AI Chips: Not As Easy As It Looks. ARD #112

Michael Parekh AI: Reset to Zero

6 hours ago

14 MIN READ

Today’s theme: AI chips — not as easy as it looks. From making the chips, to financing the chips, to making money around them. It’s all very hard, and getting harder. Three stories for the AI Tech Wave — Nvidia, SK Hynix, and Softbank — each with my Take first, then my Overall Take.

(1) Making Complex AI Semiconductor Systems Is Really Hard — Even for the Best

MP TAKE: Making AI semiconductors — the GPUs, the CPUs, the memory chips, all of it combined — is extraordinarily hard, even for the best. Case in point: Nvidia. SemiAnalysis, via CNBC, walked through the two-year ‘tick-tock’ roadmap Jensen Huang has executed across six years of this wave — Ampere to Hopper to Blackwell to Vera Rubin, and next Feynman. On a GTC slide it looks smooth and inevitable. It isn’t. These are the chips that build the AI data centers now running at $100-150 billion per tech company — trillions in aggregate, with Goldman Sachs estimating at least $7 trillion by the end of the decade, including power.

Because under the hood, each system assembles thousands — tens of thousands — of components from dozens, even hundreds, of suppliers, and Nvidia has to make it all work with its CUDA and open-source software libraries, not for one customer but for hundreds, each demanding custom configurations for training, inference, frontier models, reinforcement learning. The Kyber system-assembly component for the Vera Rubin Ultra just slipped from 2027 to 2028 on a raft of technical issues. Because demand runs so far ahead of supply through the rest of the decade, it’s not a hiccup for Nvidia’s revenue — but it is a huge operational hiccup, forcing Nvidia’s worldwide supply chain and every customer’s data-center plans to reconfigure around it. We don’t often see these hiccups; when we do, they’re notable. The lesson: the delays come with the roadmaps too.

Sources, in narrative order: CNBC — Nvidia’s next-gen AI rack system delayed to 2028 on manufacturing snags, SemiAnalysis says. Google Gemini — Nvidia’s two-year annual ‘tick-tock’ roadmap. For longtime readers: ‘Nvidia’s dense announcements at GTC 2026 worth a trillion+ by 2027’ in AI-RTZ #1031.

(2) Asian Chip Companies Keep Raising Mega Dollars

MP TAKE: SK Hynix is only the latest Asian chip company filling up its capital tanks. Reuters reports it’s positioned to raise about $28 billion — probably a bit more — in additional capital to accelerate investment through the boom, on top of the multi-hundred-billion-dollar HBM (high-bandwidth memory) fab commitments it announced last week with Samsung and the Korean government. This is the supply side of ‘RAMageddon’: memory has gone from 4-5% of the bill of materials in your computers, phones and gadgets to 40-50%, with prices climbing 30, 40, even 50% every six months or so amid real shortage.

And here’s the tell: demand for these capital raises is running 2x or more ahead of supply. Expect a lot more of them — and not just from memory companies. Google raised over $80 billion a few weeks ago; some of it funds its own TPU chips, where Broadcom and TSMC are the beneficiaries. Raising mega dollars to make more chips is core to this ‘not as easy as it looks’ theme — and we’re watching it proceed in the same tick-tock form.

Sources, in narrative order: Reuters — South Korea’s SK Hynix launches $28 billion US listing to ride the global AI wave. For longtime readers, in narrative order: ‘US hamstrung on memory chip supply via China’ in AI-RTZ #1128; ‘Semiconductors Remain a Key AI Gating Factor’ in ARD #104; ‘Push Comes to Shove on the Global Memory Crisis’ in ARD #108; and ‘SK Hynix Ready for its AI Close-Up’ in AI-RTZ #557.

(3) Softbank’s Masa Son Accelerates Into AI Bets, Chips and All

MP TAKE: Softbank’s Masayoshi Son — a veteran across multiple tech waves — is accelerating into AI, chips and all, and the FT has a detailed piece on how he’s remaking Softbank in his own image. I’ve had the chance to be in the rooms with Masa since the early PC and internet days — I helped Goldman take Yahoo public, and he was behind the Yahoo Japan entity that carved out the internet’s role in Asia. Early investor in Alibaba (hundreds of billions in profit), a 90% crash in the dot-com bust, $100-billion swings with the Saudi-backed Vision Funds — huge ups and downs, always aggressive.

Most relevant to today’s theme: he’s been an early, ~90% owner of ARM — the company whose templates and IP that Apple and dozens of others use for the systems-on-a-chip in everything from data centers to gadgets. He’s committing tens of billions to OpenAI, a big new data-center build in Europe, and much more across Asia and the rest of the world. Truly a global investor — and, most importantly, one investing in a systematic way across this wave. Because of ARM in particular, he’s a big player in this ‘chips are hard’ story.

Sources, in narrative order: FT — Son remakes Softbank in his own image. For longtime readers, in narrative order: ‘Rooting for Masa Son’ in AI-RTZ (2023); ‘Masa’s Back!’ in AI-RTZ (2024); ‘Masa Son Bets on AI Power and Chips’ in AI-RTZ #408; and ‘Arms Together on ARM Holdings’ in AI-RTZ (2023).

MP OVERALL TAKE

All three chip challenges — making them, funding them, and investing around them — have been true in almost every tech wave. This one is exponentially harder, because of the scale and the zeros involved.

For most participants at every stage, it’s really a ‘Ready, Fire, Aim’ cycle — not the ‘Ready, Aim, Fire’ systematic sequence we’d all prefer. Everyone from Meta to Google to Microsoft is investing ahead of the visibility for these markets, and ahead of the sheer technical difficulty of building everything from the chips at the bottom of the stack to the applications at the top. That’s a remarkable feature of every tech wave — PC, internet, mobile, now AI — but this one is far more complex, and the innovation is just getting started. It won’t be done in two or three years; it’ll run five, six-plus years as these systems get built and find their product-market fit.

And it all keeps coming back to the hardware. The software — large, medium and small language models, up through the applications — has to be constantly super-optimized to the hardware as that hardware is made and made available at scale, at ever-increasing pricing for now. That’s my overall take on the ‘not as easy as it looks’ theme for AI chips. Buckle in and brace ourselves indeed.

Gadget AI — The Best ‘Good Enough’ $400, 65-Inch Mini-LED ‘Dumb TV’ of 2026

MP Take: Almost all the AI podcasts, we talk about is for smarter devices — smarter glasses, smarter phones, smarter everything. But there’s a price: these products depend on our use and our data, and they resell that data. We trade our privacy for functionality — as we have for 25 years of social media, where Google and Meta made clear that we, the billions of users, are the product. Apple, at one end, made a virtue of privacy and trust. And your TV — the thing you never think about as a privacy risk — has been watching you for years: every channel, everything you do, monitored by the manufacturer and resold. That’s why the screens keep getting bigger and cheaper at Costco and Walmart. In fact, Walmart bought Vizio, one of the biggest low-cost TV makers, to monetize that data directly at the Walmart level.

So it’s notable that the senior TV analyst at The Verge reviewed the Vizio $400, 65-inch Mini-LED set — one of the few TVs you can actually run as a ‘dumb TV’, thanks to a couple of features in how Walmart set up the Vizio operating system. You genuinely cannot walk into a Best Buy, Costco or Walmart and buy a TV that does the minimum of monitoring — this is a rare exception, packing the latest mini-LED display tech at a $400 price. For the less than one percent of us who want to take some control back, it’s a very interesting read.

Sources, in narrative order: The Verge — Vizio accidentally made the best dumb TV on the market; Walmart — VIZIO 65” Mini-LED Quantum 4K QLED HDR Smart TV (2026). For longtime readers, in narrative order: ‘Scaling AI Trust is Job #1’ in AI-RTZ #382; ‘Apple’s Focus on AI & Privacy’ in AI-RTZ #372; and ‘The Glass Half Empty on AI Devices’ in AI-RTZ #872.

Questions

Q1 — What’s MP’s best strategy to minimize TV monitoring and tracking?

I don’t worry about it as much as I likely should — most of us don’t. But one approach I’ve taken: put an Apple TV streaming device on every TV, and never let the set boot on the manufacturer’s default OS — the moment that connects to Wi-Fi, a ton of monitoring kicks in. Run everything through Apple’s tvOS over HDMI instead, and you minimize a lot of the tracking (there are good guides online to tune it further). That’s in contrast to a Roku box (Roku just got bought by Fox for billions), or Google’s and Amazon’s Fire TV $20-40 streamers. Apple TV was about $100-120 until a couple of weeks ago — now bumped to $200, unfortunately, courtesy of RAMageddon. Not zero, but relatively better than the alternatives.

Q2 — What’s MP’s most wished-for feature in these TV gadgets and technologies?

A technology in the TV OS — Apple’s or anyone’s — that finally solves the ‘tyranny of choice.’ We have an explosion of programming across silos: this network, that network, free, paid. There are search engines, but no mechanism to proactively find what you’d actually like to watch. For most people it takes as much time to find something as to watch it — especially for half-hour shows. AI can play a big role here, but it’s barely begun — the solutions aren’t there yet. That’s what I’m looking forward to on the TV front: smart TV, dumb TV, or anything in between.

Wrapping up

Today’s AI-RTZ #1139 — Microsoft’s ‘Sum of the Parts’ (SOTP) Spin-Off case — a look at the move afoot to recommend a Microsoft spin-off of some of its resources to unlock stock value. It’s a theme I’ve written about for the big tech companies — Google in particular — and Microsoft has some interesting assets and opportunities here as investors lean in up and down the tech stack. No stock recommendations here — just analysis. Recommended as today’s reading.

Tomorrow — ARD 113 on AI-RTZ 1140.

Thanks for joining us today, AI Curious Folk. Stay tuned.

Full Source Reading —

For the broader context, see the canonical sources for ARD 112 — in today’s narrative order:

Clips from today

Clip 1 — AI’s Role in TV Choice Overload

Watch on YouTube Shorts

We have an explosion of programming across streaming silos, but no good way to proactively find what to watch — it can take as long to find a show as to watch it.

MP Take: What I most want is a truly horizontal, truly deep system to find what to watch fast, and re-find it after interruptions, across every subscription and search engine. AI can play a big role here — but it’s barely begun. A burning need still looking for a mainstream solution.

Clip 2 — AI Chips: Ready, Fire, Aim

Watch on YouTube Shorts

Chips are the core input of the AI Tech Wave — and making them, financing them, and investing around them is exponentially harder this cycle because of the scale and the zeros.

MP Take: For now the AI chip business is ‘Ready, Fire, Aim’ — everyone from Meta to Google to Microsoft investing ahead of the visibility. It’ll be years, not quarters, before it matures into ‘Ready, Aim, Fire.’ Buckle in and brace ourselves indeed.

Clip 3 — Privacy Trade-off for Smarter Devices

Watch on YouTube Shorts

Almost all AI investment is for smarter devices — but they depend on our use and data, and they resell it. We’ve traded privacy for functionality for 25 years; we, the users, are the product.

MP Take: Apple made a virtue of privacy and trust, and stays its most ardent big-tech champion. But for the vast majority, the trade holds — the ‘dumber,’ more private option is likely a sub-1% niche of the global market.

Clip 4 — TVs: The Hidden Surveillance

Watch on YouTube Shorts

Your TV has been watching you for years — every channel, everything you do, monitored by the manufacturer and resold. That’s why big screens keep getting cheaper; Walmart even bought Vizio to monetize it.

MP Take: To take back some control, put an Apple TV device on every set and never boot the manufacturer’s default OS — that’s where the monitoring lives. The Verge found a rare $400 Vizio you can actually run as a ‘dumb TV.’ Not zero, but better.

About AI Ramblings Daily (ARD), and AI-RTZ

Both are daily. Both are free. Both are about AI. But they’re different mediums carrying different messages.

AI-RTZ is the morning text — a deeper written take on one idea, published by at least 5 AM EST. Today: post #1139.

AI Ramblings Daily is the afternoon video + podcast — my ad hoc takes and perspective on the day’s AI issues & news flow, around 20 minutes, with short 1-2 minute clips for quick topic views. Today: episode #112.

Subscribe to either or both on michaelparekh.substack.com. They run as separate Sections you can opt into or out of.

Links used in today’s show (already embedded inline above; listed here for reference)

Take 1 — Making Complex AI Semiconductor Systems Is Really Hard:

Take 2 — Asian Chip Companies Keep Raising Mega Dollars:

Take 3 — Softbank’s Masa Son:

Gadget AI — The Best ‘Dumb TV’ of 2026:

Q1 + Q2 — Minimizing TV tracking + the tyranny of choice:

Q1 — (MP’s own strategy; no external source)
Q2 — (MP’s own wish-list feature; no external source)

Companion text:

AI-RTZ #1139 — Microsoft’s ‘Sum of the Parts’ (SOTP) Spin-Off case

(NOTE: The discussions here are for information purposes only, and not meant as investment advice at any time. Thanks for joining us here.)

Subscribe to AI: Reset to Zero for daily AI Ramblings + Sunday Bigger Picture posts.

AI Chips: Not As Easy As It Looks. ARD #112

Share

(1) Making Complex AI Semiconductor Systems Is Really Hard — Even for the Best

(2) Asian Chip Companies Keep Raising Mega Dollars

(3) Softbank’s Masa Son Accelerates Into AI Bets, Chips and All

MP OVERALL TAKE

Gadget AI — The Best ‘Good Enough’ $400, 65-Inch Mini-LED ‘Dumb TV’ of 2026

Questions

Q1 — What’s MP’s best strategy to minimize TV monitoring and tracking?

Q2 — What’s MP’s most wished-for feature in these TV gadgets and technologies?

Wrapping up

Full Source Reading —

Event 1 — Making Complex AI Semiconductor Systems Is Really Hard

Event 2 — Asian Chip Companies Keep Raising Mega Dollars

Event 3 — Softbank’s Masa Son

Gadget AI — The Best ‘Dumb TV’ of 2026

Clips from today

Clip 1 — AI’s Role in TV Choice Overload

Clip 2 — AI Chips: Ready, Fire, Aim

Clip 3 — Privacy Trade-off for Smarter Devices

Clip 4 — TVs: The Hidden Surveillance

About AI Ramblings Daily (ARD), and AI-RTZ

Links used in today’s show (already embedded inline above; listed here for reference)

Want the latest?

More like this

Momentum Monday – Does The Bounce Continue?

Adviser links: adviser burnout

$RXO Sets An All-Time High

Let’s be friends!