Advanced node manufacturing and heterogeneous integration require partnerships that span the full value chain.
I had the privilege of joining a panel at the Global Semiconductor Alliance (GSA) Tech Summit in June in Scottsdale, Arizona, titled “Collaboration Models That Actually Work.”
It was a fitting title for an event that brought together executives from across the semiconductor ecosystem, including foundries, fabless companies, equipment makers, EDA vendors, cloud providers, and systems integrators. It’s especially relevant when the industry is being forced to confront the fundamental question of its increasing complexity. As recently noted, “Manufacturing complexity is rising faster than the industry’s ability to manage it. As architectures move deeper into 3D, production disperses globally, and product cycles compress, scale alone is no longer a differentiator. Operational coherence is.”
The “Collaboration Models That Actually Work” panel during the recent GSA Tech Summit included (from left to right): Danny Wilkerson, Managing Partner of Invictus EV Innovation Technology; Ming Zhang, Vice President of Fabless Solutions from PDF Solutions; and Rich Simoncic, Chief Operating Officer at Microchip Technology.
I came away with a sharper view of where the industry stands and, more importantly, where it needs to go. Here are the three things that stuck with me most.
For most of semiconductor history, collaboration has happened within layers. Design teams collaborated with other designers. Fabs shared process learnings with equipment vendors. Fabless companies coordinated with OSATs. These were deep, productive relationships, but they were vertical silos that rarely connected all the way through. In a recent SemiconWest keynote, PDF’s John Kibarian used the term “stage-gate and crisis driven collaboration” to describe this traditional model.
What I kept hearing at the GSA Summit, and what I believe strongly, is that the era of layer-by-layer collaboration is over. The physics of advanced node manufacturing and heterogeneous integration demand something different: full-stack collaboration.
Full-stack collaboration means that the companies designing chips need to deeply understand the manufacturing process, not just as a constraint but as a co-design space. It means that foundries need to work hand-in-glove with equipment makers on process development, not just on tool qualification. It means that fabless companies, fabs, OSATs, and equipment suppliers all need to share data, align on yield strategies, and make coordinated decisions across what used to be hard organizational boundaries.
This kind of collaboration requires expertise that spans both design and manufacturing. It requires partnerships that span the full value chain, not just bilateral agreements between adjacent players, but multi-party coordination across the entire supply chain. The companies that will win in the chiplet era at sub-2nm nodes, and in advanced packaging are the ones that figure out how to do this first.
Critically, that collaboration has to focus on value. Not process metrics for their own sake, not data exchange for the sake of openness, but measurable business outcomes: faster yield ramp, lower cost per die, faster time to market. Value focus is what separates collaboration that produces results from collaboration that produces meeting agendas.
One of the most resonant themes across the Summit’s panels was how much of the collaboration gap is fundamentally a data gap. Different parts of the supply chain are generating more data than ever before, from process sensors, metrology tools, test equipment, and design simulations. This data lives in incompatible formats behind organizational firewalls, and in systems that were never designed to talk to each other.
My conviction, reinforced by what I heard in Scottsdale, is this: data is the common language for collaboration across the full stack. It is the substrate on which every cross-organizational decision rests. You can have the best partnership intentions in the world, but if your process data and my yield data cannot be correlated because we are using different lot ID conventions, different metadata schemas, or different time-stamping conventions, you cannot close the loop between what happened in your fab and what I am seeing in my test results. Additionally, as Mike Campbell, senior vice president of engineering at Qualcomm, mentioned at the 2025 PDF Solutions Users Conference, we are now operating in a context where the amount of manufacturing data has grown by a factor of six since 2022.
The industry has made real progress on common data standards: SEMI standards, STDF (Standard Test Data Format), and newer AI-ready frameworks are all pushing toward interoperability. But adoption is uneven, and the technical challenge of aligning data models is often underestimated. In practice, data scientists in manufacturing environments spend as much as 80 percent of their time just cleaning and aligning data before they can run a single analysis, according to the Pragmatic Editorial Team in “Overcoming the 80/20 Rule in Data Science.” That is a staggering inefficiency, and it is entirely a solvable problem, but solving it requires the whole supply chain to agree on what a “wafer” is, what a “lot” is, and how to track them consistently from fab to test to customer.
Several speakers at the Summit touched on this, including the Deloitte and Applied Materials presentations on AI-driven manufacturing. The common thread: AI is only as good as the data it runs on. Getting the data right is not a precondition for AI deployment; it is the deployment.
There was a lot of AI talk at the GSA Summit, as you would expect. AI for chip design. AI for yield optimization. AI for scheduling fabs. AI for security. Agentic AI. Physical AI. What struck me most was not any specific application, but a more fundamental question that kept surfacing beneath the hype: what is the right relationship between humans and AI in this industry?
Here is how I think about it. The semiconductor manufacturing process is, at its core, a physics problem. Transistors are physical objects. Process variations are physical phenomena. Yield excursions have physical causes. At the same time, semiconductor companies are businesses. Every decision about where to run a lot, how to allocate tool capacity, when to pull a reticle, or whether to accept a wafer has economic consequences.
AI is uniquely positioned to be the bridge between physics and economics. It can process sensor data across thousands of process steps, identify physical anomalies that no human engineer could track manually, and translate those observations into economic recommendations: prioritize this lot, flag this tool, adjust this process parameter. That is genuinely new capability, and it is already being deployed in leading fabs and test operations.
Here is the caveat, and I want to be clear about this: humans need to set the traffic rules. AI should drive the car autonomously, but the rules of the road, covering what constraints matter, what tradeoffs are acceptable, and what accountability structures govern the decisions, must come from people, what we called “Human Governance with AI Execution.” This is not a hedge against AI capability. It is a recognition that in a manufacturing environment, the consequences of wrong decisions are serious and often irreversible. A semiconductor fab is not a recommendation engine. Accountability cannot be delegated to a model.
The companies I see succeeding with AI in manufacturing are the ones that have been deliberate about this division of labor. They are using AI to surface signals, rank hypotheses, and suggest actions, while keeping experienced engineers in the loop to evaluate, approve, and learn from those suggestions. The companies struggling are either waiting for AI to be “ready” before deploying anything, or over-automating in ways that erode the engineering judgment that makes the AI outputs trustworthy in the first place.
The GSA Summit’s opening keynote from Anthropic’s Field CTO and several of the panel discussions reinforced this point from multiple angles. The most useful framing I heard: AI agents are becoming capable of doing tasks autonomously, but the most valuable deployments are the ones where human expertise shapes the problem definition, evaluates the output, and learns from the feedback loop. That is the model that works.
I left Scottsdale with a clearer sense that the semiconductor industry is at an inflection point that is less about any individual technology and more about how we organize ourselves to use the technologies we already have. The companies that figure out full-stack collaboration, building the data infrastructure, establishing the governance models, and deploying AI in a way that amplifies rather than replaces engineering judgment will have a durable advantage in the decade ahead. We recently demoed an example of such an approach using LLM to analyze test data.
The GSA Tech Summit is a good barometer for where the industry’s thinking is. This year, I heard less debate about whether collaboration matters and more urgency about how to make it real. That is progress. The hard work is in the implementation: getting data models to align, getting organizations to share without losing competitive advantage, and getting AI deployed in ways that actually change manufacturing outcomes.
Those are solvable problems. But they require the full stack, spanning design, manufacturing, equipment, test, and everything in between, to work on them together. I am optimistic that we are getting there.
One moment from the Summit stuck with me as a perfect illustration of all this. I ran into Rich Simoncic, Chief Operating Officer of Microchip Technology and fellow panelist, at the networking reception. We were both on the “Collaboration Models That Actually Work” panel, and the conversation flowed naturally from the stage into the evening.
Somewhere in the middle of it, Rich recalled that the last time we had spoken in person was at CES in 2019, seven years ago, when I was in the thick of building a chiplet startup. Seven years in semiconductor time feels like a lifetime ago. But he remembered it clearly. In an industry with thousands of companies, hundreds of conferences, and executives constantly in motion, that kind of continuity matters. The semiconductor world is, in the end, a small world. The relationships built across years and events are part of what makes collaboration possible when it counts. The technical infrastructure is essential, but so is the human network underneath it.
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
Leave a Reply