# Your PUE Doesn’t See What Your AI Is Worth **Author:** WEKA **Published:** October 13, 2026 ![Monochrome data center with server racks, glowing data links to a central tower, and a large foreground gauge.](https://cdn.sanity.io/images/ult5g8gw/production/1cda0e7b2b893ef3d6e60289f585b2913212ee94-1584x672.jpg) ## TL;DR Rethink how you measure AI efficiency, from the data center to the balance sheet, in this episode of Deep Geeks. - Understand why PUE stops at the building’s walls, and what sits around it. - See the five-rung ladder from watts to outcomes per token. - Reframe the AI bubble debate: It’s bubble wrap, and metrics decide what pops. Twenty years ago, [Vernon Turner](https://www.linkedin.com/in/vernon-turner-beyond-iot/) was an analyst at [IDC](https://www.linkedin.com/company/idc/) who helped build the metric every data center still lives by. Now, he’s ready to move past it: “I was there at the beginning helping to create the PUE framework. So it’s ironic that 20 years later I’m there going, ‘You know, it may have served its time.’” PUE (power usage effectiveness) isn’t wrong. It just answers a question AI leaders stopped asking. In the latest episode of Deep Geeks, host Dr. Serena Huang sits down with Vernon, now IoT and Sustainability Advisor at [Cognizant](https://www.linkedin.com/company/cognizant/), and WEKA’s [Val Bercovici](http://linkedin.com/in/valentinbercovici), who returns to the show. Vernon helped shape the industry’s efficiency metrics while Val works with neoclouds and inference providers almost daily. Together, they trace what leaders should measure, from the watts entering the building to the revenue AI generates. It’s worth your next 45 minutes. ### **A metric built for the building** PUE compares the power coming into a data center with the power that reaches the IT equipment. A score of 2.0 means half the energy goes to cooling, power conversion, and other overhead instead of compute. Over two decades, operators drove this number down, with Vernon noting hyperscalers now run at roughly 1.1 to 1.2. Here’s the problem: The curve has flattened. And PUE measures facility overhead, not whether an AI workload produces anything worth paying for. “We’re measuring the power outside coming into the data center,” Vernon said. Val treats PUE as the price of entry, not the ultimate scorecard. “A PUE of 1.5 or two or above kind of disqualifies you from even having an efficient data center conversation.” Clear that bar, and today’s real questions emerge. ### **Five levels, from watts to outcomes** Vernon’s framework keeps PUE as the foundation and builds up from it: 1. **PUE:** how well the building uses power 2. **Tokens per watt:** how much AI work this power produces 3. **Tokens per person:** who consumes AI, and how much 4. **Tokens per workload:** which applications and business units drive costs 5. **Outcomes per token:** what all these tokens truly generated Level 4 matters most to a CFO. Instead of a CIO declaring, “I’m really efficient. Look at me. I should get a prize,” they can tell each business unit exactly what applications cost in AI. Level 5 ties this cost to revenue: the tokens that built a product on one side, the tokens that sold it on the other. Val pushes in the same direction from the inference perspective. He said OpenRouter’s leaderboards now track cost per session, not just cost per token, and newer benchmarks measure cost per task. The question shifts from “How many tokens?” to “How many tokens did it take to get the job done?” ### **We measure the easy end** You probably stop at the first two levels, “Frankly, because it’s easy,” according to Serena. The dashboards already exist. Measuring by workload means figuring out which apps and which teams are contributing to it, and someone has to do the engineering just to produce the metric. AI makes this harder. Traditional workloads stayed in their lane. Vernon said AI workloads “are like crabgrass.” Agents and multi-agent systems reach into everything, so sizing them keeps shifting. His advice: Pick one or two workloads, measure them, and expect the complexity to climb. ### **AI isn’t a bubble. It’s bubble wrap.** Here, measurement stops being an ops concern and becomes a board-level priority. Vernon describes the pressure plainly: “How long can I keep going without somebody saying, ‘Stop. I want to see my ROI on my trillion-dollar data center investment”? Val rejects the single-bubble story. AI demand keeps growing with no end in sight, and this tells him people want what AI delivers. The open question is what it costs versus what customers will pay. “Instead of a bubble, I like to say we’re dealing with bubble wrap. There are going to be some bubbles that will pop. We’ll pop some ourselves. But there isn’t one giant bubble that’s going to burst here.” The missing metrics help decide which bubbles will disappear. Val describes the “golden handcuffs” of training revenue: Investors prefer stable three-year GPU contracts, so providers build for training even though inference now dominates AI spend. The neoclouds Val works with can’t build inference-optimized capacity because they’ve already sold it to training customers. His fix: Develop strong metrics that investors can use to underwrite inference while also working as financial instruments. The bubbles that pop will be those unable to show what their tokens produced. ### **Watch the full episode** This conversation is for AI strategists who answer to a CFO and for infrastructure leaders tired of PUE being the only number on the metrics slide. Serena’s takeaway: Start with one or two AI workloads, measure what they cost and what they produce, and loop in finance, procurement, security, and ops from the start. **Watch the full conversation:** [AI Broke the Yardstick. Now What?](/podcast/ai-broke-the-yardstick-now-what) or listen on [Spotify](https://open.spotify.com/episode/34exgWTXJW07HXEYd5l2jx?si=ZaVXDnzoQIG2sMR0rPhfTw) and [Apple Podcasts](https://podcasts.apple.com/us/podcast/ai-broke-the-yardstick-now-what/id1896864991?i=1000796100105). Make sure to subscribe to Deep Geeks to keep up with emerging trends in AI infrastructure. If this discussion changes how you think about AI and data, share it with someone who needs to hear it.