Then Apple Leaned In

"What must be admitted: once the mouse was visible, Apple spent real silicon following it, on purpose, release after release."

The Follow-Through, Dated

Luck ends the day a company notices. From December 2023 onward, Apple's releases point at inference with a consistency that no accident produces. Each item below is a primary source with a date; together they are the founder's "admit what must be admitted" clause, made concrete.

Date	What Apple shipped	Apple's words	What it aims at	Evidence
2023-11-28 / 12-05	MLX — an array framework for Apple silicon; PyPI 0.0.2 a week later	"MLX is designed to take advantage of" unified memory	a first-party door for inference, matching the community's	GitHub, PyPI, MLX docs
2025-03-05	M3 Ultra with 512 GB	"AI professionals can use Mac Studio with M3 Ultra to run large language models (LLMs) with over 600 billion parameters directly on device"; "M3 Ultra is built for AI"	capacity, named for models for the first time	Apple newsroom
2025-10-15	M5 with a Neural Accelerator in every GPU core; Tensor APIs in Metal 4	"directly programming the Neural Accelerators using Tensor APIs in Metal 4"	prefill compute	Apple newsroom
2025-11-19	MLX on M5: TensorOps support, macOS 26.2 required	"up to 4x speedup … for time-to-first-token"; decode "19-27% … thanks to its greater memory bandwidth"	Apple measuring its own inference, by phase	machinelearning.apple.com
2025-12-12	macOS 26.2: RDMA over Thunderbolt 5	"available starting with macOS 26.2 on Macs with Apple silicon with Thunderbolt 5"; the release notes name "distributed AI inference using MLX"	pooling Macs for models that exceed one pool	Apple technote TN3205; support page
2026-03-03	M5 Pro / Max, Fusion Architecture	"up to 4x faster LLM prompt processing than M4 Pro and M4 Max"	prefill, on laptops	Apple newsroom
2026-08-25	M5 Ultra, quad-die; Mac Studio with 512 GB again	"run massive models entirely on device"; "A cluster of four Mac Studio systems delivers up to a remarkable 3x faster AI inference than a single system"	capacity and clustering, sold as the product's purpose	Apple newsroom

Read the Table as Intent

Three things stand out. First, the vocabulary changes. The 2021 release talked about ProRes streams; the 2025 release says "over 600 billion parameters directly on device" and "built for AI". Apple began describing the pool in the workload's own units. Second, the silicon changes: per-core Neural Accelerators are a compute answer to the phase — prefill — that bandwidth does not help, and RDMA over Thunderbolt is an interconnect answer to the ceiling, both aimed at problems only an inference workload has. Third, the measurement changes: by late 2025 Apple was publishing its own inference figures by phase, attributing decode gains to bandwidth and prefill gains to the accelerators — the physics track's split, in Apple's own accounting.

None of this is evidence that Apple designed the M1 for models (the previous lessons settled that), and all of it is evidence that Apple understood, by 2025, exactly what the community had found in 2023 and chose to build toward it. That is the deliberate follow-through, and it is why the founder's phrase has three parts rather than two.

How the Household Reads It

The household bought its two 512 GB machines the month the M3 Ultra shipped, for exactly the sentence in the release. It runs Apple's framework for the same reason the community did: it is the door that reads each weight once. And it has not chained its Macs with RDMA, for a reason the big-models track states as a judgment — the direction is right and the time is not. Leaning in is what Apple did; whether to follow is a separate question, answered per workload, and the next lesson lists what Apple's follow-through has not changed.

Code

follow_through.py — Apple's inference-facing releases, by phase they address·python

#!/usr/bin/env python3
"""Apple's post-2023 releases sorted by which inference constraint each one
addresses. Every row is a dated primary source; the sorting is this quest's."""

RELEASES = [
    # date,        item,                                  constraint addressed
    ("2023-12-05", "MLX 0.0.2 on PyPI",                    "a first-party runtime door"),
    ("2025-03-05", "M3 Ultra, 512 GB",                     "capacity (the pool)"),
    ("2025-10-15", "M5 Neural Accelerators + Tensor APIs", "prefill compute"),
    ("2025-11-19", "MLX 0.30 TensorOps on M5",             "prefill compute (software)"),
    ("2025-12-12", "macOS 26.2 RDMA over Thunderbolt 5",   "capacity beyond one pool (interconnect)"),
    ("2026-03-03", "M5 Pro / Max Fusion Architecture",     "prefill compute on laptops; two dies"),
    ("2026-08-25", "M5 Ultra 1.2 TB/s, quad-die",          "decode bandwidth (+50%); die count"),
]

by = {}
for d, item, c in RELEASES:
    by.setdefault(c.split(" (")[0], []).append((d, item))
for c, items in by.items():
    print(c)
    for d, item in items:
        print(f"   {d}  {item}")

print("\nNot on the list: anything that raises the 512 GB ceiling of a single pool.")

Exercise

Run follow_through.py and note which constraint has the most releases against it and which has none. Then find one Apple release since 2023 that is NOT on the table and decide whether it belongs: does it address an inference constraint (which one?), or is it a different workload wearing AI vocabulary? Add it with your verdict.

Hint

Candidates: Apple Intelligence's on-device model, Private Cloud Compute, the Foundation Models framework for developers, Apple's Core ML Llama article. The test is the constraint: capacity, bandwidth, prefill compute, interconnect, or a runtime door. Marketing that names no constraint is not follow-through; it is vocabulary.