Mack

The huge difference: HBM, advanced packaging, high-end PCBs & optics are **tech-complexity bottlenecks** — limited by cutting-edge processes, yields, and a handful of players (TSMC, SK Hynix, etc.) that can’t ramp fast enough for exploding AI demand. MLCCs, power semis & analog are **volume/allocation crunches** on mature nodes — strained by broad demand across EVs, industrial, consumer + AI, plus raw material issues, but easier to expand with existing fabs. One is “can’t build it yet,” the other is “everyone’s fighting for share.”

The Critical Weakness of AI Computing Power… Broadcom Identifies "3 Major Bottlenecks," Supply Shortages Could Persist Until 2027 • Broadcom has identified the key bottlenecks in the current AI supply chain as: ▲lasers, ▲advanced process wafers (particularly TSMC), and ▲PCBs. Notably, lead times for the small-form-factor PCBs used inside optical transceivers have extended from approximately 6 weeks to 6 months, with relief not expected until 2027. • These small-form-factor PCBs require high-difficulty mSAP processes due to the demands of high-frequency signal processing and ultra-compact design. Because their production processes partially overlap with IC substrates for HBM, the surge in global AI server demand is paradoxically eroding PCB production capacity. Additionally, supplier qualification takes over 6 months, leading hyperscalers to lock in existing supply chains through long-term contracts. • Laser components have also been flagged as a critical bottleneck in the CPO era. Lasers for AI data centers require high-power, low-noise characteristics, and the yield rate passing stringent reliability testing remains below 30%. The number of vendors with InP (indium phosphide)-based production capabilities is particularly limited, and in CPO architectures, laser demand increases not linearly but multiplicatively, further intensifying supply pressure. • On the wafer front, TSMC's advanced process nodes are projected to hit a bottleneck by 2026. However, the real issue lies in back-end advanced packaging. The CPO era requires high-difficulty hybrid bonding technology to integrate optical and silicon dies, with throughput and yield severely constrained in the early stages. • Furthermore, TSMC must allocate the same production lines across NVIDIA, Apple, AMD, and Qualcomm, as well as Google, Meta, and OpenAI — effectively operating production capacity under a "rationing" regime. New capacity additions also face 12–18 month equipment lead times, making near-term relief unlikely. • Meanwhile, expansion at suppliers of packaging substrates, underfill, testing, optical communications, and probe cards is progressing even more slowly, causing bottlenecks to proliferate across the entire supply chain. As a result, supply chain bottlenecks and price inflation are likely to persist structurally, with the segments experiencing bottlenecks emerging as the critical nodes wielding pricing power.

TrendForce on April memory contract prices: - April DRAM prices rose 57% compared to 1Q26. - Compared to 1Q26, PC DRAM rose 40%, server DRAM rose 48%, mobile DRAM rose 80%, and consumer DRAM rose 60%. - April NAND prices rose 65–70% compared to 1Q26. - eMMC/UFS, mainly used in mobile devices and SSDs, rose around 75–80%.

Can we now put to sleep the argument that software is not a moat for neoclouds? Once you need to spend $600M+ to cover just a fraction of the software stack, that thesis gets reduced to ashes $IREN is just starting its road to building a software stack, while a company like $NBIS is already at the top That gap should deserve a valuation premium

I'm still laughing how much Swedish hate their own frontier companies so much. That they write hit pieces every day on $SIVE. This one was entertaining: Local journalists show up to an empty $SIVE administrative building uninvited. Because they can't fathom the CEO is in Silicon Valley or design team is working on US Gov CHIPS act dev in the US. And because there weren't many cars parked outside + CFO wouldn't take questions about secretive hyperscaler deal financials. They wrote a random negative hit piece. By repeating "There are several who make lasers like these and Sivers are far from alone". Several like $LITE, $COHR, $60B+ companies. and reported earlier that "CPO is nothing special, it's been around for years." While GS projects CPO going from $1B -> $91B TAM over the next two years. Even put "Plans" in quotation marks because they didn't think Sivers is supplying lasers to $JBL 1.6T LRO. IMO, $SIVE ends up as a $10B+ company next year, especially if they follow what $LITE / $COHR did with downstream IP integration to capture more of CPO module BOM. Just don't think Swedish people understand hyperscaler supply chains, concept of forward growth, or the fact that employee count doesn't equate to revenue. Transfer of control from local Swedish -> West is always appreciated, as this was a majority owned local retail company before.