In Colorado River Basin, Data Center Water Consumption is Not the Problem

Though we hear lots about data center water use, a little perspective is worthwhile. Actually, industry in general, and data centers in particular, represent a small part of water consumption in the desert-like U.S. southwest, for example.  

In the Colorado River Basin, where I live, agriculture accounts for about 74 percent of direct human water use and 52 percent of overall consumptive use (including reservoir and other evaporation), a study estimates. 

Water consumed for agriculture amounts to three times all other direct uses combined. Cattle feed crops including alfalfa and other grass hays account for 46 percent of all direct water consumption.

Another study by Landsat found that in the Colorado River basin, 52 percent of consumption was for agriculture; natural vegetation representing 19 percent of use; evaporation claiming 11 percent of the water and households plus cities and industry consuming 18 percent of the Colorado river’s water. 

 

We need to be thoughtful about data center water consumption, to be sure. But data centers are not the problem, in the Colorado River basin supporting 40 million people. In Colorado, for example, data center water consumption is a negligible percentage of total water consumption.

In Colorado, agriculture consumes 89 percent of the water used by industrial, municipal and agricultural users, for example. Consumers use no more than 20 percent of total water.

Colorado Annual Water Use by Sector
Sector	Delivery / Withdrawal	Consumptive Use	Approx. Share
Agriculture	~13–13.3 million acre-feet (AF)	~4.7 million AF	~87% of delivery; ~89% of consumptive use waterknowledge.colostate.eduWater Education Colorado
Municipal and Industrial	~1.025 million AF (combined)	~0.371 million AF	~6.7% delivery; ~7% consumptive use waterknowledge.colostate.edu
Self-Supplied Industrial	~0.168 million AF	~0.212 million AF	~1.1% delivery; ~4% consumptive use waterknowledge.colostate.edu
Data Centers (subset of industrial)	Variable; example: QTS Aurora uses ~0.5256 million gallons/year (~0.0016 AF); CoreSite Denver proposed up to 805,000 gallons/day (~2.9 AF/day, ~1,060 AF/year) Business InsiderBusiness InsiderGoverning		negligible

David Ricardo was Probably Right About Automation, Hence Right about AI

David Ricardo is an economist whose views on automation and substitution of machinery for human labor are relevant for discussions of the impact of artificial intelligence on jobs. Basically, Ricardo came to the conclusion that substituting machinery for human labor was not good for workers. Lots of workers who never study economics could figure that out for themselves. 

With regard to the impact of AI, we might find lots of cognitive workers pondering the impact as well. 

In principle, automation can increase wages, but only when accompanied by new tasks that raise the marginal productivity of labor, and when workers in the affected industries and roles are able to gain a share of productivity growth (profit sharing or some other mechanism). 

There are some early signs that what AI might be doing, in some areas such as customer service and software development, is eliminating some number of entry-level jobs. 

source: Stanford University Digital Economy Lab

Is Code the Enterprise AI Killer App?

Is code generation the first “killer app” for language models or mostly the killer app for enterprise users? 

Since 2024, one might argue, model spending on language model application programming interfaces has more than doubled. Where model development might have driven spending in prior years, inference now seems to be driving growth. 

source: Menlo Ventures 

According to Menlo Ventures, Anthropic has surged to the lead in enterprise market share, while OpenAI has lost share. Google’s Gemini growth rate also parallels that of Anthropic. 

source: Menlo Ventures 

And there seems ample reason to believe that, at least in the enterprise portion of the market, the drive to gain share fast, and invest heavily in upgraded models, has a clear customer logic. Enterprise customers do not tend to switch model suppliers once they have made a decision. But they also will upgrade based on access to better performance.

The former preference encourages the importance of gaining share fast, as once an enterprise commits to a particular platform, it will tend to stay. The latter preference explains the intense effort to produce more-capable model versions, as that drives model upgrade or switching behavior on any platform. 

So performance clearly matters for retention, not simply customer acquisition. 

source: Menlo Ventures 

So where an order of magnitude price drop might convince many consumers to use an “older” model, enterprise users tend to behave differently, preferring to buy the latest and most-capable model. 

source: Menlo Ventures 

Consumer trends are different, of course. Most consumers use the free versions of models. That has meant the monetization model is converting users of the unpaid versions to model subscriptions. But Menlo Ventures does not believe that will be the dominant form of monetization in the future, anymore than subscriptions have been the driver for search, social media or e-commerce. 

source: Menlo Ventures 

“The biggest long-term monetization opportunities won’t be subscriptions,” a Menlo Ventures report argues. “We expect rapid adoption of advertising models, transaction fees, affiliate revenue, and marketplace models.”

In the near term, others might say subscriptions are going to remain the leading direct monetization model. 

Stage	Dominant Monetization Models	Example Contexts
Today (2024–2025)	Subscription, API billing, in-app upgrades	ChatGPT, Copilot
Near-term (2025–2026)	Affiliate links, transactional agents, SaaS hybrid	AI travel planners, creative tools
Mid-term (2026–2028)	Task-based pricing, agent commissions, smart bundling	AI exec assistants, recruiting agents
Long-term (2028–2030)	Ubiquitous embedded AI, conversational ads, OEM/device-based	Voice wearables, car copilots, ambient AI

Of course, different models are going to make more sense than others, depending on the use case. What makes sense for an autonomous or embodied deployment will not make as much sense for transaction use cases or content consumption. 

Monetization Model	Consumer Trust	Scalability	Fit for Autonomous AI
Subscription	High	High	High
Transaction/Outcome Pricing	Medium–High	Medium	Very High
Affiliate/Referral	Medium	High	High
Conversational Ads	Low–Medium	Very High	High
In-app AI Features	High	High	Medium
Embedded in Devices/OS	High	Very High	High
Data-for-Access Models	Low	High	Medium

"Data Sovereignty" Might be an Illusion

As a practical matter, data sovereignty, the idea that data is subject to the laws and governance structures “only” or “exclusively” within the nation where it is collected or stored, is probably more accurately described as “data residency” in some cases and in some cases not sovereignty at all. 

Governments can lawfully obtain some data, when prosecuting major crimes, for example. And how often does any bit of data reside “exclusively” within any one political jurisdiction, in any case? 

While a country may assert that data stored within its borders is governed by its laws, in reality, data often resides in cloud infrastructure spanning multiple jurisdictions. can be subject to both General Data Protection Regulation and U.S. subpoenas under the CLOUD Act.

Mutual Legal Assistance Treaties (MLATs) and bilateral frameworks such as the U.S. CLOUD Act allow law enforcement access to data stored in other jurisdictions when investigating serious crimes.

These mechanisms sidestep national data sovereignty, creating pragmatic paths for lawful access, even if the data physically resides in a foreign jurisdiction.

If a local data center is operated by a foreign company (AWS, Google Cloud, Azure), that company may still be compelled to produce data under its home country’s laws.

So data residency becomes more symbolic than anything else, a bit of posturing, even if, under most circumstances, most data will not be subject to unusual or extraordinary access. In cases involving terrorism, money laundering, cybercrime, or child exploitation, governments often claim national security imperatives that justify sidestepping normal sovereignty considerations, and many observers might agree such practices are defensible. 

So while “sovereignty” might still hold, in practice, for most data, the protections are not absolute. 

Concept	Ideal Definition	Real-World Practice (Serious Crimes)
Data Sovereignty	Data is governed exclusively by local laws	Subject to foreign laws if provider is foreign or Mutual Legal Assistance Treaties apply
Data Residency	Data stored within borders to ensure sovereignty	Storage local; control possibly foreign
No Sovereignty	State borders don't block data access	True for intelligence ops, cybercrimes, cross-border subpoenas
Legal Access Channels	Governments access data via their own legal systems	MLATs, CLOUD Act, intelligence-sharing bypass local laws

Head Turner

 

I’m sure psychologists will have explanations for why this flight attendant turns heads (men and women, both).  I don’t think it’s just the uniform, the figure or the flight attendant mystique. Height, perhaps. Poise, certainly. I think it’s the runway model vibe.