Lessons from UltraFacility 2026 on fabs’ biggest blind spots

Over three days, operators and suppliers returned to the same barriers. Contamination cannot be measured reliably online. Infrastructure benefits are being lost in procurement and capital models that do not capture them. New technologies struggle to clear qualification because fabs want references before taking the risk.

Power Has Moved to the Front of the Facilities Agenda

Sargent & Lundy framed the shift clearly: semiconductor manufacturing “was a big fish in the energy market with first refusal on access to reliable and cost-effective power supply” but is now “being squeezed by demands from data centers.” A typical advanced fab draws 60 to 150 MW; major data-centre campuses can start at around 500 MW.

But grid capacity is only part of the resilience problem. Intel’s Josune Armas attributed almost half the severity of Intel’s power-related impacts to medium-voltage equipment failures triggered by utility disturbances, putting a significant part of the risk inside the fence.

However, while fabs are being asked to cut power, model their systems and deploy AI,much of the electrical system remains unmonitored. Armas said visibility generally stops at the low-voltage substation, and that this blind spot forces designers to stack “buffer on top of buffer” in capacity margins, building in spare capacity that better data would show is not needed. Asked which retrofit pays for itself first, Ramon Lugo of Schneider Electric pointed not to a product but to power-quality metering at every substation, which can expose sags, harmonics and losses that may have been running for years. Without context, he warned, a data lake becomes a “data swamp”.

Sub-20 nm Metrology Remains the Water Blind Spot

The industry is working towards contamination limits that conventional online monitoring cannot reliably measure. One target discussed at the conference was one particle of 10 nm or larger per 300 mm wafer layer, while current online monitoring typically bottoms out closer to 20 nm. Across the whole conference, no instrument on show put a sensor in the pipe at that scale.

Don Hatter of Intel summed up the problem: “The wafer is still the best metrology in the world for measuring defects. And that is the opposite of what it needs to be.” Incoming checks test against a list of known contaminants, so something genuinely new passes all of them, and the wafer is what finally registers it. By then, product may have to be scrapped.

That shifts attention upstream, to chemical suppliers, water-treatment systems and the components inside them. A stainless-steel pump was shown producing three to four times more particles, with online ICP-MS identifying a 316 stainless-steel signature, while a commercial polishing resin bed was shedding iron nanoparticles. Neither had failed an incoming check, because neither was an incoming stream.

The Costliest Mistakes Are in the Metrics

Some of the week’s most useful findings were about what gets measured, and where in the decision chain.

At procurement, Wojtek Osowiecki of Lam Research said he had never seen a water-consumption specification for a process tool because conventional tool cost-of-ownership models omit water and electricity. That leaves major facility operating costs outside the criteria used to buy the equipment that drives them.

At operation, Intel’s hydrofluoric waste system was judged on fluoride alone, leaving lime dosing unchecked. The excess lime raised hardness in the water reaching downstream ion exchange, limiting how much could be reclaimed. Taking a whole-system view cut the dosing without capital investment. It was one of several changes in a program that reduced the site’s freshwater withdrawal by around 23% in under a year. The fault lay not in the water but in what the system had been told to optimize.

At investment, Guru Navaneetham of onsemi described a chiller replacement rejected at CFO review despite net savings of nearly $1.5 million and a return of around 30%, while a project returning 12% was funded instead because the chiller failed on payback period. His argument was that simple payback misses asset life, savings beyond the payback date and the time value of money. His case was for judging facilities investment on net annual cost of goods sold (COGS) savings, net present value (NPV) and internal rate of return (IRR) instead.

Someone Has to Take the First-Fab Risk

Thirteen vendors pitched on Day 0. One, Membrion, could point to a multi-year deployment in a fab: two and a half years treating copper waste at a Lam Research plating tool in Oregon, running at 250 gallons per day. That exposes a familiar adoption problem. Fabs want proven references before taking on process or facilities risk, but new technologies cannot build those references unless someone goes first.

Vendors were often selling ways to reduce that exposure as much as equipment itself, including treatment-as-a-service and build-own-operate contracts. At Gradiant's Day 2 roundtable, Siva Kumar Kota described the company installing and financing a system itself, with the customer buying it back after about three years once the KPIs are demonstrated.

Clayton MacDougald of Emerald offered a simple test: ask who uses, chooses, profits and pays. His recommendation was to run fast pilots against agreed KPIs on simple contracts, and to match the validation burden to the operational risk rather than applying the same standard to every technology.