using OpenAI or Anthropic. Jain says, “The only country in the world that has produced models outside of US is China”; Stebbings adds “maybe a little bit” in France, without clearly naming the model. Stebbings insists sovereignty is unequivocally rising after the Trump administration banned Anthropic’s latest models a month ago — Europeans “cannot rely on a US individual who could ban our access to intelligence.”
• The evidence Stebbings brings from OpenRouter: the first six models by usage are Chinese; Anthropic is the first US model at seventh. Jain’s structural explanation: the US open-source community is strong in many other areas, but models “require a lot of upfront investment which is not open-source friendly” — skunkworks development can’t fund them.
• Stebbings games out regulatory capture — Sam gives Trump 5%, Trump taxes and bans Chinese open models. Jain: “I hope not… I doubt that’s going to happen” — and the argument runs the other way: “it’s paramount for the US to build” its own open models, with Nvidia and other motivated Bay Area parties already funding the push. Quickfire adds: Google is the legacy company best at AI (“though they are AI companies, so it’s unfair”); too much capital is “creating failure paths” — seed startups paying $500k for engineers Google won’t bid on; and founding “is not a sexy job… you really have to be crazy.”