Who REALLY Wins the AI Race? | Why Teams Will Get Bigger Not Smaller in an AI World | Glean Founder

Why does it matter? Because the CEOs quietly shrinking teams might be handing the race to whoever didn't

Arvind Jain built Rubric into a public company, then built Glean into one of the world's leading enterprise AI platforms. His read on what AI actually does to organizations flatly contradicts the consensus: he plans to grow from 1,000 to 5,000 people while every major CEO he knows is cutting.

• Companies that shrink using AI while rivals grow will be outproduced and beaten — same tools, more ambitious output wins
• 90% of enterprise AI workloads can already run on open source, making frontier pricing increasingly indefensible
• The AI ROI crisis in enterprises is a context architecture failure — most tokens burn before real work even begins
• Letting vendors run your AI agents means surrendering the institutional knowledge that compounds over years

Shrink your headcount with AI while your competitor grows — and they'll simply outproduce you

The competitor that keeps its team while you shrink — with access to the same AI tools — will produce more and beat you. Jain's logic is clean: take two rivals with the same toolkit. One cuts headcount to maintain current output with fewer people. The other keeps its team and uses the same AI to build a ten times better product, produce ten times more. "They're going to beat you."

Pushed back on this — doesn't more headcount slow everything down? — Jain cuts through: "That argument has always been true. That's not an AI argument." The question isn't whether lean teams move faster; it's whether shrinking to match current output loses to a rival who expands to capture more. "I don't think the world's greatest companies are going to be companies with 100 people." Watch the frontier labs — they're hiring as aggressively as anyone alive.

Glean's target is five thousand. In a room full of peers moving the opposite direction, that's either the sharpest competitive bet of this moment — or the most expensive mistake being made in enterprise AI right now.

90% of enterprise AI workloads can already run on open source — and frontier providers are priced like that's not true

Frontier model companies are valued in the hundreds of billions. The floor underneath just shifted.

"90% or greater of use cases can now be fully handled by many many different models including open source models" — not a prediction, the current reality. GLM 5.2 was "the very first time where our own team feels comfortable that now we can run majority of our workloads on that model," and that happened weeks before this conversation. Three years out, Jain expects most enterprise AI globally to run on open source.

The forcing function is cost. Enterprise AI budgets are already blowing past annual caps within months of deployment. Open source runs an order of magnitude cheaper. "The model business on its own is actually probably not as lucrative as everybody believes." OpenAI, he heard, was preparing to slash prices in response to the competitive pressure. The labs that survive will have become application companies by then — not model companies.

Enterprise AI isn't failing because models are weak — it's failing because models have to brute-force their way to the right starting point

The enterprise AI ROI problem isn't a capability failure — it's an infrastructure failure that a better model won't fix.

"Most enterprises today are rolling out AI by just throwing it into the system and connecting AI with all your enterprise systems in a rudimentary manner using MCP servers," Jain says. The model then has to "brute force its way into trying to figure out and assemble the right raw materials to complete the task." The consequence: "most of the tokens are being burnt just trying to assemble the right context for that given task." The model isn't failing because it's dumb. It's burning compute on orientation before doing any real work.

Upgrading to a smarter model in the same setup doesn't solve this — it just makes the brute-forcing more expensive. The fix is context infrastructure: making sure agents arrive at each task with exactly what they need, rather than scrambling to find it first.

Let a vendor run your agents and the institutional knowledge your company spent decades building becomes theirs

Most undocumented workflow knowledge — the shortcuts, the judgment calls, the optimizations nobody ever wrote down — is about to transfer from your employees to your AI agents. And the agents won't necessarily be yours.

"Based on doing this work over and over again you now built all these learnings that you apply in real time," Jain says. When AI agents take over those tasks, that learning accumulates inside them the same way it once accumulated in people. "All of that institutional learning is actually going to accumulate in that agent that is doing that work."

If you don't control the agent — its model, its weights, its data — you've handed that compounding advantage to a vendor. "If you don't own the learning that it actually gains over the years, then you're basically fully dependent on these AI companies." He calls it "more than technology dependence — real operational dependence." Outsourcing the agent is outsourcing the flywheel.

Glean writes nearly 100% of its code with AI — and still can't prove it's shipping faster

"Almost 100%" of Glean's code is now AI-generated. Nobody writes initial code by hand. When the host asks whether product shipping speed has actually improved — Jain pauses.

"The actual shipping speed of products has not increased even though coding speed increased significantly — because that's only a small part of overall shipping a product." Glean is shipping more features, but Jain attributes it to having a larger and more tenured team. He can't isolate the AI contribution.

The bottleneck shifted from writing to reviewing. Glean enforces human review on all AI-generated code — "probably more conservative than most." The reasoning: "you can write a million lines of code, but it becomes incredibly hard to actually maintain it and understand it over time." Write ten times faster, spend proportionally longer reviewing and maintaining — the net throughput gain may approach zero. Companies claiming AI transformed their engineering velocity might be measuring the wrong thing entirely.

The open source debate is already over — the only question left is whether your enterprise will put a Chinese model on its servers

"It's not open source versus closed source," Jain says. "The question is: are they okay with the Chinese model or not. That's the only question here."

GLM 5.2 already handles the majority of Glean's workloads on-premise — nothing sent to China, the data sovereignty concern dissolved. What remains is the backdoor worry — "What if there's a back door? Some magic back door that we don't even understand" — and competitive optics: if it becomes known your company runs CCP-adjacent AI, rivals can weaponize it.

But the call will be forced regardless. Open source is now within months of frontier capability at a fraction of the cost. "Ultimately it boils down to who's willing to be bold — the early movers will make the move first and then it'll become more normal." Organizations with a clear policy already in place will be choosing. Everyone else will be reacting.

The specialist role is becoming a liability — composite roles spanning engineering, product, and sales will define the next generation of org design

Data analysts who configure dashboards without business thinking — gone. Sourcers who find candidates but don't run the full cycle — gone. The three-way split between account executives, solution engineers, and solution architects — probably gone.

Jain sees composite roles replacing the specialization layer: someone who "acts like engineers, product managers, designers" simultaneously; on the go-to-market side, reps who can negotiate, demo, and speak to use cases in the same meeting. AI executes the narrow specialized tasks; humans hold the cross-domain judgment.

The host catches the contradiction: composite roles imply smaller teams, which cuts against growing to five thousand. Jain's answer is direct — "You have to do ten times the work to get the same amount of revenue from your customers in the future." The job titles collapse inward. The ambition expands to match.

The AI capacity dividend will compound fastest for whoever reinvests it as ambition rather than pocketing it as savings

The throughput AI creates is real. Where it goes next is the whole game. Every organization can cut costs with AI — that advantage compresses to zero as everyone does the same thing. The companies that grow their teams and raise their ambitions while rivals shrink are running a different bet: more capacity creates more output, more institutional knowledge locked inside agents they actually own, more surface area to compete. Cost savings are one-time. Better products compound. The companies banking AI efficiency as headcount reductions are cashing out early. Their rivals are reinvesting the dividend.

---

Topics: enterprise AI, open source models, AI ROI, team size and headcount, model commoditization, institutional knowledge, composite roles, frontier models, AI coding productivity, Glean, AI agents, model sovereignty, Chinese AI models, context architecture