Note: After organizing the content, I asked AI to help summarize and edit it.
The expert who participated in this call is a Samsung packaging engineer specializing in HBM hybrid bonding research and development.
With experience spanning equipment, process development, and R&D, he brings a broad perspective across the packaging stack and shared a number of practical insights throughout this discussion.
Another beneficiary of emergence of Muse and other consumer agents
Over the past three months, many market participants have been concerned that memory prices would drop rapidly after peaking, following the path seen in past consumer electronics cycles.
However, we already expressed our view in our report following our attendance at FMS in early August: we believe memory prices are more likely to settle into a longer plateau after reaching their peak.
Established suppliers lead today's OCS ramp, but integrated designs bring new competitors.
Industry and technology analysis, not investment advice · The author may hold positions in the securities mentioned Executive Summary (Preview) An optical circuit switch (OCS) changes connections between fibers without…
It can remove equipment and power consumption associated with intermediate optical-electrical conversion and packet processing.
Raising estimates on the clear personal agent beneficiary
Based on the case study library from our recently published META report, Meta Muse handles a substantial volume of high-frequency tasks, such as cross-platform price comparisons (e.g., up to 146 searches for a single tr…
Because Meta proactively provisions persistent cloud-based execution environments for user agents, relying on headless browser automation and high-frequency API orchestration, the compute workload shifts heavily back to…
Insights from 654 Muse Use Cases and Our Advertising Agency Interview
We’ve been discussing Muse’s potential to reshape Meta’s business with our institutional clients for the past two weeks, especially during our weekly calls.
We thought it was time to share that thinking here, too.
Hot Chips is a 3-day conference held at Stanford where the industry’s chip architects present their newest designs, and it covers memory, CPUs, GPUs, networking, and custom accelerators from an electrical and computer e…
Each day has its own theme: day 1 was memory, day 2 was processing units, and day 3 was systems, networking, and custom silicon.
Compute scarcity forces China’s big clouds to buy most tokens externally, creating a token-factory supply chain the U. S. never needed
China has been talking up its “token economy” in earnest this year.
In March the National Data Administration formally christened the token “Ciyuan” and started using it as a national statistic; at the end of August the Guiyang big data expo ran under a token-themed banner, with NDA chi…
A follow-up Q&A addressing the key questions we collected since our last popular deep dive.
After we published our AIDC report on 27 August, investors questioned our conversion of turbine nameplate capacity into data-center power.
We had estimated that AIDC could claim roughly 18 GW of large gas-turbine nameplate in 2027, but investors felt our haircut to delivered power was too aggressive.
The era of the HBM density peak. A focused analysis of the memory supply and demand outlook for 2028
HBM is a core technology that made the AI revolution possible.
Even though it is more expensive than other memory, chip designers put more and more HBM into their accelerators as a memory solution with high bandwidth and large capacity to process AI workloads.
On September 12, Anthropic CEO Dario Amodei published an essay, “We Must Pace the Frontier,” with a central message: “We must slow the pace at which we improve the capabilities of AI models.” OpenAI CEO Sam Altman respo…
Dario’s essay is, in part, an admission that safety engineering at AI labs has fallen behind model capabilities.
Conference takeaways reinforce a multi-year AI optics upcycle, with laser, DSP, switch and interconnect demand constrained by supply and customer qualifications rather than end demand.
In this note we summarise conference-season optics takeaways across recent investor conferences.
Our senior analysts and would be attending the OCP Global Summit in San Jose on October 12–15 and represent Funda at Semtech’s Analyst Day on October 15.
The tape last week split again: broader indices leaned soft into PPI/CPI and the next FOMC, while semis held up on company-specific news, where SOX finished higher even as mega-cap tech traded mixed.
That is the same relative setup we have been tracking since memory broke out: hardware names responding to incremental AI infrastructure demand while the index still waits on the macro prints.
The research week was dominated by compute demand from a different angle.
AI development should not be slowed uniformly because release pace, training pace, and sector readiness differ sharply across industries.
Starting this week, we are launching a new Column series where we discuss interesting topics in a more casual way than our typical research.
For this debut edition, we would like to share our view on the most widely debated topic over the weekend: whether AI should slow down its frontier expansion, and whether it would.
Goldman Sachs doubled its 2028 forecast in five months. We break down what changed, how 1.6T and 3.2T modules affect the market and, who can supply it.
Optical industry has been one of those sectors, that we’ve been covering since the start of our Substack.
Our first article was about “AI next supercycle”, going from copper, which is limited because of its physical aspects, and switching to photonics as more power efficient solution on longer distances.
TL;DR AI inference data splits into weights that are read constantly and context that is in use, versus weights that are read rarely and context that is waiting.
As the AI era arrives, the memory hierarchy is becoming more varied and more complex.
The clearest example is that over the past few months a DRAM maker, a NAND maker, and an accelerator maker each announced plans to add new memory next to HBM.
A broad AI power winner, with 2H26 revenue re-accelerating on AI PSU demand and VPD module growth
Overview We initiate coverage of Delta Electronics (2308.TW), the broadest play on AI power spanning the grid-to-core value chain.
We expect 2H26 revenue to re-accelerate on stronger AI PSU demand and explosive VPD module growth, with GPM expanding from 26Q4 on a better mix, economies of scale, and price pass-through, and OPM improving even faster.
Specific stocks related to Cerebras and SpaceX went on 100%+ runs right before these IPOs. In this article we disclose best companies you can own before Anthropic lists.
Back in June, we wrote about Anthropic proxy play, naming three biggest beneficiaries based on their balance sheet and stake in the company.
We also covered its revenue trajectory, funding rounds, and more.
G2AI US–China AI Weekly · September 7, 2026 · This Week: Astra & the CPU Second Act / The Desktop Compute Wave / US–China Power
OpenAI released GPT-6 Astra, its computer-use-first flagship, on 2026-09-03, scoring 72.6% on OSWorld 2.0 with an average single-task time of roughly 40 minutes — down 47% from GPT-5.6 Sol’s ~75 minutes.
This is not a routine flagship iteration: OSWorld 2.0 comprises 108 long-horizon tasks averaging 318 tool calls with a human median completion time of 1.6 hours, and Claude Opus 4.8, the strongest model at launch, score…
The software vs semi split that defined last week finally started to rebalance.
Software that printed what the channel had been pointing to kept working through most of the week, while the semi tape spent the first half getting tagged by rates — the 10-year pushed to 4.8% on Tuesday, the highest si…
Yield, not capacity, is what the AI build is short of. Whoever owns the loop that learns it collects the rent, until the format changes.
Along with my friend of Crack The Market, we put his top-down view of market structure and capital allocation next to my bottom-up view of how chips are actually built , to take apart the layer that has quietly become t…
owns the market structure, the economics and the investment judgment; I own the process and the physics.
Redrawing the optical investment map with technology direction, valuation, and market expectation all converted into factors.
The last Optical Investing Round Two piece covered why this is a good moment to look at optics again, and roughly how the names sort out.
This one redraws the map as of September 2026, converting the direction of optical technology, the valuation of each stock, and the expectations already priced in into factors.
#1 in data center interconnect with $10B in backlog. But the stock is at a 6-month low.
We’ve talked about Ciena in the past, and it’s one of our best performing positions in our portfolio on AutoPilot called “Photonics Is Next” , because we discovered it as early as March, at the NVIDIA GTC.
The stock is at its 6-month low, even though they just reported their record quarter.
G2AI Info is our column for exclusive news and analysis, built on conversations with people across the industries we cover.
In our AI weekly we struck a rather skeptical tone when we titled one piece “China’s AI Labs Are Selling Annualized Orders, Not Revenue.” But something notable is happening to Zhipu’s business model.
G2AI Info is our column for exclusive news and analysis, built on conversations with people across the industries we cover.
In our AI weekly we struck a rather skeptical tone when we titled one piece “China’s AI Labs Are Selling Annualized Orders, Not Revenue.” But something notable is happening to Zhipu’s business model.
Comparing Conclusions Across the Two Rounds, and What Is New This Round This round covers 7 enterprise samples plus 1 cross-client enterprise AI transformation consultant, taking the two rounds together to 19 enterprise…
Both rounds asked the same set of questions, but of entirely different companies, roughly a week apart, so differences between the rounds reflect sample composition and do not constitute a time series.
BE’s AIDC demand thesis remains supported by a ~15GW North America AI power shortfall in 2027, but the debate has shifted to delivery execution.
Executive Summary AIDC power scarcity remains the structural driver.
We estimate North America faces a ~15GW AI power shortfall in 2027, as grid interconnection, turbine lead times, and downstream power infrastructure constrain how much announced capacity can actually be energized.
5 of 6 hyperscalers as customers, and 88% AEC share. Company breakdown and Q1 earnings analysis.
is one of just 4 stocks in our hyper-concentrated $18M Asymmetrical Bets fund, which holds at any given time the handful of stocks we believe have the most asymmetrical upside.
After selling off hard today with macro pressure, dropped another 10% after it released its earnings.
Sales Execution Sales partner growth is generally measured by Consumption, which is highly correlated with Revenue, because SNOW primarily evaluates partners on Consumption.
A few contacts use consulting revenue where Consumption data are unavailable.
Optics are moving inside the tray as scale-in links target memory-class bandwidth, roughly 10x scale-up per GPU and about 100x scale-out
Core conclusion Optics are moving inside the tray, and the memory tier it targets is the largest bandwidth tier in the system, roughly ten times the scale-up bandwidth per GPU and about one hundred times the scale-out.
The capital is already moving, with $2 billion each from NVIDIA into Lumentum and Coherent, $3.25 billion from Marvell for Celestial AI, and a revenue ceiling of up to $120 billion on the Google and Marvell warrant.
The market spent the week trading NVIDIA’s print first, then putting rates and crowding back into the price.
The numbers are still fast: $96.2 billion of revenue, $89.0 billion from data center, and $108 billion of Q3 guidance that assumes no China data-center compute.
The market spent the week trading NVIDIA’s print first, then putting rates and crowding back into the price.
The numbers are still fast: $96.2 billion of revenue, $89.0 billion from data center, and $108 billion of Q3 guidance that assumes no China data-center compute.
Our strategies returned 18.4% this month after momentum bounced back from July. Here’s what delivered and how were thinking about the market.
Hey everyone, One goal of ours for the rest of the year is to bring you guys more visibility and transparency into our funds/performance to help you get maximal value out of our subscription and stay convicted in core t…
As many of you know, one of the most interesting parts about our publication is that we have over $55M AUM on Autopilot , an app where thousands of retail investors follow our thematic portfolios trade by trade, without…
The biggest shift in tech this week is the divergence between software and semi/hardware performance.
In our software previews this quarter (e.g., CRM, NOW), one meaningful change we noticed was a shift in enterprise customers’ tone.
To recap, over the past few quarters, when asked about the impact of LLMs on incumbent SaaS systems, only channel partners defended the value of the system of record (SoR), while customers were hesitant.
cHBM has started to create new design value for memory vendors. NVHBM is the structure through which NVIDIA pulls part of that value into a common architecture.
NVHBM Is Not a New HBM.
It Is a Shift in Who Owns the Design NVIDIA disclosed NVHBM on August 26, 2026.
cHBM has started to create new design value for memory vendors. NVHBM is the structure through which NVIDIA pulls part of that value into a common architecture.
NVHBM Is Not a New HBM.
It Is a Shift in Who Owns the Design NVIDIA disclosed NVHBM on August 26, 2026.
Over the past few months, enterprise adoption of AI has become one of the market’s central debates.
From the coding-agent surge early this year to the tokenmaxxing discussion in May-June to the more recent open-source shock, whether—and when—enterprise AI can take the baton from coding and become the next growth engin…
On August 27, 2026, Anthropic released MHS (Model Hardware Standard) as a research preview — a shared specification for letting AI agents safely operate physical equipment, promising to compress lab and factory hardware…
The launch lands on top of a steepening curve: annualized revenue climbed from roughly USD 9.0 billion in December 2025 to USD 65.0 billion by late July 2026 — 7.2x in seven months , a compound monthly growth rate (CMGR…
On August 27, 2026, Anthropic released MHS (Model Hardware Standard) as a research preview — a shared specification for letting AI agents safely operate physical equipment, promising to compress lab and factory hardware…
The launch lands on top of a steepening curve: annualized revenue climbed from roughly USD 9.0 billion in December 2025 to USD 65.0 billion by late July 2026 — 7.2x in seven months , a compound monthly growth rate (CMGR…
From how an AI GPU data center actually works to the technology, economics and investment approach for each of the three companies.
Disclaimer This material is general research opinion based on public information and is not investment advice tailored to any individual.
Future contracts, capacity targets, ARR and secured power are conditional figures, and actual outcomes may differ depending on financing, permitting, utility schedules, supply chains, customer acceptance and demand.
From how an AI GPU data center actually works to the technology, economics and investment approach for each of the three companies.
Disclaimer This material is general research opinion based on public information and is not investment advice tailored to any individual.
Future contracts, capacity targets, ARR and secured power are conditional figures, and actual outcomes may differ depending on financing, permitting, utility schedules, supply chains, customer acceptance and demand.
Nvidia's flawless $96.2B quarter beat all expectations; the real risk lives off-balance-sheet — $108.5B guarantees, becoming AI's lender of last resort.
Nvidia’s quarterly revenue rose from $46.7 billion to $96.2 billion in twelve months — a 106% increase that cleared all three expectation layers at once: the company’s own $91.0 billion guidance midpoint, the $92.2 bill…
I have covered every Nvidia report since the Hopper cycle began, and I cannot recall a quarter in which the whisper number itself was beaten.
Nvidia's flawless $96.2B quarter beat all expectations; the real risk lives off-balance-sheet — $108.5B guarantees, becoming AI's lender of last resort.
Nvidia’s quarterly revenue rose from $46.7 billion to $96.2 billion in twelve months — a 106% increase that cleared all three expectation layers at once: the company’s own $91.0 billion guidance midpoint, the $92.2 bill…
I have covered every Nvidia report since the Hopper cycle began, and I cannot recall a quarter in which the whisper number itself was beaten.
North America’s 2027 AI demand could reach twice its deliverable data center capacity - why power access sets the pace, and what it means for xAI, CoreWeave, Nebius, and IREN.
Executive Summary North American AI data center demand in 2027 runs to roughly 35GW on an IT basis, against 16.5-23.4GW that can actually be delivered.
Where our two independent supply methods overlap, that gap is 1.8-2.1x, and it is the number this whole report is built around.
Is the market wrong, yes I think so. I increased my position today with 30% more shares.
We’ve now seen Klarna and Smart Eye both fall brutally after reports that were quite good.
This kind of drop used to only happen when a company announced a share issue with a 35% discount, or when a fraud was uncovered, or when a major customer walked away.
There was some spicy news this morning about a spicy chip from a spicy frontier lab who will be the last presenter today at Hot Chips.
Mamma mia follow me on on X: x.com/jasonschips subscribe for an extra 500Tbps memory bandwidth Contents Samsung LPDDR5X Processing in Memory XCENA MX1 CXL Computational Memory Device Thor Ultra: Ethernet NIC Chip NVIDIA…
主题存储/HBM光互连
原文明确标的未从公开摘要识别
来源健康度
每个来源单独记录成功、失败与最近采集时间
Global Semi Research指定关注 · 存储、先进封装、光互连、AI 算力与中国半导体产业链
正常09月23日 16:44 本轮 feed 条目 20 条
Anders Storm当前跟踪 · Sivers、光互连、CPO 与 AI 超级周期
正常09月23日 16:44 本轮 feed 条目 20 条
FUNDA / FundaAI当前跟踪 · AI 基础设施、软件与算力投资研究
正常09月23日 16:45 本轮 feed 条目 20 条
Jason's Chips当前跟踪 · 半导体、AI 投资、光互连与芯片经济学
正常09月23日 16:45 本轮 feed 条目 20 条
Jimmy狐狸当前跟踪 · 美股半导体、光通信、AI 资本开支与个股研究
正常09月23日 16:45 本轮 feed 条目 20 条
Latent Value当前跟踪 · 光互连、CPO 与半导体供应链工程验证
正常09月23日 16:45 本轮 feed 条目 2 条
The Midwestern Investor当前跟踪 · 技术与半导体公司、冷却、光学与设备
正常09月23日 16:45 本轮 feed 条目 15 条
Damnang当前跟踪 · 半导体技术、光学、电源与公司研究
正常09月23日 16:45 本轮 feed 条目 20 条
Asymmetrical Bets名单匹配刊物 · 光通信、测试、连接器与小型科技公司研究
正常09月23日 16:45 本轮 feed 条目 20 条
Semiconductor Business Intelligence已有公共源 · 半导体供需、公司、行业事件与存储价格