Romeo Esparrago on Apple's August bombshells (videos) - 'A Studio is overkill, agreed, from a personal purchase budget & use case POV. So far, with my M4 Pro Mac Mini, the cloud AI agents I use (Comet, Runway, Suno) are pretty good. Even better, when I choose the processing to ‘computer’, it processes faster. Both off-cloud & time-to-output speed on my Mini is helping me conserve my monthly Ai credits = $ saved.'
on Apple's August bombshells (videos) - 'Hi, Fred, Another earlier discussion with my stepson may reveal both an issue for Apple and a possible path forward. “I was researching the plausibility for Apple to be at the forefront of this “trustless peer to peer LLM/AI compute network” (which clearly would be revenue generating). There is actually a constraint and I haven’t yet figured out an effective way around it. NVIDIA, AMD and Intel hardware has something called a “Trusted Execution Environment” (TEE, for short), which is cryptographically verifiable execution of AI compute in a fully confidential/private way (i.e., RAM and bus are all encrypted and the hardware owner can’t even see what is being run or computed). Apple hardware has something called a “Secure Enclave” which I thought could handle that, but it turns out it can’t be extended to LLM inference. It only can be used for secure key exchange. Apple needs to extend its Secure Enclave hardware capabilities so it works with MLX and the unified memory systems. At that point, you can have fully private local inference where the hardware owner won’t know what is being computed or what messages are being passed back and forth. Then, the install base of all compatible Macintosh machines could be leveraged in a peer to peer fashion for a “world computer” AI — trustless, secure, private, permissionless AI. (We even could add further capabilities like individual fine-tuning, etc., but that’s later.)” Later, he posted this: “Also, ARM (what M4 and M5s are built around) has a spec called Realms, that allow lower privileged software (e.g., an application or VM) to protect its content and execution from attacks or monitoring by higher-privileged software such as the OS or hypervisor. It’s optional and Apple hasn’t announced support for that functional standard. As soon as Apple supports ARM Realms, distributed, peer-to-peer, trust-less, private/confidential AI is possible with the Mac user base providing the compute fabric.”'
on Apple investors: Mark your calendars for 9/9/26 - 'Now here’s an interesting question I posed to Gemini: “how much would it have cost Apple to buy enough DRAM memory in Q1 2026 to supply 200 million iPhones for fiscal year 2027?” “Purchasing enough mobile DRAM in Q1 2026 to supply 200 million iPhones for 2027 would have cost Apple an estimated $14.4 billion to $18 billion for baseline configurations, climbing up to $24 billion if accounting for Pro model specifications. The calculation depends on the specific baseline component pricing and rumored hardware memory maps from the first quarter of 2026: The Cost Breakdown Framework • The Baseline Q1 2026 Price: Before the massive Q2 2026 price spike hit the industry, conventional mobile LPDDR5/5X DRAM sat at a baseline contract price of roughly $10 per GB. If applying Apple’s historic scale and bulk-purchasing power, an estimated high-volume contract rate would average around $8 to $10 per GB. • The 2027 iPhone Memory Profile (iPhone 18): Supply chain data reveals that Apple’s lower-end 2027 iPhones (such as the base iPhone 18 and 18e) are mapped to feature 9GB of DRAM (utilizing an A20 chip with 1.5GB x 6 dies), while premium Pro models require 12GB of DRAM. Simulated Purchase Scenarios Allocation Strategy RAM per Phone Total Volumetric RAM Needed Total Estimated Q1 Cost All Base-Model Allocation 9 GB 1.8 Billion GB $14.4B – $18.0 Billion All Pro-Model Allocation 12 GB 2.4 Billion GB $19.2B – $24.0 billion ⚠️ The Strategic Impact of the 2026 Memory Crisis Had Apple waited past Q1 2026 to negotiate those allocations, the financial exposure would have doubled. By mid-2026, mobile DRAM pricing surged by nearly 100% quarter-over-quarter, hitting upward of $19.30 to $21.00 per GB. According to analysis from JPMorgan and the Financial Times, this ongoing AI infrastructure supply squeeze means that memory is on track to account for up to 45% of an iPhone’s total component costs by 2027 (up from just 10% historically). This massive cost shift is precisely why Apple has aggressively lobbied to open access to alternative Chinese memory suppliers and why global RAM production capacity is already fully sold out through 2027.” “how about 230M iPhones of all types, and use the Feb-March 2026 months for memory costs?” “Securing enough mobile DRAM during the February–March 2026 window for a full multi-tier production run of 230 million iPhones would have cost Apple an estimated $19.3 billion to $24.8 billion. Locking in orders during these exact months represents a critical “last chance” window right before the absolute worst of the 2026 memory crisis upended the industry. The Volumetric RAM Map Industry analysis indicates Apple is moving away from uniform 8GB capacities across the board. The upcoming generational lineup splits the 230 million hardware allocations into two specific configurations: • The Baseline Allocation (9GB per phone): The standard iPhone models are mapped for 9GB of LPDDR5X DRAM to handle localized iOS 27 Apple Intelligence tasks. • The Premium Allocation (12GB per phone): The high-end Pro, Pro Max, and upcoming Fold models require 12GB of DRAM. Purchase Cost Breakdown (Feb-March 2026) During February and March 2026, TrendForce tracked mobile contract prices rising rapidly, solidifying standard high-tier baseline procurement costs at $10.00 per GB. Even factoring in Apple’s typical high-volume market leverage, the sheer size of the order means the absolute lowest bulk discount tier would likely sit at $8.00 per GB. Depending on how Apple splits its manufacturing volume between entry-level and Pro units, the financial breakdown scales as follows: • Scenario A: Balanced 50/50 Product Mix (115M Base / 115M Pro) -> Total Volumetric RAM: 2.42 Billion GB -> Estimated Total Cost: $19.32 Billion to $24.15 Billion • Scenario B: Premium-Heavy 40/60 Product Mix (92M Base / 138M Pro) -> Total Volumetric RAM: 2.48 Billion GB -> Estimated Total Cost: $19.87 Billion to $24.84 Billion Why the Feb-March Timeline was a High-Stakes Gamble Had Apple delayed this massive capital expenditure by even a few weeks, the price tag would have broken the bank: • The Q2 Explosion: Immediately following March, mobile DRAM contract prices suffered a historic 93% to 98% quarter-over-quarter surge, driving baseline component costs past $19.00 per GB. • The Rejection of Discounts: Suppliers like China’s memory giant CXMT began flatly rejecting Apple’s attempts to negotiate lower bulk prices, as AI data centers simultaneously bought out entire production pipelines at maximum market rates. Because Apple had to absorb these massive baseline procurement increases, it explains why outgoing leadership publicly warned Wall Street that price hikes are coming to consumer hardware to maintain Apple’s near-50% gross margins.'
on Premarket: Apple was red, turned green - 'The Nvidia earnings results should boost the entire market tomorrow. $64 billion in operating income and they forecast revenue growth of 70%. They will soon be making $100 billion or more per quarter.'
on Apple's August bombshells (videos) - 'Joseph, it’s great to hear the assessment from your stepson. Please ask him to compare with Nvidia’s announced, but not yet shipping, RTX Spark with unified memory. Since it’s not shipping, it’s OK to wait, or to just get good guesses. The developer perspective matters. Nvidia offers CUDA, but saddles it on to a Window system – ugh. Apple offers Metal, MLX and CoreML running on elegant, secure MacOS.'
on Premarket: Apple was red, turned green - 'Re: #4. News seems to indicate that those infilled early orders were upgraded to the new models – for free!'
on Premarket: Apple was red, turned green - '@Joseph, Low again at ~29M. Much of that was in the last few minutes. Weird.'
on Apple's August bombshells (videos) - 'Only if you -need- that power. I have an M1 Pro and an M4 (air). Load averages on those machines remains really low. It’s rare for me to keep more than 2 cores busy. And the only time since I’ve bought the M1 Pro that I ever heard the fans come on was with a local AI program that tried to index a rather large photo collection (so I could say, “Show me all the photos with boxcars.”) That took 36 hours, but that was expected.'
on Apple investors: Mark your calendars for 9/9/26 - 'Man, I have not been drooling like this over a Fall event season for Apple in a long time. Smart hubs. Folding phones. New Airpods. Awesome Siri AI. The most powerful desktops. The new iPhones are almost overlooked. Maybe even “One More Thing?” I suspect I will be utilizing Apple Upgrade a lot in the coming months.'
