Harry opens by introducing Arvind Jain, co-founder of Rubrik (which IPO'd successfully) and now CEO of Glean, an enterprise AI company. Glean started as an internal search company β "a Google for your work life" β and evolved into an AI platform that Jain describes as "a superset of ChatGPT, Claude, Gemini all combined into one product experience... a co-worker for your employees connected to all of your company's context."
Harry references Palantir's Alex Karp, who said on CNBC that the largest enterprises in the world are more skeptical than ever of frontier model providers, and asks Jain if he agrees.
If you believe that majority of the work we do today is going to be done by an agent that is fully powered by one of these frontier model companies, then in some sense you've transferred a lot of your operations to these technology providers. This is more than technology dependence β this is real operational dependence.
Jain elaborates on why this matters more than it looks: institutional learning that used to accumulate in employees' heads over years of doing a task β undocumented tweaks and optimizations β now accumulates inside the agent doing the work instead. "If you don't own the learning that it actually gains over the years, then you're basically fully dependent on these AI companies to get your work done." He frames this as the central question facing enterprises today: how to use AI technologies while retaining control over the compounding learnings that should belong to the enterprise.
Asked whether enterprise customers are rotating away from frontier model providers toward open source, Jain says yes β "we are at a real inflection point with open source." No enterprise wants to be locked into a single model provider; everyone wants control of their destiny and the ability to use many models.
Jain says AI costs exploding past annual budgets within a month or two has really accelerated the desire for open source, "coupled with the fact that we now have really good models in open source." Asked what customers actually care about β cost, or data ownership/on-prem control β Jain says right now the drive is coming from cost. The earlier fear of model companies training on enterprise data has largely subsided as long as the contract terms are right.
90% or greater of use cases can now be fully handled by many many different models including open source models. GLM 5.2 is the very first time our own team feels comfortable that we can run the majority of our workloads on that model.
Jain clarifies that the real question isn't open source versus closed source β it's whether enterprises are comfortable with a Chinese model. Harry pushes back: if you own the deployment, run it on-prem, and share nothing back to China, why wouldn't you be okay with it? Jain's answer is that it's mostly comfort and paranoia β fear of an undetectable backdoor, and reputational risk of being known to use a Chinese model that competitors could weaponize against you. "It again boils down to who's willing to be bold... the early movers will make the move first and then it'll become a more normal thing."
On pricing, Jain argues the model business on its own β even setting aside open source β faces real competitive pressure across labs, and rumors that OpenAI was planning to drastically cut prices in response. "The model business on its own is probably not as lucrative as everybody believes. But these companies now have a lot more things β they're no longer model companies only."
We've been telling customers that majority of enterprise workloads will actually be on open source models in 3 years for sure.
Harry says every one of Jain's investors told him he had to ask this question: does Jain worry that Anthropic will do to Glean what it's done to Figma, legal, or health β moving in and cannibalizing the business?
I think we should be careful about what they've actually done for Figma or the legal space or finance space. They are launching these vertical packs but I think they're quite shallow. I don't actually know of people who are moving their workload entirely from Figma to Anthropic β it's actually sort of net new, expanding the market.
His example: designers still use Figma, but non-designers who aren't experts on the primary tool are now able to do some of that work with Claude. That's market expansion, not cannibalization, in his view. Pressed on whether he worries about Anthropic's enterprise push more broadly, Jain concedes they already face that competition every day β customers constantly ask "Claude can also connect to enterprise systems through MCP, so what's different about Glean?" and he has to explain why context is actually hard to build well. He adds a striking admission: Claude Co-work/Desktop's primary use case has always been question answering β "the largest application or use case for AI in the world today is in fact information seeking and question answering" β which means Anthropic arguably started competing with Glean before others did.
On first-mover advantage: "It's actually very advantageous but it's a thing that helps you, it's not going to carry you." Glean gets credit for being the first enterprise AI company, first to bring RAG into the enterprise, first to build conceptual semantic search β brand and the right to compete, even though Glean is now much smaller than the giants OpenAI and Anthropic have become.
For almost all other AI companies that are not doing frontier model training, they should see the model companies as a huge asset, not a competition. Everything Anthropic, OpenAI, and Google are doing, as well as innovation in open source β that's great news for us. They've allowed us to deliver a product we could never build without that help.
Harry asks directly: aren't we seeing the ultimate commoditization of the model layer, given the pace of new releases and open-source/Chinese providers? Jain says yes, from the enterprise use-case perspective β that's one of Glean's core value-adds, picking the right (often cheaper open-source) model for each task automatically.
The conversation turns to Microsoft, described by Harry as having built a phenomenal business on a "70% as good" product bundled with a nice sticker for enterprises.
For us they are one of our most significant competitors, and the bundling strategy actually works. But AI is moving towards consumption-based models. Once you move towards consumption there's no inherent bundling advantage β I let the users choose where they want to do their work, and I only have to pay for that particular unit of work.
Harry counters that consumption-based pricing might not matter for large enterprises: getting 15 vendors approved through compliance is a much bigger headache than approving Microsoft as a single bundled vendor β a vendor management problem companies didn't have before. Jain concedes the point is true but says companies that have been on the other side of "the Microsoft onslaught" mostly cite pricing as the main killer β "it's hard to compete with free."
Asked who's the fiercer competitor β Microsoft or the frontier labs β Jain says it's too early to tell, but Microsoft is formidable: in prospecting conversations, "we are a Microsoft customer and we already have co-pilot" comes up far more often as an objection than "I've embraced one of the lab products and there's nothing else I'm going to do."
Returning to Alex Karp's "AI is not working" comment, Harry asks how enterprises should think about ROI. Jain says there are real pockets of value realization β customer support is the clean example, where a support agent resolving 10 cases a day now resolves 12 because of AI, a concrete productivity measure.
Coding is murkier. "The majority of AI spend right now is on coding, and the practice has changed β most developers now actually use AI to write code, they're not writing it by hand anymore." But most companies report that actual shipping speed of products hasn't increased even though coding speed has, because coding is only a small part of shipping a product overall.
What percent of Glean's code is written by AI now? It's probably about almost 100%. Nobody's actually writing the initial code by hand anymore. But we actually enforce human reviews β you cannot generate tons of AI code and just check it into the repos. We're probably more conservative than most other companies.
He describes an internal debate about eliminating code review altogether now that the bottleneck has shifted from writing code to reviewing it β many companies are doing exactly that, letting AI-generated code go straight into the repo. Glean chose not to, willing to "pay the cost of reviewing the code" because the writing part is so much faster now.
Most enterprises today just throw AI into the system, connect it to everything via MCP in a rudimentary manner, and let the model brute-force its way into assembling the right raw materials to complete the task. In this mode AI is super slow and very costly β most of the tokens are being burnt just trying to assemble the right context. You have to invest around it, so it can work faster at a lower cost.
On the push to have every employee try to "replace themselves with AI," Jain calls it the wrong goal β "you're giving too much credit to AI... it's just not ready right now." He challenges Harry to name one job AI can fully replace, using an EA as the example: it can do the majority of the tasks, but not "that final intangible."
Harry notes that Glean is over 1,000 people, and asks how many Jain expects in 5 years. "Hopefully 5,000. We're going to grow." Harry pushes: he sits with the biggest CEOs in the world and every single one is shrinking teams.
Take Coca-Cola and Pepsi. One decides to shrink, the other keeps more people. Both have full access to the same AI tools. If you shrink to do the same amount of work with fewer people, your competition can choose to elevate and build a 10x better product with more people. They're going to be larger. They're going to beat you.
Harry counters that post-COVID layoffs of 15-20% made companies faster, and that more people slow everything down β an argument that predates AI. Jain agrees teams can slow each other down, but insists "people are also your asset" and points to the model labs themselves hiring aggressively as evidence.
Asked about composite roles β engineers who are also PMs and designers, salespeople who can also demo and architect solutions β Jain says this generalization away from specialization is something he's driving hard inside Glean itself. Harry notes this logically implies smaller teams. Jain agrees but reframes: "you have to do 10 times the work to get the same amount of revenue from your customers in the future. You have a much smaller team to deliver the same amount of work you used to deliver before. We're just forced to do more."
We had a 15-person on-call team whose job was to triage every production issue. We built an agent that now takes care of 95% of those issues automatically. But it's doing that at a cost which is questionable β we were spending a million dollars a month on that particular agent, more than the cost of the humans it replaced.
Harry references Ben Horowitz's $300M Anthropic spend being 3.7% of developer salaries, calling that "relatively small." Jain says that number doesn't seem high on its own, but notes open source can already do the same work for a tenth of the cost, and that historically technology cost and labor cost were never put in the same sentence β "this is the first time we're hearing that hey, I would rather have fewer humans and more tokens... I just feel like this is not how technology works. The models are supposed to get cheaper and cheaper."
We saw something bizarre β in the last 6 to 9 months every model actually increased their per-token price. If you go back 15 months everybody thought token price was going to just keep falling like before. We don't know what happened here.
Harry immediately offers his own theory: "they needed to prove that they were good businesses before they went public." Jain's bet remains that AI gets much cheaper over time; Harry counters that if it does, "these already loss-making businesses which prop up our entire global economy... are very threatened."
On token budgeting internally, Jain admits Glean "did probably what most companies did, which is we didn't do anything" β letting people figure out what they can do with the tech. There's a power law: some people spend $10-15K in tokens a month, others $20. Everyone has embraced basic information-seeking/Q&A use cases; advanced use cases remain limited to about 5% of the employee base.
On recruiting, Jain says β surprisingly β that hiring got easier compared to the SaaS peak, because big tech employers stopped growing headcount and kept laying off (he cites Meta). But top AI/ML talent is more fiercely contested than ever, and pay scales have completely changed, forcing even startups to pay $300-500K for a great developer. "The $2 million seed round just doesn't go anywhere," Harry notes; if a founder needs to hire four people, that's $6 million right there.
On sovereign models: Jain says the desire was actually stronger a year ago than now, as many nations realized building their own frontier model wasn't realistic and became comfortable letting local enterprises use OpenAI or Anthropic. Harry pushes back, citing the Trump administration banning Anthropic's latest models and Europeans realizing they can't rely on a US individual who could cut off their access to intelligence β arguing the trend is unequivocally rising. Jain agrees there are no results yet but notes it's only been a month.
You're right that we don't have open models from the US, but it's not because open source as a movement is weak in the US β it's actually quite strong. Models require a lot of upfront investment which is not open-source friendly. Open source software has historically been skunk works with no funding, and they still built something. You couldn't build models that way.
Harry notes that on OpenRouter, Anthropic was the first US model at #7 β the top six were all Chinese β and asks if the US should just accept that. Jain says being able to run inferencing in a contained environment makes people comfortable, but the US "absolutely won't feel okay" with that trend long-term, and points to Nvidia and others investing to promote US open-source model development.
Quick fire round:
- Biggest advice for someone studying CS today: "It's fine to study it. Don't get too worried because of what other people are telling you."
- Legacy company that's adopted AI best: Google ("though it's kind of hard/unfair to put them in that category since they're an AI company").
- What he'd most like to change about the startup ecosystem: too much capital available, creating unsustainable structures β e.g. a startup on a seed round paying half a million dollars to one engineer, "while Google is not, and Google knows they don't need to buy talent like that."
- On startup exits being harder: disagrees it's gotten harder overall β "it's easier to build a startup and get a good exit from it these days than it used to be in the past."
It's not a sexy job. It's actually one of the most stressful things and you really have to be crazy. Almost all of that glamour, money, and respect people imagine β irrelevant. You have to be truly mission-oriented to survive as a founder.
On the Series C being priced expensively when Glean barely had a business (sub $2-5M revenue, valuation north of a billion): Jain says it was less about the money and more "a statement to be made to prospective employees" β validating to the market that Glean was building something special. Asked if employees care who the investors are: "Absolutely β I promise you if you have [Sequoia, DST, etc.] great candidates suddenly want to talk to you a lot more."