X Follow-Graph Synthesis

Excerpt from GFHK’s monthly call: Memory sector commentary Customers are pushing back strongly against price increases approaching 30%, leading us to slightly lower our forecast for DRAM price growth in the third quarter. By contrast, we are becoming more positive on NAND. Demand for KV cache offloading continues to exceed expectations, while there is also an emerging trend of substituting expensive DRAM with NAND. We also have a positive view on SK hynix’s second-quarter results. We expect revenue of KRW 85 trillion and a gross margin of 63%. $SKHY $MU $SNDK

Daily News Recap The Information: ASML Plans Price Increases for Chipmaking Equipment, Despite TSMC Resistance Article Summary: - ASML is considering raising the prices of its EUV systems, citing surging demand from the AI and memory semiconductor markets. It has also notified some customers that it plans to increase DUV equipment prices by 10%. - TSMC, ASML’s largest customer, is pushing back against price increases for both EUV and DUV systems. However, some Chinese semiconductor companies have accepted the higher DUV prices. - ASML has raised its 2026 revenue outlook to €43 billion–€45 billion and plans to increase annual production capacity for two of its key systems by 30% this year. - TSMC faces a dilemma: ASML’s equipment is essential for expanding its AI chip and advanced CoWoS packaging capacity, but the company also wants to keep costs under control while securing additional systems. - China is likely to be hit hardest by the DUV price increases because export restrictions prevent Chinese companies from purchasing EUV systems and certain advanced DUV systems, while alternative suppliers remain limited. Jukan’s Take: TEL and other equipment makers are already pursuing aggressive price increases. In meetings with the media and sell-side analysts, TEL said it is seeking to raise equipment prices by as much as 50% over the long term. Lead times across the semiconductor equipment industry have already increased significantly. Given the circumstances, it is only natural that toolmakers would want to raise prices. Still, it is somewhat surprising that TSMC is already pushing back. TSMC appears to be extremely focused on protecting its margins. The Information: Apple Hunts for AI Chip Acquisitions - Apple is actively searching for acquisition targets, including semiconductor startups, to strengthen its in-house AI server chip capabilities. - Its current M2 Ultra-based servers have limitations when processing large AI models, forcing Apple to run some features of the new Siri on Nvidia chips hosted in Google Cloud. - Delays to Apple’s next-generation AI server chip, “Baltra,” have also fueled concerns that the company’s semiconductor design expertise remains too heavily focused on mobile applications. - Apple is seeking to reduce its dependence on Nvidia through M5 and M7 Ultra-based server chips and a joint development project with Broadcom. However, the M7 Ultra-based product is not expected until 2029. - Given Apple’s recent consideration of major acquisitions, potential changes to its cash policy, and an eventual CEO transition, the company may be more willing than in the past to pursue a large-scale acquisition of a semiconductor or AI company. Jukan’s Take: The delay to Baltra, combined with Apple’s search for semiconductor startup acquisitions, raises the question of whether Apple is effectively admitting that it lacks sufficient expertise in server chips. ETNews: Samsung’s Pyeongtaek Plant 5 Poised to Place More Than KRW 10 Trillion in Semiconductor Equipment Orders - Samsung Electronics has nearly completed the selection of equipment suppliers for Phase 1, the first production line at its Pyeongtaek Plant 5 (P5), and is expected to begin issuing purchase orders shortly. Supplier selection for Phase 2 is also underway, meaning the total order value is expected to significantly exceed KRW 10 trillion. - P5 will be a massive fab featuring six cleanrooms across three floors, with a target of increasing production capacity by more than 50% compared with Samsung’s existing Pyeongtaek fabs. - Its main products will be advanced DRAM, including HBM. Samsung is also considering adding foundry lines depending on market conditions. - The entire facility is scheduled to become operational by 2028, but Phase 1 could begin operations as early as next year if equipment deliveries proceed smoothly. - To address the global memory shortage and expanding AI infrastructure demand, Samsung has asked equipment suppliers to shorten their delivery lead times. - Because equipment purchases account for approximately 70%–80% of total fab construction costs, domestic and international semiconductor equipment companies with strong technological capabilities could be major beneficiaries. iNews24: Omdia Says HBM Prices Will Begin Declining in 2030, With New Fabs as the Key Variable High-bandwidth memory prices are expected to begin declining in 2030. However, supply shortages are likely to persist until new production facilities come fully online.

It would be pretty funny if Hynix missed the quarter immediately after their $7b (or whatever it was) US ADR listing and super bullish roadshow where they made fun of Micron for agreeing to price ceilings in their LTAs. Semianalysis well above consensus for this Q, KIS spec sales evidently below. Time will tell!

SK Hynix has reportedly removed price caps in its long-term memory supply agreements, unlike Micron, which will allow spot price increases to be fully reflected in its contract pricing. $MU $NVDA $AMD

China’s government is working with state-run Huawei and state-invested SwaySure to set up a massive DRAM complex in Shenzhen with 140,000 wafers-per-month (wpm) capacity (12-inch) to manufacture 28nm DRAM, media report, adding former TSMC fab director Steve Liu has been hired as CEO, and the former president of Japan’s Elpida Memory (bought by Micron), will be head strategist for the project. SwaySure markets itself as a niche memory maker, but supply chain reports say it is mainly a ‘shadow fab’ for Huawei, as are Pengxinwei (PXW) and Pengxinxu (PXX), which are undergoing massive fab building currently. SiCarrier is Huawei’s main fab partner. #semiconductors

[Exclusive] Samsung Officially Denies U.S. ADR Plans—but Launches Behind-the-Scenes Review Samsung Electronics has reportedly begun internally reviewing scenarios for issuing American depositary receipts (ADRs) in the United States, even examining the practical procedures involved. Although the company has officially denied reports that it is considering such a move, working-level staff have begun assessing the potential costs and benefits and the necessary procedures at the direction of senior management. Early Stage of Examining Possible Structures and Procedures According to semiconductor industry sources on July 15, Samsung Electronics executives recently instructed relevant departments to examine whether a viable structure could be created for issuing ADRs in the United States. Finance, investor relations and other relevant teams are consequently identifying the tasks and procedures that would fall within their respective areas if the company were to pursue a U.S. listing. The review is understood to be a preliminary study assessing feasibility, rather than an indication that the company has decided to list or determined the timing or size of an issuance. Each department is examining the requirements for a U.S. stock-market listing and the internal preparations that would be necessary. As part of this process, Samsung Electronics staff have reportedly requested information from SK hynix, which recently issued ADRs and listed on Nasdaq. The request was intended to gather information about the preparation process and practical experience of SK hynix, which completed its U.S. listing ahead of Samsung. This suggests that Samsung’s internal review has progressed beyond examining the relevant rules and procedures on its own to studying the experience of a company that has actually completed such a listing. Bloomberg reported on July 14, local time, that Samsung Electronics had held preliminary discussions with investment banks about a potential U.S. listing through ADRs. It said the discussions were at an early stage, with no specific decisions made regarding underwriters or the issuance structure, and that they might not ultimately lead to a listing. Immediately after the report, Samsung Electronics stated, “We are not considering a U.S. stock-market listing through the issuance of ADRs.” SK hynix Also Distanced Itself from Speculation Before Going Public with Its Plans Some observers, however, argue that Samsung Electronics’ official denial does not necessarily mean that the possibility of a future ADR issuance has been completely ruled out. Companies often refrain from disclosing preliminary reviews of overseas listings or large-scale securities issuances because transaction terms and schedules can change before formal procedures such as board approval and regulatory disclosures begin. When speculation about a U.S. listing by SK hynix first emerged, the company also maintained that no specific decisions had been made. It later proceeded through its internal decision-making process and filings with the U.S. Securities and Exchange Commission, gradually formalizing its ADR issuance and Nasdaq listing. Until the company publicly announced its listing plan, it distanced itself from market speculation and disclosed the initiative only after beginning the formal process. On July 10, SK hynix issued ADRs at $149 per share, raising approximately $26.5 billion, or about 40 trillion won. This represented the second-largest U.S. stock-market listing on record, behind SpaceX, which raised $85.7 billion through an initial public offering last month, and the largest ever by a foreign company. “ADRs can be structured in a variety of ways depending on whether new shares are issued and which listing method is used,” a semiconductor industry official said. “My understanding is that following the success of SK hynix’s U.S. listing, Samsung Electronics also began internally examining the structures available to it and the procedures that would be required.”

@KevInvestingYT Scale up CPO on GPU is delayed as we said. ¯⁠\\⁠_⁠(⁠ツ⁠)⁠_⁠/⁠¯ Not my job to care about stocks. My job is to track the supply chain accurately

Lots of talk about UHP lasers for CPO, and not enough about what makes it so hard to implement. This article fixes that. We will establish the basics of high power laser physics, and in the future, examine LITE, COHR, etc. https://t.co/fXlxMe0VXn https://t.co/jdy1EiafFw

He literally brought a sheep in his Mercedes convertible. What a legend.

Aehr’s latest results exceeded my expectations across all three key validation points I outlined in my $AEHR report published last March. I believe this earnings release carries significance far beyond the quarterly results themselves. AEHR’s bookings, customer transitions into volume production, and capacity expansion plans serve as powerful leading indicators of where optical production is heading before that demand is fully reflected in the revenues of optical component and module suppliers. In this article, drawing on AEHR’s results and what I have been hearing from industry contacts in Silicon Valley, I explain in detail what this quarter really means, where we currently stand in the optical production cycle, and how I believe investors should position themselves across the optical value chain over the next 12 to 24 months. I am hopeful that this earnings release could serve as a catalyst for renewed momentum in the optical theme, particularly as more optical companies report their results throughout July and August. Full article: https://t.co/ImnCsafZuG

@manifesting93 @SemiAnalysis_ We said scale up CPO on the GPU. Not CPO generally.

First \Mass-Produced\" Silicon Photonics \"Taiwan’s second-largest contract chipmaker

Nvidia $NVDA says a Vera Rubin NVL72 rack paired with a Groq 3 LPX rack can deliver up to 35X higher token throughput per MW, creating up to a 10X revenue opportunity for AI factories versus Blackwell. Meanwhile, acquiring 1 TB of HBM through Blackwell GPUs costs more than 5X as much as adding 1 TB of DDR5. The I/O Fund’s latest two-part analysis breaks down why offload engines may become a key solution for scaling memory independently of GPUs. https://t.co/uPYR85mgXx

Inference processing power in $NVDA Rubin is 5X higher than Blackwell, but HBM bandwidth rises only 2.8X, leaving the memory wall firmly intact. As hyperscalers move from AI buildout to AI monetization, they have to use memory more efficiently. Our latest piece breaks down why offload engines are emerging as a key solution for improving tokens per watt. https://t.co/tb9aTVB6Q2 $GOOG

UPDATE: This is a fascinating post by @satyanadella building on Alex Karp’s point about AI sovereignty. Satya names it the Reverse Information Paradox: enterprises don’t just pay for intelligence with money — they pay again by feeding frontier models their proprietary knowledge, corrections, traces, and evals. That institutional know-how then compounds inside the provider. Karp diagnosed what technical customers actually want: control over their compute, models, data stack, and alpha — to own the means of production rather than have it transferred. Satya explains why the current regime structurally does the opposite, and what enterprises must do about it: establish a real trust boundary with private evals, proprietary learning loops inside the tenant, decoupled orchestration, and the explicit right to fine-tune on their own outputs. That’s how your alpha compounds for you instead of leaking to the model layer. https://t.co/TtQzBxidx6

@EricLDaugh Day 1. Much bigger things coming. But this is the magical unlock that makes everybody a direct owner & participant in our capital markets rather than increasingly dependency on the govt. This is economic freedom! EVERY child a shareholder. ������������������

After listening to $AEHR ’s latest earnings call, I’ve become even more convinced of my bullish view on $AMD.

“Every wafer shifted to feed Nvidia’s H100s [my comment: and primary Blackwells] or AMD’s Instinct accelerators cuts into supply for DDR4. That deliberate reallocation throttles supply, which in turn forces system builders and distributors to stockpile what’s left…” “Major OEMs—including ASUS, MSI, and other leading motherboard and PC manufacturers—engaged in panic buying of RAM stocks.” “This pricing surge is more about supply being deliberately throttled than demand suddenly spiking.” “Micron stops selling memory to consumers as demand spikes from AI chips, noting that memory was facing “a global shortage” and that AI accelerators consume large quantities of memory—Nvidia’s GB200 carries 192GB per GPU and Google’s Ironwood TPU requires 192GB of HBM, against the roughly 16GB in an ordinary laptop.” — Garciaguirre v. Samsung Electronics, filed June 25, 2026.

Legacy Media types are calling this Alex Karp interview a “crash-out” so that’s your first clue that he is actually saying something extremely insightful. He is articulating what real “AI safety” looks like in the enterprise. Not abstract alignment research or certification by a government-run DMV for AI. Real AI safety for businesses is the ability to control their own data, model weights, and compute — so a frontier lab can’t hoover up their proprietary knowledge and turn it into their next product. As Karp explains, technical customers want “control over their compute, their models, their data stack, and their alpha. They want to know they own the means of production, and it’s not being transferred to someone else.” Don’t think that can happen? Just look at Figma. According to The Information, Anthropic “blindsided” its then-business partner with the launch of Claude Design. Figma’s founder said Anthropic had not been “consistently honest” with them. Anthropic’s chief product officer had even served on Figma’s board until three days before the launch of Claude Design. Figma’s stock has fallen sharply this year while Anthropic’s valuation has surged. This isn’t an isolated example. Anthropic has launched Claude Science, Claude Security, Claude Legal, and of course Claude Code — each expanding into categories previously served by companies building on top of their models. The pattern is consistent: watch where value is being created, then move in directly. Dominate the model layer, then use that position to capture the most lucrative verticals. Dario has argued that open source models powerful enough to compete with Anthropic are “dangerous.” But dangerous to whom? Not to enterprises that want to retain control over their data and workflows. Dangerous to a business model that benefits from customers having few real alternatives at the model layer. As Karp exposes, true enterprise safety isn’t trusting that a lab’s future roadmap won’t include your business. It’s retaining the ability to choose — at the model layer — who gets to see and use your alpha.

Narrative violation: A new study of 21,559 firms in the U.S. finds that “companies that adopt AI tend to grow faster following adoption”. “Firms making the largest AI investments grow employment by roughly 10% following adoption, while low-intensity adopters see no statistically significant change.” “Entry-level headcount rises 12% for high-intensity adopters.” “Gains emerge gradually and are broad across roles, including engineering, sales, administration, and customer service.” “The results counter predictions that AI adoption will lead to broad job loss.” The study is based on observed AI spending from Ramp card and bill pay data linked to Revelio Labs workforce records.

I really don’t think there is a problem here at all. Plenty of states excited about the opportunity for new investment and progress. Flat out ban and presser keeps people from wasting time. Play on. All good.

@RepPatHarrigan Turns out Americans want their kids to be shareholders on Team America! ������������������

https://t.co/Yh1mCoPGAn

I talked with @RyanGreenblatt today about his and his coauthors' AI 2040 work; very fun and productive at least for me. I won't quote him but I think where I personally landed was as follows * AI is indeed a policy emergency and our policymakers and institutions need to fully wake up; the AI 2040/2027 folks have done a lot to address this * But for me AI safety is as much an epistemic emergency -- in which we need to rapidly build institutions that allow us to observe, orient, decide, and act around AI safety risks and opportunities -- as it is a policy emergency * Unlike the AI 2040 folks who believe they can predict societal effects like mass unemployment and dangerous goal-driven superintelligence with enough confidence that they can recommend policy now, I see the next decade, epistemically, as a very dangerous speed run of the uncertainties and dislocations of, say, the 1865-1945 period in the US that nobody in 1865 could have predicted * To wit, could anyone have predicted the role of the internal combustion engine at the beginning of that period in the emergence of tank and air warfare that happened between 1914-1945? With transformative technologies, the problem isn't to predict stuff that's impossible to predict, it's to build institutions able to adapt to the unpredictable in ways that serve society * Another analogy I like is to the Ukraine/Russia war in which they're speedrunning robotics and AI innovation within a complex adaptive competition that rewards agility and resists prediction * Anyways, there are indeed places we know enough and need immediate policymaking; rules around who can train on what data and who gets compensated, on model distillation, on where we can export our most powerful GPUs, and rules that encourage or even require AI diffusion and adoption (in the case of cyberdefense) * But I would still contend that we have no idea what the 5-10 years are going to look like. We need agile and powerful institutions capable of acting within shorter time horizons here. If recursive super-intelligence and misalignment and mass unemployment emerge as the major risks, such institutions can then deal with them as they move over the horizon of actionability * @sebkrier has a good essay in the same vein as this (I think?) arguing we need to 'muddle through' (the essay is way better than that and I'm not doing it justice) -- https://t.co/PUo7nwJwdP -- I think muddling through actually means building the best AI safety OODA loop in government and civil society ever * The UK AISI and US CAISI are examples in embryo of part of what we need. But if they are to be what we need they need to become not just safety labs experimenting on pre-release checkpoints of models but also safety observatories concentrating information about AI's safety effects out in society; trusted sources of truth for policymakers so they can act way faster * Today, because we aren't in the regime I'm describing, we're experiencing costly safety false positives based on policy confusion * For example, in cybersecurity, the bans/delays/restrictions on Mythos/Fable/GPT-5.6 were widely seen as ridiculous by much of the cybersecurity defender community (who understand access to these models benefitted defenders more than attackers) and this community is now exploring adopting Chinese open weights models into our supply chains as a backup in case future US model don't let us ask about factoring products of primes. * Let's not keep making these kinds of mistakes! If you've read this far, thanks, and thanks @RyanGreenblatt for the really helpful discussion

The mega bull case for AI infrastructure would be *if* market share shifted away from certain frontier labs with 90%+ inference margins toward cheaper models, whether open-source or closed. It would increase the ROI on AI spend for end customers by increasing intelligence per dollar, which would drive incremental token demand. Margin dollars would effectively get redistributed from the frontier labs to AI infrastructure providers. The infra winners would be those with the lowest per token cost and the winners at the model layer would be those with the highest token efficiency. There are many reasons Jensen is so focused on open source, but this is likely the most important one as I think he is probably less worried about a monopsony these days. Lower margin % at the model layer = more margin $ at the infra layer all else equal. With SpaceX and Meta being vertically integrated and possessing the #3 and #4 models respectively it is more possible than ever. Note that Grok 4.5 is ahead of Fable for some useful tasks at a much lower cost, so ranking them #3 is conservative. This is not happening yet. Cheap, mostly open source tokens are likely the majority of volume today but the majority of economic value is still accruing to the most intelligent models. Might change though. We will see.

A tragic reminder - get your CAC scan - the mammogram for the heart. Cardiac arrest / heart disease is the #1 killer & largely preventable. $150 & 15 mins to save your life. In honor of Sen. Graham we should immediately cover w insurance & launch a natl ad campaign. ������ @DrOz

At some level the bull case for both humans and trad software is intelligence per watt.

In case you’re wondering what happened to Prof Jiang he just popped up on my instagram stories feed as part of a paid fashion brand ad. Well deserved considering how poorly his Iran predictions have held up. https://t.co/OtiBZjt4oy

\I believe that burn in is by far the fastest growing segment in all of semiconductor test right now. And wafer level burn in in particular.\" Gayn Erickson $AEHR CEO. Indeed."

TSMC 2nd quarter (Q2) conference on Thursday (7/16), media report expectations: Q2 revenue NT$1.27 trillion (already known, US$ not known) -Net profit seen up 59% to NT$632.6 billion (US$19.6B), 10th straight quarter of growth -Gross margin could exceed 68% -Operating margin seen hitting new high -R&D expense seen hitting record high due to advanced process investments Q3 revenue forecast seen +9%-12% vs Q2 in USD terms (some 15% estimates) Full Year -Raise full year revenue to mid-30s% increase from current ‘above 30%’ growth, most bullish scenario sees 40% -Unlikely to raise 2026 capex (some see $58B-$60B) but could discuss plan for next 3-years 5-year CAGR forecast 2024-2029: -AI semiconductor revenue CAGR for 2024–2029 seen raised to 70%-80% from mid-to-high 50% now TSMC Official Q2 Guidance -Revenue US$39.0B to $40.2B -Gross Margin: 65.5% to 67.5% -Operating Margin: 56.5% to 58.5% -Exchange rate forecast NT$31.7 per US dollar. What investors want to know: -Overall AI demand health -CoWoS packaging capacity expansion -N2/N3 ramp and pricing -Is memory shortage hurting demand for logic chips? -Update on global construction projects -Update on advanced processes and packaging (CoPoS, SoIC, etc) R&D -TSMC view on intensified competition from Intel, Samsung -Will TSMC raise prices again in 2027? (5%-10% increase expected) -Gross margin projections for 2026 and 2027 -Capex strategy for next 3-years -Semiconductor industry outlook Links: https://t.co/jRoAHHj7IB https://t.co/tUdnIEcJ0n https://t.co/8bA7qCjV4q https://t.co/HOIoQDCtl7 https://t.co/0GTALbLkiM https://t.co/WOgHeoAhVA https://t.co/htUJGlFjSR https://t.co/ZPOmnZMCRf https://t.co/bJozxXJDKE

I try to spend most of my time finding the best secular growth tech product cycles that will crush earnings over the next 12 months like Dell and Credo, but I loved this piece on IBM's mainframe potentially being disrupted by AI. Finally! My question is what does this mean for Computer Associates at Broadcom? ������

Or he failed to keep it together (Anthropic, Thinky, Core Automation, etc) and bloviated so much about his tech that everyone saw it as existential and it created completion from all the hyperscalers. If he played his cards right I'd bet there'd be no Anthropic or Meta TBD or XAI

Every morning when you wake up you should have a clear view on three things: A) AI lab ARRs exiting 2027 - I’d guess buyside is somewhere around $250b exiting ‘26 and $450b exiting ‘27. I believe the latter number is closer to a trillion. B) hyperscaler capex in ‘27 through ‘30. At this point even the retards understand that sellside estimates are perennially too low and every quarter we play the fun game of sellside raising estimates +10% into the print and then another +10% post print when invariably company mgmt makes clear they will build the machine god or die trying. More important is having a clear view on when or if there is a peak in ‘28 or an ever escalating cadence of spend. (Demand for intelligence is unlimited) C) Revenue per GW - self explanatory. Spend time on the neoclouds even if you don’t plan to trade them. Understand why Silicon Index on BBG isn’t fully representative of compute supply-demand etc. If you can’t recite exactly how a memory spike, scaling laws halting, a breakthrough from Deepseek or aliens showing up on the White House lawn will impact these three things while you’re blindfolded you have work to do and aren’t equipped to trade AI related names. Good luck homies ������

If there was any doubt that Anthropic was playing the pull-up-the-ladder Machiavellian regulatory capture playbook, there is none now. Cutthroat business sharks.

Instead of looking at AI chip spot prices, look at what companies are actually paying. xAI's IPO disclosure gives some insights: - Anthropic pays $5.3/GPU-hr (~2x spot) - Google pays $11.5/GPU-hr (~3 to 5x spot) That's the premium for speed, cluster size, etc.

Apple Sues OpenAI for IP Theft, AI Price War Breaks Out https://t.co/s7Do2IirUT

@RobotChocobo @ren_stocks I own no public equities nor does SemiAnalysis We have a strict compliance program + department Fluidstack reporting is seperate from my interest, but yes weve discussed facts incl Anthropic deals or datacenter progress inclusong some delays https://t.co/9jYaAHL4M9