What Nobody Tells You About AI and Your Water Supply

Last week during a workshop on AI tools in the classroom at Pace University, a faculty member asked about AI and water use. I expected the question. The workshop was about classroom applications, not infrastructure, so I kept the answer short and moved on. Here’s a longer answer.

Data centers consume water directly, through cooling towers that evaporate it to remove heat. US data centers directly consume an estimated 17 to 19 billion gallons a year, a figure projected to climb toward 60 to 110 billion gallons by 2030. The bigger draw is indirect. Generating the electricity these facilities use consumed roughly 211 billion gallons of water in 2023, mostly through cooling at power plants. A data center's water footprint depends as much on how its power is generated as on how its servers are cooled.

Location compounds the problem. Two thirds of new data center construction since 2022 has landed in areas already under water stress. Municipal systems supply nearly all of that water, and the infrastructure upgrades to meet rising demand often fall on residential ratepayers rather than the companies building the facilities. Transparency has not kept pace either. Most operators disclose only partial or aggregate figures, so the public rarely sees facility level numbers.

The fixes already exist - sort of. Closed loop and liquid cooling systems cut water draw at the facility to near zero. Siting near hydro, nuclear, or wind power removes most of the indirect draw tied to electricity generation. The Massachusetts Green High Performance Computing Center in Holyoke, built in 2012 on hydro power from the Connecticut River, has been proving this model for over a decade. But.... scale matters here. MGHPCC runs on a 15 to 19 megawatt connection with about 10 megawatts dedicated to computing. A single hyperscale AI campus now under construction can draw a gigawatt or more, one hundred times that load. Holyoke proves the cooling and power sourcing approach works. It does not prove that approach scales directly to a facility built for AI though.

Regulation is starting to catch up. State legislatures introduced more than 200 data center bills in 2025 and enacted over 40 of them across 21 states. Most of what passed addressed ratepayer protection and siting incentives rather than water directly. Roughly 30 of the proposed bills targeted water consumption specifically, and only a handful of those were enacted. South Carolina and Kansas are considering mandates requiring closed loop cooling. California, Iowa, and Michigan are pushing water use disclosure requirements. Virginia is taking a different route, tying grant funding to the use of reclaimed wastewater instead of potable water in cooling systems. None of this is settled yet. Water legislation is still catching up to where energy legislation was two years ago, and disclosure requirements are starting to follow the path carbon reporting took a decade earlier.

I told that faculty member the honest version this week: the technology to fix this exists, and it works today in Holyoke. Scaling it up is an engineering problem, and engineers solve those. What is missing is the will to require it before the next facility breaks ground.

The Classical Engine Inside Every Quantum Computer

Before writing this, I realized NVIDIA Ising already showed up in Chapter 12 of the book, the one on AI-assisted calibration and error correction. That chapter only had NVIDIA's side of it. A few days after that post went up, IEEE Spectrum ran a piece by Edd Gent that talked to Q-CTRL, Riverlane, IBM, and Google about the same problem, and the picture got more interesting. This post is what Chapter 12 looks like with that reporting folded in.

I have spent the past several months writing about quantum computing. Topological qubits, neutral atoms, silicon spin, photonics. Each one has its own physics, its own failure modes, its own argument for why it will win. Putting the book together, I noticed something missing from every chapter: what runs the thing once you turn it on.

The answer is classical computers. A lot of them. A recent IEEE Spectrum piece by Edd Gent lays out why, and it changes how I think about the supply chain side of this industry too.

Qubits are not reliable the way transistors are. A transistor comes off the line and works, billions of times, with no babysitting. A qubit drifts. It decoheres. It needs constant tuning just to stay usable, and that tuning happens in two stages. The first, called “bring up,” determines the frequency each qubit resonates at, how long it holds its quantum state, how sensitive it is to control pulses, and how strongly it interacts with its neighbors. Every one of those numbers affects error rates and how the qubit responds to control signals.

Done by hand, bring up still requires someone with a PhD and can take days or weeks, according to Jay Guilmart, lead product manager at Q-CTRL. The process resists scripting because each step depends on the result of the one before it. So Q-CTRL built software that examines each measurement, diagnoses what failed, and decides whether to proceed, repeat, or back up, rather than running a fixed sequence. Calibration does not end once a machine is running either. Parameters drift over time, so the system needs “runtime recalibration” to nudge things back into spec. But every cycle spent recalibrating is a cycle not spent running a circuit. As Guilmart puts it, if recalibration eats your uptime, the high fidelity you maintained is worthless.

Then there is error correction. Quantum information gets spread across many physical qubits to form a “logical qubit,” so that errors in individual qubits can be caught and fixed without destroying the encoded information. Measuring a qubit directly collapses its state, so error detection works indirectly through “parity checks,” which compare pairs of qubits to see if they agree. The pattern of agreements and disagreements is called a syndrome, and classical algorithms called decoders read that syndrome to locate errors.

This has to happen fast. Superconducting and silicon spin qubits hold their state for only microseconds to milliseconds, so decoding has to finish inside that window or the algorithm stalls. That rules out general-purpose processors. Decoders run on FPGAs or ASICs, chips built specifically for speed, according to Jerry Chow, CTO of quantum-centric supercomputing at IBM.

This is where the AI argument gets interesting, and where the field splits. In April, Nvidia released two models aimed at this problem. One uses a vision-language model to read calibration plots and feed an AI agent that decides how to adjust the processor. The other uses a convolutional neural network to catch the simple, localized errors that make up most faults, passing only the harder cases to a traditional decoder, for roughly a 2x speedup. Sam Stanwyck, Nvidia's director of quantum product, makes the case for AI on inference speed: models take time to train, but once trained they run fast and parallelize across chips as qubit counts grow.

Marco Ghibaudi, VP of engineering at Riverlane, pushes back on a different axis. Running anything on a GPU adds latency, even with massive throughput. His framing: you can have a very fat pipe, but if it is also a very long pipe, the data still arrives late. Riverlane's approach has been to shorten the pipe itself and make every stage faster, rather than chase raw throughput.

IBM's Chow lands in a similar place on calibration. He thinks AI shows real promise for understanding new architectures or unfamiliar circuit types, where you do not yet know what you are looking for. But for a well-characterized device where you are hunting small deviations from a known baseline, simpler physics-informed methods are cheaper and faster. Google's Adam Zalcman frames the two approaches as complementary rather than competing: neural networks are good at finding hidden patterns in syndrome data that handwritten decoders miss, so Google is building architectures that can run both, including its AlphaQubit 2 model.

Andi Gu, a Harvard PhD student working on AI decoders, takes the long view. He expects the “bitter lesson” to apply here eventually, the same pattern that played out in other parts of AI: general learning methods, given enough data and a large enough model, outperform hand-built algorithms over time. The obstacle right now is latency. Gu's group is working on shrinking AI decoders enough to fit on an FPGA, though smaller models trade off some accuracy, and finding that balance is still unresolved.

What stands out reading this against the QSCA work is how much of the “quantum workforce” conversation undersells this layer. When people talk about quantum jobs, they tend to mean physicists and people who understand qubits. But calibration software, decoder design, FPGA and ASIC engineering, and the AI models being built to assist both are squarely classical computing jobs. Guilmart's own warning, that calibration overhead will “blow up” past a thousand qubits and that current techniques will not scale, is as much a hiring problem as a technical one. Nobody in the article claims to have this solved, and Guilmart says plainly that no one is winning this battle today.

Every qubit type I have written about, no matter how exotic, sits on top of a rack of classical silicon doing the unglamorous work of keeping it alive. 

*****

Download Quantum from the Ground Up, a free PDF now available at gordostuff.com. This first edition covers posts through June 2026 and runs 50 pages across 19 chapters. It is written for someone entering the quantum workforce, or seriously considering it, who has a technical background but has not taken a graduate course in quantum mechanics.

The Internet in Seven Stages: Where We've Been and Where We're Going

The internet started as a research project connecting computers in the 1970s. Today it connects 5 billion people, countless devices, and increasingly intelligent systems. Mallik Tatipamula and Vint Cerf (the co-inventor of the internet) recently mapped its evolution into seven distinct phases in IEEE Spectrum, each building on the previous ones.

Phase 1: The Original Internet The foundation began with computer networks using common protocols. Early applications like email and file transfer proved that standardized connectivity could democratize information access. The World Wide Web in the early 1990s added URLs, HTTP, and browsers, transforming a research tool into a global platform.

Phase 2: Mobile Internet Smartphones in the early to mid-2000s made connectivity portable. The internet moved from desktops into pockets, enabling social networks, mobile payments, ridesharing, and on-demand services. The app economy emerged, putting services at our fingertips constantly.

Phase 3: Internet of Things Sensors, appliances, vehicles, and city infrastructure joined the network. The IoT created a nervous system linking physical and digital worlds, enabling smart farming, remote healthcare monitoring, and optimized manufacturing.

Phase 4: Internet of AI Agents (Current) We're now entering Phase 4, where AI agents can perceive, reason, act, and collaborate. Digital agents like coding copilots and workflow orchestrators operate in software. Physical agents like autonomous vehicles and industrial robots function in both digital and physical environments. Value comes from networked intelligence, not isolated systems.

Phase 5: Internet of Senses (Future) Networks will transmit touch, taste, and smell alongside audio and video. Haptic wearables will let shoppers feel fabric texture online. Doctors will examine patients remotely using haptic gloves. Smart cities will sense traffic and crowd movement directly through their networks using ISAC (Integrated Sensing and Communications).

Phase 6: Ubiquitous Internet (Future) Terrestrial and non-terrestrial networks (cellular, Wi-Fi, satellites, high-altitude platforms) will merge into one unified global system. Connectivity will extend everywhere: remote villages, oceans, skies, orbit, and cislunar space.

Phase 7: Quantum Internet (Future) The final phase will use quantum entanglement and teleportation to create ultra-secure channels and connect distributed quantum processors. Quantum sensors will achieve unprecedented precision. The Quantum Internet won't replace classical networks but augment them.

Each phase extends connectivity's reach. The internet evolved from moving data packets to becoming the intelligent, resilient fabric supporting our digital future.

Read the full article at IEEE Spectrum: https://spectrum.ieee.org/history-of-internet-7-phases

Why Your AI Assistant Might Be the Security Problem

You trust AI without thinking about it. Your voice assistant orders groceries. Your bank's AI approves transactions. Your phone's facial recognition unlocks with a glance. These systems work so well that you forget they can be fooled.

Here's the problem: hackers aren't breaking into these systems anymore. They're tricking them. And your firewall can't stop it.

Traditional hacking breaks through walls. Someone steals your password, penetrates the firewall, accesses the database. AI manipulation is different. The AI already has access to sensitive data. It already has permissions to take actions. Security tools see normal activity. They don't know the AI is being tricked into making bad decisions with data it's allowed to see.

Data Poisoning: Teaching AI the Wrong Lessons

In 2019, a factory's AI predicted when machines needed maintenance. It suddenly started failing. Equipment broke down without warning. Critical repairs were missed. Routine maintenance happened on perfectly fine machines.

Hackers had accessed the sensors monitoring the equipment. They didn't corrupt everything at once. They made tiny changes:
temperature readings slightly off, vibration data nudged higher, performance metrics tweaked. Each change looked like normal sensor drift.

The AI learned from this poisoned data for months. It learned that warning signs meant nothing. By the time anyone noticed, the AI's entire understanding was wrong.

Think about your spam filter. It learns from emails you mark as spam. What if someone slowly trained it to ignore phishing emails? Your bank's fraud detection learns from transaction patterns. What if someone gradually normalized suspicious behavior? You'd never notice until it was too late.

Data poisoning works because AI systems are designed to adapt and learn. That's their strength. It's also their weakness.

Adversarial Inputs: Making AI See Things Wrong

A hospital's AI reads MRI scans and flags potential problems. Doctors started noticing odd mistakes. The AI saw tumors that weren't there. It missed ones that were.

Someone had tampered with the images before the AI analyzed them. They changed a few pixels here, adjusted brightness there. Doctors looking at the same scans saw nothing unusual. The AI saw something completely different.

This is like putting trick glasses on someone. They're looking at the same thing you are, but they see it wrong. Except you can't tell the AI is wearing trick glasses.

Your phone's facial recognition works the same way. Researchers printed specific patterns on glasses frames that caused the AI to misidentify people. The AI looked at the face, processed the features, and confidently returned the wrong answer. The system worked perfectly. Someone just learned its blind spots.

Adversarial attacks craft inputs designed to confuse AI while appearing normal to humans. The AI isn't broken. The input is the problem.

Model Inversion: Talking AI Into Breaking Rules

Banks use AI to answer customer service calls. The AI can check your balance, transfer money, verify transactions. It needs these permissions to help you.

Now imagine someone calls repeatedly, testing how the AI responds. Can they get it to summarize information about other customers? Will it generate reports it shouldn't? Can they phrase questions that make it leak data?

The hacker isn't breaking into the database. They're not stealing passwords. They're just asking questions cleverly enough that the AI gives away information it shouldn't.

Users of smart home voice assistants reported targeted advertisements based on private conversations. Investigators found that attackers had extracted sensitive information by reverse engineering the AI's responses. The assistant wasn't hacked in the traditional sense. The AI model itself became the vulnerability.

Your voice assistant works similarly. Researchers embedded commands in audio that sounded like noise to humans. Your phone heard "order 50 pizzas." You heard static.

Speed makes this dangerous. Hackers can use their own AI to attack yours, testing thousands of question variations per minute. Human security analysts can't keep up.

Hidden Backdoors: Secret Triggers Embedded in AI

A corporation used voice recognition for secure building access. They discovered that unauthorized people could enter by speaking a specific phrase. The phrase acted as a trigger embedded in the AI during training.

The AI worked normally for everyone else. Only that exact phrase granted access regardless of who spoke it. The company had purchased the AI model from a third-party vendor. Someone had planted the backdoor during training. Traditional security testing wouldn't catch this. You'd need to test millions of potential inputs systematically.

Scale makes this terrifying. One corrupted AI model affects millions of users simultaneously. When someone embeds a backdoor, every copy of that AI inherits the problem. Your voice assistant might be fine today, but an update could push the vulnerability to everyone overnight.

The Fake CEO Call: All Four Methods Combined

A company executive got a call from his CEO. Urgent matter. Need to transfer money immediately. The voice sounded exactly right: same accent, same tone, same speaking style. The executive sent the money.

The CEO never made that call. AI cloned his voice from YouTube videos and generated the conversation in real time. This happened in 2019. The technology is far better now.

This attack used adversarial inputs (synthetic voice designed to fool recognition systems) combined with model inversion techniques (analyzing how voice AI responds to craft convincing fakes). As deepfake technology improves through data poisoning of detection systems and potential backdoors in voice processing AI, these attacks become harder to detect.

You can clone someone's voice from a few seconds of audio. You can generate fake videos on a laptop. When money is involved, you can't trust what you hear or see anymore.

What You Can Do

Verify financial requests through multiple channels. Boss calls asking for money? Text them. Email them. Walk to their office. Don't rely on voice alone.

Use multi-factor authentication everywhere. Biometrics aren't enough when AI can fake voices and faces. Combine password plus phone plus fingerprint.

Question unusual AI behavior. When your voice assistant does something weird, when your bank's AI makes a strange decision, consider manipulation rather than malfunction.

Understand what AI can access. Voice assistants that can order products need payment information. AI customer service can view your account. Banking AI can transfer funds. Minimize what you share.

Review privacy settings quarterly. AI companies regularly update their systems. Disable features you don't use.

The Bottom Line

Most companies don't realize this yet. They apply old security measures to new problems. They assume their AI is protected because their network is protected. It isn't.

The attacks are already happening. The defenses are still being figured out. And your AI assistant doesn't know it's being fooled.

These threats are documented in detail in Mountain Theory's white paper "Emerging Threats to Artificial Intelligence Systems and Gaps in Current Security Measures" by Michael May and Shaun Cuttill. The research analyzes real world incidents and identifies critical gaps in current security frameworks that leave AI systems vulnerable to manipulation. Read the full white paper at https://mountaintheory.ai/emerging-threats-to-artificial-intelligence-systems-and-gaps-in-current-security-measures/

Pushing the Limits with AI-Integrated Online Engineering Courses

Online engineering courses have spent two decades trying to prove they could match classroom instruction. Personally and based on my experience - when built the right way - I know they can. But now.... AI integration forces a harder question: can they exceed it?

The traditional model relies on static content delivery. Students watch recorded lectures, complete assignments, and wait for feedback. AI changes the timeline. Students get immediate responses to questions, instant code reviews, and real-time troubleshooting assistance. The delay between confusion and clarity shrinks from days to seconds.

Consider circuit analysis. A student builds a simulation, gets unexpected results, and stops. Previously, they posted to a forum or waited for office hours. Now they describe the problem to an AI assistant that walks through their schematic, identifies the error, and explains why the voltage divider calculation failed. The learning happens in the moment of need, not after the moment passes.

This shifts the instructor role. You become the designer of AI-assisted learning experiences rather than the primary content source. Your expertise matters more, not less. You create the problems AI helps students solve. You build the scaffolding AI uses to guide discovery. You intervene when AI explanations miss the mark or when students need human judgment about design tradeoffs.

The data tells you things classrooms never could. Which concepts cause repeated AI queries? Where do students get stuck despite AI assistance? What questions reveal deeper misunderstandings? You see learning patterns across entire cohorts in real time.

Personalization becomes practical at scale. AI adapts problem difficulty based on student performance. It recognizes when someone needs a simpler explanation or a more complex challenge. It suggests prerequisite reviews when knowledge gaps appear. Each student gets a version of the course tuned to their current understanding.

Assessment changes fundamentally. Take-home exams become meaningless when students can query AI for solutions. You need problems that require synthesis, judgment, and creativity. Design challenges with multiple valid approaches. Optimization tasks where students must justify their choices. Projects that integrate concepts across the curriculum. AI becomes a tool students must learn to use effectively, like MATLAB or CAD software.

The limits matter. AI makes factual errors. It generates plausible-sounding nonsense. It cannot replace hands-on lab experience or teach professional judgment. Students need to know when AI helps and when it hinders. That metacognitive skill becomes part of the curriculum.

Cost drops while quality rises. You eliminate textbook expenses. Students access powerful tools without licensing fees. AI handles routine questions while you focus on complex guidance.

The technology moves faster than accreditation. ABET criteria assume traditional delivery models. Program reviews ask about contact hours and lab facilities. You need documentation showing that AI-assisted online courses meet outcome requirements. Early adopters provide the evidence later programs will need.

Engineering education has spent decades moving online. AI integration represents the next boundary. Courses that use it well will outperform traditional formats on learning outcomes, student satisfaction, and cost efficiency. The question is not whether to integrate AI, but how quickly you can do it effectively.

The limits are being pushed. Some will break.

What I Tell Students About AI and Their Careers

Yesterday I wrote about my experience meeting with first year engineering students at Wentworth Institute of Technology. Here’s a bit more on what was discussed. 

AI always comes up. Whenever I talk to students - AI - it's something we are all concerned about. We worry about career security, whether we will be replaced by a bot, how fast it is moving. I have my own opinions on this from both a career perspective and a classroom perspective. Here's my take.

 

AI is now standard equipment. You'll use AI to draft reports, run design iterations, and analyze data. These tools are already normal day to day for many engineers. I like to compare AI to CAD. In the 1980s, senior engineers worried that CAD would eliminate drafting jobs. It did. But it created more engineering jobs because projects got cheaper and faster. Engineers who learned AutoCAD (arguably the first CAD program to gain widespread adoption is the 1980s) early had an advantage over those who clung to drafting tables. The same pattern applies now.

 

Computers can't make decisions that matter. AI suggests solutions based on parameters you provide. You decide which parameters count. You recognize when outputs look correct but fail in reality. You know when code conflicts with physics. You take responsibility when designs fail. Software doesn't do that.

 

Hands-on work resists automation. Site inspections, equipment troubleshooting, and field verification require presence. Civil, mechanical, and construction engineering involve messy reality. Sensors lie. Materials behave unpredictably. You verify assumptions with your hands and eyes.

 

Communication grows more valuable. Clients need translation between technical reality and business needs. Regulators need convincing. Teams need coordination. AI generates text; you read people and adjust strategy accordingly.

 

Choose your specialization carefully. Deep technical knowledge in stable domains (structural analysis, thermodynamics, electromagnetics) pairs well with computational tools. You provide expertise; computers handle calculations. Broad systems thinking also works. You connect disciplines; software optimizes within constraints you define.

 

Skills that will protect your career with reference to AI:

·       Learning new tools quickly

·       Critical evaluation of automated outputs

·       Client and stakeholder management

·       Hands-on troubleshooting

·       Ethical decision making

·       Cross-disciplinary thinking

 

Many engineering faculty disagree on this topic. Now for the fun stuff. Some professors ban AI tools in their courses. They have concerns about academic integrity and skill development. I understand the concerns and have the same ones, but this stuff is not going away. Employers today expect engineers to use AI tools day one, so I've shifted (and continue to shift) my courses to incorporate AI into assignments and labs. Students need practice evaluating AI outputs and knowing when to trust them. Learning these skills in school sure beats learning them under deadline pressure at a first job.

 

Advice to students. Use AI tools in your coursework when allowed. Learn their limitations through experience. Discover where they fail. If a professor bans AI, respect that rule. But seek out courses that teach you to work with these tools effectively.

Your generation will work alongside AI throughout your careers. Good engineers get better with better tools.

How 6G Will Improve on 5G in the Same Spectrum

A couple years ago, I spent some time developing 5G wireless technology content for faculty to use in their classrooms. Here comes the next generation.... 6G…. and (of course) AI is playing a major role in network management.

The transition from 5G to 6G represents more than just a numerical increment—it's a
fundamental rethinking of wireless network design. Many people believe new spectrum bands are needed for meaningful improvements between generations but the situation is more complex. The radio waves used by 5G in its sub-6 GHz and mmWave 24-40 GHz bands can be utilized more efficiently by 6G technology to achieve significant performance improvements. These enhancements would represent fundamental changes to network reliability and capacity and intelligence rather than minor adjustments without needing costly spectrum license purchases. The innovations would emerge from rethinking both signal transmission physics and network management intelligence. 

Advanced MIMO Systems: 6G technology could implement massive arrays with 1,000+ elements to generate extremely precise beams which reduce interference and boost capacity beyond the 5G maximum of 64-128 antenna elements. 

Smarter Waveforms: The waveform technology in 6G would surpass OFDM by implementing adaptive waveforms which modify their patterns according to environmental conditions. The system functions like an automobile which adjusts its body shape to achieve better aerodynamics during specific situations. 

AI Network Management: Like recent advancement in 5G, 6G networks will employ AI to forecast user activities and data requirements so they can distribute resources before users initiate their requests. 

Cell-Free Architecture: Your device would establish simultaneous connections with multiple transmission points which work together to provide seamless coverage throughout the network. The 6G radio technology functions as an environmental sensor to help the network optimize signal paths through its ability to detect physical obstacles and movement patterns. 

When We'll See It: Based on historical wireless technology evolution patterns and current industry roadmaps, commercial 6G networks are likely to launch around 2030-2032, following a progression similar to previous generations:

·      Initial technical requirements and vision documents are already being developed (2023-2025)

·      Research and standardization work will accelerate through 2026-2028

·      The first official 6G specification (3GPP Release 20 or 21) is expected around 2028-2029

·      Early trial deployments would follow in 2029-2030

·      Commercial availability would begin in leading markets by 2030-2032

Several major telecommunications companies including NTT DoCoMo, Samsung, and Huawei have published 6G whitepapers targeting 2030 for initial deployment. Various international research initiatives like the EU's Hexa-X project and China's national 6G promotion group are working toward this timeline.

6G's improved features would achieve 2-3 times better efficiency without spectrum expansion but the largest benefits would emerge from incorporating additional high-frequency bands.