Quantum from the Ground Up: First Edition Now Available

I have been learning and writing about quantum computing on gordostuff.com since October 2025. The posts started as a way to make sense of announcements as they happened: IonQ hitting 99.99% two-qubit gate fidelity, Microsoft unveiling Majorana 2, IBM running a 12,635-atom protein simulation, Apple releasing a 50,000-step formal proof of its post-quantum cryptographic library. Each one went up as a standalone post. After eight months and seventeen posts, it made more sense to put them together in one place.

The result is 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.

Download: Download PDF

The book covers six qubit fabrication platforms in sequence: superconducting circuits, trapped ions, photonics, neutral atoms, silicon spin qubits, and topological qubits. Each chapter describes the physical mechanism, the fabrication or trapping method, current performance numbers, and the specific engineering problems that remain unsolved. The qubit chapters are followed by sections on hybrid classical-quantum protein simulation, AI-assisted hardware calibration, post-quantum cryptography, and the hardware landscape as it stands in mid-2026.

Two chapters address the workforce directly. The first maps the adjacent skill sets the quantum supply chain actually needs: semiconductor process engineering, cryogenic systems operation, RF electronics, machine learning for hardware calibration, and cryptographic implementation. None of those require a physics PhD. The second chapter maps the full degree pathway from a two-year associate degree through a PhD, with specific programs, what each credential prepares you to do, and salary ranges at each level. The field has a real workforce shortage. The shortage is not only at the PhD tier.

The plan is to update the book quarterly as new posts are published. The field moves fast enough that a quarterly cycle makes sense: slow enough to let any given announcement settle, fast enough to stay current with actual engineering progress. This edition covers October 2025 through June 2026. The next update will incorporate posts from Q3 2026.

Seventeen posts, fifty pages, fifteen images, one dark navy cover. The posts were worth writing individually. They are more useful together. Download free at Download PDF.

Support This Work

The book is free and will stay free. If you find it useful and want to support future editions, you can contribute at ko-fi.com/gordostuff.

Beaming Power From a Cessna

Back in 2023 I remember reading this writeup on Caltech's Space Solar Power Demonstrator. They beamed a small amount of microwave power between panels on a satellite in LEO. Neat physics, but I have heard the space-based solar pitch for thirty years and the engineering never closes. Launch mass, spectrum rights, pointing accuracy, regulatory approval for a multi-megawatt beam over populated areas. Something in that list always kills it.

Then I read this Spectrum article about Overview Energy. They flew a Cessna turboprop over Pennsylvania at 5,000 meters in 70-knot crosswinds and held a power beam on a ground receiver the whole time. Watts, not kilowatts, but the first time anyone has done it from a moving platform at altitude. They beamed near-infrared light, not microwaves. That choice is the part worth looking at.

Microwaves have two problems for ground power delivery. The good bands between 2 and 20 GHz are already spoken for. 5G, GPS, satellite links, weather radar, military. You will not get a license to transmit megawatts in any of them. The second problem is beam spread. Diffraction limits how tight you can focus, and at microwave wavelengths the spot on the ground is kilometers across even from a big aperture. That means a rectenna farm the size of a small city.

Optical wavelengths change the math. Same aperture, beam spot drops from kilometers to meters at GEO range. No spectrum fight either, since power transmission at optical frequencies is not allocated the way RF is. The receiver question gets interesting too. Overview wants to drop the IR onto existing utility-scale PV farms. Silicon is not tuned for a single IR line so you lose efficiency at the receiver, but you also skip building a new class of ground infrastructure and skip the permitting fight that comes with it. Whether the efficiency hit is worth the deployment savings depends on numbers we have not seen yet.

Now the scaling, which is what every engineer reading this already knows is the hard part. A working demo at the bench is one thing. A working demo a thousand or a million times bigger is a different machine, with different failure modes, different cost curves, and different second-order effects that did not exist at small scale. Heat dissipation that is trivial in a benchtop laser becomes the dominant design problem in a megawatt source. Vibration modes that did not matter in a fixed lab fixture wreck pointing accuracy on a 100-meter deployed structure. Manufacturing tolerances that were acceptable on one unit are not survivable across the thousands of components a flight article needs. Cost per watt that pencils at the prototype scale almost never holds when you push three or four orders of magnitude. This is the part of engineering that nobody outside the field appreciates and everyone inside it has been burned by.

Apply that to Overview. The Cessna run was watts. DARPA's July 2025 demo pushed 800 watts across 8.6 km for half a minute. Overview's roadmap calls for megawatts from GEO by 2030 and gigawatts later. Five to nine orders of magnitude. Continuous operation instead of a 30-second pulse. Pointing accuracy in the microradian range from 36,000 km up, on a satellite that heats and cools every orbit and flexes accordingly. The Cessna proves they can track a target from a moving platform. GEO needs the same trick done a thousand times better, with thermal management for a source running at a million times the optical power.

The spacecraft itself is the other hard part. To collect useful sunlight you need a big aperture. A big aperture does not fit in a rocket fairing, so it has to fold for launch and deploy in orbit. JWST showed that can work. JWST also showed how close that kind of mission comes to dying. Add debris survivability over a 20-year design life, station-keeping fuel budgets, and disposal at end of life, and you have a spacecraft program as risky as the beam.

So where does that leave me. Still skeptical, but less than I was. The 2023 Caltech work was a physics check. The Cessna flight is a systems check. Source, beam control, pointing, tracking, and a PV-style receiver, all running together on a moving platform in real weather. If Overview gets a LEO demonstrator up and lands a kilowatt on the ground from a folded-then-deployed aperture, people will be paying close attention.

What Jensen Huang Thinks Smart Looks Like Now

Years back I had a student who was terrible at exams. Multiple choice, timed problems, placement exams, all of it. He barely squeaked by in lecture. In lab he was someone else. He would stand next to a partner who was building a circuit and just watch. Then he would point at a cap and ask what happens if it fails open. He was usually right before the thing ever got powered up.

He got a field tech job after he graduated. Two years in, customers were calling and asking for him by name. He still cannot tell you the formula for a low-pass filter off the top of his head. He can tell you, on the phone, that the problem you think you have is not the one you actually have.

I thought about him this week reading about Jensen Huang on Jodi Shelton’s A Bit Personal Podcast. She asked him who the smartest person he has ever met is. He would not answer. He said the question itself is the wrong one now.

Huang’s point: the kind of smart we have always rewarded, the technical problem solver, is becoming a commodity. He used software programming as the example. For decades that was the smart-person job. Now it is the first thing AI is doing well.

So what does he think smart looks like going forward? Three things stacked. Technical understanding, human empathy, and the ability to pick up on what nobody is saying. That last one he called seeing around corners. It comes from data, analysis, first principles, life experience, wisdom, and reading the people in the room. People who have that mix tend to spot problems before they happen.

This is part of why I started doing oral exams in my Circuits class. A written test will tell me a student got the right number. It will not tell me whether they know why they used mesh instead of nodal, what a time constant actually means, or where the sign error came from. Sit them down and make them talk through it and you find out fast. 

Most of what engineering school grades on is the part AI is taking over. Timed problem sets. Multiple choice. Take-home work that gets judged on how polished it looks. The stuff Huang is talking about, reading a situation, asking the question nobody else thought to ask, knowing something is off before the numbers say so, is harder to grade. It is also what is going to matter.

My old student would have bombed every test Huang is calling outdated. He read a lab bench the way some people read a chessboard. Twenty years ago that made him a good tech. Today it makes him the kind of engineer you cannot swap out for a chatbot.

The Oral Exam Experiment Worked

Last fall I posted that I was dropping homework from the grade book and adding an oral portion to every exam in the spring. Students were running Engineering homework problems through Gemini and handing in solutions they could not explain when I asked. I was grading a chatbot, so I stopped.

Spring semester is over. It worked. The oral portion runs about ten minutes per student. They pick one problem from their written work and walk me through it. Why mesh and not nodal. What the time constant tells you about the circuit. Where the negative sign came from. I learn more about what a student actually understands in that ten minutes than I used to learn from a semester of graded homework.

A new Lumina Foundation-Gallup study says 57% of US college students use AI in their coursework at least weekly, and one in five use it every day. At the same time, 53% say their school discourages or prohibits it. Daily use is highest among men and among business, tech, and engineering students. The students avoiding AI mostly cite ethical concerns and school policy, so the ones following the rules are falling behind on a tool they will use the rest of their careers.

My position on AI in education is simple. If we are preparing students for the jobs they are about to take, AI has to be in every class. Every engineering job they walk into will expect them to use these tools well. A program that prohibits AI is training students for a job market that no longer exists. The work is not to keep AI out of the classroom. The work is to teach students how to use it, where it fails, and when to check it against first principles.

That still leaves the assessment problem. If students use AI on everything, how do you know what they understand? You change how you measure them. Oral exams catch what written work cannot. In-class paper problems catch it. Hands-on labs, where a student wires a circuit on a breadboard, takes scope measurements, and explains what they are seeing, catch it cold. Take-home essays graded on polish do not catch anything anymore.

The AI can solve the circuit. It cannot explain why this student chose the loop they chose, and it cannot wire the breadboard when the lab is due at five. That is what we should be assessing, and that is the work employers are hiring for.

The Wrong Engineering School Is Closing

Hampshire College is closing down. I live a few miles from the campus in Amherst, Massachusetts. It is a beautiful piece of real estate, and watching it go dark will sure be a waste. So here is a better use for it: reopen it as an engineering school built on the same pedagogical model that made Hampshire unusual for 56 years.

Hampshire ran on a simple premise. Students learn more when they direct their own education. No required courses, no grades, narrative evaluations instead of transcripts. It produced Ken Burns, among a long list of graduates who learned to frame problems from scratch. Then it ran out of money. At $29,683 per semester, with 625 students and a $26.5 million endowment, the finances were never going to hold.

The finances failed. The pedagogy did not.

Engineering education runs on the opposite assumption. Students cannot be trusted to direct their own learning until they survive two years of calculus sequences, physics surveys, and weed-out courses designed less to teach than to filter. ABET accreditation requires programs to demonstrate prescribed curricular coverage across a fixed sequence, and schools optimize for compliance. Roughly half of engineering students switch majors or drop out, most citing the culture of years one and two. The students who make it through are good at structured problem sets. Whether they can identify a problem worth solving is a separate question the curriculum rarely addresses. The NAE's analysis of engineering education and workforce pathways has long flagged a gap between what programs produce and what employers actually need: graduates with strong professional and problem-framing skills alongside technical ones.

A Hampshire-style engineering school would look different. Students identify a technical challenge in the first semester and spend four years building toward a solution, pulling in mathematics, circuits, materials, software, and fabrication as the work demands. Faculty serve as advisors. Credentialing comes from demonstrated competency, not accumulated credit hours. The model produces engineers who can operate without a defined rubric, which is the actual job description for most engineers after around year three. Olin College of Engineering in Needham, Massachusetts has run a version of this since 2002 and its graduates are consistently sought after. The model exists and it works. 

The timing matters. AI handles the structured problem set. Tools like Gemini and Claude write the code, run the simulation, and check the math. The irreplaceable skill is knowing what problem to run. Engineers who are excellent inside a defined problem space and lost outside it are the engineers whose value will compress fastest. Quantum, biotech, and climate tech all require graduates who can work across disciplines without a map. The standard curriculum produces the wrong graduate for all three.

The objections are real: accreditation, licensure, employer acceptance. All solvable, given enough will. The campus infrastructure is already there, sitting on several hundred acres in the Pioneer Valley. Hampshire's library, labs, and residential buildings are not going anywhere quickly. The harder problem is finding people with the will to try something genuinely different rather than build another school that looks exactly like the others.

Hampshire believed every student was capable of running their own education, even if they occasionally used that freedom poorly. Engineering education has never believed that. Maybe we should try.

44 Million Teachers: AI Can Save Time, But It Can’t Save the Profession

Throughout K-8 I had an excellent academic experience. Off the top of my head - Mrs. Hebert, Mrs. Elsden, Mrs. Halla, and Mr. Valliere, my first male teacher in fifth grade, followed by Mr Pasqualini, Mr Crean, Mr Bash, all ran their classrooms like the work mattered. My mother taught 8th grade English with the same conviction. Then came 9th grade.

My high school had lost its accreditation due primarily to over crowding. The city responded with double sessions while a new school was being built. Eleventh and twelfth grades ran from around 7 am to noon, ninth and tenth ran from around 1 pm to 6 pm. The teachers worked hard, and looking back I can see how much effort they put in. But half the students had mentally left the building before they walked in the door. Teaching into that kind of indifference is exhausting in a way that effort alone cannot fix. It sucked. By the time we moved to the newly constructed high school for 11th grade, I had pretty much checked out. I wanted to just get through it and move on to college. The light bulb didn’t go on until I got there, where I felt challenged and found out I could excel, just like I had in grades K-8.

I’ve spent over forty years on the other side of that equation, teaching courses like circuit analysis, photonics, and robotics. So when I read Ben Gomes’s recent interview in Forbes, his argument sure made a lot of sense. Gomes is Google’s Chief Technologist for Learning and Sustainability, and he spent 21 years building Google Search. His point: the biggest problem in education is motivation, and AI cannot solve it. High-achieving people are almost never unlocked by an algorithm. They are unlocked by a person, usually a teacher who made them feel the work mattered. Once that happens, tools can accelerate everything. Without it, nothing moves.

What those double-session years showed me is that teacher motivation and student motivation are not separate problems. The teachers were trying. The system had stripped away a lot of the conditions that make student motivation possible, and no amount of individual effort fully compensates for that. Faculty working hard into a wall of disengagement will not hold together indefinitely. That is not an argument against AI tools. It is an argument for taking teacher retention seriously as the central issue.

A six-month pilot with Northern Ireland’s Education Authority found teachers using Google’s AI tools saved an average of 10 hours per week. Google has committed $50 million in AI education grants and is building training materials for teachers across the United States. Those are real gains. But Gomes frames the stakes correctly: there is a projected worldwide shortage of 44 million teachers. That gap exists because the profession burns people out before they finish a career. If AI recovers enough time to make the job sustainable, it addresses the shortage at the source. If that recovered time just gets absorbed by more administrative load, nothing changes.

Gomes also makes a point about what education should teach as AI handles more of the mechanics. Programming syntax matters less. Conceptual thinking matters more. How do you decompose a problem? How do you think about abstraction? Those questions don’t disappear because a tool can write the code. In engineering education I see this directly. The students who do well aren’t the ones who memorized the most procedures. They’re the ones who understand why the procedures work.

Mr. Valliere, my Mom and the rest didn’t teach me content I can still recite. They taught me that learning was worth doing. Two years in a broken system came close to erasing that, despite the genuine effort of the people in front of the room. College restored it. The difference, every time, was whether the conditions existed for learning to take hold. AI tools that give teachers back their time and energy are worth every dollar. The goal should not be a faster classroom. It should be keeping the people in it.

Google Quantum AI Expands to Neutral Atoms. I’ve Covered Both Platforms. Here’s the Context.

Source: Building superconducting and neutral atom quantum computers | Google Quantum AI | March 24, 2026

Google Quantum AI announced last week that it is adding neutral atom computing to its existing superconducting program. I covered both platforms in my qubit series earlier this year, so this is a good moment to connect the dots.

I wrote about superconducting qubits in January and neutral atom qubits in February. Google’s announcement is a practical illustration of why both matter, and why no single platform has won.

Superconducting systems have scaled to circuits with millions of gate and measurement cycles, each running in about a microsecond. Neutral atom arrays have reached roughly 10,000 qubits, but their cycle times run in milliseconds. The tradeoff breaks cleanly along two axes: superconducting scales more readily in circuit depth; neutral atoms scale more readily in qubit count. Google is betting that running both in parallel gets to commercially useful hardware faster than doubling down on one.

Key tradeoffs between superconducting and neutral atom qubit platforms.

To lead the neutral atom effort, Google hired Dr. Adam Kaufman from JILA and NIST in Boulder, Colorado. He retains his CU Boulder faculty appointment. Boulder is a credible home for this work: it hosts the NSF Q-SEnSE Institute, the National Quantum Nanofab, and the U.S. EDA Quantum TechHub. Google also noted its continued collaboration with QuEra, the neutral atom startup whose researchers built much of the foundational methodology.

Google also said it expects commercially relevant quantum computers based on superconducting technology by the end of this decade. Adding neutral atoms to the portfolio is not a hedge on that timeline; it is a way to cover problem types that each architecture handles differently.

If you want the technical foundation for what Google announced, my January and February posts are a great place to start.

Writing Is How I Learn

Writing is how I learn. Not a side effect of learning, the mechanism itself. It's only taken me 60 years or so to realize. Yeah it took me a little longer than it should have to figure out, partly because my mother was an English teacher. Growing up with that in the house, writing felt like an assignment, something to be graded and corrected. I guess I avoided it for years.

Now I cannot imagine working without it.

The habit really started in college. I was never a yellow highlighter. I took detailed notes in class, then went back and rewrote them, filling in gaps, looking up anything I had not fully understood. I was not studying. I was writing my way to comprehension. I just did not recognize it as writing at the time.

Looking back at over 800 blog posts since 2005, I cannot identify a single topic I learned in school, at least not directly. Along the way I wrote five textbooks. Each one forced the same process at a longer scale: find the gaps, trace the logic, write until it holds. What came after was anything but familiar. But the foundation school set has mattered more than I appreciated at the time.

Circuit analysis taught me how to trace cause and effect through a system. Signals and systems gave me a way to think about information in motion. Mathematics gave me a tolerance for abstraction. Physics gave me an instinct for what is physically possible and what is not. None of those subjects mapped directly to telecommunications, networking, cybersecurity, AI, or quantum computing. But without them, I would have had nothing to connect the new ideas to. Every emerging technology I have written about made more sense because of something I learned in a classroom decades earlier, even when the connection was not obvious at first.

My first real test was the transition from the plain old telephone service network to internet protocol. POTS was settled territory: copper pairs, circuit switching, predictable behavior. IP emerged as none of those things. The protocols were still being written, the standards contested, and the people producing the documentation were often the same people building the systems. There was no textbook that had caught up. Writing about it forced me to work through the logic myself, trace the signal path, understand why packet switching broke the assumptions that circuit switching had held for a century. Reading alone would not have gotten me there.

Each technology that followed emerged the same way. Networking protocols solidified and cybersecurity emerged alongside them, then ahead of them. AI emerged from research labs into practice before most organizations knew what to do with it. Quantum computing is emerging now, still settling on its own vocabulary, its own benchmarks, its own honest assessment of what it can and cannot do. I came to each as an outsider working from preprints, conference proceedings, vendor white papers, and conversations with people who were themselves still figuring it out.

That turned out to be the ideal condition for writing. When there is no established explanation to defer to, you have to build your own. For me, the act of building it is where the learning happens.

Every draft reveals what I actually understand and what I have been skimming over. Missing causal links surface immediately. So do circular definitions I had mistaken for insight. Quantum is no different. I understood qubits loosely until I wrote a series on qubit technologies for a general audience. Six posts in, I understood the field in a way that a month of reading had not produced.

My mother would note that I came around eventually :)

This is pretty much how I learn. Researchers call it elaborative encoding: building understanding by reconstructing information in your own words rather than receiving it passively. The VARK model would place me in the read/write category, though writing as a cognitive tool goes beyond simple preference. It is the mechanism by which I connect new ideas to existing foundations.

Not everyone learns this way, and that is a really important point. Some people think through conversation. Others need to build something physical, draw a diagram, or hear an explanation out loud before it clicks. The method matters less than finding the one that works for you and committing to it. For me, writing is that method. It always has been, even when I did not want to admit it.

Writing also serves readers who learn differently. A well-constructed post gives the visual learner a structure to follow, the read/write learner direct access to the logic, and the reflective learner something to push back against. I write to learn, but if the post does its job, someone with a completely different learning style picks something useful out of it too. That possibility has kept me writing for 21 years.

If you work in a technical field and you are not writing, start. Pick something you half-understand and write until you fully do.

This past week that came full circle. Someone at the Quantum Supply Chain Accelerator site walkthrough at STCC Technology Park mentioned my quantum computing posts. It was an encouragement to keep writing.

AI Runs on GPUs - It Stalls on Pipes

Based on: Rebuilding The Foundation: Why AI Infrastructure Needs To Change  |  Will Eatherton, Cisco  |  March 17, 2026

Every AI infrastructure conversation eventually lands on power. Megawatts per rack, cooling costs, grid capacity, carbon footprint. The power story is real, but we can’t forget about bandwidth.

I’m an old telecom guy, having spent over 30 years in telecom and networking, including running National Science Foundation Centers of Excellence focused on communications infrastructure. Bandwidth constraints are not new. What's new is the scale and the speed at which AI workloads are exposing them.

A 200,000-GPU (Graphics Processing Unit) cluster can consume 435 MW (Megawatts) of critical IT power. Of that, 17 MW goes to optical transceivers alone, just to move data between chips. When you scale to a million GPUs, the transceivers by themselves consume roughly 180 MW. That's not a power problem. That's a data movement problem that shows up on the power bill.

Cisco's Will Eatherton made the case this week that the real bottleneck in AI infrastructure has shifted from compute to data movement. GPU procurement still dominates the conversation. Networking, storage, and security are where the constraints are actually forming.

Bandwidth

Training large models requires clusters of tens of thousands of GPUs exchanging data continuously. The industry has settled on 102.4 Tbps (Terabits per second) switching silicon as the baseline for serious deployments. Traditional pluggable transceivers hit a wall at 800G and 1.6T speeds. The DSP (Digital Signal Processor) in each transceiver consumes up to 30W per port; at 200G channels, electrical loss reaches roughly 22 dB (decibels) before the signal reaches fiber. Two approaches address this.

Linear-drive Pluggable Optics (LPO) removes the DSP and lets the host ASIC (Application-Specific Integrated Circuit) drive the optical module directly, cutting per-link power by up to 50%.

Co-Packaged Optics (CPO) goes further by integrating optical engines onto the switch package itself, dropping electrical loss to 4 dB and per-port power to 9W. CPO eliminates the transceiver and DSP entirely, embedding electronic-to-optical conversion onto the switch ASIC.

Nvidia's Quantum-X InfiniBand CPO switches, entering production in 2026, deliver 115 Tbps across 144 ports at 800G. Broadcom's Tomahawk 6 (TH6-Davisson) ships 102.4 Tbps with full CPO. IDTechEx projects the CPO market will grow at a 37% CAGR (Compound Annual Growth Rate), exceeding $20 billion by 2036.

Topology

Scale-up (NVLink within a rack) and scale-out (InfiniBand or Ethernet across a data center) are both approaching practical limits. The next phase, scale-across, federates compute across geographically distributed locations into a single pool. Telecom engineers solved a version of this problem decades ago with distributed switching and ATM (Asynchronous Transfer Mode) traffic engineering. AI adds a harder constraint: gradient synchronization across a WAN (Wide Area Network) requires low, symmetric latency that wide-area networks were not built to guarantee. It breaks the latency-symmetry assumption in standard collective communication libraries such as NCCL (Nvidia Collective Communications Library), requiring deep-buffer routing, topology-aware all-reduce algorithms, and control planes that make traffic decisions based on path characteristics, not just throughput.

Nvidia's Spectrum-X Ethernet CPO platform targets this scale-across problem, combining switching and routing in a single solution with deep buffer support and integrated in-network computing via SHARP (Scalable Hierarchical Aggregation and Reduction Protocol).

Figure 1. Scale-across WAN topology: two GPU clusters connected via deep-buffer routers across a WAN, with shared DPU/SmartNIC security enforcement.

Storage

AI training creates a mixed-access pattern: large sequential reads across petabytes of training data, burst checkpoint writes during fault recovery, and sustained KV-cache (Key-Value cache) write pressure as context windows grow. RDMA-based (Remote Direct Memory Access) protocols, including RoCE (RDMA over Converged Ethernet) and NVMe-oF (NVM Express over Fabrics), cut storage latency from milliseconds to microseconds. Idle GPUs cost the same as active ones. When ingestion starves GPUs of data or checkpoint bursts block training progress, accelerator cycles go idle. Storage has to be designed into the architecture from the start, not bolted on.

Security

Model weights represent hundreds of millions of dollars in training cost. Protecting them requires hardware-based trust, confidential computing, and network segmentation. SmartNICs (Smart Network Interface Cards) and DPUs (Data Processing Units) now enforce zero-trust policy at line rate, isolated from the host OS (Operating System), handling IP filtering, session tracking, and rate limiting without CPU (Central Processing Unit) involvement. Multi-tenant inference clusters must maintain customer separation while meeting latency SLAs (Service Level Agreements), adding another layer of security complexity that traditional perimeter models were not designed to handle.

Organizations that get interconnect, storage, and security right will have capacity that GPU-focused competitors cannot replicate from a single cluster. Those that don't will rent infrastructure from the ones that do.

Source: Cisco Blogs: Rebuilding The Foundation — Why AI Infrastructure Needs To Change

CPO Market Data: IDTechEx — Co-Packaged Optics (CPO) 2026-2036

Nvidia CPO Technical Detail: Nvidia Developer Blog — Scaling AI Factories with Co-Packaged Optics

CPO Technology Overview: EDN — Where Co-Packaged Optics Technology Stands in 2026