It’s been another restful couple of weeks in AI. Euphoria over Anthropic’s S-1 and their apparent cash efficiency of what is a pretty stunning business was replaced by resurgent concern of the unit economics of the build-out, as well as the implications for the growth drivers of the S&P 500. Then this morning we layered on the results of the midterms, and the confirmation of the historic democratic unpopularity of AI and data centers in the US. The fun & games will continue in the coming weeks and months, no doubt.
The advent of Anthropic’s public market debut combined with new large question marks about how AI’s rails will get built feels like the start of a new chapter in a very long and important book. So it’s a reasonable time for us to set out our key thoughts on what’s going on with water & AI, and how we see it playing out. There are more facts on the field, and the implications for water, and for BIV’s strategy in the market, are substantial.
We’re going to take this in two sections. First, the physical aspects of AI. A lot of holes in the ground, a lot of cooling, and a lot of fear (and loathing). Second, the implications of AI in water. What does this platform shift mean for the water industry? What isn’t obvious from the outside? A future piece will look at the implications of all this for running companies (including a venture capital firm) in the water space, but that’s a separate area of consideration.
So, to the physical.
The build out of compute capacity is inevitable. There is too much demand for it not to be. John Collison at Stripe (not a hyperbole merchant) says that they’re at 1% of opex on tokens, and have a clear pathway to 20%. His prediction is that opex spend from corporates will be c.10% of total opex. The likely path of demand cannot be understated. It’s worth tracking the price elasticity of demand, but for now it seems that rental prices are growing even as token prices go down because of the overall weight of demand (but it’s a complicated picture). Even attenuation of demand by an order of magnitude will mean we will need much more data center capacity. The only questions are where and how.
The build out of compute is inherently unpredictable. Now even more so. Exit polling had 56% of voters cite AI and data centers were a crucial or major factor in their vote. Moratoria votes are likely to go 41-0 in counties across the country. The balance of Governorships poses material problems for developers, even if they can solve their issues with power and talent availability. There are multiple overlays of uncertainty - siting, power availability, legal challenges (the Finns are angry), massive community opposition, even demand over time. Then there is the problem of the neoclouds, and whether or not their capital needs are actually fundable. What gets built and when is increasingly unpredictable.
Hyperscalers & neoclouds may be getting the message. Hyperscalers plus the NRDC, Sierra Club, League of Conservation voters announced the “Responsible Data Center Collaborative”. The sooner we skip to the builders just paying directly for grid (water & energy) upgrades, energy & water efficiency, large community benefits, the better. The degree of wealth creation at the labs vs the impacts of data centers locally is an untenable and unwise imbalance, and a (functionally irrelevant) increase in the compute price tag to benefit communities is the only way to get these things built.
Water is volumetrically irrelevant, but politically expedient. Water volumes can be material locally, but it tends to be for data centers that have already been built. New designs are 90% more water efficient, and even fully built out, data centers will consume a fraction of what is used by golf courses, for example, immediate local hydrology on the ground notwithstanding. Semiconductors and power are much bigger water issues (and as such represent extremely large markets).
Power is valuable enough that minimal-water-use cooling will be less common than claimed. Even with this technology, water will be used where possible if freely available, because it’s very cheap. Power is the true bottleneck, and efficiencies there are likely to be prioritised over efficiencies in water use.
The opportunities in “picks and shovels” are both myriad, and very popular with the investment world. This means competition is high, deals are both crowded and expensive. Nevertheless interesting opportunities are presenting themselves, from coolant monitoring, to volumetric water offset purchasing. It will be interesting to watch the market price these opportunities in the coming months - our sense is that the scale of demand will overcome any concerns over the eventual scale of the build out.
The system integrators and engineering consultants are doing very, very well indeed. Everywhere. The third quarter earnings calls were remarkable. Jacobs’ data center backlog doubled year-on-year. CBRE’s infrastructure services rose 45% from Q2. Vertiv’s net sales were up 61% year-on-year. The list is long. There are interesting opportunities for new entrants in conditions of large backlogs, especially if they can build themselves up with a fundamentally different operational and margin profile.
So are the equipment suppliers. Our great friends and colleagues at Xylem saw a 200% increase in data center orders in Q3 (300% in Q2). Everyone else associated with the market also did well.
We have forgotten the pace of the non-AI industrial build-out. We’re still building an amazing amount outside of the data center (and related) boom. Manufacturing construction spending is still 2x what it was in 2019, even after a 30% slide from the late 2024 peak. The right way to think about this is that we are rebaselining industrial water demand in the US.
Once we’ve built all of this stuff, how will all that compute be used in helping our water & wastewater systems function faster, better, cheaper? It was a true throwback to revisit this piece I wrote in 2017 which feels almost impossibly quaint ($5.4bn in AI VC investment that year 😱) It’s been interesting to see our progress against the issues highlighted. This is what seems clear to me:
Water is gradually solving its data problem. In that 2017 article, the real barrier to actual application of AI was the depth and quality of water’s data. We’re still a long way behind (40% AMI penetration vs 80% in energy, for example). Everything new that is installed naturally becomes a data source. We’re starting from a long way back - shout out to the operators still doing chlorine residual testing by hand, for examples. My Partner Christine did a great summary on the workflow ramifications of this data expansion.
AI in the pursuit of marginal gains. This was excellently articulated by Antoine Walter in his forthcoming TFM episode. The reason we started BIV was because of the density of smart people who were choosing to start companies to solve water problems. Water is looking like an increasingly vibrant talent anthill, with people crawling over each problem, trying to solve it. AI has supercharged those people, and it’s also a useful lens to highlight problems that can be deleted next. The best way to think about it is that any workflow that can be overtaken by AI has to be and will be. We’re going to transform utility and industrial operations one pain in the ass at a time.
AI as core to the virtual integration of utilities. We have talked a lot in the US about the weight that small systems have to carry on minimal resources, and the benefits of integrating those systems to achieve some semblance of scale and the advantages that go along with it. AI is already helping with that. It will be a superpower for contract operators, a major emerging force in the industry. Previously, contract operators would max out at 8-10 systems per operator, but with tools like Ziptility and Waterly they can double that, while also moving to a proactive footing. This kind of help is vital because we’re going to have 27,500 fewer operators by 2031. Tech has to help pick up the slack. TeamSolve help operators in their second week function like they’ve been in the job for two years. Ziptility now makes grant applications about 95% easier, for example. Current Software allows small utilities to process billing on cutting edge software, and manage their overall financial picture in a way that is unrecognisable.
You may have to not call it AI. We just found out how unpopular AI is. But more importantly, a great rule of sales in markets like water is that nobody cares what’s in the box, especially when they hate the AI label. They just care what the box does. Does it solve their problem? Is it easy to use? Is it reliable? Does it create other problems down the line? Focusing on the functionality rather than the technology will be important with a lot of customers. Daupler nailed this, and now serve 240 communities with a massive and growing pipeline.
Most organisations are going to have a transition problem. Pareto will abound where 20% of people in any organisation will be fluent in AI, and they’re going to be producing 80% of the output. Management will face a choice about the others. Especially as the AI-fluent will likely want to stop being stymied by the hold-outs. This is 100% behind Microsoft’s recent employee buy-out. There are going to be some important management choices for a lot of incumbents. This is on top of having to deal with upstart competitors that don’t have the overhang of legacy operations, and are operating far leaner and faster to deliver the same (or improved) value propositions. As my Partner Steve says, it’s definitely a revolution vs an evolution and companies need to act accordingly.
Hardware is a good place to build. All hardware can now be rebuilt from first principles, and there will be some surprises ahead, especially in the largest markets. Opportunities abound across the hardware universe in water, with the built-in advantage that it can’t be vibe-coded. Between materials science advancements, 3-D printing, design capabilities, comms and interconnection capabilities, robotics - some seriously large product verticals are now in play.
We’re about to find out more about software moats. Some very interesting questions about the legacy software platforms. How does this affect the SCADA market? Will someone finally be able to challenge Esri? It is worth noting the strength of the language from the mild-mannered Fred Wilson at USV when talking about their new fund - “Obliterate, don’t Automate”. It applies across the product ecosystem. Everything is on the table. The funeral of the SaaS model is likely to be overblown, however. People will just not want to maintain their own software in the end, and will pay someone else to do so. It’s fun for a while, but you end up in a mess and specialisation dictates that it’s just optimal to outsource after a point.
It will be a superpower to smart people. We love the visual metaphor of swarms of ants crawling all over every problem in this $1.6trn market to make whatever they can faster, better, cheaper, and with better outcomes for customers, companies, and the environment. The degree to which entrepreneurship in water is becoming more feasible and accessible is tremendously exciting. The energy is only going one way, and we’re absolutely at the bottom of the mountain here.
As you the entrepreneur think through which part of the puzzle to aim at, you’ve never been in a better position. The markets are huge and growing, incumbents will find it hard to respond, inorganic growth is a must, PE interest is at record highs. It’s worth remembering the things that don’t change (here and here, for our take) as well as leaning into the things that have, that are changing the game on the field.
It will be fascinating to see where this all goes next. Thanks for reading.
