The frame running through every item today:

Building multi-gigawatt AI data centers is getting HARDER.

The financing side has accelerated to historic speeds — hundreds of billions per year, moving toward trillions. But the physical-world layer — power generation, transmission infrastructure, transformer supply, GPU coherence at scale, local siting permissions — is racing to keep up, and increasingly NOT keeping up. The headline numbers on capex announcements get the attention; the realities of breaker capacity, switchgear lead times, and county-board hearings get the deployment timeline. Today’s three items map the friction.

Three Key Takes today:

(1) Half of Planned AI Data Centers Waiting for Power. The headline number is striking — roughly half of the planned US AI data center builds have been delayed or cancelled, with the bottleneck cited as power infrastructure, transformer supply, and parts coming from China. Capex announced ≠ capex deployed. The reporting: Tom’s Hardware — Half of planned US data center builds delayed or cancelled; AI build-out flips the breakers. The FT angle: FT — Data center delays choke AI expansion. The transformer crunch: Bloomberg — US data center boom relies on hard-to-find electrical equipment. Standing thesis from the AI-RTZ archive: AI-RTZ #407 — The AI Power Grab.

MP Take: “Despite rapid acceleration of AI data center financing and funding, the technical challenges of building and then maintaining AI data centers will remain challenging through the decade. Especially due to rapidly changing AI demand profiles — from training-driven loads to inference-driven applications like AI search, chatbots, and AI Agents. Leaves room for tech and AI companies up and down the AI Tech Stack to raise prices. Including at Nvidia and TSMC levels.

The build target keeps moving as customer-facing demand re-shapes what each gigawatt actually needs to serve. Power, transformers, switchgear, and parts pipelines are physical-world inputs that don’t respond to capex acceleration the way GPUs do. A gigawatt data center is at least a $50 billion expense, and power turbines take months to build with very precise tolerances. The capex willingness is necessary but not sufficient.”

(2) xAI’s Challenges to Big AI GPU Installations. xAI’s run at gigawatt scale is showing the operational reality of the new computer. Two recent pieces from The Information surface what the rest of the industry is about to learn — multi-GPU coherence at this scale is genuinely hard, and the “fast and cheap” data center build that xAI piloted is now revealing hidden costs. The reporting on GPU coherence at scale: The Information — xAI shows it’s tough using a lot of GPUs at once. The hidden-costs piece: The Information — xAI’s fast, cheap data center build hidden costs. Standing thesis on the gigawatt buildout: AI-RTZ #481 — Power Plays Begin to Build Gigawatt.

MP Take: “The big tech companies have a lot of teething pains ahead as they truly scale to multi-gigawatt data centers. These massive computers need constant, low-latency communications between tens of thousands of GPUs running in parallel — and the failure modes get harder, not easier, as the cluster scales.

The workload is rapidly evolving — hardware and software combinations have to keep up with fast-changing customer needs at both the consumer and enterprise levels. xAI is the early field test of what the rest of the cohort is about to encounter at the gigawatt scale they’ve all committed to. Elon Musk’s team tends to go around AND through obstacles, and they’re finding that this stuff is not only art, but there are hidden costs that get baked in to doing some of this on a more sustained basis.”

(3) Local Community and Regulatory Pushback to AI Data Centers. The third axis of friction is getting under-reported but is increasingly material — local community and state-level regulatory opposition to AI data center siting. Power pricing, water draw, and resource crowding are translating into county-board hearings, zoning fights, and state-level legislative pushback that federal AI policy can’t override. The reporting: NY Times — AI data centers face local opposition as construction accelerates. Standing thesis on AI’s need to play nicer with neighbors: AI-RTZ #503 — AI Need to Play Nice With Other.

MP Take: “Local community and regulatory pushback is on the rise to AI data centers, due to a variety of drivers — ranging from local power pricing to resource crowding on water and grid capacity. Despite attempts to govern these uses from a federal perspective, the state and local political realities will continue to be a strong headwind to AI data centers through the decade.

This is not a phase. The siting fight is the new permitting fight — and just like the older permitting fight on roads, refineries, and pipelines, it gets resolved at the county and state level on a case-by-case basis. National AI policy can’t pre-empt municipal water rights or state utility commissions. Build cycles get longer because of this; expect that to be the story for years.”

Plus: Gadget AI — Operating Systems and Software Integrations Need to Change for AI Agents. Our computer and smartphone operating systems, plus the broader software-integration layer (including emerging standards like MCP), will need to evolve rapidly to host AI Agents at scale. The current OS surfaces were designed for human users, not autonomous-agent users. Source: Axios — Agents, AI software, model context protocol. Standing thesis on AI Agents needing their own internet: AI-RTZ #1023 — AI Agents Increasingly Need an Agent-Native Internet. The daily-use anchor: AI-RTZ #1054 — Working Out Daily With AI and AI Agents.

MP Take: “This is a multi-year effort to evolve operating systems and an internet designed for humans to accommodate AI Agents at scale. The current desktop OS, mobile OS, and web stack assume human users with hands, eyes, and attention. Agents have none of those. They need machine-readable affordances, structured tool calls, deterministic protocols, and reliable identity.

MCP and the agent-native internet ideas point in the right direction, but it will take a lot longer than currently assumed. Today’s internet over the last 25 years was designed and built for humans. Most websites assume any bot knocking on the door is malware — so they don’t let it in. OpenAI and Anthropic with cloud code, codex, and MCP protocols are doing herculean things despite today’s operating systems, not driven by them.

Expect this layer to be a decade-long re-platforming, not a 12-month sprint — and expect the early years to feel like the late-1990s web before standards settled. Hybrid is the shape of the AI era (per yesterday’s Ep 68 Gadget AI take on Blackberry QNX) — and the OS / agent-protocol layer is where that hybrid stitching happens.”

Bonus — today’s AI-RTZ companion #1077 covers ‘AI Godfather’ Yann LeCun’s common-sense take on AI thus far — a deeper read on LeCun’s framing of where current LLM approaches hit ceilings, and what a more grounded, embodied, common-sense AI architecture might look like as a next-generation alternative.

Closing Questions —

• MP’s negative AI Agent surprise of the week? Claude Cowork still forgets its core instructions on various projects day to day — despite rigorous efforts to have it memorize and record those instructions in its files. Even with standing playbook files, project-specific context, and per-session pre-flight reads built into the workflow, the system reliably drops core context at predictable points and has to be re-prompted. It’s like having a human assistant or intern who shows up as a blank slate every day. The memory architecture isn’t there yet. These tools are very, very new — Cowork didn’t exist a few months ago. So this is not a criticism, just a status report. Lot more work to do to make this stuff daily reliable for regular users.

• MP’s positive AI Agent surprise of the week? MP’s getting a lot of utility out of the FREE tiers of Google Gemini, OpenAI ChatGPT, and Perplexity. A couple of friends and relatives asked — “I don’t want to spend $200 a month, what can I do?” So MP took some computers, didn’t log in to the top-tier accounts, and used the free tiers. Particularly Google Gemini, ChatGPT, and Perplexity are relatively useful for a lot of basic things people might want to do at the free tier. By contrast, Anthropic Claude Cowork still needs MORE paid compute credits to handle MP’s daily workflow — the per-token usage profile of Cowork is meaningfully heavier than the chat-style alternatives. Anthropic has even been testing potentially taking Cowork out of the $20 tier entirely. So that may be the one you have to pay for; the others are surprisingly capable on the free tier. Stay tuned.

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

Clips from today’s episode

Short — AI Agents Want Your Steering Wheel Early adopters of AI agents are finding they need a separate computer for AI agent work versus their own work — because they’re clashing. It’s like two drivers trying to grab the same steering wheel in a car. Not a long-term workable solution. AI agents need a lot more control of your device. Operating systems need to redesign for AI Agents at the device level — multi-year effort that will take longer than currently assumed.

Short — Today’s Internet Was Built for Humans Today’s internet over the last 25 years was designed and built for humans. Most websites assume any bot knocking on the door is malware. OpenAI and Anthropic with cloud code, codex, and MCP protocols are doing herculean things despite today’s operating systems — not driven by them. Operating systems need to be redone aggressively for AI Agents. Apple, Microsoft Windows, Android — plus new players. OpenAI may build its own AI smartphone to get around these issues. Hybrid is the shape of the AI era; the OS layer is where the hybrid stitching happens.

Short — AI Data Centers Face Growing Local Opposition AI data centers are facing accelerating local opposition for both construction and operation — driven by noise, pollution, water draw, power pricing, and resource crowding. Federal AI policy is pro-industry, but state and local regulators are getting tougher. The siting fight is the new permitting fight — county and state level case-by-case, just like roads, refineries, and pipelines. National AI policy can’t pre-empt municipal water rights or state utility commissions.

Short — AI Agents: Blank Slate Every Day Claude Cowork is helpful — saves 2-3 hours a day. But every session, the software needs to be reminded of preferences, projects, and how MP likes to do specific things. It’s like having an intern who shows up as a blank slate every day. They’re getting better, but you have to remind that intern every day. The memory architecture isn’t there yet — these tools are very, very new. Lot more work to do to make this stuff daily reliable for regular users.

Ep 69 scope: 1 Main + 4 Shorts — MP’s pre-recording scope decision. No Segments or Hooks (matches Ep 62/63/65/66/67/68 4-clip pattern).

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 #1077 — Yann LeCun’s common-sense take on AI thus far.

AI Ramblings Daily is the afternoon video + podcast — my ad hoc takes and perspective on the day’s AI issues & news flow, around 16 minutes, with short 1-2 minute clips for quick topic views. Today: episode #69.

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 — Half of Planned AI Data Centers Waiting for Power:

• Tom’s Hardware — Half of planned US data center builds delayed or cancelled

• FT — Data center delays choke AI expansion

• Bloomberg — US data center boom relies on hard-to-find electrical equipment

• AI-RTZ #407 — The AI Power Grab

Take 2 — xAI’s Challenges to Big AI GPU Installations:

• The Information — xAI shows it’s tough using a lot of GPUs at once

• The Information — xAI’s fast, cheap data center build hidden costs

• AI-RTZ #481 — Power Plays Begin to Build Gigawatt

Take 3 — Local Community and Regulatory Pushback:

• NY Times — AI data centers face local opposition

• AI-RTZ #503 — AI Need to Play Nice With Other

Gadget AI — OS / software integrations need to evolve for AI Agents:

• Axios — Agents, AI software, model context protocol

• AI-RTZ #1023 — AI Agents Increasingly Need an Agent-Native Internet

• AI-RTZ #1054 — Working Out Daily With AI and AI Agents

Companion text:

• AI-RTZ #1077 — Yann LeCun’s common-sense take on AI thus far

The frame running through every item today: Nvidia (reporting 5/20/26), remains the Global AI ‘picks-and-shovels’ constant in the background — directly and indirectly providing AI infrastructure to everyone else.

The interesting question for the next 12-24 months is which of the other six Mag 7 companies have the best positioned distribution surfaces, balance sheets, and AI go-to-market motions to convert capex into compounding returns. Today’s Q1 prints across the cohort make the answer clearer than it’s been in months: Apple, Google, and Amazon — Reasons below.

Three Key Takes today:

(1) Apple’s Strong Q2 Demand and AI Positioning. Apple reported a strong Q2, driven by iPhone, Mac, and China demand — with global supply-chain discipline doing real work to power the result. The reporting: WSJ — Apple boosts starting price for Mac mini after AI demand surge. The cash-strategy shift: The Information — Apple’s cash strategy shift. Standing thesis on Apple’s AI bounce: AI-RTZ #978 — Apple’s AI Bounce Picks Up Steam.

MP Take: “The quarter was driven by iPhone, Mac, and China demand — and global supply-chain discipline did the heavy lifting on margins. None of the other Mag 6 come close to Apple’s unique global hardware ecosystem differentiation.

The combination of Apple Silicon, the iPhone installed base, the Mac refresh cycle, and the still-maturing services layer gives Apple a distribution surface that no other Mag 7 can replicate. Big AI, software, and services distribution opportunities are all still ahead — Apple is just getting started on the AI-monetization side, and the hardware base is the runway.”

(2) Even Apple Not Immune from Memory and Chip Constraints. The flip side of Apple’s strong quarter is showing up in the Mac mini pricing — Apple raised the starting price to $799, citing AI memory demand surge. The reporting: WSJ — Apple boosts starting price for Mac mini after AI demand surge · MacRumors — Mac mini now starts at $799. Standing thesis on Apple’s supply-chain lock-in: AI-RTZ #1010 — Apple Supply Chain Lock-in Tech. And on Apple’s distribution flex: AI-RTZ #1040 — Apple Flexes Its Distribution.

MP Take: “Apple is still the BEST positioned amongst peers on this issue — and especially well positioned on any rebound opportunity. The shifting cash strategy also signals rising flexibility on Supply Chain dynamics,. Especially vs peers who are increasingly cash-flow strapped as they invest hundreds of billions in AI infrastructure.

Apple has the balance sheet, the supply-chain leverage, and the time horizon to absorb the memory and chip squeeze better than anyone in the Mag 7. The Mac mini price hike is a signal.”

(3) Google and Amazon Stand Out This Quarter vs Other Non-Apple Peers. Google and Amazon both delivered standout Q1 prints among the non-Apple Mag-6, and the structural read is that they’re each just hitting their AI stride. The reporting: FT — Alphabet/Google Q1 2026 · WSJ — Amazon AMZN Q1 earnings report 2026 stock. Standing thesis on Google’s AI-fied search call from August 2023: AI-RTZ — Google’s AI-fied Search. And on Amazon AWS as the AI ‘underdog’ from 2023: AI-RTZ — Amazon AWS AI ‘Underdog’. The Apple-Google AI distribution partnership: AI-RTZ #965 — Apple Goes With Google Gemini. Google’s TPU-driven AI tokens lead: AI-RTZ #727 — Google Ramping Up Intelligence.

MP Take: “Google and Amazon are both just hitting their AI stride — and more than the other Mag 7s, each has a unique consumer AND enterprise footprint. Even relative to Apple, which is less enterprise-focused.

Microsoft suffers from minimal consumer exposure, Meta from too much consumer focus and ad revenues. Google is the right ‘Goldilocks’ AI play — especially with its Apple distribution partnership running through Gemini in the iPhone.

Amazon is unique amongst the Mag 7 with close equity AND business partnerships with BOTH OpenAI and Anthropic — and a uniquely deep distribution to global businesses, large and small. Amazon is to global enterprise distribution what Apple is to global consumer distribution at scale — a unique blend of hardware and software services. And Google sits in the middle of both.

The other Mag 7s — Microsoft, Meta, and Tesla — are not even close on this type of positioning. And Nvidia, of course, is the key provider of AI infrastructure to them all, directly and indirectly.”

Plus: Gadget AI — Robotaxis early but ramping across the US. The mainstream robotaxi rollout is finally getting interesting — Waymo expanding, Tesla still ‘vibe-coasting’ on the promise, and Uber emerging as the most strategically-positioned demand-distribution funnel for ALL the robotaxi players. Source: WSJ — See how the robotaxi industry is taking off across the US. Standing thesis: AI-RTZ #761 — Tesla ‘vibe-coasts’ on AI Robotaxis.

MP Take: “Google Waymo is still the most impressive Mobile AI Gadget out there. Robotaxis remain a ‘to-be-proven’ product on a mainstream rollout basis, and the US is still behind China in this area.

Uber remains the most interesting partner for ALL robotaxi ventures — because of their demand distribution funnel, their ability to mass-govern the rider experience, and their eventual ability to service fleets. The robotaxi winners over the next 24 months are likely to be the ones who pair the best AI driving stack with the best demand-side aggregator. That’s the Uber play in a sentence.”

Bonus — today’s AI-RTZ companion #1073 covers the Elon Musk / OpenAI ‘showcase’ trial — the morning text reads it as a strategic showcase rather than a legal verdict, and walks through what the AI industry should actually take away from week one of testimony.

Closing Questionss —

• What’s MP’s key con on today’s robotaxis? Directing them in edge cases — at departures and arrivals at airports especially. The middle of the trip is fine. The drop-off and pick-up are where today’s robotaxis still struggle, and that’s exactly where rider expectations are highest and where the ‘magic moment’ has to land. No one to help load and unload luggage.

• What’s MP’s key ‘magic’ moment in today’s robotaxis? None yet. None of them have hit the 100% satisfaction of that first Uber ride back in the 2000s — the moment that flipped the consumer category overnight. Perhaps by year-end in some of the cities where they’re being deployed. TBD.

  Stay tuned.

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

Clips from today’s episode

Short — Apple Still Best Value Even With Price Hikes Apple raised the Mac mini starting price from $599 to $799. So is Apple still the value leader? Yes. The quarter was driven by iPhone, Mac, and China demand. Global supply chain discipline did the heavy lifting on margins. None of the other Mag 6 come close to Apple’s unique global hardware ecosystem differentiation.

Short — Mac Mini Now $799: AI Memory Crunch Hits Apple Apple raised the Mac mini starting price to $799 citing AI memory demand surge. The 256GB and 512GB configs got pulled from the Apple store entirely. Apple is still the BEST positioned amongst peers on this issue. Apple has the balance sheet, the supply-chain leverage, and the time horizon to absorb the memory and chip squeeze better than anyone in the Mag 7. The Mac mini price hike is a signal — not a wound.

Short — Apple’s Supply Chain Edge Beats Mag 7 Peers Apple has decades of practice using its huge cash hoard to lock in supply years in advance from TSMC, Samsung, SK Hynix, and Micron. Apple plus Nvidia are the two best-positioned hardware-and-software configuring companies in this AI tech wave.

Short — Google + Amazon: The Goldilocks Mag 7 AI Plays Only Google moved up nearly 10% on Q1 — about $300B in market cap added. Amazon also stood out. Both are just hitting their AI stride. Google is the right ‘Goldilocks’ AI play, especially with the Apple distribution partnership running through Gemini in iPhone. Amazon is unique amongst the Mag 7 with close equity AND business partnerships with BOTH OpenAI and Anthropic.

Ep 67 scope: 1 Main + 4 Shorts — MP’s pre-recording scope decision. No Segments or Hooks (matches Ep 62/63/65/66 4-clip pattern; Ep 64 expanded to 5 for fuller financial-debate material).

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 #1073 — On the Elon Musk and OpenAI ‘Showcase’ Trial.

AI Ramblings Daily is the afternoon video + podcast — my ad hoc takes and perspective on the day’s AI issues & news flow, around 17 minutes, with short 1-2 minute clips for quick topic views. Today: episode #67.

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

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

Take 1 — Apple’s Strong Q2 + AI Positioning:

• WSJ — Apple boosts starting price for Mac mini after AI demand surge

• The Information — Apple’s cash strategy shift

• AI-RTZ #978 — Apple’s AI Bounce Picks Up Steam

Take 2 — Apple Memory & Chip Constraints:

• WSJ — Apple boosts starting price for Mac mini after AI demand surge

• MacRumors — Mac mini now starts at $799

• AI-RTZ #1010 — Apple Supply Chain Lock-in Tech

• AI-RTZ #1040 — Apple Flexes Its Distribution

Take 3 — Google + Amazon Stand Out:

• FT — Alphabet/Google Q1 2026

• WSJ — Amazon AMZN Q1 earnings report 2026 stock

• AI-RTZ — Google’s AI-fied Search (Aug 2023 call)

• AI-RTZ — Amazon AWS AI ‘Underdog’ (2023 call)

• AI-RTZ #965 — Apple Goes With Google Gemini

• AI-RTZ #727 — Google Ramping Up Intelligence (TPU lead)

Gadget AI — Robotaxis ramping in the US:

• WSJ — See how the robotaxi industry is taking off across the US

• AI-RTZ #761 — Tesla ‘vibe-coasts’ on AI Robotaxis

Companion text:

• AI-RTZ #1073 — On the Elon Musk and OpenAI ‘Showcase’ Trial: [AI-RTZ-1073-URL — capture once published, swap before posting]

The frame running through every item today: “AI Infrastructure sees strong incumbent AND startup investments.” The Mag-7 capex story (Take 1) sets the incumbent baseline. But two of today’s three items are about startups — ex-Twitter CEO Parag Agrawal’s Parallel (Tale 2), and Masa Son’s Roze (Take 3) —

Here are today’s Three Key Takes:

(1) Post-Mortem on Mag 4-of-7 Q1 — AI Capex Continues to Climb. Four of the Mag 7 reported Q1 last night — Google, Microsoft, Amazon, Meta — and AI capex guidance came in slightly ahead of expectations across the board. The architecture read: Implicator AI — Cloud growth gives Big Tech more time to keep spending. Today’s morning companion: AI-RTZ #1072 — Big Tech Continues to Load on AI Infrastructure Capex.

MP Take: “Only Google is up. The other three are down. Overall AI capex number guidance is slightly ahead of expectations. Meta is at $145B on the high end vs earlier $135B for 2026. Google and Microsoft at $190B each now. Amazon highest still at $200B-plus.

Microsoft continues to monetize enterprise. Google surprises on the upside in Cloud and core AI businesses, especially Gemini. Meta continues to fund via its ad businesses. And Amazon is REALLY getting started leveraging its market-leading AWS global franchise. Particularly with OpenAI post revised Microsoft deal. All have a lot of headroom to continue to outperform in July Q2 and beyond.”

(2) ex-Twitter CEO Parag Agrawal’s New Parallel AI Startup. Parag Agrawal — Twitter’s last CEO before Elon — has raised another $100M for Parallel AI, an agent-to-agent internet startup now valued at $2B+. The reporting: Implicator AI — Parag Agrawal raised $100M; now he has to build the web’s second user. Standing thesis on AI agents needing their own machine-to-machine internet: AI-RTZ #1023 — AI Agents Increasingly Need an ‘m2m’ Internet. And on a better internet for both AIs and humans: AI-RTZ #823 — A Better Internet for Both AIs and Humans.

MP Take: “It’s a timely move by this team in an opportunity just getting started. Expect a lot more players soon, especially on the startup front.

It’s an area where the incumbent Big Tech public and private companies are too focused and invested in the core opportunity around LLM AIs and AI Data Centers. So the opportunity for a parallel AI Agents-to-Agents ‘m2m’ internet is a secondary focus for them for now — and that’s exactly the gap a startup like Parallel can run through.”

(3) Softbank Masa Son’s ‘Roze’ by any name, AI Infrastructure ‘Startup’. Softbank is reportedly planning a US listing for Roze — a new AI Data Center and robotics mega-startup — at a $100B+ valuation on a 2026 timeline. The reporting: FT — Softbank plans to list new AI Data Center and robotics mega AI startup in the US. Standing context on Softbank’s prior Stargate commitments with OpenAI: AI-RTZ #1006 — OpenAI’s Difficulties Scaling Up.

MP Take: “Roze is Softbank’s ‘fast-follow’ version of the SpaceX/xAI/Tesla Optimus humanoid robot strategy — an Elon ‘Mini-Me’ strategy. Masa this time on AI as major FOMO investor vs prior tech waves, where he was early and invested in prime global positions.

He’s taking his traditional big swings, but the current haste in his approach is in contrast to prior strategies that have worked for him over the last three decades and tech waves. The Roze project needs a lot more definition and clarity, especially if it’s truly on a 2026 IPO timeline at a $100B+ valuation.”

Plus: Gadget AI — GenZ opinions of AI continue to hit new lows, per the latest round of consumer surveys. Source: The Verge — Gen Z and AI. Standing thesis on young people weary of AI gadgets beyond job-loss fears: AI-RTZ #872 — The Glass Half Empty on AI Devices. On AI Trust as the key driver of mainstream success: AI-RTZ #382 — Scaling AI Trust is Job 1. And on the joy gap: AI Ramblings Daily #58 — AI in US Being Sold with Fear, Not Joy.

MP Take: “This is partly due to today’s AI being in its ‘mainframe’ enterprise stage — and not a true consumer product like the internet (AIM), social (Facebook), or mobile (Snapchat) eras, where young people were enticed by ‘killer apps’ right at the beginning. Also, the AI industry is too enamored with their own ‘AGI-pilled’ products and plans to empathize with the social stigma and antipathy to today’s AI offerings on the consumer side.

Most people, especially young ones, will get most of the basic AI advantages via free offerings from Google and Apple. The ‘hot’ AI coding and enterprise ‘mainframe’ AI apps are not focused on their core needs.

Latest numbers highlight the gap between AI companies building and selling what they want to sell vs understanding what users — especially young ones — want to buy. As mentioned above, AI needs a more joyful set of value propositions — not just ‘learn AI for your job or for a job.’ It’s a global challenge. And it’ll take longer than currently anticipated by all.”

Bonus — today’s AI-RTZ companion #1072 covers Big Tech continuing to load on AI Infrastructure Capex — the morning text drills deeper into the Mag-4 Q1 prints framing today’s show.

Closing Qs —

• What are the non-Mag-7 earnings results that caught my attention in Q1 thus far? Samsung. Samsung’s quarter was above expectation, especially driven by AI memory demand. They’re expecting the memory shortage to worsen in 2026 — just like SK Hynix. Sources: CNBC — Samsung Q1 earnings on AI memory chip demand · Nikkei Asia — Samsung warns memory shortage to deepen next year as 2027 orders come in. The Korea-memory thread continues from Ep 65.

• What AI product from an unexpected source caught my attention this week? Netflix. Netflix rolled out a vertical-video ‘clips’ feed in its mobile app — US and global. Source: The Verge — Netflix vertical video feed. Classic example of TikTok-style consumer AI surfacing in places you don’t expect — and a quiet reminder that the consumer-AI battlefield isn’t just OpenAI vs Google. Stay tuned.

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

Clips from today’s episode

Short — Gen Z’s Apathy Toward AI GenZ sentiment on AI keeps slipping. AI is in its ‘mainframe’ enterprise stage — not a true consumer product like the internet (AIM), social (Facebook), or mobile (Snapchat) eras, where young people were enticed by ‘killer apps’ right at the beginning.

Short — AI Agents: The Next Internet Infrastructure Parag Agrawal’s Parallel AI raised $100M at a $2B+ valuation for an agent-to-agent internet. Big Tech is too focused on LLM AIs and AI Data Centers. The ‘m2m’ internet for AI agents is a secondary focus for them — that’s exactly the gap a startup like Parallel can run through.

Short — Masa’s AI IPO: A Mini-Me Strategy Softbank’s Roze is a ‘fast-follow’ of the SpaceX/xAI/Tesla Optimus humanoid strategy — an Elon ‘Mini-Me’ play. Masa on AI as major FOMO vs prior tech waves where he was early. The 2026 IPO timeline at $100B+ needs more definition and clarity.

Short — AI: Fear-Driven Adoption AI is being sold through fear, not joy. The industry is too AGI-pilled to empathize with the social stigma and antipathy to today’s AI offerings on the consumer side. Latest numbers highlight the gap between what AI companies build vs what users — especially young ones — want to buy.

Stay Tuned.

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 #1072 — Big Tech Continues to Load on AI Infrastructure Capex.

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 #66.

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

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

Take 1 — Mag 4-of-7 Q1 AI Capex:

• Implicator AI — Cloud growth gives Big Tech more time to keep spending

• AI-RTZ #1072 — Big Tech Continues to Load on AI Infrastructure Capex

Take 2 — Parag Agrawal’s Parallel AI:

• Implicator AI — Parag Agrawal raised $100M; now he has to build the web’s second user

• AI-RTZ #1023 — AI Agents Increasingly Need an ‘m2m’ Internet

• AI-RTZ #823 — A Better Internet for Both AIs and Humans

Take 3 — Softbank’s Roze:

• FT — Softbank plans to list new AI Data Center and robotics mega AI startup in the US

• AI-RTZ #1006 — OpenAI’s Difficulties Scaling Up

Gadget AI — GenZ Opinions of AI:

• The Verge — Gen Z and AI

• AI-RTZ #872 — The Glass Half Empty on AI Devices

• AI-RTZ #382 — Scaling AI Trust is Job 1

• AI Ramblings Daily #58 — AI in US Being Sold with Fear, Not Joy

Q1 — Samsung Q1 / Memory Boom:

• CNBC — Samsung Q1 earnings on AI memory chip demand

• Nikkei Asia — Samsung warns memory shortage to deepen next year as 2027 orders come in

Q2 — Netflix Vertical Video Clips:

• The Verge — Netflix vertical video feed mobile app UI

Companion text:

• AI-RTZ #1072 — Big Tech Continues to Load on AI Infrastructure Capex

The frame running through every item today: “AI Infrastructure investment and pricing dynamics remain the main show for this quarter and far more.”

Tonight’s Mag 7 reports — Amazon, Meta, Google, Microsoft after the close — are just the latest of many updates ahead.

Today’s three items each pressure-test a different layer of the AI capex story: AI compute commitments, AI silicon roadmaps, and AI memory supply.

Three Key Takes today:

(1) OpenAI Evolves Beyond Stargate. OpenAI is shifting from building and owning AI infrastructure to working with multiple partners to build it on their behalf — capex via renting, not owning. The reporting: FT — OpenAI evolves beyond Stargate. Standing context on OpenAI’s earlier scaling difficulties: AI-RTZ #1006 — OpenAI’s Difficulties Scaling Up.

MP Take: OpenAI has made a marked shift from building and owning its AI infrastructure to working with many partners to build and rent AI capex. Provides multi-gigawatt and multi-hundred-billion-dollar capex financing flexibility — especially against the impending mega-AI IPO in 2026. Takes reliance off Softbank.

The original Stargate framing was a single-partner commitment. First with Microsoft, then replaced by Softbank. The new picture is a portfolio of AI infrastructure relationships (Oracle, CoreWeave et al), that reduce concentration risk on both sides — OpenAI gets latitude on cost and capacity ramps, partners get to compete for slots in the buildout.”

(2) Google Rolls Out 8th-Generation TPU for AI Training and Inference. Strong update from Google, but the GPU race is far from over. The architecture read: Implicator AI — Nvidia still wins per chip; Google changed the fight; Sony changed the benchmark. Standing thesis on Google vs Nvidia silicon: AI-RTZ #1065 — Google’s Latest TPU Chips Improve.

MP Take: Nvidia’s GPU roadmap remains in the lead — Blackwell, Vera Rubin, Feynman, and beyond. AI GPUs now seeing 1-to-1 CPUs for inference, vs 1 CPU to many GPUs before. That ratio shift opens the door for a range of new incumbents and new entrants.

Intel has been a clear beneficiary. Memory companies continue to benefit — SK Hynix, Samsung, Micron, and others. The TPU update is real, but the AI-silicon stack is broader than the headline GPU-vs-TPU framing suggests. AI compute is not a zero-sum game.”

(3) Memory Companies Make South Korea’s Stock Market Bigger Than the UK. Korea has surpassed the UK to become the world’s 8th-largest stock market — driven almost entirely by SK Hynix and Samsung’s memory franchises riding the AI capex wave. The market-cap story: CNBC — South Korea overtakes UK as world’s 8th-largest stock market. The memory-boom narrative: FT — South Korea’s memory boom. Standing backcat across the bottleneck arc: AI-RTZ #1030 — Global memory shortages, AI-RTZ #918 — Memory ramping up as the next AI bottleneck, AI-RTZ #951 — Memory chip supply constraints, AI-RTZ #557 — SK Hynix ready for its AI close-up.

MP Take: It’s notable that South Korea has surpassed Japan, in particular. Memory remains a multi-year AI bottleneck that is likely to remain a supply constraint through the decade at least. Despite higher-than-historic volatility due to questions around supply ramps by the key players.

The Korea-passes-UK milestone is a pricing-and-bottleneck story dressed as a market-cap headline. Watch the supply ramps — that’s where the volatility lives.

Plus: Gadget AI — Apple is reportedly adding Siri natively into iOS and iPhone Photos this June. Source: Bloomberg — Apple’s Siri-in-Photos plan. Standing thesis on Apple’s Siri-Gemini integration potential: AI-RTZ #965 — Apple Goes with Google Gemini. Standing thesis on Apple’s vertical-stack supply-chain advantages: AI-RTZ #1010 — Apple Supply Chain Lock-In Tech.

MP Take: Apple continues to play the AI opportunity with its core ecosystem and vertical-stack advantages. Especially evident in bottom-up AI models — large and small — running on Apple Macs, iPhones, and other devices, with opportunities for Apple’s world-leading App Store developers to leverage AIs of all types via Apple operating systems.

Watch for key Apple software initiatives at WWDC this June — especially around AI MCP and App Intent protocols focus, among other initiatives.

Bonus — today’s AI-RTZ companion #1071 covers OpenAI’s Stormy Weather — the IPO debate, Microsoft re-cut, Elon trial, Stargate evolution, all under one weather-system frame. Different lens, same throughline: AI infrastructure is the main show, and OpenAI’s repositioning is the most-watched piece of it.

Closing Qs —

• What’s the key Apple AI product MP is most looking forward to in 2026?

  • A standalone Siri App with Gemini AI integration. That’s the unlock — Siri as a true AI surface on every Apple device, with Google’s brain behind it. Apple’s 2.5B+ device users plus the App Store/Services franchise plus a credible AI brain via Gemini — that’s the integration that actually changes the daily-use experience for the largest installed base of mainstream consumers.

• What app in current daily use will it most likely replace?

  • Perplexity. Once Siri-with-Gemini lands at the OS level on iPhone, the daily-use case can shift due to ubiquity across Apple devices. Stay tuned.

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

Shorts Clips from today’s episode

Short — AI Compute Is Not a Zero-Sum Game Google’s 8th-gen TPU is out. Nvidia’s roadmap stays in the lead. But AI compute isn’t zero-sum — the 1-to-1 GPU/CPU shift opens the door for incumbents AND new entrants. Intel benefits. Memory companies (SK Hynix, Samsung, Micron) continue to benefit.

Short — Siri Comes to iOS Photos Apple is reportedly adding Siri natively into iOS and iPhone Photos. The next visible step in surfacing Apple’s AI ambitions. Watch WWDC this June for AI MCP and App Intent focus.

Short — Standalone Siri App with Gemini AI The Apple AI product MP is most looking forward to in 2026. Apple’s 2.5B+ device users plus the App Store/Services franchise plus a credible AI brain via Gemini — that’s the integration that changes daily-use experience for the largest installed base of mainstream consumers.

Short — OpenAI Shifts From Build to Rent OpenAI’s marked shift from building and owning AI infrastructure to working with many partners to build and rent AI capex. Multi-gigawatt and multi-hundred-billion-dollar capex flexibility, especially pre-IPO. Takes reliance off Softbank.

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 #1071 — OpenAI’s Storm Weather.

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 #65.

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

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

Take 1 — OpenAI Beyond Stargate:

• FT — OpenAI evolves beyond Stargate

• AI-RTZ #1006 — OpenAI’s Difficulties Scaling Up

Take 2 — Google’s 8th-Gen TPU vs Nvidia:

• Implicator AI — Nvidia still wins per chip; Google changed the fight; Sony changed the benchmark

• AI-RTZ #1065 — Google’s Latest TPU Chips Improve

Take 3 — Memory Boom Passes UK:

• CNBC — South Korea overtakes UK as world’s 8th-largest stock market

• FT — South Korea’s memory boom

• AI-RTZ #1030 — Global memory shortages

• AI-RTZ #918 — Memory ramping up as the next AI bottleneck

• AI-RTZ #951 — Memory chip supply constraints

• AI-RTZ #557 — SK Hynix ready for its AI close-up

Plus — Gadget AI on Apple Siri + iOS Photos:

• Bloomberg — Apple’s Siri-in-Photos plan

• AI-RTZ #965 — Apple Goes with Google Gemini

• AI-RTZ #1010 — Apple Supply Chain Lock-In Tech

Bonus — today’s AI-RTZ companion:

• AI-RTZ #1071 — OpenAI’s Storm Weather (swap placeholder with public URL after publish)