Every memory investor owning San Disk , Micron , Samsung, SK Hynix etc. I talk to is asking the same question: how much longer can memory stocks keep running, and how do I get out before the turn?

Their questions are valid. Most of the revenue growth in this cycle isn’t coming from volume. It’s coming from price. That is exactly what late-cycle memory has historically looked like - massive price hikes, followed by oversupply, price correction and everything turning violent. Memory cycles do not gently roll over. They snap.

The real question isn’t if the cycle turns. It’s whether you’ll see it before the crowd does. I’ve spent the last few days building an AI workflow to do exactly that. Here’s what I landed on.

The Tool I Picked (and Why the Chinese Models Disappointed Me)

I tested Perplexity, Qwen, DeepSeek, and Claude for this specific use case. I included Qwen and DeepSeek deliberately. My hypothesis was that the Chinese models might surface data or reporting the Western models would miss, given how much of the memory supply chain runs through Asia. But it was not the case.

The Chinese models didn’t add the edge I was hoping for. I prompted in English and my theory is that when you prompt in English, the models query English websites and returns English-language results only. There might be alpha sitting in local-language websites. Explicitly asking the model to search local websites and publications might help surface those results. I explored this idea in a previous piece:

Between perplexity and Claude: Claude was clearly the best for this workflow. Its citations are tight, it’s willing to dig into SEC filings and earnings transcripts, and it doesn’t hallucinate quotes the way the others sometimes do. It also double checked all the data points when prompted and updated itself using official Government sources unlike perplexity.

With a clear llm winner for my memory dashboard, I built two complementary workflows: one qualitative, one quantitative.

Workflow 1: Track the Industry Chatter (With Receipts)

The first workflow is simple but disciplined: every few days to every couple of weeks, I ask Claude to surface exact quotations from CEOs, industry executives, sell-side analysts, SEC filings, and earnings call transcripts on the state of memory.

Specifically: Is supply tight or loosening? Are prices rising or falling? What’s being said about capacity constraints? What’s happening with inventory? Please provide exact quotes with source.

The key discipline: ask for exact quotes with sources, not summaries. Summaries get hallucinated. Direct quotes with citations are auditable.

To date, virtually every quote that has come back has been bullish:

• Sanjay Mehrotra (Micron CEO): “DRAM requirement in advanced AI accelerators has now doubled year-over-year. Without more memory, without faster memory, AI just cannot scale up.”

• Choi Joon-yong (SK Hynix HBM Head): Projected 30% annual growth in HBM market through 2030, driven by “strong end-user demand for AI capabilities.”

Tight supply. Rising prices. Capacity sold out into 2027.

But there’s one crack in the wall.

Samsung warned in August 2025 that HBM3E supply could outstrip demand growth, hinting at potential near-term oversupply in older-generation stacks. That’s the only bearish signal I’ve found in the entire wall of bullish industry commentary. But we are moving to the next generation of memory chips (HBM4), so we can ignore this.

A practical note on getting better results: A lot of the most valuable commentary lives behind paywalls: sell-side research, niche trade publications, paid newsletters. If you have access to that material, paste it directly into Claude’s context window and let the model cross-reference it against the public data. The model becomes dramatically more useful when you feed it what the open web doesn’t index.

Workflow 2: The Export Data Dashboard That Tracks the Cycle in Near-Real-Time

This is the more powerful of the two workflows, because it gives you a quantitative signal instead of a qualitative one.

Here’s an insight most investors miss: a huge amount of valuable economic information is hiding in plain sight in import/export data, particularly customs data published by Asian governments. I started prompting Claude to find the freshest, highest-frequency data on DRAM, NAND, and HBM shipments, and turned it into a live dashboard.

The data source I converged on: Korean monthly semiconductor export data.

Why Korea? Because roughly 80–85% of South Korea’s total semiconductor exports are memory, and Korea is home to Samsung and SK Hynix, the two largest memory producers on Earth. When Korean semiconductor exports accelerate or decelerate, you’re watching the global memory cycle move in near-real-time, with maybe a two-to-four-week lag.

Here’s what the absolute dollar amount and year-over-year growth in semiconductor export value looks like over the most recent months:

This is not a topping pattern. This is an accelerating curve.

The cycle is still expanding, and the rate of expansion is itself increasing. As of right now, the data says one thing clearly: still in the boom phase of the memory supercycle

Link to the full dashboard

Link to my prompts for creating this dashboard if it interests you

When to sell you memory stock?

The whole point of this exercise is to spot the inflection before the market prices it in.

When the cycle turns, multiple trend lines will start declining simultaneously. The dashboard tracks several key turn signals: moving average crossovers, consecutive months of declining year-over-year growth, month-over-month reversals, and exports falling below their six-month average.

But here’s the critical framework: if the trend lines are declining and industry chatter is bullish, follow the trend lines.

Industry statements might continue the bullish statements for various motivated reasons but data would never lie.

What I’m Building Next

This framework is a strong foundation, but there are clear ways to sharpen it:

Layer in pricing and inventory data. Combining the export numbers with DRAM spot pricing, channel inventory checks, and company guidance would give a much fuller picture of where we are in the cycle.

Go deeper on Asian trade flows. The Korean export data tells you what’s leaving Korea. The next step is tracking where it’s going: more granular import data from China and Taiwan, where the memory ships for GPU and server assembly.

Make it live. Right now, the Claude artifact doesn’t auto-refresh. I’m building a live version for the Singularity Research website so subscribers can watch the same numbers I’m watching, updated in real time, without having to rebuild the workflow themselves.

The Bottom Line

We are currently in the accelerating phase of the memory supercycle. But historically memory cycle cracks

The point isn’t to predict the exact day. Nobody can do that. The point is to build the system now, so that when the data turns, you see it on the first of the month, not three months later when the stocks are already cut in half.

Build the dashboard. Track the quotes. Refresh monthly. That’s the whole edge.

Thanks for reading. Subscribe for free to receive new posts and support my work.

Until next time,
Singularity!

Meanwhile, the companies that will 10–100× over the next three years are hiding in plain sight — in memory fabs, laser factories, and power equipment backlogs that most investors have never heard of.

I just taught the first live class of my free AI Investing Bootcamp — a course focused on investing in AI stocks from a technical perspective, covering the AI fundamentals every investor needs to know. Future classes will go deep on AI tools, workflows, prompt engineering, and hands-on sessions with Claude Code, Codex, and more with the selected group of 10 people.

Thanks for reading AI tools for investing! Subscribe for free to receive new posts and support my work.

This post is the core thesis from Class 1.

If you only remember one thing: the bottleneck is the investment thesis.

First, the basics — what actually happens inside a data center

AI has two phases. Understanding the difference is the entire foundation.

Training is where a model learns. It reads the internet, books, and code over weeks or months. It costs $100M to $1B+. It happens rarely — maybe once every year or two per frontier model.

Inference is where the model works. Every time you ask ChatGPT a question or Claude writes code, that’s inference. It happens billions of times daily and it’s growing exponentially.

Training and inference are fundamentally different businesses with different hardware, different economics, and different investment implications. Nvidia’s Mellanox acquisition in 2020 was a training play. Their reported interest in inference-focused companies tells you where the puck is going.

Both require massive capital expenditure. Both are bottlenecked. And those bottlenecks are where the opportunity lives.

Bottleneck 1: The Memory Wall

Here’s the number that should keep GPU investors up at night.

Since 2010, compute power (raw FLOPS) has grown 60,000×. Memory bandwidth? About 100×. Interconnect speeds? Roughly 30×.

A modern GPU can do trillions of calculations per second. But it spends most of its time waiting — waiting for data to arrive from memory. The chip is literally starving.

This isn’t a bug. It’s physics. And it creates an enormous investment opportunity.

Michael Dell said it plainly at a Bank of America event in April 2026: AI memory demand is surging 625× from 2022 to 2028. That’s not a typo. Per-accelerator memory is growing from 80 GB to 2 TB (25×), and accelerator deployment is expanding roughly 25× as well. Multiply them together: 625×.

Only three companies on Earth manufacture HBM (High Bandwidth Memory): SK Hynix, Samsung, and Micron. New fabs take four years to build. Major new capacity doesn’t arrive until 2027–28.

Do the math on a 625× demand surge with a four-year supply lag.

Bottleneck 2: The CPU Shortage Nobody’s Talking About

GPUs get all the headlines. But CPUs are quietly becoming the next constraint.

Why? Two reasons.

First, synthetic data generation. Models need new training data, and one of the best sources is synthetic — CPUs running real-world tasks like browsing websites, shopping on Amazon, and navigating apps to generate datasets at scale. Thousands of CPU instances running 24/7.

Second, agentic inference. When AI runs code, opens browsers, executes scripts, and coordinates multi-step tasks, all of that runs on CPUs. Every agentic action multiplies CPU demand.

The old ratio was roughly 1 GPU to 2 CPUs. In the agentic world, it’s heading toward 1 GPU to 8 CPUs. AMD, Intel, and ARM all stand to benefit from a constraint most investors aren’t pricing in yet.

Bottleneck 3: The Optics Revolution

This one is almost entirely off most investors’ radars.

At 800 Gbps, copper cables hit their physical limits — heat corrupts the signal. At 1.6 Tbps (where next-gen data centers are headed), copper simply cannot function. It burns.

The solution is Co-Packaged Optics (CPO) — lasers through fiber, packaged directly with the switch chip. Lower latency, lower power, smaller footprint. Deployment of 1.6T systems begins Q4 2026.

The market numbers are staggering. Goldman Sachs projects the optical networking TAM growing from roughly $15B in 2026 to $154B by 2028. That’s approximately 10× in two years.

But here’s the real kicker: there’s a massive supply shortage. Global demand for Indium Phosphide substrates — critical for 1.6T transceivers — is approaching 2 million units. Production capacity? 600,000. That’s a 70% supply gap and the rerating of Sivers Semiconductor

Companies to watch: Lumentum, Coherent, Marvell, Broadcom, Ayar Labs.

Bottleneck 4: The 800V Power Problem

Previous-generation GPUs (A100, H100) ran on standard 48V power distribution. Existing data center infrastructure worked as-is.

New GPUs — the B200, GB200, and upcoming Vera Rubin — require 800V.

That’s not an upgrade. It’s a complete teardown.

Every transformer, switchgear unit, busbar, panel, and power distribution unit in the data center must be replaced. The entire power delivery chain from grid to chip must be rebuilt. Existing infrastructure cannot be retrofitted.

This creates demand across the full electrical supply chain — from grid-level transformers down to board-level capacitors (MLCCs) and inductors. ABB, Schneider Electric, Vertiv, Eaton, Murata, TDK — all of these become critical links.

Bottleneck 5: Power Itself

Data centers can’t wait four years for grid connections. They need power now.

Each GPU generation demands more. A100 clusters needed roughly 6 kW per GPU and 40 MW per campus. B200 and GB200 clusters need 15 kW per GPU and 300 MW per campus. Vera Rubin and beyond? 20+ kW per GPU, 500 MW+ campuses.

Hyperscalers are getting creative. Natural gas fuel cells (Bloom Energy) are being deployed as immediate backup because they face far fewer permitting hurdles than new grid connections. Backup generators from Siemens, GE, and Caterpillar have five-year backlogs.

One important note for investors: ignore “data center cancellation” headlines. Hyperscalers routinely file permits in three or four states simultaneously, knowing most will be denied. When three out of four get cancelled, that’s not a demand signal — it’s how the process works. The one that gets approved still represents massive spend.

The revenue is real

The most common bear case is that AI spending is irrational — that companies are pouring billions into infrastructure with no return.

Look at the numbers. Anthropic went from $100M in revenue in 2023 to $1B in 2024 to $10B in 2025, reaching a $44B valuation by April 2026. Hyperscalers have committed up to $9T in cumulative CapEx. New data centers take four years to build.

The cycle is self-sustaining for at least two to three more years. The early warning signs to watch: API rate limits being removed (supply catching demand), CapEx guidance deceleration in quarterly earnings, and satellite-based data center tracking from firms like SemiAnalysis.

The framework

Here’s the simplest investing framework I can give you:

Follow Jensen Huang. Whatever he says are the specifications of the next Nvidia GPU — follow that supply chain. Every new spec creates demand for components that only a handful of companies can supply.

The bottleneck categories: high-power lasers, silicon photonics, 800V power systems, MLCCs and inductors, HBM and CoWoS packaging, liquid cooling, grid transformers, gas turbines, and advanced PCBs.

2026 data center CapEx is on track to exceed $1T. Component demand growth across these categories ranges from 35× to 100×.

What’s next

This was just Class 1 of the Bootcamp — the technical foundation. Future classes cover mastering AI prompts and workflows, building investing workflows with AI, hands-on sessions with Claude Code and Codex, and ultimately building your own personalized investment dashboard. I would all the slides and learnings through the email.

Subscribe to AI Tools for Investing to follow along.

Thanks for reading AI tools for investing! Subscribe for free to receive new posts and support my work.

And if this post was useful — share it with one investor friend who’s still only looking at Nvidia.

The real money is in the bottlenecks.

Coming back to our substack saved post question. We all are guilt of saving Substack posts thinking “I’ll read this later”… and then never revisit them.

I’m guilty too — I only have 7 saved right now. But one reader just told me he has 400 and feels completely overwhelmed. They asked me to suggest an AI workflow to make his life easier.

Here's a quick workflow to turn your graveyard of saved posts into an organized knowledge map — in under 10 minutes.

The AI-Powered System to Turn Your Saved Posts Into Pure Insight

Here’s the exact 6-step process I use to organize, understand, and actually use everything I save:

Step 1 — Go to substack.com/inbox/saved and scroll until every saved post is visible on screen. I only have 7, so no scrolling needed — but if you've got hundreds, keep scrolling until they're all loaded.

Step 2 — Right-click anywhere on the page → click Inspect. You should see the browser's developer tools panel open up.

Step 3 — Find <div id="entry"> in the elements panel → Right-click it → Copy → Copy Element.

Step 4 — Paste that copied element into any LLM (we used Claude — see our prompt here) and ask it to extract the Substack article URLs. You can provide an example URL to guide it. Here's what we got back:

https://ragingbullinvestments.substack.com/p/a-royalty-on-market-madness
https://yianisz.substack.com/p/indium-phosphide-inp-the-quiet-bottleneck
https://nikhs.substack.com/p/sandisk-2qfy26-earnings-the-confession
https://thequalityinvestor.substack.com/p/roko-deep-dive
https://crackthemarket.substack.com/p/optical-and-networking-supercycle
https://mispricedassets.substack.com/p/building-an-institutional-grade-portfolio
https://stocknarratives.substack.com/p/a-4x-in-4-years-surfing-the-data 

Step 5 — Head to NotebookLM → Create a New Notebook → Add Sources → Paste all the URLs from Step 4.

Notebooklm has a limit of 50 urls in the free plan. If you have more links, you can consider buying the paid plans or splitting your effort into two notebooks.

Step 6 — Once NotebookLM generates its auto-summary, click Mind Map in the right column for a visual overview of everything you've saved.

Figure - this is the mind-map of my saved substack posts

The result? A bird’s-eye view of all your saved reading — organized by theme. You can click into any category, ask follow-up questions, and actually learn from the content you saved months ago. I suggest you give this a try.

If you have questions about using AI for workflows like this, drop a comment or email me — I’m always happy to dig in and share what I find.

Hope this helps you turn saved posts into actual edge.

Like + share with a friend who’s drowning in tabs.

Until next time, Singularity

I’ve been running this AI newsletter for a while now, writing about AI tools and concepts for investing.

Thanks for reading AI tools for investing! Subscribe for free to receive new posts and support my work.

But here’s the truth: a lot of the really powerful stuff just doesn’t land the same way in text. Some ideas need to be shown, not just described. Some tools only click when you see them used live. Some questions only come up when you’re actually trying things together.

So I’m trying something new. I’m running a small, completely free live class series — just 5 spots. It’ll be 6 to 8 sessions, about 90 minutes each. We’ll cover everything from the basics to the advanced stuff I wish I knew earlier:

- What AI actually is and what you need to know

- How to write prompts that actually work

- Killing hallucinations and getting reliable outputs

- Connecting your own data so the AI uses the right information

- Deep dives into Claude — Projects, Desktop, Code, and Chrome extensions that actually save time

- Running your own openclaw instance

- Real workflows that tie everything together

The goal isn’t to make you an expert in every tool. It’s to give you enough hands-on experience that you can start figuring out your own use cases and actually grow with AI.

Who I’m looking for:

- People with different backgrounds

- Different experience levels — complete beginners to folks already playing with AI

- People who can commit to most of the sessions

In exchange for a free seat, I just ask for your candid thoughts at the end of each class and at the end of the series.

If you’re interested, just reply to this email with: Your name - Your current experience level with AI - One sentence about why you want to join

I’ll pick 5 people from the replies and get back to you with the schedule. Looking forward to hearing from you

Best,

Singularity

Thanks for reading AI tools for investing! Subscribe for free to receive new posts and support my work.

Hallucinations.

Thanks for reading AI tools for investing! Subscribe for free to receive new posts and support my work.

Made-up numbers. Fake citations. Confident-sounding garbage that could quietly wreck your portfolio.

The brutal truth? In finance, one wrong figure isn’t just annoying - it’s expensive.

The even better truth? Almost all hallucinations are completely preventable.

I’ve refined a system that cuts them by ~95% in real-world use. No fluff. No theory. Just what actually works when the stakes are real.

Here’s the exact playbook:

1. Ground Everything in Your Own Data (This One Move Is a Game-Changer)
Before you type a single prompt, paste in your source material into the model prompt - SEC filings, earnings transcripts, 10-Ks, investor decks, whatever you have.
Then hit the model with this line:
“Use ONLY the documents and data I have provided. Do not draw from your training data.”

Do this first and you eliminate the majority of hallucinations instantly.

2. Pick the Right Model for Financial Work
From heavy daily use, Perplexity is currently the most accurate for financial data, with Gemini a very close second. Both crush real-world numbers and context.
But models evolve fast - what wins today may lose in three months. Re-test every quarter.

3. Force It to Ask, Never Guess
Add this to every important prompt:
“If you don’t have the data, ask me for it. Do not estimate or assume.”

This single instruction stops the model from inventing numbers to fill gaps.

4. Demand Sources and Make It Show Its Work
Tell it explicitly:
“Cite every single claim with the exact official source (SEC filings, company IR page, etc.). No exceptions.”

When the model knows it has to prove every statement, it gets dramatically more careful.

5. Demand Exact Quotes When Accuracy Matters Most
For high-stakes analysis, add:
“Respond using direct quotes and exact wording from the source material only.”

Verbatim reproduction forces the model to stay glued to reality instead of paraphrasing from memory.

6. Make It Verify Before It Answers
One magic line:
“Verify all citations and sources before giving me your final answer.”

It sounds basic. It’s ridiculously effective.

7. Use Chain-of-Thought Prompting
Add:
“Before your final output, think step-by-step and ensure there are no logical fallacies or unverified assumptions.”

This turns the model into its own fact-checker. This may not be necessary in thinking models, only in zero-shot prompting models.

8. Always Use a Thinking Model for Serious Analysis
Skip zero-shot models. Use a thinking/reasoning model (Claude with extended thinking, OpenAI o1/o3, etc.).
They reason step-by-step before answering - which kills confident errors.
Pro tip: Claude Sonnet with extended thinking is often more than enough and saves you a fortune in tokens compared to the bigger models.

9. Connect Live Financial Data via MCP (The Underrated Power Move)
Stop relying on the model’s outdated training data. Plug it directly into real-time sources using Model Context Protocol (MCP) servers.

Two options I actually use:

• Financial Model Prep — Excellent for earnings calls and full financial datasets. Paid plans with earnings calls start around $129/month.

• DefeatBeta API — Completely free and open-source. Covers earnings calls, financials, stock prices with impressive accuracy. Docs + MCP integration here: https://github.com/defeat-beta/defeatbeta-api

(Enterprise options like Fiscal exist with native Claude/Gemini support, but expect enterprise pricing.)

10. Use Few-Shot Prompting for Perfect Output Format
Want a specific analysis style or table format? Give the model 1–2 perfect examples of input → ideal output before you ask your real question.
It adapts instantly and consistency skyrockets.

The Bottom Line

The above points can summarized into the following sentences that you can add in your next prompt -

Use only the data and sources I have provided and not your training data, cite every claim with its exact source with prefernce to official sources like SEC or company website, include direct quotes where possible, verify all citations before responding, ask me if you don't have enough information rather than assuming, and before giving your final answer ensure there are no logical fallacies, unverified assumptions, or gaps in your reasoning.

You should save it as a skill or project and make it easier to reuse them

Along with the above line, please chose the right llm and connect to an finance MCP. You’ll eliminate ~99% of hallucinations in your financial prompts.

The core principle is dead simple: the more you constrain the model to verified, sourced, real-time information — and the more you force it to show its reasoning — the less room it has to invent anything.

Finance is one of the highest-stakes arenas for LLMs. Spending an extra five minutes on your prompt is the cheapest insurance you’ll ever buy.

If this helped, forward it to someone doing investment research with AI.

And reply with what’s worked (or failed) for you — I read every response.

Thanks for reading AI tools for investing! Subscribe for free to receive new posts and support my work.

Until next time,
Singularity