In this conversational video, Robert Ta and guest Jonathan dissect Uber’s rapid depletion of its AI budget within four months, which led the company to impose a $1,500 per‑engineer cap on AI tool spend. They trace the root cause to a "token maxing" mindset—leaderboards and short‑term incentives that push teams to consume as many AI tokens as possible without clear ROI. The discussion expands to a broader framework of levers—people, process, and tools—highlighting how executives must align incentives (carrots and sticks) and implement fine‑grained cost governance. Practical recommendations include per‑team spend caps, real‑time alerts, attribution of token usage, extensive logging and evaluation of token costs, and internal knowledge‑sharing of optimization techniques. The overall message warns CTOs and engineering leaders that unchecked AI spend can slow feature delivery and erode value, but disciplined governance can enable faster shipping without bankrupting the organization.
Chapters
Uber blew its entire AI budget in 4 months. Now engineers are capped at $1,500. Here's what happened and what matters for engineering leaders shipping with AI coding tools. You'll be able to stop flying blind on AI tool spend if you listen to this and just think these things through, and perhaps build a cost governance layer that lets your team ship faster without bankrupting the company or getting too addicted to tokens. Uh, Jonathan. This has been the hottest piece of news in the last week. And I feel like every summer there is some sort of like angst about like, oh, is AI, what's going to happen? There's a new constraint or whatever. And I think we're seeing supply and demand sort of play out in equilibrium over time. I'd love to get your take on this news around Uber blowing its entire AI budget in 4 months, especially from a CTO perspective. Did you kind of see this coming? What do you think enterprises have to think about going forward?
Yeah, I don't know that I saw the timing coming, but preceded this was the token maxing philosophy, you know, and it seemed like there was token maxing leaderboards within enterprises. So within for-profit enterprises, there were token maxing leaderboards and there must have been a belief that maxing tokens was going to lead to ROI in a shorter time frame. The bills come in and I think it's like, okay, where's the ROI? And there's not strong answers that CTOs could be accountable for and say, yeah, it's, it's working and you're getting ROI and, you know, we are, we're fulfilling our duty to shareholders, a little bit of like the hype cycle hitting, um, the balance sheet happening here.
Yeah.
Um, yeah, it's really interesting, right? Like, if you're an executive at these, at these enterprise companies that need to stay ahead, you're looking at this huge data center spend. You're like, this whole thing's happening, we have to be part of it. Like, what is rational, right? And I think this is maybe just like, you know, part of a rational set of incremental decisions that the company executives are making, you know, in the like inefficient optimization problem that is capitalism. So I don't know, what do you think?
Yeah, no, I love what you just said. It's an inefficient optimization problem within capitalism because I think these companies, right, there's this multifaceted intersection of levers that an executive has to play. And I think about it from the lens of people, process, and tools. And within people, now there's probably like people slash agents, right? Because it's not just people running companies anymore, it's also people, using agents. And then a lot of these times, especially in certain surface areas, like we now have autonomous agents running really long loops towards goals. So the, with those levers, the way I think about it is how can you incentivize the behaviors you want towards getting your tokens to be used by your humans in your organization, applied through processes, applied with tools like coding agents to then capture market share. Like, ultimately, the goal is you get more revenue per employee. Now, in this new world where we— you're seeing a lot of companies are cutting headcount, right? And they're cutting what the Cloudflare CEO was talking about in terms of the measures, one of the operational sides of the business, like internal audit and finance and ops and stuff like that. And so the question then becomes, if you can unlock the value and velocity of somebody like one of your senior ICs, give them tokens, and they turn those tokens into dollars hitting the balance like, that's a really great dream. The hellish nightmare is also what Uber just saw, which is it didn't turn into clear economic value. In fact, our feature delivery actually slowed down. My take on this, because we do AI DLC consulting ourselves, like, we found ourselves in a position where we're servicing enterprise, where we're taking our own expertise. Jonathan's a 12-time founder, entrepreneur. I'm on my second tech startup. With Clarity by Epistemic Me here with Jonathan, building, um, AI personalization and the future around alignment and things like that that we talk about on the pod. And we're helping companies figure out these levers because it's complex and there's also people culture. One thing I'll draw on there on the people culture side, because part of the token maxing conversation as well is, oh, that's just clearly like incentivizing the behavior you want short-term without a long-term thought. Right? Like, it's like, go max out your tokens. Like, Facebook, Amazon was— were doing this from what I've heard in industry, and they're just like, go, go put the numbers up, right? But do the numbers translate to the balance sheet? Like, uh, actually market share, or, uh, at the very least, leading indicators of business success like, uh, product adoption metrics or, or marketing metrics? So I think executive leaders, just to round out the thought, really have to think about the carrots and sticks, just to keep it simple. I think management tactics around carrots and sticks, everybody could conceptualize that. Just thinking about the carrots and sticks more. And I think this is a time where we're seeing organizations realize that their budget cycles should be treated really fine-grained, like cloud compute budgets, like alerts, caps per team attribution from day one. So you know who's like, you know, maxing things out. And then the other very real thing is We're all still learning how to build better products, AI-native, and optimize our token costs and optimize our context windows. So even within an organization, there's a certain level of taste applied to your own domain, uh, for different workflows, whether it be coding or, or go-to-market or whatever. So share that knowledge internally, share the skills and commands that optimize token costs, like do evals around token costs, like log the crap out of that and do evals around that and get that into a learning group.
