Semiconductors & Advanced Manufacturing
The week's headline in chips: TSMC just announced its next three generations of technology and broke ground on a US packaging plant — all while signaling it can get there without the world's most advanced (and most contested) chip-printing machines. Meanwhile, Samsung's workforce is heading for the picket lines, with AI memory supply hanging in the balance.
TSMC Keeps Moving Forward Without the World's Priciest Tool
TSMC (Taiwan Semiconductor Manufacturing Company — the Taiwanese foundry that manufactures most of the world's advanced chips, including those designed by Nvidia, Apple, and AMD) dropped a significant roadmap update this week: 3 new process nodes are coming, and the company broke ground on a packaging facility in Arizona. But the detail that's drawing the most attention is what TSMC won't need to get there.
All of these new technology generations, TSMC says, can be manufactured without high-NA EUV — that is, High Numerical Aperture Extreme Ultraviolet lithography machines, the newest generation of chip-printing equipment made by ASML, the Dutch company that holds a global monopoly on this technology. Each machine costs well over $300 million, and ASML can produce only a handful per year. They are also the subject of aggressive US export-control pressure aimed at restricting China's access — which makes any dependency on them a geopolitical and supply-chain risk simultaneously.
By publicly committing to a technology roadmap that doesn't require these machines, TSMC is doing two things: signaling manufacturing flexibility at exactly the moment when access to equipment is most politically fraught, and reassuring customers that the AI chip supply chain isn't hostage to the world's scarcest tools.
The Arizona packaging groundbreaking is the other half of the story, and arguably just as important. Packaging — the process of assembling multiple chips into a single module — has become a strategic chokepoint in AI hardware. Techniques like CoWoS (Chip on Wafer on Substrate — a method that stacks high-bandwidth memory directly alongside an AI processor on the same substrate, dramatically improving how fast data moves between them) are currently supply-constrained. Building more packaging capacity on US soil addresses both the bottleneck and the political pressure to onshore critical semiconductor infrastructure.
Memory Under Pressure: A Strike, a Shortage, and IBM's Shrug
Two stories this week trace the fault lines in the memory chip market — and they pull in opposite directions.
More than 30,000 Samsung workers took to the streets this week, with 18 days of planned strikes set for May. Samsung is the world's largest maker of DRAM (Dynamic Random Access Memory — the workhorse memory inside computers and servers) and a critical supplier of HBM (High Bandwidth Memory — the specialized, stacked memory chips that sit directly on AI accelerators like Nvidia's Blackwell GPUs, where they feed data to the processor fast enough to keep it from sitting idle). Samsung has been fighting to catch up to rival SK Hynix in the HBM race. A prolonged strike would hit a company already under competitive pressure at exactly the wrong moment — and potentially ripple into AI hardware availability if HBM supply tightens further.
IBM, meanwhile, told investors it's "unconcerned" about the memory shortage's impact on its supply chain and is seeing continued progress in its infrastructure business. That's a notable data point: a large enterprise technology company is signaling it isn't feeling acute pain from the shortages that have dominated analyst commentary. Whether that reflects IBM's purchasing power, its product mix (its mainframes and storage systems use different memory configurations than AI servers), or actual loosening in the overall market is unclear. But it's a useful counterweight to the most alarming supply-crunch narratives.
The Race to Move Data With Light
Marvell Technology — a semiconductor company that makes networking, storage, and data-infrastructure chips, and an increasingly important player in custom AI chip design — acquired Polariton Technologies, a startup focused on silicon photonics. The financial terms weren't disclosed.
Silicon photonics is the technology of using light (photons) rather than electrical signals (electrons) to move data between chips, between servers, and across data center racks. As AI models have scaled to hundreds of billions of parameters, a growing share of the computational bottleneck isn't raw processing speed — it's bandwidth: how fast you can feed data to the processors. Optical interconnects can move more data, faster, with less heat than traditional copper-wire alternatives, and at the scale of a modern AI data center, those differences compound dramatically.
Every major chip company and hyperscaler (the industry term for the giant cloud providers — Amazon, Microsoft, Google, Meta — that own the world's largest data centers and AI infrastructure) is now racing to control more of this interconnect stack. Marvell's acquisition of Polariton is a bet that silicon photonics components become as strategically valuable as the chips they connect. It won't be the last deal in this space.
What to Watch
TSMC's combination move — new process nodes, US packaging capacity, and a technology path that sidesteps dependency on the rarest equipment — is the week's most strategically significant signal. It's a statement of resilience at a moment when tariffs, export controls, and geopolitical uncertainty are running high. The Samsung labor situation is the near-term wildcard: if strikes extend through May and HBM shipments slip, the effects will be felt in AI hardware delivery timelines within a quarter. And the Marvell/Polariton deal is a preview of what the next phase of chip M&A looks like — not just buying compute, but buying the infrastructure to move data between it.
TL;DR - TSMC announced 3 new chip generations and a US packaging facility groundbreaking — and explicitly said it doesn't need the world's most advanced (and geopolitically fraught) chip-printing machines to get there - 30,000+ Samsung workers are planning 18 days of strikes in May, threatening supply of the high-bandwidth memory that AI accelerators depend on; IBM says its own supply chain is fine - Marvell bought a silicon photonics startup, signaling the next arms race in AI hardware is about moving data with light — not just computing faster
Compiled from 1 source · 20 items
- Data Center Dynamics (20)