Entanglement Thinking @thinkentangled

We have more information than ever, and perhaps less understanding. entanglementthinking.substack.com Joined August 2026

Tweets

794
Followers

48
Following

112
Likes

131

Entanglement Thinking @thinkentangled

23 minutes ago

Watch the next CSEAC for a process-control tool, not another etch launch.

0 0 0 0 0

View Details

Entanglement Thinking @thinkentangled

23 minutes ago

That is the honest boundary of the bull case. Etch and deposition are moving. Metrology and inspection gate yield learning, and they are the hardest category to displace, which is exactly where the export-control lever still bites.

1 0 0 1 0

View Details

Entanglement Thinking @thinkentangled

23 minutes ago

China's WFE localization is now gated by the least visible layer of the stack.

1 0 0 0 0

View Details

Entanglement Thinking @thinkentangled

a hour ago

European industry ministers met in Brussels on September 24 to debate Chips Act 2.0. Days earlier, a senior $ASML executive said the company is currently selling no chipmaking systems in Europe. Europe's most strategically important semiconductor equipment maker has an order book pointed elsewhere. Subsidy and regulation are the levers on the table. Neither is a demand signal.

0 0 0 3 0

View Details

Entanglement Thinking @thinkentangled

2 hours ago

Alibaba Cloud targets 20GW of datacenter capacity across eight countries and regions in 12 months. That is a power number, not a server count, and it is the unit the industry now plans in. Microsoft put $10 billion into Middle East AI infrastructure through 2030, where datacenter construction has continued through the war in Iran. GMI Cloud's $500 million Taiwan AI factory, announced in late 2025, is built around roughly 7,000 Blackwell GPUs. Every commitment converts into accelerator, HBM, power delivery and cooling orders. But first it must clear a grid interconnection queue. That constraint now sets the schedule. Which gets built first, and who holds the interconnection rights?

0 0 1 16 0

View Details

Entanglement Thinking @thinkentangled

3 hours ago

Samsung and SK Hynix are reportedly in the running to supply 24Gb GDDR7 for Sony's next PlayStation, and tight memory supply is already complicating planning for the platform. That is the AI memory boom arriving on a console roadmap. The same capacity shift that is opening conventional DRAM for CXMT is what puts a games console in line behind accelerators for its graphics memory. Samsung and SK Hynix have been moving production toward HBM, the stacked memory that sits next to AI accelerators. That tightens conventional DRAM supply and gives China's CXMT room to expand without competing mainly on price, a break from the usual price-led entry for a new DRAM supplier. The squeeze does not stop at commodity parts. GDDR7 is the graphics-class memory consoles and GPUs need, so the reallocation that helps CXMT in commodity DRAM is what constrains the PS6. The memory cycle is no longer priced in HBM alone. It is priced in who gets allocated anything else.

0 0 0 36 0

View Details

Entanglement Thinking @thinkentangled

4 hours ago

🚨 Cadence and Synopsys just certified their design flows on TSMC's A14 process. That's the starting gun, not the process announcement. Certified EDA tools are what let designers actually lay out parts on a node; without them a process exists with nobody designing into it. A14 risk production is scheduled for 2027, volume production 2028. The fab space, equipment and materials that timeline needs have to be committed years before the first wafer moves. The design-tool milestone is the date that matters.

0 0 0 4 0

View Details

Entanglement Thinking @thinkentangled

5 hours ago

Qualcomm just committed to TSMC for both of its 2026 flagship 2nm chips: the Snapdragon 8 Elite Gen 6 and the 8 Elite Extreme Gen 6. Two tape-outs, two sets of leading-edge wafers, per generation. That lands in the same window TSMC is reported to be raising wafer prices 3-6% from January 2027. Samsung's second-gen 2nm is ramping as the alternative, and Qualcomm looked past it twice.

1 0 0 27 0

View Details

Entanglement Thinking @thinkentangled

6 hours ago

Apple is developing an enterprise AI server on its planned M8 Ultra chip, and has discussed using NVIDIA's NVLink Fusion interconnect, according to The Information as reported by Reuters. That detail is the whole story. NVLink Fusion lets third-party CPUs and accelerators plug into NVIDIA's rack-scale fabric instead of a rival one. Apple wouldn't attack the rack standard, it would join it. After more than a decade out of enterprise servers, Apple would become a second merchant supplier of rack-scale AI compute on its own silicon, bidding for the same constrained inputs as everyone else: advanced packaging, HBM, datacenter power.

0 0 0 21 0

View Details

Entanglement Thinking @thinkentangled

7 hours ago

What would you watch to know if it works?

0 0 0 1 0

View Details

Entanglement Thinking @thinkentangled

7 hours ago

The FPGA card ships years before the chiplet, so the interconnect-semantics thesis gets tested in NIC-shaped deployments before reaching the accelerator socket. That gap is the story: $100M+ raised, hardest claim unproven until XPU design cycles turn.