▶ YouTube Summaries

← All summaries
Theo#▶ Watch

Apple Mac Price Hikes from the RAM Crisis

2026-06-29 Mon ⏱ 22 min t3dotgg

The RAM/NAND-flash crisis (driven by AI GPU demand) has forced Apple to hike Mac prices hard and fast — a maxed-out 128GB M5 Max that cost ~$7K now runs ~$10K. Theo explains why Apple silicon's unified memory still makes Macs a strong value for local LLM inference, walks through what caused the price spike, and gives concrete buying advice (act now on refurb deals before they vanish).

Why Mac specs don't compare to a normal PC

  • Apple's SoC puts CPU, GPU, video encode/decode, and RAM on one die; RAM is unified — shared between CPU and GPU.
  • This is why people cluster Mac Studios for AI: a lot of unified RAM can hold large models the GPU can use directly.
  • Example: a 5090 has only 32GB VRAM. Models that fit (e.g. Qwen 3 27B) run far faster on the 5090 than on a MacBook. But once a model exceeds ~30GB (e.g. GPT-OSS-120B, ~60GB), normal GPUs must offload and crawl — Theo's 128GB MacBook runs it at ~100 tokens/sec.
  • Rule of thumb: small model that fits a normal GPU → use the GPU; big local model (>30GB) → a high-RAM MacBook is one of the best values, even post-hike.
  • antirez built a simple C runtime for DeepSeek-4 Flash that plugs into environments like PyTorch and runs fully offline on Mac RAM.

Alternatives for big local models

  • Framework Desktop (AMD Strix Halo, Max Pro 395, shared LPDDR5): 128GB now ~$3,500, 64GB ~$2,000, refurb 64GB ~$1,700 (no SSD included). Good non-Apple option for large local models.
  • Nvidia DGX Spark (128GB, 8-core ARM): Theo calls it "kind of garbage" — only worth it if you specifically need CUDA with lots of RAM.

Disk speed as a second value prop

  • MacBook SSD hits 12,000–15,000 MB/s writes vs ~7,000 MB/s for a typical PCIe 4 NVMe — approaching DDR4 RAM speeds.
  • Apple solders top-tier chips onto the board/processor, which is why their storage/memory could command a premium before.

What caused the price spike

  • 32GB DDR5 kit went from ~$100 last year to $500–600; still rising.
  • NAND flash used for SSDs and the chips GPU makers need overlap; AI buildout (OpenAI, Nvidia) is soaking up allocation, starving consumer RAM and SSD supply.
  • Apple kept prices flat for a year via legacy supply deals; those are now expiring and being renegotiated at much higher cost.
  • Rumor: Samsung planned a 60% hike, opened at 100% expecting Apple to haggle down — Apple immediately accepted the 100% to secure allocation.
  • Apple is even petitioning the US government to buy memory from CXMT, a blacklisted Chinese supplier — a sign of how desperate allocation has gotten.
  • Compute is shuffling everywhere: Google selling TPUs to Meta, Meta renting GPUs from SpaceX, etc.

Buying advice

  • MacBook Neo: was a great deal at $600, now starts at ~$700 — less compelling; a discounted older M3 Air (16GB/512GB ~$800) is a better basic machine. Don't pay full MSRP for an Air right now.
  • Refurb store restocked 128GB M5 Max options: ~$5,800 for 128GB / 2TB, ~$3,700 for the 2TB nano-texture — "not great, but not terrible." Get the nano-texture display.
  • Refurb catches: no Apple Card payment plan, not all countries, can't stack with student discount; same warranty as new.
  • 64GB M5 (~$3K) is a strong pick if you run lots of Docker but not huge models.
  • Mac mini / Mac Studio for dedicated inference/agent work: basically no good deals right now; base mini runs hot (pegged at 100°C / 100% CPU on agent work).
  • Theo admits he panic-bought another MacBook — expects the market to get worse before it gets better.