The Stack Is Becoming the Product
15 October 2026
Dear readers,
Last week, I suggested that the interesting action in technology was moving outward from individual AI models toward the larger systems around them.
Well, apparently technology companies were listening to me😊.
Not really, of course. But this week provided rather more evidence for the same idea.
Across semiconductors, scientific research and even quantum computing, one theme kept appearing: Having a powerful component is increasingly not enough (if it ever was, that is a different question!). Anyway, what matters is how that component fits into a larger system.
A new AI chip needs an entire computing infrastructure around it. A semiconductor breakthrough has to survive the painful (but technically super interesting) journey from laboratory idea to high-volume manufacturing. An AI scientific assistant becomes much more interesting when it can interact with real laboratory instruments. And in quantum computing, building better individual qubits may eventually matter less than connecting them reliably (to be honest, this has been the case for a long time, but we will eventually dive deeper into quantum computing at a later stage so, enough said for now).
In other words, the technology stack itself is becoming part of the innovation.
That is worth paying attention to.
1. Microsoft Is Turning Custom AI Silicon into Infrastructure, Not Just a Chip
Microsoft announced on 12 October that its Azure Maia family will move toward multi-gigawatt deployment beginning in 2027, with Maia intended to become a primary inference system across Microsoft services. Maia 200 is already handling production workloads, while Maia 300 is planned for 2027.
There are two ways to look at this.
The obvious one is that hyperscalers want more control over their AI chips rather than relying completely on external accelerator suppliers.
That is true, but I think the bigger story may be architectural. At this scale, designing a processor is only one part of the job. The chip has to work with memory, networking, software frameworks, data centers, cooling systems and enormous amounts of electricity.
Once deployment reaches gigawatts, the meaningful unit is no longer really the processor. Rather, it is the AI factory.
This is an important change in the competitive landscape. In traditional computing, companies could often specialize in one layer of the stack. Increasingly, large AI operators have an incentive to optimize several layers simultaneously:
Model design affects silicon, which affects memory and networking. Those, in turn, affect rack architecture. And rack architecture affects cooling and electricity.
And, as we keep discovering: Build this one and that one becomes the next bottleneck😊.
The AI hardware race therefore looks less and less like a simple contest over whose chip can perform the most calculations and is becoming a systems-engineering contest.
Looks like everyone now wants to play the “Elon Musk game”!
2. Semiconductor Innovation Is Becoming More Collaborative, as Well we More Software-Driven
Two announcements this week illustrated another fascinating transition:
Applied Materials opened its new $5 billion EPIC Center in Silicon Valley on 12 October, designed to bring chipmakers, equipment manufacturers, system companies and researchers together to develop semiconductor manufacturing technologies in parallel rather than sequentially.
Then, on 14 October, NVIDIA announced new tools and collaborations aimed at what it calls AI-defined fabs, using accelerated computing and AI for materials simulation, manufacturing simulation and chip inspection. Partners include Applied Materials, Cadence, KLA, Lam Research, Siemens and Synopsys.
Now, semiconductor manufacturing has always been astonishingly complicated. Nothing new there. But the development model itself is changing.
Historically, one company might develop a material, another would design equipment around it, another would integrate the process, and eventually somebody would discover that the whole thing did not quite behave the way everyone hoped (been there).
At atomic dimensions, that sequential approach becomes painfully slow. So, the industry is trying to shorten the loop:
Design → Simulate → Manufacture → Measure → Learn → Redesign
And AI is starting to enter several stages of that loop. This is significant because one of the least glamorous but most important questions in technology is “Can somebody manufacture this repeatedly, economically and with acceptable yield?”. In other words, a key question is “Can someone do this at scale, and do it economically?”.
Semiconductor progress increasingly depends not only on better transistor ideas but on shortening the learning cycle that turns those ideas into reliable manufacturing processes.
And it is fair to claim that this may prove just as important as the transistor architecture itself.
3. Autonomous Science Is Beginning to Look Like a Closed Learning System
Danaher and NVIDIA announced a collaboration on 14 October to combine AI with laboratory instruments, automation, scientific workflows and experimental data. Their stated goal is to create a continuous learning cycle in which models help determine what experiments should be performed and experimental results then improve the models.
Now I find this one particularly interesting as we often talk, rather loosely, about “AI for science,” but that phrase covers several very different things to the point of losing its true meaning. For example, I do not doubt that building an AI system that reads and summarizes scientific papers is useful. I can certainly use such summaries (this is not just theory, I actually do that quite often) but, all things considered, I think a system that actually proposes hypotheses is even more interesting (oh, there are a number of ways to do this. The language I am developing, Chilán, tries to do just that, in a slightly different way but that is another topic, for another day…).
Anyway, an AI system that proposes a hypothesis, interacts with laboratory equipment, collects experimental results, learns from them and chooses the next experiment begins to look like something qualitatively different. This means that we can now see the learning loop resembling the following (interestingly enough, this is strikingly similar to what I described to my students in the Neural Network and SVM class 2 days ago, when I introduced the learning cycle in the broader context of Machine Learning):
Question → Model → Prediction → Experiment → Measurement → Updated Model → Better Question
Scientific research has always worked through feedback loops, of course. That is certainly not new. So, there is no claim that we have suddenly invented the scientific method.
However, what might change soon is the speed and scale of the loop. If AI can help choose the most informative experiment rather than merely analyze results afterwards, then the scarce resource may gradually shift from human analytical time toward laboratory capacity, reliable instruments, high-quality biological data and experimental validation.
There is, however, an important warning: A machine-generated hypothesis is not a scientific discovery! And neither is a simulated result the same as an experimental result.
And let’s not forget that a promising biological mechanism is certainly not a medicine.
Th bottom line is that this is one of those cases where my conclusion remains as: Very interesting. But show me the experimental evidence first, dude.
4. Quantum Computing Is Starting to Encounter Its Networking Problem
On 9 October, IonQ reported an experimental quantum interconnect linking a trapped-ion qubit with a solid-state quantum memory through photons, achieving more than 1,000 entanglement events per second. The company describes the result as a step toward distributed quantum computing.
We should be careful here because this is a company-reported technical milestone, and quantum computing remains particularly good at producing impressive numbers whose ultimate economic meaning is not immediately obvious😊. Still, the underlying problem is important.
If useful quantum computers eventually require very large numbers of qubits, endlessly enlarging a single processor may not be the only (or even the best) way to get there.
Classical computing learned this lesson long ago. We did not build one unimaginably enormous processor. Instead, we learned how to connect processors, memory and networks into larger systems.
Quantum computing may ultimately require something similar:
Quantum Processor → Quantum Memory → Quantum Interconnect → Distributed Quantum System
If so, interconnect performance becomes another potential bottleneck (and you know what they say, a bottleneck is an opportunity yet to be explored!). Again, this does not mean IonQ has solved distributed quantum computing. Nope! But it does tell us something useful about where engineering attention is moving.
I have been watching this space for a while now and it does seem that the quantum question is slowly expanding from “how good is your qubit?“ to “how do you build an actual computer out of lots of them?“. That sounds like tangible and exciting progress to me.
The Connection: We Are Moving from Components to Systems
Taken separately, these developments concern custom processors, semiconductor manufacturing, scientific laboratories and quantum networking. However, taken together, they point toward the same underlying transition:
For much of the current technology boom, attention has focused on individual capabilities, be it a smarter model, or a faster chip, or a better qubit, or even a more accurate biological prediction.
But real-world value appears when those individual components become part of reliable systems. So perhaps the more useful technology stack now looks something like the following:
Capability → Integration → Control → Deployment → Measurement → Feedback
In our examples:
- Microsoft is integrating silicon into enormous computing systems,
- Applied Materials and NVIDIA are tightening the feedback loop between chip design and manufacturing,
- Danaher and NVIDIA are trying to connect AI reasoning directly to experimental science, and
- IonQ is confronting the problem of connecting quantum components into larger machines.
Different technologies, to be sure, but similar problem!
What I Am Watching Next
First, whether hyperscaler custom silicon materially reduces the cost of AI inference or simply moves the bottleneck somewhere else in the infrastructure. This is one of those things that is easier said than done!
Second, whether autonomous-science platforms begin producing independently validated discoveries rather than merely impressive demonstrations of research automation.
At the end of the day, verification rules!
Sources & Further Reading
- Microsoft: Azure Maia multi-gigawatt deployment. Microsoft: Azure Maia expansion
- Applied Materials: EPIC Center. Applied Materials: $5 billion EPIC Center
- NVIDIA: AI-defined semiconductor fabs. NVIDIA: AI-defined fabs
- Danaher: Collaboration with NVIDIA on autonomous science. Danaher: Autonomous Science collaboration
- IonQ-Quantum memory-enhanced interconnect. IonQ: Quantum interconnect milestone
