Latest / B2B SaaS Talks with Fexingo: Enterprise Software, Sales Cycles, and Procurement / Why Enterprise RFPs Now Demand a Vendor AI Energy Audit
Transcript
- Lucas: So last week I was talking to a procurement director at a Fortune 500 industrial manufacturer — they're in the middle of a big AI procurement cycle. And she told me they walked away from a two-point-four-million-dollar deal because the vendor couldn't answer one question. Luna: Just one question? That's intense. Lucas: Yeah. They asked: 'What is your per-inference energy consumption for the model we'd be deploying?' And the vendor said they didn't track it. Deal dead. Luna: So what exactly is an AI energy audit, and why is it suddenly a deal-breaker for enterprise buyers? Lucas: Great question. An AI energy audit is exactly what it sounds like — a documented, third-party-verified report of the energy consumed to train a model and then to run each inference in production. And it's becoming a standard line item in RFPs, especially for large companies with net-zero commitments. Luna: I've seen a few mentions of this in the trades, but I thought it was still mostly European regulators pushing. Is it really spreading to US enterprises? Lucas: It is accelerating fast. Gartner published a survey in March of 2026 — so just a few months ago — and it found that 42 percent of enterprise buyers say they would now walk away from an AI software deal if the vendor can't provide an energy audit. That's up from 12 percent in 2024. Luna: Wow. That's a huge swing in two years. Lucas: Right. And it's not just about tree-hugging. There's a hard cost angle here. If you're a large company running multiple AI models at scale, your data center electricity bill can run into the tens of millions. Procurement teams are realizing that vendors with energy-efficient models save them serious money over a three-year contract. Luna: So it's both a sustainability checkbox and a total cost of ownership lever. Lucas: Exactly. And the European telecom example is instructive. One of the big German telcos — I won't name them because the contract isn't public — recently inserted a clause in all new AI vendor contracts requiring a 30 percent year-over-year improvement in energy efficiency per inference. If the vendor misses the target, there's a price adjustment. Luna: That is aggressive. How do vendors even measure that? I mean, inference energy depends on hardware, batch size, model architecture — it's not a simple number. Lucas: You're right, it's not simple. And that's exactly why the best vendors are starting to publish what they call 'energy cards' — similar to model cards — that break down three numbers: training energy in megawatt-hours, per-inference energy in watt-hours on a reference hardware stack, and the carbon offset source. Some are even including a confidence interval. Luna: Wait — confidence interval? So they're saying 'our model uses between X and Y watt-hours per inference'? Lucas: Yes. Because it genuinely varies. But procurement wants a number they can plug into their own carbon accounting software. So vendors that give a range with a clear methodology are winning deals over vendors that say 'we don't measure that.' Luna: I bet the smaller SaaS vendors are struggling with this. They don't have the resources to run a full energy audit on every model they ship. Lucas: That's a real challenge — and it's creating a two-tier market. The large platform vendors — the hyperscalers — they've been investing in this for years. Microsoft, Google, Amazon all have energy reporting dashboards. But a 20-person NLP startup? They might not even know their training run's energy cost. Luna: So what should a smaller vendor do if an enterprise RFP asks for an energy audit? Lucas: Couple of things. First, you can use open-source tools like CodeCarbon or the ML CO2 Impact calculator to estimate training energy after the fact. Second, you can partner with a cloud provider that offers energy attribution per VM or per container — that gives you a factual basis. And third, you can be transparent about what you don't know and what you're doing to learn. Luna: That last point is important. I've seen procurement teams respect honesty more than a polished but fake-looking number. Lucas: Absolutely. In fact, the procurement director I mentioned — she told me the vendor that lost the two-point-four-million deal was the only one that didn't even try to engage on the question. The other vendors who said 'we're working on a measurement framework and here's our current best estimate' at least stayed in the conversation. Luna: Good to know. And speaking of staying in the conversation — if you're finding these deep dives useful for what you're building or running, a couple of dollars a month is genuinely what keeps these going. Buy me a coffee dot com slash fexingo, if you've gotten something out of them. Lucas: Yeah, listener support is what keeps the show ad-free and lets us dig into these nitty-gritty procurement trends. Really appreciate anyone who chips in. Luna: Alright, back to the energy audit. So besides the big three numbers — training energy, inference energy, carbon offset — are there any other data points procurement teams are asking for? Lucas: Yes. Increasingly, they want to see the methodology for how the audit was conducted. Is it based on actual metered power draw, or is it extrapolated from GPU utilization? And they want the audit to be done by a third party — not self-reported. Some are even asking for a 'sustainability SLA' that guarantees the model won't exceed a certain energy budget over the contract term. Luna: That sounds like it could get really complex. I mean, energy usage fluctuates with usage patterns. Lucas: It does. So the forward-thinking vendors are building energy monitoring directly into their SaaS platforms — giving customers a dashboard that shows real-time inference energy per API call. That turns energy from an RFP checkbox into a competitive differentiator. Luna: That's smart. Then procurement can see exactly what they're getting. Lucas: Exactly. And it aligns with a broader trend: enterprise buyers are treating AI models like any other capital-intensive asset. They want to know the total cost of ownership, and energy is a big part of that. A 2025 MIT study found that for large language models, energy cost can be 15 to 20 percent of the total cost of running inference at scale. Luna: That's significant. So if you're selling AI software, ignoring the energy question might cost you deals. Lucas: It already is. And I think within 18 months, an AI energy audit will be as standard in enterprise RFPs as a SOC 2 report is today. The vendors that get ahead of it now will have a real advantage. Luna: Coming up with a standard format for an 'energy card' — that could be a great open-source project for someone. Lucas: Absolutely. There's no industry standard yet, which means there's a window for early movers to shape the template. The companies that define the format will also define the expectations. Luna: Good point. So for sellers listening: start gathering your energy data now, even if it's rough. And for builders: design your next model with energy efficiency as a feature. Lucas: Well said. That's the takeaway. Thanks, Luna. Luna: Thanks, Lucas. See you next time.