The Strongest Case Against Apple's Design

"아빠는 원래 monolithic에 반대하던 사람이야. 그건 지금도 그렇고. 그냥 애플이 쥐밟았다고 생각해. 하지만, 인정할 건 인정해야 하니까."
— the founder, 2026-09-15. ("I was against monolithic design from the start, and I still am. I think Apple simply stepped on a mouse. But what must be admitted must be admitted.")

Stated at Full Strength

A strawman would be easy here and would insult the position. So here is the case against a sealed, single-vendor package, made as well as the household can make it, every item labelled as the judgment it is.

Capacity is chosen once, by you, on the day you pay. There is no second slot. If your workload grows past the pool — and in an era when the models double in size between purchases, it will — the only upgrade path is another whole machine. The modular PC lets a buyer be wrong about the future cheaply; the sealed package charges full price for the correction.
The vendor's roadmap becomes your roadmap. When memory gets cheap, or faster, or denser, a PC owner benefits the week a new module ships. A Mac owner benefits when Apple ships a new Mac, at Apple's cadence, at Apple's price. Between generations, the ceiling does not move: 512 GB was the top of the range in March 2025 and is still the top in the M5 Ultra of September 2026.
There is no repair path, only a replacement path. A failed memory chip on a PC is a module. On a Mac it is a logic board, which is to say the computer. A sealed package turns every component failure into a machine failure.
The erased boundary erased the isolation with it. One memory pool means one arbitration layer, and one process holding the GPU can starve every other. The discrete-GPU world paid for its boundary with copies and got a fence in return. This household has lived with the side effect across four hardware generations; the GPU track tells it in full.
The machine freezes while the world moves. A purchase fixes one year's capability for the three years or so a machine is worth running. Every model release after that day is a free upgrade to someone renting compute and a depreciation charge to someone who bought it. A modular machine softens this — swap the part that fell behind — and a sealed one cannot.
A sealed package prices itself. With no substitution inside the box there is no competitor's part to price against, so the price is set by what the buyer will pay for the whole object, and a generation that raises bandwidth by half and leaves the ceiling where it was can still ask more. The honest price gauge, in the household's experience, is the used prior generation: two used 512 GB machines against one new one is a comparison the new one rarely wins on capacity, redundancy or known bugs.

None of these is a claim that the design does not work. All of them are claims about what it costs to own. That is the founder's position, and it is not a fringe position: it is the layered-industry position, held by most of the people who built the industry, and the referee in lesson four agrees it will be right again.

What Must Be Admitted

And then the other half of the founder's sentence. The household that holds every objection above runs its entire local AI stack — retrieval, embeddings and reranking, image generation, media transcoding, a local fallback rung for its coding agent — on nine Apple silicon Macs, two of them at the 512 GB ceiling. It does so because in the regime the referee named, nothing else fits the job at any price a household pays. What must be admitted is threefold. The design is right for the regime: integrated memory is the correct architecture for a bandwidth-bound workload that does not yet fit anywhere. The mouse was real: Apple built this for a phone before the workload existed, and then — this is the part fans skip and critics deny — followed through on purpose, with a framework, a 512 GB configuration, per-core accelerators and a die-to-die fabric aimed squarely at the accident. And the costs above are paid knowingly, by people who can price them, which is a different thing from being fooled.

Both halves ride together for the rest of this quest. When a later lesson says unified memory wins, it is not forgetting this one; when a later lesson names a cost, it is not forgetting the referee.

Code

sealed_costs.py — each cost, its level, and whether Apple could remove it without changing the architecture·python

#!/usr/bin/env python3
"""The case against a sealed package, as data. 'level' is from track one
(die / package / industry); 'architectural' means removing the cost would
mean giving up unified memory itself, not merely changing a policy."""

COSTS = [
    # cost,                                  level,      architectural?  what would remove it
    ("capacity fixed at purchase",            "package",  True,  "a socket — which the bus cannot afford (lesson 3)"),
    ("vendor roadmap = owner roadmap",        "industry", False, "a second source for the package; Apple chooses not to"),
    ("no repair path, only replacement",      "package",  True,  "replaceable modules; same socket problem"),
    ("no isolation between GPU claimants",    "package",  False, "an arbitration policy in macOS; not a hardware limit"),
    ("machine freezes while models move",     "industry", False, "true of any purchase; softened by modular upgrades"),
    ("no competitor's part to price against", "industry", False, "a UCIe/CXL-style open package; a business decision"),
]

print(f"{'cost':40} {'level':9} {'architectural':13}  what would remove it")
for cost, level, arch, fix in COSTS:
    print(f"{cost:40} {level:9} {'yes' if arch else 'no':13}  {fix}")

arch = sum(1 for c in COSTS if c[2])
print(f"\n{arch} of {len(COSTS)} are the architecture itself; {len(COSTS) - arch} are choices Apple could reverse.")

Exercise

Run sealed_costs.py. Then write the strongest case FOR the sealed package in the same six-row format — each row a benefit, its level, whether it is architectural, and what the modular alternative would have to give up to match it. Finish by marking which rows in the two tables are the same fact seen from opposite sides.

Hint

At least three pairs are mirror images: fixed capacity ↔ a bus a socket could not carry; no GPU isolation ↔ zero-copy sharing; no competitor's part ↔ the processor, memory and OS tuned as one object. When a cost and a benefit are one fact, neither side gets to keep only its half.

Stated at Full Strength

What Must Be Admitted

Code

External links

Exercise

Progress

Comments 0