Hello world, it’s Wednesday, August 12th.
TSMC’s CoWoS yield has reportedly crossed 99%, though ABF substrate and memory shortages mean AI chip bottlenecks aren’t going away yet. CXMT hit 90% yield on DDR5, which is putting real pressure on Apple’s memory sourcing decisions. Foxconn posted a 35% quarterly profit jump on AI server demand, and US power consumption is on track for record highs in both 2026 and 2027 as data center buildouts collide with local grid politics.
Let’s get into it. — Austin & Vik
Be sure to check out the Semi Doped podcast on YouTube or your favorite podcast player!
TSMC CoWoS Yield Clears 99% but ABF and Memory Keep AI Chips Waiting
TSMC’s 5.5-reticle CoWoS yield has reportedly crossed 99%, removing what had been one of the sharpest constraints on advanced AI packaging capacity. The foundry isn’t resting there: it has cut its 3DIC development cycle to one year while actively working to contain CoWoS material risks. Yet TSMC itself has flagged ABF substrates and memory as the next bottlenecks in the AI chip supply chain, a candid admission that solving one layer of the stack just exposes the one beneath it.
Nvidia currently controls roughly 60% of TSMC’s packaging queue, while Microsoft’s Maia 300 program is seeking 300,000 units from the foundry. That concentration makes the ABF and memory constraints acutely felt across the industry. To expand its packaging footprint, TSMC is reportedly pursuing two AUO fabs valued at over TWD 30 billion to add FOPLP and CoPoS capacity, with one Longtan site potentially supporting 1.4nm production as well. The acquisitions, if completed, would broaden TSMC’s options beyond CoWoS, though the ABF and memory supply gaps sit upstream of any packaging format TSMC might choose.
Behind the demand:
Camtek, which supplies inspection tools for advanced packaging lines, reports orders topping $600 million and expects its advanced packaging revenue to grow 70% from Q1 to Q4.
TSMC and Sony have confirmed a $4.7 billion joint venture in Japan targeting image sensor chips, a separate packaging-adjacent expansion outside the AI CoWoS queue.
Austin: Supply-demand imbalances have different timespans. Substrates should clear first as an ABF line is a few hundred million dollars and 18–24 months of laser drills and platers, no EUV allocation required. A DRAM fab is $15–20B+ and three years.
CXMT DDR5 Yield Hits 90%, Pressing Apple on Memory Sourcing
Chang Xin Memory Technologies has crossed a yield threshold that few expected so soon. According to Chosun Ilbo and corroborated by Digitimes, CXMT’s DDR5 production yield on its 17nm process node now exceeds 90%, sitting just two percentage points behind Samsung. The company also entered the MSCI index in a record 14 days after its listing, the fastest inclusion in the index’s history. That combination of manufacturing maturity and financial-market legitimacy is what’s pushing CXMT into conversations it wasn’t part of a year ago.
Apple, facing an AI-driven DRAM supply shortage, has been testing CXMT memory in iPhones and MacBooks, according to reporting from multiple outlets, though Apple World Today expressed skepticism about how far those tests have progressed. The sourcing decision isn’t clean: U.S. export controls and Apple’s own supply-chain compliance requirements create what one report describes as dual barriers that complicate any shift toward Chinese suppliers. CXMT’s PC adoption, meanwhile, may stay confined to China and emerging markets in the near term, limiting the pricing pressure it can exert on Samsung, SK Hynix, and Micron outside those regions.
In related news:
SK Hynix is reportedly reviving a NAND expansion inside China as YMTC and other rivals race to meet AI storage demand.
Apple’s parallel effort to source NAND from YMTC faces similar geopolitical and compliance snags as its CXMT discussions.
AI Data Center Power Demand Sparks Moratoriums, Deals, and Gas Plant Fights
The EIA projects US power consumption will hit record highs in both 2026 and 2027 driven by AI infrastructure, and the political fallout is arriving faster than the substations. Chicago Mayor Brandon Johnson signed an executive order imposing a data center moratorium while new rules are drafted, joining smaller municipalities like Salix, Iowa in pumping the brakes. Amazon’s plan to build a 7.65 GW gas plant in Texas drew particular fury: CleanTechnica reported it would become the single largest pollution source in the country, and Pecos County residents have already begun voicing objections to the broader campus. Virginia’s grid is straining too, with Dominion pushed deeper into costly spot power markets by the Northern Virginia buildout.
Hyperscalers are responding with a mix of community money and regulatory compliance. Meta launched a $1 billion fund to support data center host communities and agreed to Texas grid and water standards that Governor Abbott announced publicly. Wistron’s CTO put the constraint plainly: power, not GPUs, is now the real bottleneck. Alpha Compute struck a $55 million deal to acquire Pennsylvania land and gas rights for a new campus, illustrating how directly operators are now chasing fuel supply, not just real estate.
Behind the demand:
Michigan’s AG is targeting the OpenAI-Oracle data center deal on grounds that haven’t been fully detailed publicly.
Crusoe is testing Aalo microreactors as a nuclear power source for AI data centers.
AIG’s CEO says the construction boom is maxing out property and casualty insurers.
Austin: I live in Iowa and have never heard of Salix. Turns out its tiny and has a population of 295 people according to the 2020 census.
Someone in the industry or some PAC needs to get on top of this narrative ASAP.
Traditional media is going to take the headline “datacenter moratorium” from the agenda of a tiny rural town hall meeting and run with it.
And all those little stories are snowballing into one big out-of-control one.
Who will step up and be a thought leader and control the narrative?
Foxconn Profit Surges 35% as Taiwan ODMs Race to Add AI Capacity
Foxconn reported a 35% jump in quarterly profit driven by AI server demand, beating analyst forecasts and confirming that the assembly tier is capturing real margin expansion as hyperscaler spending holds. Foxconn assembles AI servers for Nvidia, and Bloomberg reported the result as a direct read-through on sustained AI infrastructure spending. The numbers arrived alongside a separate but telling move from rival Compal, which formally opened its Daxi AI server manufacturing center in Taoyuan, adding dedicated floor space for rack-level integration at a moment when physical build-out has become the binding constraint.
Compal’s Daxi plant effectively doubles the company’s AI server production capacity, according to Digitimes, and the company is targeting AI at 30 to 40% of total revenue by 2027. The bottleneck has shifted, too. Wistron’s CTO told Digitimes that power delivery, not GPU supply, is now the primary constraint limiting AI data center expansion. Wistron is also moving on that front, planning a 10MW AI computing center in Tainan to extend its own manufacturing footprint beyond pure assembly.
On an adjacent front:
Compal’s Daxi facility sits in Taoyuan’s Daxi District, chosen for proximity to existing logistics and supplier networks in northern Taiwan.
Foxconn’s Q2 result beat consensus forecasts across the board, with AI server revenue growth outpacing the company’s own prior guidance.
Wistron’s Tainan computing center will carry a 10MW power capacity, reflecting how power infrastructure is now sized into facility planning from the start.
Austin: I love that Wistron is standing up 10MW of AI compute to help it’s own operations, plus “academic research at National Yang Ming Chiao Tung University, and national projects” per TaiwanNews.
When the ODM is dogfooding the product, you know the value is real.
Sector Watch
Compute & Edge
Credo Technology scheduled its Q1 fiscal 2027 earnings call for September 1, 2026, covering results for the quarter ended August 1, 2026. (Credo)
Ambarella will report its Q2 fiscal 2027 earnings on September 3, 2026, at 1:30 p.m. Pacific Time. (Ambarella)
IBM and Together AI sign a $240 million multi-year deal to deploy an Nvidia HGX B300 inference cluster on IBM Cloud for open-source AI workloads.
Memory
ChipMOS posted its highest quarterly revenue since 2014 in Q2 2026, up 28.7% year-over-year, swinging to a profit of NT$1.28 per share from a NT$0.75 loss. (ChipMOS)
Samsung and SK Hynix shares rally as Temasek report boosts investor optimism on HBM demand outlook. (Bloomberg.com)
Nanya Technology is set to become ASML’s newest EUV customer, signaling a DRAM technology upgrade at the Taiwanese maker. (Bits&Chips)
Optics & Networking
Lumentum reported Q4 and full fiscal year 2026 results. (Lumentum)
Tower Semiconductor and OpenLight expand the PH18DA InP-on-silicon photonics platform with a Cadence PDK, enabling integrated laser, modulator, and amplifier designs. (Tower Semiconductor)
Google announces three new subsea cable systems across the Americas, linking Panama, the Dominican Republic, and Chile. (Bloomberg Tech)
Foundry
Samsung and TSMC both push High-NA EUV adoption back to the 1nm generation around 2030, deferring the costly machines from current nodes. (TrendForce)
Rapidus 2nm ambitions anchor Japan’s $430 billion 2040 national chip roadmap, with the government betting heavily on a single domestic foundry. (digitimes)
aweXome Ray raises 7.5 billion won in a Series B bridge round to accelerate commercialization of its EUV pellicle membrane technology. (The Elec)
Packaging
Avaco launches a pilot production line for glass substrate through-glass via (TGV) processing, advancing a key enabling technology for next-generation advanced packaging. (The Elec)
Silicon Labs posts Q2 2026 revenue of $228 million with strong earnings growth, signaling recovery in its low-power wireless IoT business. (Silicon Labs)
ASPEED Technology showcases open server management and security architectures at OCP APAC, advancing BMC integration for rack-scale AI servers.
Edge
Silicon Labs posted Q2 2026 revenue of $228 million, reporting strong earnings growth in its low-power wireless IoT business. (Silicon Labs)
AMD pitches its heterogeneous SoC architecture as the better fit for robotics and physical AI workloads, directly challenging Nvidia GPU dominance.
Silicon Labs posts Q2 2026 revenue of $228 million with strong earnings growth in its low-power wireless IoT business. (Silicon Labs)
Policy & Trade
US Congress presses the administration to close TSMC-linked loopholes that allow sanctioned Chinese entities to access advanced chips indirectly. (Businesskorea)
South Korea sovereign wealth fund prepares to deploy more than 1 trillion won in AI and strategic-industry investments starting next year. (Bloomberg Tech)
Adani Group secures an $800 million data center loan after US prosecutors dropped charges, unlocking fresh capital for Indian AI infrastructure (Bloomberg.com)