Why Investors Today Count on Kilowatts

Dave: Am I doing everything I can to get the most intelligence out of the kilowatt hours that I'm drawing? How much power am I pulling off the grid, and am I getting the most out of that? Because we only have a limited number to go around. The demand for intelligence seems to be unlimited. The supply of kilowatts is very much limited. So yeah, we've got to be doing everything we can at all layers.

Serena: Dave, welcome to Deep Geeks. It's so great to have you today.

Dave: Thank you for having me.

Serena: Let's talk about your work at Generation Investment. It's at such a unique intersection of growth and sustainability, as you mentioned. As you have conversations with these AI companies, what patterns are emerging for you? What are the strongest companies doing differently in the portfolio?

Dave: That's a great question. Just to quickly back up — Generation is an investment firm that was founded a couple of decades ago on the thesis that if you understand long-term sustainability trends — the energy transition and how we decarbonize our grids, access to affordable healthcare, access to financial inclusion — deeply, you'll not only help make the world a better place, but you'll find better ways to invest and make more money for your clients. That's fundamentally our philosophy and approach.

In terms of how that's working out in our portfolio, I think we see two sides of AI.

Dave: One is where AI can be used to fundamentally address real-world problems. Take healthcare, for instance. We have a number of healthcare companies, and you can now see a path where, through AI triage, AI note-taking, and automated follow-up of appointments using agents, you could get to the point where zero clinician time is spent on administration and all of it is spent on clinical care.

That's hugely transformative in a world with an aging population and healthcare becoming increasingly expensive. We're seeing this right now — one of my portfolio companies, AlayaCare, is deploying agents across the home care and home health agencies it works with, freeing up huge amounts of clinician time and improving access to care.

So on one side, we see an incredible set of opportunities around how AI can transform healthcare, financial inclusion, or energy. On the other side, we see both opportunities and challenges in dealing with the risks of AI. We're absolutely focused on how you make sure AI is energy efficient — and that's one of the reasons we invested in WEKA, really around how you improve tokens per kilowatt hour. There's also a bunch of other risks around governance, privacy, and compliance that we think are a huge opportunity.

So we're bullish on how AI can solve humanity's challenges, and we also think there's a huge role for people like us in making sure that's done in a sustainable way.

Serena: The AI note-taking has freed up clinicians' time — I've seen that even in my recent visit with a doctor. He asked for my permission to use it, and I said, "Funny you should ask — I work in AI, so please, please do use it." But something I pay attention to is whether AI actually makes them focus on the patient more, as opposed to being even more distracted and multitasking.

I've had a few friends who are skeptical and anxious about the privacy aspect, and I tell them you can always opt out — but pay attention to whether they're focusing more on you as a patient, because they should be able to as a result of AI.

Dave: That's right. I think in healthcare in particular, the promise of AI is technology that allows clinicians to be more human — so the administrative, data-input, data-extraction work can be done by machines, freeing humans to do what humans do best, which is care for other people. That's a hugely important design principle when I think about how AI will be applied to healthcare.

Serena: I remember sitting through startup pitches two years ago, and when people said, "We're building AI," that was enough for the whole deck almost. A lot has changed since then. Where do you see investors focusing these days? What are they asking now that they weren't 18 or even 24 months ago?

Dave: Nearly all of our portfolio companies have always been deploying some form of AI or machine learning, and most have built basic versions into their products. What we've seen over the last 18 to 24 months is people building the harnesses and scaffolding for true agentic solutions — things that don't just auto-summarize or run a simple optimization algorithm, but can do sustained work. These agents are now live in the field working with customers.

So the question moves from "can you build an agent-like platform that works in demos" to "can you deploy it with customers and manage the change management of that?" AI as a bolt-on to the way people currently do things doesn't realize its full potential. It's probably an overused analogy, but putting electricity into a factory that was run on steam power didn't change the factory — you needed to change how it worked. Similarly, if you deploy agents with your customers, they're going to have to change some of the ways they work, and that requires change management.

So now we're in the age of not just "can you build them" but "can you deploy them, and can you help your customers get the value out of them that they're looking for?" The question isn't "can I build and sell the agents" — it's "what's the real-world impact happening on the back end?" That's what investors are really looking for now.

Serena: There's such a gap, right? A few years ago we heard a lot about pilots, and in our prep call you mentioned companies moving from experimenting to deploying. As you look at your portfolio now, what tells you they've actually moved from pilot to real deployment?

Dave: Ultimately this is about what the customers say — what I say is relatively irrelevant. Are you making those customers' lives better? Take an example from the energy sector: one of my companies, Volue, is deploying AI to improve the function of control rooms in the power sector, making them more efficient and doing optimizations.

The question becomes: does it work? Do your customers see the value? Do they think, "Wow, this is something I want more of"? It's really about ultimate customer satisfaction and their own ability to measure hard ROI. If you're selling into innovation budgets, that's not where it's at — it needs to be selling into real business leaders who see that this reduces their risk, reduces their cost, or ideally grows their revenue by optimizing the business further.

As a sustainability-focused investor, we're always asking the business ROI question, because that's ultimately how our companies sell to their customers and make money. But we're also asking the impact question — is it improving the lives of clinicians, improving medical outcomes, reducing waste and loss on the grid? These impact questions go hand-in-hand with the customer ROI questions.

Serena: Everyone is talking about how to get ROI on AI these days, and we sometimes think in the AI world, ROI isn't just return on investment, but return on intelligence. When you hear a company explain the value they're creating with AI, what convinces you they're creating real business value and not just AI activity? Can you give a concrete example?

Dave: The simpler the ROI is to explain, the more real it is. If there's a complicated logic chain required to explain it to customers, it's probably not going to land. Take one of our portfolio companies, Octopus Energy — the leading electricity retailer here in the UK. They've been deploying AI through their software platform, Kraken, for two or three years to improve how customer service messages are handled. They have two North Star metrics: does it improve the efficiency of their team, and does it improve the happiness of their customers? If it does those two things — which they've proven it does — that's ROI.

That's really where the best form of it comes down to: what does this business exist to do, and does this improve the metrics that matter most? It becomes harder to measure when you're talking about things like internal productivity around speeding up R&D — we've moved from lines of code as a metric, to commits as a metric, and now toward features and business-value features. It's harder there, but when you have a company with a very clear mission — how do we make our customers happier while being more efficient — if you hit both of those, you know you're winning.

Serena: I don't know if every company has truly moved away from lines of code to commits — I think we're still getting there. I've talked to a lot of leaders recently who are trying to get help measuring AI ROI, and I see most metrics are around AI activity: lines of code, prompts, even tokens. If you don't have any activity, you're not going to get ROI — fair — but the connection to things like productivity, efficiency, and innovation is harder to measure. Do you see that shift really happening, or is this a lack in the enterprise where we're still thinking too much about activity and not connecting to business outcomes yet?

Dave: The simpler the activity, the easier it is to measure. When you're talking about relatively easy-to-measure activities — say, scheduling an appointment, and whether it was the right appointment — you get relatively fast feedback loops, and that's clearer. It's harder as you move toward complicated multi-stage tasks where you can't measure a single outcome.

I think there's going to be an adoption curve. At the start, you let everyone try stuff, so you can't have any view of anything. I think everyone's now at the point where they're letting their teams experiment. Then it's going to be about best-practice sharing, and honing in on what processes need to change, and whether we've incentivized our people the right way to achieve the right outcome.

This has been true for engineering leadership forever — it goes back to the book The Mythical Man-Month. If you force overly bureaucratic measurement of productivity within engineering organizations, you end up incentivizing the wrong things, because what you actually want to incentivize is business value. So this is really about: are you giving people access to the right tools? Are you helping them share best practices? Are you getting out of their way so they can use those tools as efficiently as possible? Do you have the right people — because the right people always want to achieve the right outcomes.

If you create too many false forms of measurement, you end up with people grading to the test. Whereas if you have incredibly empowered individuals who believe in the mission and want to succeed, that's going to be a far better outcome than any measurement chart will produce.

Serena: I like what you said a lot — a few things stood out, so I'll summarize for our listeners. One is the focus on simplicity — the KISS principle, when it comes to metrics for AI adoption or business outcomes and ROI: it has to be simple and easy to explain and understand. The other is your focus on customers — are they happier, better off as a result of what you're doing? Because if not, frankly, nothing matters. You could be doing a lot more with AI, but if the customer impact isn't there, it doesn't really matter.

Dave: And I do think this is where, if you have a really clearly mission-driven company that knows what change it wants to see in the world — whether people join for that mission and are incredibly motivated by it — that alignment means you'll get way more out of these tools than if you're trying to force adoption. We think winning cultures are an incredibly important part of it, and winning cultures aligned with strong missions are always better. In a world of AI enablement where your people can do more than they ever could before, that combination of mission and culture is going to be even more powerful than it was before.

Serena: Definitely. For engineering leaders listening, what you said really stood out — are we incentivizing the right behaviors, do we have the right people, are we building the right culture? I love the focus on humans, because we talk so much about AI, but we know humans need to be in the loop, and we also worry what happens to us as AI gets more capable. As far as building winning teams, do you have real examples of leadership practices that create the right incentives and environment for that winning culture?

Dave: That's a great question, and it's core to how we think about and assess companies. Very often when we're doing diligence, we'll ask to speak to the chief people officer — and it's amazing how rarely other investors do that. It's always surprised me when people say, "You're the first person that's wanted to speak to our chief people officer."

Chief people officers are incredibly important — obviously you also need a CEO who can lead from the front on culture, so people function isn't a replacement for that. But it's an unbelievably important part of due diligence: do you have companies that simultaneously have a winning culture — people showing up because they want to win, because they believe they can win, and hold very high standards for themselves — while also having a sufficiently kind, mission-oriented culture, so they're doing things for more than themselves?

That's what we look for in companies. I'm blessed to be on the board of a number of companies where we have that together. But that's the challenge — sometimes there's a view that says because a company is mission-driven, it needs a culture that's less focused on winning, with a lower bar. From our perspective, no. If you want to change the world, transform the lives of millions of people, save gigawatts of power for the grid — you need a higher bar for yourselves than you would for something less important.

Serena: Well, I hope our chief people officer listeners are lighting up as much as I am — I'm going to send this episode to a few of them who'd really appreciate it. To get to know your culture, talk to the chief people officer. There's no one better who can tell you the truth of what's happening inside the organization.

Serena: You've mentioned energy and sustainability, and that's what potentially holds companies back from making progress with AI — it can become a constraint. From your history investing in energy technology and software, where we're now seeing energy become the constraint that determines success, are you seeing companies actually plan for this? How are they approaching it?

Dave: It's a great question, and it's really core to how I came to AI — through the energy problem. This was probably six or seven years ago, talking to people thinking about the sizes of cluster they might end up building — back then for autonomous vehicles or computer vision — and seeing how much power consumption was going to increase. My somewhat naive view at the time was that wind and solar were increasingly cheap, so the power problem was kind of solved. And the reduction in the cost of renewables is one of the wonders of the world — but you have to connect them all, and we don't have a grid that can move at the speed of renewables or the speed of AI.

We became increasingly concerned about the grid build-out as a constraint on AI. When we were speaking to utilities back then and asked, "How are you getting ready for this AI power wave that's coming?" they'd say, "We don't see anything yet." That's what got us super interested, back around 2020–2021, in the energy problem of AI.

We think about it at two levels. One: how do we rapidly, but in a low-carbon way, get more capacity onto our grids? Two: how do we optimize the tokens per kilowatt hour? We've got a constrained number of kilowatt hours — a constrained amount of power we can put into this — so how do we get more intelligence out?

Back then, we called this thesis "green data." When we announced our investment in WEKA, it was all about green data and the question: if you've got GPUs burning power but not doing useful calculations because the data isn't fed in fast enough, that's wasteful. If you can feed more data in and massively improve the efficiency of your cluster — in the experience of one WEKA customer we spoke to during diligence, taking a training run from two weeks down to four hours — you're optimizing your intelligence per kilowatt hour. And now, moving into the world of inference and WEKA's augmented memory grid, we're seeing the same thing happen. For us, this is an incredibly important North Star for anyone building large-scale machine learning or AI applications: am I doing everything I can to get the most intelligence out of the kilowatt hours I'm drawing?

Serena: That's a very high bar, Dave. So the metric I hear you say is intelligence per kilowatt — is that right?

Dave: Intelligence per kilowatt hour, yes. For how much power I'm pulling off the grid, am I getting the most out of that? Because we only have a limited number to go around, and the demand for intelligence seems to be unlimited. The supply of kilowatts is very much limited. So we've got to be doing everything we can at all layers — what WEKA's doing in the augmented memory grid, what WEKA's doing in NeuralMesh, and other solutions — to get the maximum intelligence we can out of the resources we're drawing.

Serena: Such a good reminder — a lot of companies are probably wondering where that metric is on their scorecard. It's probably not shown to the board regularly, I'd bet. But investors are going to start asking, because there could be a constraint on how much we can grow. What are your practical recommendations for companies that want to optimize this metric if they haven't really thought about it yet?

Dave: This is the flip side of tokenomics — many people have talked about this extensively. In many ways, maximizing resource efficiency is the same thing as maximizing return on cost, so it's many of the same things others have talked about elsewhere.

Ultimately it's about the economic concept of the binding constraint — looking through your stack and asking: do I have the best power efficiency in the physics of my data center? Increasingly, people do. But am I running storage that's optimally feeding data in? Am I running networking that's efficient? How am I moving toward photonics, which will reduce the energy intensity of networking significantly? How am I thinking about the use of KV caching and other techniques — one of our portfolio companies, Redis, has done a bunch of work to make sure you only hit the model when you need to. How do you do smart model routing to make sure you're using the right model for the right task? It's all of these elements, and as power becomes more and more the bottleneck, we're going to have to ask ourselves this question at every level of the stack.

Serena: That was very helpful and practical for anyone just getting started. When I run into people who are concerned about energy and ask what they can personally do, it's — at the personal level, are we using the smallest model possible for the task? A great parallel to what you just said at the company level, but we can also individually try to minimize energy consumption to some extent.

Dave: That's right. But I do think the onus really is on companies deploying AI to be the ones doing it. We're very bullish on the transformative effect AI can have — we're not people who say, "Try to reduce your usage." No — intelligence can help solve the world's problems, so we should have more of it. But we have to do that within physical constraints, and that's about not being wasteful — looking through everything we're doing when architecting our clusters to make sure we're doing the most.

This isn't going to be optional. If you look at the demand for data centers coming onto the grid in places like PJM in the US, and the amount of power we can build and interconnect, it's very clear we're going to be at a power constraint — we already are. If you look at the capacity auctions run by PJM, they've massively increased the price of power capacity. We're getting to levels that clearly show demand outstrips supply. So I don't think companies are going to have a choice. Consumers are going to demand intelligence in a variety of ways, the grids won't be able to build fast enough, and everyone's going to need to come back to: how do I get the absolute maximum efficiency out of the systems I'm building? Otherwise they're going to lose to competitors adopting more efficient technologies and getting more out of the kilowatts they have.

Serena: Excellent reminder — I like the focus on the responsibility being on companies to make the right decisions and get the most intelligence, because that demand is only going to keep going up.

Serena: Let's talk about the future for a minute. As you think about the next 18 months, what do you think the AI investment landscape will look like?

Dave: The really hard thing about being an investor is you have to try to predict the future, and predicting the future is very, very hard. Our companies are generally some of the most advanced in terms of adopting AI, because those are the kinds of founders we back — probably in the top 1% for how much they're using AI to improve internal processes. That will cascade across companies of all sizes. I think we'll see more and more companies over the next 18 to 24 months deploying AI in R&D, customer support, and the ways our companies already are — a mass growth in the use of AI inside the enterprise, both for internal efficiency and for building better products.

You may see some of our companies' spending on AI for internal use start to plateau as they think about which are the right models to use, when to use open source, learning the optimization. It wouldn't surprise me if spend continued to go up from here, but it wouldn't surprise me if it flattened — I don't think it'll keep growing at the 20% month-over-month it was a few months ago. I don't think that's sustainable. But more and more companies in the mainstream corporate world will start using AI the way our companies already are.

The second thing: the power issue is going to become front and center. Power shortages will be real — you can see this in any interconnection queue — and this is going to force the big data center builders to get creative about things like funding demand response, like Google just did with a company called Voltus, or speeding up the interconnection process for bringing new renewables and batteries onto the grid.

If we don't solve the power problem, two things happen. One, AI gets constrained. Two, we risk losing our social license to operate, where local communities say, "You can't just take the power and force the price up — if you do that, we're going to resist." You're seeing this across the US right now with backlash against new data center builds. So data center developers are going to have to become grid developers and help fund the grid upgrades we need, or there's going to be a real social-license-to-operate problem.

Serena: Absolutely — find ways to give back to the community. We heard earlier in our series that data centers in some European countries convert waste heat and share it back with the community, so there's free heat for everyone in the city. In different parts of the world, because regulations and incentives are different, that's not always top of mind. But I like the reminder to think about getting ahead of the power constraint.

Serena: If you could wave one magic wand and pick one metric you'd like to see from AI companies to create long-term value, not just hype, what would it be?

Dave: Can I have two?

Serena: Sure — but you have to pick the primary.

Dave: I think there's two things, because there's what you build and there's how you build it. In what you build, it's really about customer satisfaction, however you measure that — are you improving the lives of your customers and achieving your mission by doing so? That'll be a different metric depending on whether you're a healthcare company or an energy optimization company, but fundamentally: are you helping your customers in a way that makes the world a better place and achieves your mission, and are they happy with that? That's the first one.

And in how you build it — it's intelligence per kilowatt hour. With the resources I have, am I doing the maximum I possibly can to get the most intelligence out of the fewest resources?

Serena: I like that. So: customer satisfaction, and intelligence efficiency — can we call it that?

Dave: Yeah, sure.

Serena: Thank you, that's very helpful. All right, everyone listening — those are the two magic metrics to bring to your next meeting.

Serena: Dave, let's end with some advice for tech founders who might be listening — whether they're building an app to coach employees or one that helps clinicians take notes. If they're listening today, what's one thing they should get right before going into a meeting with an investor?

Dave: Ultimately it's about why your company exists and what its higher purpose and higher mission is. There's no point building companies that don't impact the world in a positive way. That doesn't mean they shouldn't make money — they have to make money — but you're not going to inspire your team, your customers, or yourself if there isn't a reason you're doing what you're doing. That reason is both to make money and create value for your shareholders, and to have a huge impact on the world. For me, it's about being really clear about what's the driving force that's going to make you and your team want to work incredibly hard — and to me, that always comes best from mission-driven companies.

Serena: I couldn't agree more. The startup founders I've seen who are really successful have that really strong mission, and it's what lets them bring people along — and when things get hard, get themselves out of bed asking, "Why am I still doing this?"

Dave: Yeah, because it's definitely going to be hard. That's the only thing you can promise any entrepreneur — there are going to be moments of deep suffering. In those moments, why are you doing what you're doing, and does it motivate you more than the pain? If it doesn't, there's no point.

Serena: Does it motivate you more than the pain — that's the bar.

Serena: Thank you so much for all the insights you've shared today, Dave. Any last word you want to leave with us?

Dave: I just wanted to say thank you — to the whole WEKA team. It's been quite a journey so far working with WEKA. There's a huge wave of opportunity coming for everyone building the right companies in the right way in the AI wave, and being on the right side of history is hugely valuable. I'm excited to be on the journey.

Serena: Thank you, and I really appreciate the behind-the-scenes view as an investor — we so rarely get this kind of opportunity. Really, thank you for your time today, Dave.

Dave: Thank you.

Serena: Thanks for listening to Deep Geeks. A huge thank you to my guest today, Dave Easton. If today's episode made you think differently about how AI gets built or powered, share it with someone who needs to hear it. Find Deep Geeks on Spotify, YouTube, or wherever you get your podcasts. Until next time.