$5016.T JX Advanced Metals is investing ¥120B to increase its InP substrate production capacity by 10x - Nikkei

Didn't think much of it when Trendforce first posted about CoPoS/glass core substrates since they have been talked about in recent months Conceptually the square panel makes sense - less wastage/warpage, better dielectric/surface properties Then I realized, a lot of the existing front-end workflows rely on spinning the (round) wafer to achieve uniformity (e.g. in polishing, photoresist coating/deposition) and process consistency. How does that work for a square panel? I found an old video from @IanCutress that explains why chips are square and wafers are round m.youtube.com/watch?v=Rhs_Nj…

Interesting

Chanced upon this startup - Sapient Intelligence - who introduced a Hierarchical Recurrent Model (HRM) about a month ago to pit against standard transformer models (by that referencing to all modern LLMs) Their HRM model works on a dual-timescale architecture that decouples deliberation into two layers: strategic and deliberation  As an empirical proof of efficient pretraining, they built a 1B-param model (HRM-Text) trained from scratch in 1.9 days on 40B tokens, 16 GPUs (not sure which) and a $1500 budget, with pretty remarkable benchmark results Similar playbook as Manus: China-born founders, set up base in Singapore with a hub in Palo Alto Elon Musk reportedly tried to recruit them with multi-million pay packages to join xAI when they were still at Tsinghua Uni Brain Lab, which they turned down. At that time they were the only group other than DeepSeek exploring RL for LLMs Interesting company to watch

I did some foundational work on prompt engineering earlier this year Little did I expect this process to be used someday to jailbreak models (Mythos & Fable 5) and doing so using agents to conduct systematic attacks The whole process runs on iterative testing loops and rubrics Looks like I should revisit those notes again

Origin ...

I don't disagree with what @dylan522p has said but Intel does not need to raise all debt, or equity, at once. They have balance sheet flexibility. They can even consider hybrid alternatives - convertible note, preferred stock, etc which are less dilutive than straight equity with better credit treatment My previous post was to point out that when cost of equity is much higher than cost of debt, any sensible CFO will only undertake a dilutive equity raise as a last resort - when the company is almost breaching debt covenants, have reached the limits of their debt headroom, or in the worse case close to a default event, which is clearly not the case for Intel The inherent issue in proposing for Intel to raise massive equity capital today is that we are asking them to use a permanent capital solution to address a transitory supply constraint issue, which should resolve over time as ramps happen over stages The argument that their stock trades at a premium now is also subjective, as one's view of intrinsic values, calculated using a myriad of parameters and assumptions, may vary from another One alternative that Intel can pursue, like TSMC, is to negotiate with their customers (if not already done so) to put up prepayments to secure fab/packaging capacity. These prepayments help part-finance, together with existing cash flows, any capex requirements and would be positive for Intel from a ROIC standpoint. In my view, this is also a de-risking move on the part of Intel so they do not end up with redundant spending or capacity Lastly, as we have seen how Intel has been raising equity, it is mostly from the ecosystem partners/USG, of which the value creation/political goodwill generated far outweighs dilutive impacts But noted all points, and always good to have a healthy debate

I'd recently been offered a look at a nuclear start-up mid-stage funding opportunity so am doing some work particularly on the 11 advanced reactor projects (including SMRs/micro-reactors) selected by the US DOE for their reactor pilot program, fuel types (UO2/LEU, LEU , HALEU, TRISO), and coolant/heat pipe technologies Very interesting space given the grid infrastructure crunch in the US driven by data center demand but still trying to figure out how/at what value to underwrite any company at this point given that commercial criticality won't be until 2028/2029 (maybe someone out there may know - do reach out!) Antares' Mark-0 was the first project reported a few days ago to have achieved initial zero-power criticality ahead of the July 4 deadline set by President Trump in his May 2025. This is a key milestone since the test confirms that the reactor can operate safely and establishes a basis that would allow subsequent reactors to produce electricity starting next year On this note, Japan and the US held talks this month toward a general agreement for Japan to invest up to $40B in small modular reactors (SMRs) built by $GEV GE Vernova and $6501.T Hitachi. A $25B investment in $SMR NuScale Power is also being considered Link below: asia.nikkei.com/business/ene…

Intel's current cost of equity: ~10% Intel's longest duration bond (maturing in 2066) currently yields at ~6.2% Their latest multi-tranche (5-40Y) bond issue in April were heavily oversubscribed (~7.7×) and priced at almost near par to secondaries As a former capital markets banker, I am not sure if I agree with this thesis

$HON Honeywell's equity story seems to be under-appreciated, in my view The company has been undertaking portfolio actions - with the major ones being the full spin-off of Solstice Advanced Materials (in Oct 2025), IPO of Quantinuum (last week) and the upcoming full spin-off of Honeywell Aerospace (later this month) Once the spin-off of Honeywell Aerospace is complete, the remaining company (RemainCo) shall be rebranded as Honeywell Technologies, with a streamlined portfolio comprising Process Automation & Technology (PA&T), Building Automation (BA) and Industrial Automation (IA) Markets seems to perceive the RemainCo as a mature business with low-to-mid-single-digit revenue growth (albeit with >20% adjusted EBIT margin), but overlook the criticality of some of their businesses One segment that stands out to me is their LNG business (under PA&T) With the acquisition of Air Products' LNG process technology & equipment business in Sep 2024, Honeywell not only has an end-to-end integrated LNG business spanning pre-treatment (which they have ~40% global share), NGL recovery and liquefaction, but also owns the proprietary technologies underlying all of Qatar's natural gas primary infrastructure (through Qatar Energy) That is, all 14 existing LNG process trains in Ras Laffan, were fabricated using AP-X / AP-C3MR equipment and technology Despite transitory impact from the Middle East conflict (0.5% on 1Q revenue and ~1% expected on 2Q revenue), the damage to two of the LNG trains (S2-4 and S3-6) from Iran's attack will mean Qatar will need to order parts and equipment to facilitate the full repair of these trains, which may take up to 5 years Moreover, Qatar has not stopped their planned expansion in North Field, which is the largest non-associated (i.e. independently of crude oil) natural gas field in the world For the expansion in North Field East (NFE) and North Field South (NFS), all six LNG process trains (4 mega trains totalling 33mtpa in NFE and 2 mega trains totalling 16 mtpa in NFS) are going to be constructed using the AP-X LNG process technology The reason why Qatar favours the AP-X proprietary technology for its mega trains is because they have 100% relative efficiency (i.e., they deliver the exact output using the planned parameters) - which demonstrates how critical and entrenched this business segment is Honeywell Technologies' Investor Day later today would be an interesting event to watch

As a follow-up to my previous post, it looks like there is some nuanced difference between "launched/announced" and "mass production/deployment" of 800VDC power racks/systems The named ecosystem partners that have "launched/announced" 800VDC racks/systems: Delta Electronics: deltaww.com/en-US/products/d… Flex: investors.flex.com/news/news… LITEON: liteon.com/en/news/press-cen… Lead Wealth: lead-wealth.com/product-deta… Megmeet: megmeet.com/products/info.ht… As for "mass production/deployment" timeline, an article from Taipei Times which was published last Thursday quoted the following from Delta Electronics: • 3Q 26: planned small-volume shipments of its 800VDC power racks to Nvidia, with these systems to be deployed first in Nvidia's data centers before being introduced to other customers later Mass-production and deployment timing would "depend on the rollout schedules of the latest AI platforms" • Also in 3Q 26: /-400VDC systems slated for mass-production "Many CSPs" are adopting this architecture, including a major "search engine company" that has deployed the technology (i.e., Google) When asked whether the 800VDC architecture would become the dominant next-generation power-delivery standard for AI data centers, Jensen replied: “Yes.” (but he didn't say when) When asked whether Delta would be the primary supplier of the architecture, Jensen replied: “I have no idea.” (lol) Article link: taipeitimes.com/News/biz/arc… While it's really up to one's interpretation of what Jensen said with some ambiguity, given that Delta's 800VDC racks are only going to be shipped in small batch first to Nvidia for their own deployment (and Delta is supposedly the lead supplier), there are grounds to believe that 800VDC adoption by Nvidia's customers will hinge upon these factors: 1) validation, 2) cost, 3) space configuration (since these power sidecars are supposed to sit in the white space) and 4) timing of Kyber rack rollout (with >1MW total rack power, to make the transition meaningful)

SemiAnalysis published an institutional note with a sub-section on the 800VDC power architecture being delayed. I don’t have access to it so I cannot comment on its contents, but I have gathered data points that I think could explain why 1. The case for 800VDC power architecture is strongest when/if total rack power need breaches 1MW – we are not there yet As I pointed out in an earlier post, Flex introduced a new 110kW 3RU power shelf for Vera Rubin NVL72 at COMPUTEX last week. These units, which are integrated into the rack, contain AC-to-DC rectifiers that convert AC input (of 345 to 480 VAC range) into 52VDC output at full load, with power distributed to rack-level payloads through a DC-DC bus architecture The latest iteration of Vera Rubin NVL72 rack has a Max P TDP of about 227kW, and incorporates 4 x 3RU power shelves for a total of 440kW rack power (4 x 110kW) The overprovisioning of rack power (~2x) is to handle power transients/spikes (especially when systems transition rapidly from idle to active state). The redundancy also ensures that the rack achieves high power availability (99.9% for AI training, 99.999% for AI inference, as stipulated by OCP) In this setup, since PSUs are integrated within the rack, there is no issue with copper busbar/overload As we move to Vera Rubin Ultra, each rack is expected to contain the same number of GPUs (72 Rubin Ultra GPUs). Assuming a per-GPU TDP of 3.6kW, total rack TDP could be over 300kW (72 x 3.6kW total GPU TDP CPU networking/switch trays overheads). When Jensen mentioned that rack power could be up to ~600kW for Vera Rubin Ultra, I believe the figure includes overprovisioning (as above). This also assumes that it is not configured in a Kyber rack design. In fact, Nvidia mentioned about the Polyphe prototype that was used in a fully functional GB200-based multi-rack NVL576 scale-up system using 8 MGX NVL racks Link: developer.nvidia.com/blog/nv… Now, things would be different with a Kyber rack design. In the same article linked above, Nvidia said that Kyber will first be introduced with Vera Rubin Ultra as a standalone NVL144 system. The number of GPUs per rack in a Kyber rack = 144, so total GPU TDP = 3.6kW x 144 = ~520kW. Accounting for system peripherals and overheads, and with overprovisioning, total rack power would most likely be >1MW. This is where 800VDC power architecture makes a difference Kyber was initially targeted for commercial launch in 2H 2027. With 800VDC power architecture reportedly being postponed to beyond 2028, this means that Kyber rack production is likely being delayed 2. Availability of 400V power architecture as an alternative While Nvidia is championing the 800VDC power architecture, three companies (Microsoft, Meta and Google) have been working on a /- 400VDC alternative, codenamed Diablo 400, which is available now and actively ramping Link: opencompute.org/documents/oc… Why 400V? It leverages on already mainstream EV technologies in charging/battery infrastructure This architecture, which features a disaggregated power rack (sidecar) to deliver /-400 VDC to an IT rack, allows scalability (to >1MW) with option to parallelize power racks together. This is an important point since it may further push out the commercial timeline for 800VDC, especially when supply chain for 400V is already established and mature In my view, for 800VDC to ramp, we also need the technology for solid-state transformers (SSTs) to develop The engineering challenge for SSTs today is to simultaneously achieve high power density and efficiency. Current SSTs have achieved power densities of ~0.1MW/m3 at 98%-98.5% efficiency – performance that is broadly comparable to conventional line-frequency transformer (LFT)-based interfaces but has not yet matched LFT efficiency curves above 99%. The target that was set at the recent APEC was 1MW/m3 with 30% lower losses than today’s SOTA SSTs While SSTs are currently in early-stage development and more costly than their LFT counterparts (at ~2x the price), the good news is that the SST ecosystem is broadening, with participation from pure-play startups (DG Matrix, Amperesand, Heron Power), PV manufacturers, UPS makers and transformer makers. The industry is also converging on the 250kW power level for the building-block single SST module. My guess on when SSTs are in good form? Probably after 2028, which coincides with the timeline set out in point one above

Looks like my maiden note on optics/CPO is going to age well lol Recap on my closing remarks - "Finally, where do people see CPO maturity & reliability? In a recent vote (somewhere that the panel saw) … 70% of the people voted for year 2030. There will be challenges over the next year or two" I do my own research/channel checks for two reasons: 1 - I don't always trust broker research (and absolutely not everything on X lol) Everyone has an inherent bias when it comes to shaping narratives Even financial models can be and are easily manipulated because they depend heavily on assumptions 2 - Doing your own research is more fun and fulfilling

$FLEX Flex announced three new products at COMPUTEX last week which directly & indirectly supports Nvidia's AI servers: Power shelf conductive energy storage system (CESS) intermediate bus converter (IBC) 1 / 110kW power shelf for Nvidia Vera Rubin NVL72 Designed to provide rack-level power distribution. Supports modular power architectures, including future 800VDC deployments, as well as power disaggregation strategies Power shelf is air-cooled, has slots for 1 x power supply controller (PSC) and up to 6 power supply units (PSUs) PSUs and PSC are hot-pluggable (i.e., no/minimal downtime during maintenance/servicing) 6 PSUs x 18.4kW = ~110kW DC output power (max) with AC operating input voltage range of 345 VAC to 480 VAC. Each PSU comes with a power pulsation buffer device to manage AC input current swing and EDPP non-linearity Output voltage = 52V at full load that distributes power to rack-level payloads/components through a DC-DC bus architecture PSC module incorporates power & thermal management functions and telemetry/monitoring through Redfish $3665.TW BizLink supplies both the input and output connectors (by default) 2 / 30 kW Capacitive Energy Storage System (CESS) This is designed with $7220.T Musashi Seimitsu’s hybrid supercapacitors (HSCs) CESS modulates power fluctuations associated with AI/HPC workloads through energy charges/discharges during large electrical transients Lifespan is longer that existing battery energy storage systems (BESS) with multi-million charge/discharge cycles This new 30kW variant is an upgrade from the existing 16kW rated version 3 / Intermediate bus converter (BMR317) This third-gen IBC is a compact power module that steps down high primary voltage (rack/busbar) to lower intermediate bus voltage (from 40-60V input to 5-7.5V output at a 8:1 fixed ratio), and complements voltage regulator modules (for conversion to 0.5-1.8V to power processors) in a two-stage conversion pathway. The two-stage architecture uses less copper and allows shorter high-current paths at the board/package level Two variants supporting peak capabilities of up to 2kW and 2.5kW are being introduced: the 800W version which is available now, followed by the 1kW version to be released in 2H 2026 (Attached picture of the 3RU power shelf)

$6501.T Hitachi Within a week, Hitachi has announced two strategic partnerships/collaborative initiatives - Intel: solutions and processes optimization development in foundry tool, quantum computing, energy, custom silicon & edge computing Google Cloud: expansion of their existing strategic alliance to accelerate physical AI deployment, delivery of autonomous next-gen cybersecurity solutions and enhancement of their HMAX platform with the integration of Gemini Enterprise

Given the recent discourse on NPOs, reposting my notes from a NPO expert panel session in early May In no particular order, the companies represented include: 1. Ciena 2. Coherent 3. Ranovus 4. Alibaba Cloud 5. Tencent 6. Huawei Founding members of the Open CPX MSA include: 1. Ciena 2. Coherent 3. Marvell 4. Molex 5. Samtec 6. Terahop Since then, Intel, TE Connectivity, Accton and Nexthop AI have joined this consortium as contributing members As for NPO commercial adopters, they will be firmed up close to OFC 2027 which is in March next year. From what I gather, the Chinese chip makers/hyperscalers seem to be slightly ahead of the curve Also, NPO will likely fill a niche of its own and is not going to be a bridge between pluggables and CPO. This provide optionality to all stakeholders in the ecosystem

Over the weekend, China unveiled the world's first energy hub with prefabricated substations and transformers that power data center clusters and computing facilities globaltimes.cn/page/202606/1… First launched in Qingdao, the country plans to roll out this power hub model in national-level data center facilities and regional computing power centers in 2H this year With prefabrication, construction cycle gets cut down to 5 months (from over 12-18 months) with construction costs reduced by 80% and land use by 30% This new hub also allows for 100% green energy connection, with redundancies that enable each equipment to be connected to 3 independent power sources, ensuring uninterrupted power supply. The systems also integrate advanced control modules that manage electricity-to-compute power matching, computing load and grid transients Expect this development - coupled with heavily subsidized and excess energy capacity - to drive down token costs even further in China

[Nikkei Asia] Japan looks to replace up to 5 aging nuclear reactors by 2040s asia.nikkei.com/business/ene… "The plan includes replacing two to five reactors carrying a total capacity of 2,200 to 5,500 megawatts by the 2040s. The high-end figure equals about 20% of the 31,000 MW capacity of Japan's existing reactors, excluding those slated for decommissioning. The draft also calls for replacing an additional nine reactors by the 2050s, giving Japan 11 to 14 replaced reactors by then, with a combined capacity of 12,700 to 16,000 MW."