Apple's critical weakness

ELI5/TLDR

Apple’s M-series chips are, on paper, the best hardware money can buy for data science. They have unified memory, dozens of cores, and a neural engine that costs a fraction of equivalent Nvidia kit. But almost nobody in data science uses them, because the entire ecosystem runs on Python, which is single-threaded, interpreted, and physically incapable of using what Apple built. Kiraa argues the gap isn’t hardware or even MLX — it’s the missing Swift-native version of Pandas, NumPy and the rest of the analytical toolkit.

The Full Story

The thesis under the thesis

Kiraa has been beating the “Apple wins AI eventually” drum for a while. The argument: frontier models and cloud compute will commoditise. What can’t be commoditised is the relationship with the user, which Apple already owns. This video is the caveat — there are real software gaps holding Apple back, and CUDA isn’t even the worst of them.

Apple instead focuses on something far more valuable, which is the relationship with the user.

The hardware case is already won

Traditional setup: CPU does general work, GPU is bolted on for parallel work, data shuttles between them, and you wait. Apple silicon collapses that into one shared memory pool — CPU, GPU and neural engine all reading from the same address space. A Mac Studio with 192GB of unified memory costs less than the specialised gear it replaces.

For analytical workloads — millions of rows of financial data, anomaly detection, the kind of thing where memory bandwidth and core count matter — this is structurally superior architecture.

The Python tax

So why isn’t business running its analytics on Mac Studios? One word: Python. The data science world is built on it, and Python is interpreted, single-threaded, and was never designed for performance. The community has papered over this for decades by writing the heavy libraries (NumPy, Pandas) in C and Fortran and wrapping them in Python. It works, sort of. You still pay the Python overhead, and you still can’t use multiple cores properly.

You bought a team of V12 engine, and the software is running like a single cylinder lawnmower.

On a Windows laptop you wouldn’t notice. On an M-series chip with a dozen-plus performance cores, you’re leaving most of the silicon idle.

Swift as the unlock

Swift is compiled, fast, and talks directly to Metal for GPU work and the neural engine for ML. Kiraa’s team built a Swift-native version of a Python data workflow and ran them side-by-side on the same data. Swift came in over 80x faster. He admits the example is loaded — a Python expert can optimise specific tasks heavily — but the structural point holds. Swift sees the hardware. Python doesn’t.

The CFO version: a 3-day Python job finishes in 1 hour in Swift. That isn’t a speed-up, it’s a change in what’s possible. Overnight batches become real-time. A desktop replaces a cloud cluster.

The actual problem is the ecosystem

The real barrier is the libraries that everyone depends on, and they’re locked into a language that can’t take advantage of what Apple built.

MLX is fine for training and inference, but data scientists spend most of their day on data wrangling, not training models. The missing piece is a Swift-native Pandas, NumPy, and the rest of the analytical toolkit. Kiraa is building a Pandas port himself — using the same low-level C libraries but linking through Metal for GPU acceleration — and is considering open-sourcing it.

His pitch to Apple, if anyone there is listening: fund or seed a few critical libraries. The hardware advantage is so large that even a partial port changes the economics. iPad-native analytics apps become possible. Smaller companies stop renting cloud compute. Auditable, deterministic processing replaces agents “running amok and spitting out rubbish”.

Key Takeaways

• Apple’s moat in AI is the user relationship, not the model layer. Frontier models and cloud will commoditise.
• Two software gaps hold Apple back from a developer-side win: CUDA (Nvidia’s lock-in) and Python (the data science default).
• Unified memory + 192GB at consumer prices is structurally superior for analytical workloads — but only if your software can use it.
• Python’s single-thread, interpreted nature wastes most of an M-series chip’s parallelism even when libraries are written in C.
• Swift + Metal benchmarks 80x faster than Python on Kiraa’s test workload; CFO-translation is “3 days becomes 1 hour”.
• The bottleneck is the missing Pandas/NumPy equivalent in Swift — not the language, not the hardware. Whoever builds that ecosystem unlocks the platform.
• A working Swift data-science stack would shift workloads from cloud back to local Mac Studios and enable iPad-native enterprise analytics.

Claude’s Take

The hardware-versus-ecosystem argument is correct and well-known to anyone who’s tried to do serious numerical work on a Mac. The 80x number is the part to be careful with — that’s a hand-tuned Swift+Metal demo against unoptimised Python on a workload Kiraa picked. The honest comparison would be against Python with Polars, NumPy with proper BLAS, or a JAX-on-Metal setup, not vanilla Pandas. The structural point — that Python can’t fully use M-series silicon and Swift can — survives the discount. The dramatic multiple does not.

The bigger blind spot is on Apple’s incentives. Apple has had a decade to seed a Swift data-science ecosystem and chose not to. MLX shipped in 2023 and got real traction; the analytical layer hasn’t, because Apple doesn’t sell to enterprise data teams and has no obvious revenue path from doing so. Saying “Apple should fund this” is true the same way “Microsoft should have made a great phone” was true. Capability isn’t the constraint. Strategic interest is.

Worth watching for the framing of where the moat actually sits. The “Apple wins eventually” call is a real bet, not a slogan, and the Python diagnosis is the most concrete version of the bear case I’ve seen on this channel. 7/10 — clear argument, useful frame, oversold benchmark.