Samsung, TSMC, Intel and SK Hynix Have Never Shared an R&D Lab. A $5 Billion Building in California Just Put All Four Under One Roof.
Applied Materials, the equipment maker that sells machines to nearly the entire chip industry, opened a new $5 billion R&D campus in Sunnyvale, California on October 12 — and for the first time, it's hosting engineers from Samsung, SK hynix, TSMC, Intel, Micron and Broadcom inside the same shared cleanrooms. These six companies spend billions trying to out-build each other every quarter. But on one specific problem — stacking chips together atom-tight instead of soldering them — Applied Materials is betting none of them can solve it fast enough alone, so it built them a building to stop trying separately.
Run the race: solving a packaging problem solo vs. pooling the lab
Set how many rival teams pool their knowledge inside a shared lab (1 = working alone, 6 = Samsung, SK hynix, TSMC, Intel, Micron and Broadcom all pooling, like at EPIC), then press Run to step through 16 R&D cycles. Both lines model the same hybrid-bonding defect problem closing toward "solved" — the gold line is one team working alone, the blue line is your chosen number of teams pooling discoveries. Illustrative learning-curve model with assumed rates, not Applied Materials' real defect data or timeline.
Ready to run. Pick how many teams pool knowledge, then press Run to step through 16 R&D cycles.
The plain version
Think of a cutting-edge AI chip today less like one slab of silicon and more like a club sandwich: a logic chip on the bottom, stacks of memory on top, all fused together so tightly that signals can hop between layers almost instantly. That fusing step — called "packaging" — has quietly become a bigger bottleneck for how many AI chips the world can actually build than the fancy machines that print the circuits in the first place. Chipmakers can etch transistors faster than they can stack and bond the finished pieces together without cracks, warps, or misalignment at a microscopic scale.
Normally, Samsung, TSMC, Intel, SK hynix, Micron and Broadcom guard their factories like bank vaults — they're each other's biggest competitors, and almost nothing leaves the building. But stacking chips this precisely — bonding bare silicon to bare silicon with no glue layer at all, aligned to within a sliver of a strand of DNA — is such a hard manufacturing problem that no single company's R&D team can crack every piece of it fast enough alone.
So Applied Materials, the company that builds the machines nearly all of them buy, just opened a $5 billion research campus in Silicon Valley — and for the first time, it's putting engineers from these rival companies to work in the same shared cleanrooms, on the same packaging problem, at the same time. They're still bitter rivals on the finished product. But on this one upstream manufacturing step, everyone would rather pool the hard part than each spend years solving it in isolation — because right now, that shared bottleneck is capping how many AI chips any of them can actually ship.
The expert version
Advanced packaging has been shifting from soldered "bump bonding" — chiplets connected through solder microbumps typically spaced 40–55 micrometers apart — to "hybrid bonding," which fuses the copper pads and dielectric surfaces of two wafers or dies directly, with no solder or adhesive layer, at pitches now below 10 micrometers and trending toward sub-2-micrometer pitches in some HBM and SoIC designs. Interconnect density scales roughly with the inverse square of pitch, so halving the bonding pitch can roughly quadruple vertical connections per square millimeter — exactly what's needed to stack logic dies under HBM memory fast enough for AI accelerators.
The catch is that hybrid bonding has almost zero tolerance for error: wafer warpage, surface particulates, or planarity deviations of a few nanometers can void a bond or short adjacent pads. Fixing that requires co-optimizing surface-activation chemistry, CMP polishing, sub-nanometer flatness metrology, and the bonding/dicing tools themselves — a combinatorial problem spanning multiple vendors and each chipmaker's own proprietary process recipe. Historically each company ran that validation loop separately inside its own fab, a cycle that can take 12–18 months per iteration and requires redundant pilot lines.
Applied Materials' EPIC (Equipment and Process Innovation and Commercialization) Center, opened October 12, 2026 in Sunnyvale, is built to shorten that loop: roughly 55,700 square meters, including about 16,200 square meters of cleanroom space, at a total investment raised from an original $4 billion (announced 2023) to $5 billion as customer commitments grew. Confirmed participants include Samsung Electronics, SK hynix, TSMC, Intel, Micron, Broadcom, Advantest, Screen, Kioxia, and bonding-equipment specialist Besi. On October 1, 2026, Applied Materials and Besi — partners since a 2020 joint lab in Singapore, and more tightly tied since Applied's 9% equity stake in Besi in April 2025 — expanded that collaboration specifically into die-to-wafer, die-to-die, and die-to-panel hybrid bonding and thermo-compression bonding platforms aimed at 3D-IC scaling. The bet: parallelizing process R&D across competing customers inside one shared tool set cuts the number of redundant validation cycles the industry collectively has to run — relevant given TSMC's advanced-packaging (CoWoS) lines are still described as sold out heading into 2027 even as capacity is targeted to roughly double by 2028.
Why it matters for tech + supply chain: Packaging, not the lithography machine everyone talks about, is now the real pacing item for AI chip output — and the rivals who'd normally never cooperate just admitted none of them can fix it alone.
Why it matters for tech + supply chain: Advanced-packaging yield and throughput, not wafer-level lithography, is the binding constraint on AI accelerator output right now; pooling process R&D across competing customers inside one shared facility is a structural bet that collective iteration speed beats proprietary secrecy once a bottleneck threatens everyone's shipments at once.