Cycle on Sep 14, 2026

I will start by listing the active research threads and watches for this topic to understand what has been tracked previously and what specific angles we need to address or resolve.

I will run a web search to check for updates on the Micron Taiwan union strike vote, which was scheduled to take place in September 2026.

This is a massive development. Micron offered up to 68 months' pay in bonuses to avert the strike in Taiwan. Let's fetch the RTE, Reuters, or Yahoo Finance article to get the exact details and quotes. I'll search for the Yahoo Finance article or RTE article first, or fetch `https://www.rte.ie/news/business/2026/0911/1591142-micron-workers-in-taiwan/`. Let's fetch that.

Interesting. So Micron offered up to 68 months' pay (equivalent to 5 years and 8 months of pay, with minimum NT$1.7m cash compensation) for fiscal 2026, but the Taoyuan union *rejected* the proposal on September 11, 2026, saying it "sidestepped" the discussion of a structured profit-sharing system (they want 15% of operating profits distributed quarterly). They are still moving toward a strike vote if negotiations fall apart. Let's check the status of SK Hynix labor negotiations as well. One of the watches is: "SK Hynix and Union Wage/Bonus Agreement Resolution". Let's search for "SK Hynix union wage bonus agreement September 2026" or similar.

Let's fetch the article from Human Resources Online or Aju Press on the revised deal to see if it has been finalised. Let's do a search for the latest on "SK hynix union vote September 2026" to see if the union voted on this revised tentative agreement.

This is highly relevant! The vote is scheduled for September 15th and 16th (which starts tomorrow, since today is September 14, 2026). Let's fetch `https://www.humanresourcesonline.net/sk-hynix-revises-wage-deal-boosts-cash-share-of-performance-bonuses-to-50` to read the details.

Excellent! We have two major union/labor developments: 1. Micron offered Taiwan workers up to 68 months' pay (5 years and 8 months of pay, minimum NT$1.7m cash compensation) for fiscal 2026 to avoid a strike. However, the Taoyuan union rejected the proposal on September 11, 2026 because it sidestepped discussions of a structured quarterly profit-sharing system (they want 15% of operating profits). Government-mediated negotiations are ongoing, and if they fall apart, the union will vote on whether to strike (an internal survey showed 80% backing strike action). 2. SK Hynix revised its tentative wage agreement to pay performance bonuses in a 50/50 split of cash and company stock (up from 40% cash / 60% stock in the rejected late August deal). The vote on this revised proposal is scheduled for September 15th and 16th, 2026. Let's check if there are any updates on Samsung's labor union merger (NSEU into the broader Office Workers' Union on October 5, 2026). Let's search for "Samsung National Samsung Electronics Union merger October 2026" or similar.

No new updates on the Samsung NSEU merger yet, as the date is set for October 5, 2026. Let's now turn our attention to the market and financial data. Let's call `market_view` for Micron (`MU`), SK Hynix (let's check if we can get a view for SK Hynix, or if we need to search or use ticker `000660` or similar, or just focus on Micron `MU` first), Samsung (`005930` or similar), and the sector. Let's check `market_view` with kind='ticker', id='MU'. Let's do that.

Wow, look at Micron's numbers. TTM Revenue is $90.27B (+345.7% YoY). Gross margin is 72.6%. Operating margin is 80.4%. Net income for the quarter ending May 31, 2026 was $28.24B (on revenue of $41.46B). This is an absolute massive late-cycle boom! Wait, let's look at the news in the market view: - "SK Hynix, Samsung Shares Dip In Korea After DeepSeek Debuts AI Tech That Uses Less Memory – Micron, SanDisk Hold Up" (2026-09-12) -> This is highly interesting! DeepSeek debuted a new model (V4.1 Flash?) that requires significantly less HBM, raising concerns about HBM demand. Let's research this. - "An Intel-Backed Startup Says It Can Beat HBM. Is Micron’s AI Memory Boom Vulnerable?" (2026-09-11) -> Kepler Computing emerged claiming new memory architecture can outperform HBM. Let's do a web search on DeepSeek's new AI model and its impact on HBM demand, and Kepler Computing. Let's do that.

This is a very important story. Let's fetch the Yahoo Finance article `https://finance.yahoo.com/technology/ai/articles/deepseek-cut-kv-cache-hbm-011603128.html` or the Stocktwits/Tradingview article to get the exact details. Let's do a search for the exact text of the Yahoo Finance article or fetch `https://finance.yahoo.com/technology/ai/articles/deepseek-cut-kv-cache-hbm-011603128.html`. Let's try to fetch that. If it fails, we'll try another.

This is an incredibly rich article! Let's break down the key points: - **DeepSeek's V4.1-Flash** (released on September 10, 2026) claims to cut High-Bandwidth Memory (HBM) requirements for its **KV cache** (which stores attention state during inference) by **75%** (requiring only one-quarter as much) and SSD capacity for KV cache by **87.5%** (requiring only one-eighth as much) compared to its previous generation. - The comparison is specifically for KV cache (attention state during inference), not the entire model weights or training. However, because inference is where sustained token growth occurs, architectural improvements in KV cache efficiency can act as a demand shock to memory-intensity forecasts. - The memory-shortage thesis assumes AI usage and memory demand rise together, but this unit-efficiency breakthrough shows that relationship is weaker. - This has rattled Asian memory stocks: Samsung fell 3.5% and SK Hynix fell 2.2% in South Korea on September 11, 2026, though US-listed Micron and SanDisk held up better in overnight trading. - Let's look at the financial numbers cited for Micron and Sandisk in their latest quarters: - Micron's **Cloud Memory revenue** reached **$13.77 billion** in its latest quarter with an **83% gross margin**, while **Core Data Center revenue** was **$11.52 billion** at an **87% margin**. It also shipped more than **$1 billion of HBM4**. - Sandisk's quarterly data-center revenue rose 103% sequentially to $2.98 billion, and pricing generated roughly two-thirds of the company's sequential revenue increase. - This shows extraordinary late-cycle margins and pricing power that embed scarcity. If model designers broadly cut bytes required per token, this pricing power could quickly reverse. Wait! Let's also check Kepler Computing. The article mentioned "An Intel-Backed Startup Says It Can Beat HBM. Is Micron’s AI Memory Boom Vulnerable?" Let's search for "Kepler Computing" memory HBM September 2026.

This is absolutely fascinating! Kepler Computing emerged from stealth on September 10, 2026, with a massive **$468 million** (or $470 million) funding round. Its core technical proposition: - Alternative to HBM and SRAM using **Ferroelectric RAM (FeRAM / FeDRAM)** and a novel **3D stacking technique** that places memory directly on top of logic. - It bypasses the need for cutting-edge EUV lithography tools, meaning it can be manufactured in older, less constrained fabs (e.g., using GlobalFoundries' 28-nanometer process). - It claims to achieve higher density and performance than conventional HBM and SRAM, potentially disrupting the HBM supply crunch. - They are sampling in **2026** and targeting production in **2027**. Let's fetch `https://endroid.com/2026/kepler-computing-hbm-memory-euv-bypass-468-million/` or `https://overclock3d.net/news/memory/kepler-aims-to-end-the-ai-memory-crisis-with-custom-memory-tech/` to get more technical detail and direct quotes. Let's try `https://endroid.com/2026/kepler-computing-hbm-memory-euv-bypass-468-million/` first. If that fails, we can fetch another.

This is a massive development. A $468 million funding round (including $245M CHIPS Act funding, $50M from GlobalFoundries, plus Intel Capital, AMD Ventures, Baillie Gifford, and Gates Frontier Fund) for Kepler Computing, which emerged from stealth on September 10, 2026. Wait, let's look at the dates: the article says "emerged publicly in 2024" in one paragraph, but the news was published "September 10, 2026" and mentions "emerged from stealth with a $468M funding round". Let's verify if there is any other source to double check the timing. Actually, the article title is "HBM Shortage Hack: Kepler’s $468M Play to Bypass EUV" published "September 10, 2026". The WIRED article was also published around September 10, 2026 (or September 9, 2026). This is a fresh exit from stealth with a huge funding round. Wait, let's look at the market dynamics. The global memory market is currently experiencing an extraordinary boom, with Micron's latest quarter (ending May 31, 2026) showing $41.46 billion in revenue (up 345.7% YoY) and a net income of $28.24 billion! Its gross margin is 72.6%. But we have two major potential disruptions/threats to this "permanently sold out" and "unbounded pricing power" narrative: 1. **DeepSeek's V4.1-Flash** (released September 10, 2026) claims to cut HBM requirements for its KV cache by 75% and SSD requirements by 87.5% through architectural efficiency. Since inference is where sustained token growth occurs, this unit-efficiency breakthrough could act as a demand shock to memory-intensity forecasts. 2. **Kepler Computing** emerged from stealth on September 10, 2026, with $468 million in funding (including a $245 million CHIPS Act grant) to build a 3D-stacked ferroelectric memory alternative to HBM and SRAM that bypasses EUV lithography entirely and can be manufactured on GlobalFoundries' 28nm nodes. They are sampling in 2026 and targeting production in 2027. Let's do a search on TrendForce contract pricing updates for DRAM and NAND Flash in late Q3 2026. This will give us a view of whether the "unprecedented structural shift" and "pricing soaring" is starting to show any signs of cooling or if it's still extremely strong. Let's search for "TrendForce DRAM contract price September 2026" or "TrendForce NAND contract price Q3 2026".

Let's look at this! TrendForce raised its Q3 and Q4 DRAM contract price forecasts! - For Q3 2026, PC DRAM price forecasts were raised to a **15%–20%** QoQ increase (up from 8%–13%). - For Q4 2026, PC DRAM price forecasts were raised to a **3%–8%** QoQ increase (up from 0%–5%). - Server DRAM contract prices are expected to rise by **13%–18%** QoQ in Q3 2026. - The DRAM Global Hi-Tech Industry Research Report from September 9, 2026 says: "With rock-bottom supplier inventories, average prices rose 53–58%, pushing the supply shortage and price hikes through 2027." Let's fetch `https://www.trendforce.com/presscenter/news/20260709-13140.html` to see the details on Server DRAM and LTA (Long-Term Agreements) capping price increases. This will help us understand the structural dynamics of these supply-cycle mechanics.

Excellent! Let's see what else we can find. We have a lot of great information here. Wait, let's look at the active threads and see if we can resolve any of them. The threads are: 1. `01a056dc-ee70-792d-9ba9-166eccfd4623` - Micron Taiwan Union Strike Vote and Resolution. - We have definitive news that Micron offered up to 68 months' pay in bonuses to avert a strike in Taiwan, but the Taoyuan union rejected the proposal on September 11, 2026. This means the strike threat is still active and negotiations are ongoing. - We should record this event! Wait, is the watch resolved? No, because the strike has *not* been averted yet, and the strike vote is still a possibility if negotiations fall apart. So we should record an event on this thread, but keep it open (`resolves: false` or `resolves: true` depending on if the watch is ongoing or one-shot. Wait, the watch description is: "Fires if Micron's Taiwan unions officially vote to strike or if a final collective bargaining agreement is reached, averting the strike." Since neither has happened yet—the union rejected the proposal and negotiations are ongoing—the watch hasn't fully resolved, but we can record an event to update the status, or keep it open. Actually, let's check if there are other threads we can resolve or update). 2. `01a00ec4-04a8-7d1c-8372-6e220f8231d5` - SK Hynix and Union Wage/Bonus Agreement Resolution. - Under the revised tentative agreement, SK Hynix agreed to pay performance bonuses in a 50/50 split of cash and company stock (up from 40/60). - The union is scheduled to vote on this revised proposal on September 15 and 16, 2026. - This watch is also active and the vote is tomorrow. So it hasn't resolved yet (the final agreement is pending the vote), but we can record an event or update the thread. 3. `019f1270-a8f7-7130-94c6-149ad655c192` - Micron FQ4 2026 Earnings and SCA Cash Deposit Verification. - The fire date is September 25, 2026. Let's check if Micron has reported FQ4 2026 results yet. Today is September 14, 2026, so the earnings release is likely in late September (around September 24-25, 2026). This is still in the future. 4. Let's look at the other watches. None of the other watches (e.g., Samsung Taylor Fab 2, SK Hynix Yongin Phase 1, Nanya Fab 5A EUV, NVIDIA Rubin Ultra specifications, NAND Flash supply-demand balance, etc.) have fired yet as their dates are later in 2026 or 2027. Wait, let's write up some very strong findings based on what we've discovered: 1. **Micron Taiwan Union Strike Threat & Unprecedented 68-Month Bonus Offer (slug: `micron-taiwan-union-strike-threat-bonus-dispute` or update `dram-contract-pricing-skyrockets-to-record-highs-q1-2026` or a new slug)**. - Wait, let's check if we have any existing notes that we can update. - Let's read the current notes using `read_notes` first to see their exact slugs and titles. This is a critical step to ensure we reuse slugs correctly! Let's do that.

This is excellent. We have existing notes that we can update! The relevant existing notes are: 1. `micron-taiwan-labor-strike-threat` ("Micron Taiwan Labor Dispute: Strike Threat Escalates Over 83-Month Bonus Demand") 2. `sk-hynix-labor-dynamics-stock-heavy-bonuses` ("SK Hynix Union Rejects Wage Deal Shifting 60% of Bonuses into Stock") 3. `dram-contract-pricing-skyrockets-to-record-highs-q1-2026` ("DRAM and NAND Contract Pricing Soars as Memory Approaches 68% of CSP CapEx in 2027") - we can update this with the latest TrendForce Q3/Q4 DRAM contract price forecast hikes and server DRAM LTA dynamics. 4. Let's create a NEW note for **DeepSeek's V4.1-Flash KV Cache Breakthrough & Kepler Computing's $468M Stealth Exit** which represent structural demand/supply-side threats to the memory cycle's "permanently sold out" narrative. Wait, should we put them in one note or two notes? The rules say: "For each distinct theme or story worth persisting, call write_note — ONE note per theme, never a combined write-up." - Theme 1: DeepSeek's V4.1-Flash KV-Cache memory efficiency breakthrough and its implications for HBM demand intensity. Let's call it `deepseek-kv-cache-efficiency-hbm-demand-threat`. - Theme 2: Kepler Computing's $468M stealth exit and EUV-bypass FeRAM 3D-stacking architecture as a long-term threat to the HBM supply bottleneck. Let's call it `kepler-computing-feram-euv-bypass-hbm-disruption`. - Theme 3: Update `micron-taiwan-labor-strike-threat` with the 68-month bonus offer and union rejection. - Theme 4: Update `sk-hynix-labor-dynamics-stock-heavy-bonuses` with the revised 50/50 tentative wage deal and upcoming Sept 15-16 vote. - Theme 5: Update `dram-contract-pricing-skyrockets-to-record-highs-q1-2026` with the TrendForce Q3/Q4 price hike revisions and LTA (Long-Term Agreement) dynamics. This is a very clean and comprehensive set of notes! Let's write them. First, let's update `micron-taiwan-labor-strike-threat`. Let's draft it. Wait, let's review the exact quotes and links for Micron's Taiwan labor dispute: From the RTE article: - Quote 1: `"Taiwan direct labour team members will receive total rewards equivalent to 35 to 68 months of pay, with minimum cash compensation of NT$1.7m" ($53,700) for fiscal 2026, Micron said in a statement.` - Quote 2: `In a statement, the union representing workers at Micron's Taoyuan plant rejected the company's proposal today, saying it had "sidestepped" discussion of the bonus system. "We are seeking a system that allows for a reasonable sharing of profits when the company is profitable and provides institutional safeguards during economic cycles," the union said. "If the company fails to respond to these demands, the union will continue to move towards a strike."` - URL: `https://www.rte.ie/news/business/2026/0911/1591142-micron-workers-in-taiwan/` Let's call `write_note` for `micron-taiwan-labor-strike-threat`. Let's state our intent and write it.

Now let's update `sk-hynix-labor-dynamics-stock-heavy-bonuses`. Let's review the details for SK Hynix: - In late August 2026, SK Hynix union members rejected a tentative wage agreement (by a narrow margin of just 25 votes) because it paid 60% of performance bonuses in company shares and 40% in cash. Workers were concerned about stock price volatility and the stock-heavy compensation structure. - In response, SK Hynix and the union renegotiated and reached a revised tentative agreement in early September 2026. - Under the revised deal, the performance bonus payout ratio was adjusted to an even **50% cash and 50% stock** split. - It also brought forward deferred bonus payments linked to last year's performance and expanded eligibility for housing loan support. - The company-wide union vote on this revised proposal is scheduled for **September 15 and 16, 2026**. - Quotes: - `"Under the revised tentative agreement, the payout ratio for performance bonuses will be adjusted from 40% cash and 60% company stock to an even split of 50% cash and 50% stock."` - `"The union is expected to conduct a company-wide vote on the revised proposal from 15 to 16 September."` - URL: `https://www.humanresourcesonline.net/sk-hynix-revises-wage-deal-boosts-cash-share-of-performance-bonuses-to-50` Let's call `write_note` for `sk-hynix-labor-dynamics-stock-heavy-bonuses`. Let's state our intent and write it.

Now let's update `dram-contract-pricing-skyrockets-to-record-highs-q1-2026`. Let's rename or keep the title, but update the content to reflect late Q3 2026 dynamics. Wait, let's look at the existing note's title and slug: Slug: `dram-contract-pricing-skyrockets-to-record-highs-q1-2026` Title: `DRAM and NAND Contract Pricing Soars as Memory Approaches 68% of CSP CapEx in 2027` Let's see if we can update this note with the latest TrendForce Q3 and Q4 DRAM contract pricing forecast hikes and the server DRAM LTA (Long-Term Agreement) dynamics. Let's review the details from TrendForce's July 9, 2026 press release and subsequent September reports: - TrendForce raised its Q3 2026 PC DRAM price forecast to a **15%–20%** QoQ increase (up from 8%–13%). - For Q4 2026, PC DRAM price forecasts were raised to a **3%–8%** QoQ increase (up from 0%–5%). - Server DRAM contract prices are expected to rise by **13%–18%** QoQ in Q3 2026. - A key structural dynamic: US-based Cloud Service Providers (CSPs) have entered into multi-year **Long-Term Agreements (LTAs)**, which restrict suppliers from raising prices for these clients. - As a result, the primary source of server DRAM price increases in Q3 2026 is shifting toward customers *without* LTAs, as well as incremental supply sold outside LTAs to existing LTA customers. - Despite some CSPs having LTAs, they still have a strong incentive to build inventory in anticipation of a severe server DRAM shortage expected in 2027. - TrendForce's initial estimates predict that total RDIMM bit supply will grow by only 15–20% YoY in 2027, significantly lagging the projected growth in server CPU shipments. - Quote: - `"Consequently, TrendForce predicts that server DRAM contract prices will rise by 13–18% quarter-over-quarter in 3Q26. However, as the market remains undersupplied, memory suppliers may continue revising quotations upward throughout the remainder of this quarter."` - `"During the latter part of the second quarter, major memory suppliers began providing customers with preliminary supply guidance for 2027. TrendForce’s initial estimates predict that total RDIMM bit supply will grow by only 15–20% YoY, significantly lagging the projected growth in server CPU shipments."` - URL: `https://www.trendforce.com/presscenter/news/20260709-13140.html` Let's call `write_note` for `dram-contract-pricing-skyrockets-to-record-highs-q1-2026`. Let's state our intent and write it.

Now let's write a new note for DeepSeek's KV-cache memory efficiency breakthrough. Let's call the slug `deepseek-kv-cache-efficiency-hbm-demand-threat`. Let's review the details from the Yahoo Finance article on September 12, 2026: - DeepSeek released its **V4.1-Flash** model on September 10, 2026. - The core technical claim is that V4.1-Flash requires **75% less HBM** (one-quarter as much) and **87.5% less SSD capacity** (one-eighth as much) for its **KV cache** (key-value cache) than its previous generation. - The KV cache stores the attention state during inference. - The qualification is crucial: this is specifically for the KV cache during inference, not the entire model weights or training. - However, because inference is where sustained token growth occurs, a dramatic architectural reduction in KV cache memory intensity can act as a significant demand shock to long-term memory-intensity and capacity forecasts. - The memory-shortage thesis assumes a linear relationship between AI usage (tokens) and memory demand. DeepSeek's breakthrough shows this relationship can be decoupled by algorithmic efficiency. - This rattled Asian memory stocks on September 11, 2026, with Samsung Electronics falling 3.5% and SK Hynix falling 2.2% in Seoul. - The financial backdrop shows why the stakes are so high: Micron's Cloud Memory revenue reached **$13.77 billion** in its latest quarter with an **83% gross margin**, and Core Data Center revenue was **$11.52 billion** at an **87% margin**, while shipping over **$1 billion of HBM4**. Sandisk's data-center revenue rose 103% sequentially to **$2.98 billion**, with pricing generating roughly two-thirds of that sequential increase. These extraordinary margins and pricing power reflect extreme scarcity; if unit efficiency drops precipitously, that scarcity—and pricing power—could collapse. - Quotes: - `"Its September 10 release says V4.1-Flash needs one-quarter as much HBM and one-eighth as much SSD capacity for its KV cache as the previous generation."` - `"The comparison covers KV cache, which stores attention state during inference, and it is against DeepSeek's preceding architecture. It does not eliminate memory used for model weights or training. Still, inference is where sustained token growth can turn an architectural improvement into a demand shock."` - `"Extraordinary margins and pricing power embed scarcity. If model designers broadly cut bytes required per token, demand can grow rapidly without matching today's memory-intensity forecasts."` - URL: `https://finance.yahoo.com/technology/ai/articles/deepseek-cut-kv-cache-hbm-011603128.html` Let's call `write_note` for `deepseek-kv-cache-efficiency-hbm-demand-threat`. Let's state our intent and write it.

Now let's write a new note for Kepler Computing's $468 million stealth exit and EUV-bypass FeRAM 3D-stacking architecture. Let's call the slug `kepler-computing-feram-euv-bypass-hbm-disruption`. Let's review the details from the Endroid News article and other sources: - Kepler Computing emerged from stealth on September 10, 2026, with a massive **$468 million** late-stage funding round. - This includes a **$245 million** commitment from the US Department of Commerce under the CHIPS Act, **$50 million** from GlobalFoundries (with a manufacturing partnership in Singapore and Vermont), and backing from Intel Capital, AMD Ventures, Baillie Gifford, and Bill Gates' Gates Frontier Fund. - Core technical proposition: **3D-stacked ferroelectric memory (FeRAM / FeDRAM)** alternative to HBM and SRAM. - The major innovation: **bypasses Extreme Ultraviolet (EUV) lithography entirely**. Instead of relying on ASML's highly constrained and expensive EUV tools ($150M+ per unit, 2-3 year lead times), Kepler achieves equivalent density using existing **28-nanometer process nodes** in partnership with GlobalFoundries. - It targets both SRAM (cache memory embedded in CPU/GPU dies) and DRAM-based HBM (discrete stacked memory). - They are currently operating "mini fabs" (pilot production) in Singapore and Burlington, Vermont, to demonstrate manufacturability and yields. - They plan to sample chips in **2026** and enter production in **2027**. - Quotes: - `"Kepler Computing emerged from stealth with a $468M funding round and a claim to bypass EUV lithography entirely for memory chip production, potentially disrupting the HBM supply crunch that’s strangling AI infrastructure buildout."` - `"The company targets two memory classes: SRAM (cache memory embedded in CPU/GPU dies) and DRAM-based HBM (discrete stacked memory for data centers). Both currently require advanced process nodes that demand EUV. Kepler claims it achieves equivalent density using existing 28-nanometer process nodes in partnership with GlobalFoundries."` - `"Government backing signals this isn’t venture theater—it’s industrial policy. The DOC specifically framed the grant around “innovative 3D and ferroelectric technologies,” validating Kepler’s core technical premise before commercial production proves viability."` - URL: `https://endroid.com/2026/kepler-computing-hbm-memory-euv-bypass-468-million/` Let's call `write_note` for `kepler-computing-feram-euv-bypass-hbm-disruption`. Let's state our intent and write it.

This is excellent! We have updated 3 existing notes and created 2 new notes, capturing the latest developments in the AI memory cycle. Let's review our threads and watches to see if we can resolve or record events on any of them. The threads/watches we should check are: 1. **Micron Taiwan Union Strike Vote and Resolution** (`01a056dc-ee70-792d-9ba9-166eccfd4623`): - Status: Micron offered up to 68 months' pay in bonuses to avoid a strike, but the Taoyuan union rejected the proposal on September 11, 2026. Mediation/negotiations are ongoing. This is an ongoing development, but the strike has *not* officially started or been averted yet. - Let's record an event on this watch to capture this major development, but set `resolves: false` since the final resolution (strike vote or final agreement) is still pending. 2. **SK Hynix and Union Wage/Bonus Agreement Resolution** (`01a00ec4-04a8-7d1c-8372-6e220f8231d5`): - Status: SK Hynix revised its tentative wage agreement to pay performance bonuses in a 50/50 split of cash and stock. The union is scheduled to vote on this revised proposal on September 15 and 16, 2026. - Let's record an event on this watch as well, but set `resolves: false` since the vote is tomorrow and we don't have the final result yet. Wait, let's call `record_event` for both of these watches to keep our telemetry up to date. First, let's call `record_event` on the Micron Taiwan Union Strike Vote and Resolution thread (`01a056dc-ee70-792d-9ba9-166eccfd4623`). Let's state our intent first.

Now let's call `record_event` on the SK Hynix and Union Wage/Bonus Agreement Resolution thread (`01a00ec4-04a8-7d1c-8372-6e220f8231d5`). Let's state our intent first.

This is fantastic! Let's double-check if there are any other watches we should register. Wait, we registered some major new developments: 1. DeepSeek's V4.1-Flash KV-cache memory efficiency breakthrough. 2. Kepler Computing's EUV-bypass 3D FeRAM architecture targeting 2027 production. Should we register a watch for Kepler's 2027 production or qualification milestones? Yes, that would be a great forward-looking watch! Let's register a watch for Kepler Computing. - Watch Type: `dated_catalyst` or `tripwire` or `thesis_breaker` or `live_process`. Let's make it a `dated_catalyst` for 2027-06-30 (mid-2027, around when they target production). - Let's call it "Kepler Computing 3D FeRAM Production or Qualification Milestones". - Description: "Monitor Kepler Computing's progress in qualifying its 3D FeRAM memory chips with major chip designers (e.g., AMD, Intel) or entering commercial production, scheduled to begin in 2027." - Cardinality: `one_shot` - Fire Date: `2027-06-30T00:00:00Z` - Topic ID: `019e8ec8-ffb3-71d7-981c-e48354ab25e7` - Cycle ID: `9c31cb65-1004-46c9-953a-754c0a41a59c` Let's call `record_watch` to register this. Let's state our intent first.

This is great. Let's review if there are any other actions to take. We have: 1. Called `list_threads` to understand prior threads and watches. 2. Gathered and confirmed evidence using web research and market view. 3. Updated existing notes (`micron-taiwan-labor-strike-threat`, `sk-hynix-labor-dynamics-stock-heavy-bonuses`, `dram-contract-pricing-skyrockets-to-record-highs-q1-2026`) and created new notes (`deepseek-kv-cache-efficiency-hbm-demand-threat`, `kepler-computing-feram-euv-bypass-hbm-disruption`) ensuring all are grounded in concrete quotes and links. 4. Recorded events on the active labor threads to update their status. 5. Registered a new forward-looking watch for Kepler Computing. We are now ready to call `finish_cycle` and conclude this research cycle. Let's do that.

I have successfully completed the research cycle for this topic. Here is a summary of the findings and updates recorded in this cycle: ### 1. Labor Friction Reaches Unprecedented Heights (Updates to Existing Notes) * **Micron Taiwan Strike Threat Escalates Despite Record 68-Month Offer (`[[micron-taiwan-labor-strike-threat]]`)**: * **What Happened**: In a bid to avert a strike by its 12,000 workers in Taoyuan and Taichung, Micron proposed an eye-watering bonus package for fiscal 2026, offering direct labor team members between **35 and 68 months of pay** (with a minimum NT$1.7 million / $53,700 cash floor). * **The Turn**: On September 11, 2026, the Taoyuan union **rejected the offer**, criticizing the company for "sidestepping" discussions of a structured, permanent profit-sharing system (the unions are demanding a one-off 83-month bonus and a permanent 15% quarterly operating profit-sharing mechanism). * **Status**: Government-mediated negotiations are ongoing, but the union is moving toward an official strike vote (which has 80% preliminary backing from members). A strike would immediately disrupt 50-60% of Micron's global fabrication capacity. * **SK Hynix Management Capitulates on Cash vs. Stock Split (`[[sk-hynix-labor-dynamics-stock-heavy-bonuses]]`)**: * **What Happened**: Following a narrow 25-vote rejection of an earlier contract in late August 2026 (which paid 60% of performance bonuses in stock), SK Hynix and its labor union reached a **revised tentative agreement** on September 10, 2026. * **The Turn**: The performance bonus payout ratio has been adjusted to an even **50% cash and 50% stock split**, alongside greater flexibility for workers to choose higher stock allocations in 10% increments if they prefer. * **Status**: A critical company-wide union vote is scheduled for **September 15 and 16, 2026**, to finalize the deal. ### 2. Market Contract Pricing Continues to Surge (Update to Existing Note) * **TrendForce Hikes Q3 & Q4 Contract Price Forecasts (`[[dram-contract-pricing-skyrockets-to-record-highs-q1-2026]]`)**: * **What Happened**: TrendForce revised its contract pricing forecasts upward in late Q3 2026. PC DRAM price growth was raised to a **15%–20% QoQ increase** for Q3 (up from 8%–13%) and **3%–8% QoQ** for Q4. Server DRAM is projected to rise **13%–18% QoQ** in Q3. * **The Mechanics**: While major U.S.-based Cloud Service Providers (CSPs) have entered into multi-year **Long-Term Agreements (LTAs)** that cap pricing, the burden of price hikes is shifting entirely to non-LTA customers and incremental supply. Even LTA buyers are aggressively hoarding inventory due to an anticipated severe server DRAM shortage in 2027, driven by HBM wafer cannibalization. ### 3. Structural Threats to the "Permanently Sold Out" HBM Narrative (New Notes) * **DeepSeek's Algorithmic KV Cache Memory Squeeze (`[[deepseek-kv-cache-efficiency-hbm-demand-threat]]`)**: * **The Breakthrough**: On September 10, 2026, Chinese AI pioneer DeepSeek released its **V4.1-Flash** model, claiming its architecture reduces the High-Bandwidth Memory (HBM) required for its KV cache (attention state during inference) by **75%** and SSD capacity by **87.5%** compared to the previous generation. * **The Cyclical Risk**: The AI memory supercycle assumes a linear relationship between token generation and memory demand. DeepSeek's algorithmic unit-efficiency breakthrough threatens this core assumption. If widely adopted, this decoupling could act as a severe demand shock to late-cycle pricing power, threatening the extraordinary margins of suppliers (e.g., Micron's Cloud Memory segment currently operates at an **83% gross margin** on $13.77 billion in revenue; see `/markets/MU/2026/09/14`). * **Market Impact**: Asian memory stocks plummeted on September 11, with Samsung Electronics falling **3.5%** and SK Hynix falling **2.2%** in Seoul. * **Kepler Computing Exits Stealth with $468M to Bypass EUV Lithography (`[[kepler-computing-feram-euv-bypass-hbm-disruption]]`)**: * **The Disruptor**: Silicon Valley startup Kepler Computing emerged from stealth on September 10, 2026, with a massive **$468 million** late-stage funding round—including a **$245 million** US CHIPS Act grant, **$50 million** from GlobalFoundries, and backing from Intel, AMD, and Bill Gates. * **The Technology**: Kepler claims its proprietary 3D-stacked ferroelectric memory (**FeRAM / FeDRAM**) architecture can bypass ASML's highly constrained Extreme Ultraviolet (EUV) lithography systems entirely, achieving equivalent HBM and SRAM density on mature **28-nanometer process nodes** in partnership with GlobalFoundries. * **Timeline**: The company is operating pilot "mini fabs" in Singapore and Vermont, targeting chip sampling in **2026** and commercial production in **2027**. If validated and scaled, Kepler could severely disrupt the pricing power of the "Big Three" memory suppliers by bypassing the industry's most expensive physical bottleneck. *** ### Telemetry Updates * **Recorded Event** on the Micron Taiwan Strike Watch (`01a056dc-ee70-792d-9ba9-166eccfd4623`) to capture the union's rejection of the 68-month bonus offer (remains open/active). * **Recorded Event** on the SK Hynix Wage Deal Watch (`01a00ec4-04a8-7d1c-8372-6e220f8231d5`) to capture the revised 50/50 tentative agreement and the upcoming union vote on September 15-16, 2026 (remains open/active). * **Registered a New Watch** for Kepler Computing's 3D FeRAM Production or Qualification Milestones (`01a09ef7-b160-78d5-8825-90545948a8bd`), targeting mid-2027.