TSMC chief executive C.C. Wei used his Thursday shareholder meeting in Hsinchu to tell investors what every AI hardware buyer already suspected. Customer demand for leading edge silicon is so far ahead of supply that even the company’s $165 billion US buildout will not close the gap for years. “Customer demand is so high, and we can only support so much,” Wei said, according to The Verge’s recap of the meeting. “We are doing our best to ensure TSMC does not become a bottleneck.” Coming from the foundry that fabricates nearly every leading edge accelerator shipped by Nvidia, AMD, Apple, and Broadcom, that is the strongest demand signal the industry has heard this year.
Why the $165 billion Arizona bet does not fix 2026
TSMC has been producing 4nm parts in Arizona since January 2025, and Wei reiterated that three more US fabs, two advanced packaging facilities, and a research and development center are committed, bringing total announced US spend to $165 billion. That is the largest single foreign direct investment in US manufacturing history. The catch is timing. A greenfield fab takes four to five years from groundbreaking to qualified high volume output, and advanced packaging, which is the actual choke point for Blackwell, Rubin, and the MI400 family, takes nearly as long to scale at parity yield. Wei said closing the gap with American demand would take “a very long time,” which in foundry language means the 2026 and most of 2027 supply pictures are essentially fixed today.
Energy and water in Taiwan, where the bulk of leading edge capacity still sits, are also capped. Wei did not break out CoWoS packaging capacity in detail, but TSMC has previously guided that it is doubling that capacity each year through 2026 and still cannot meet booked orders. That is the line item that gates how many H200, B200, and GB200 systems actually ship.
Pricing: steady increases, not the DRAM shock
Wei was equally direct on price. He said he would “like” to raise TSMC’s wafer prices, but ruled out the abrupt step changes that have hit memory in recent quarters. DRAM contract prices have moved up double digits quarter over quarter at Samsung and SK Hynix, and enterprise SSDs are now on multi month allocation. TSMC prefers contracted, steady increases that protect long term customer relationships. For hyperscalers buying multi year wafer allocations, that is meaningful but small. Even mid single digit annual hikes on N3 and N2 wafers translate into billions across an Nvidia or AMD product line, and almost all of it eventually shows up in the cloud bill of every enterprise buyer.
The Verge writeup also flags Deloitte’s projection that semiconductors become a $1 trillion industry by 2027, anchored almost entirely by AI workloads. The math only works if pricing power keeps drifting upstream, toward the fab and the memory makers, and away from the model providers and the clouds reselling capacity.
Memory is the second squeeze, and it is already biting
The supply story is not just logic. HBM3e and HBM4 capacity at Samsung and SK Hynix is sold out through 2026, and the spillover into commodity DRAM and NAND has already pushed consumer SSD prices up sharply this quarter. Retail and consumer hardware buyers across Europe, including the likes of Currys, Fnac Darty, and Otto, have started flagging margin pressure on PC, console, and smart device categories tied to memory inflation. Anyone running personalisation, search, or vision inference on commodity GPUs is now exposed on two axes at once: the accelerator itself and the HBM stacked on top of it.
What this means for buyers signing 2027 AI infra contracts
Our read for operators: stop treating GPU capacity as an elastic, on demand utility and start treating it like a fab allocation, because that is what it is becoming. Reserved instance pricing on H100 and H200 capacity at AWS, Azure, and Google Cloud is already trading at a 30 to 50 percent premium to last year’s spot equivalents, and that gap will widen through 2026 if Nvidia cannot ship into demand. The buyers who lock in 2027 commitments in the next two quarters, at known utilisation, will pay materially less than the buyers who wait for clarity. Clarity is not coming.
Concretely, we are pushing clients to do three things now. First, sign 2027 reserved capacity for any workload with predictable utilisation above roughly 60 percent, even if the discount looks thin today, because the alternative in twelve months is paying spot into a tighter market. Second, build cost aware routing into every agent and inference path, so that premium frontier models are reserved for the requests that actually need them and the long tail falls back to mid tier or open weight models on cheaper silicon. Third, treat the build versus buy question on inference honestly: at current GPU economics, owning H200 capacity through a colo partner pays back inside 14 to 18 months for sustained workloads above a few hundred concurrent sessions, and the lock in risk to a single hyperscaler is the larger long term threat. The teams that win the next eighteen months are the ones that already stopped assuming compute would get cheaper on schedule.
Geopolitics and the foundry alternatives
Wei’s comments land while Washington keeps tightening export controls on advanced chips bound for China and while Taiwan’s grid is under continued strain. Intel Foundry and Samsung Foundry remain the two plausible release valves at advanced nodes, but neither is a 2026 story. Intel’s 18A is ramping with a short external customer list, and Samsung’s 2nm yield trajectory is still behind TSMC’s. Custom silicon from AWS Trainium2, Google’s TPU v6, and Microsoft’s Maia is real and growing, but most of it is also fabricated at TSMC, which means it competes for the same wafers and the same CoWoS slots as Nvidia.
The four signals to watch into Q3
Nvidia’s next earnings call commentary on Blackwell and Rubin allocation, which will reveal how TSMC is actually rationing CoWoS capacity across customers.
Microsoft and Google capex guidance, already at record levels, and whether either re rates upward again to chase scarce supply.
Any concrete sign that Intel Foundry 18A or Samsung 2nm is absorbing meaningful AI accelerator volume from a top five buyer.
HBM4 qualification timelines at SK Hynix and Micron, because memory, not logic, is the harder ceiling through 2026.
The single date that matters is Nvidia’s August earnings call. If management guides Blackwell supply up and lead times down, the squeeze is loosening earlier than Wei suggested and reserved capacity premiums will compress. If Nvidia instead extends allocation language into the first half of 2027, the cost of waiting to sign your AI infra contracts will be visible in your own P and L by the end of this year.
