Apple approves drivers that let AMD and Nvidia eGPUs run on Mac — software designed for AI, though, and not built for gaming

Apple approved third-party drivers that enable AMD and Nvidia eGPUs to run on Apple silicon (Thunderbolt/USB4) for AI LLM workloads without needing SIP workarounds. Tiny Corp sells the red v2 (4x AMD 9070XT) for $12,000 and the green v2 Blackwell (4x RTX Pro 6000 Blackwell) for $65,000 and plans an exabox in 2027 with 720 RDNA5 AT0 XL GPUs (~1 exaflop) for about $10M. The custom driver is aimed at AI training/inference (not gaming), and rising Mac demand for AI workloads has tightened supply—Apple removed the 512GB Unified Memory Mac Studio option and the 256GB model saw a $400 price increase.

Analysis

Opening macOS to high-performance external GPUs creates a reframing of where AI development spending happens: instead of exclusively buying rack/cloud capacity, a portion of incremental spend shifts to endpoint+peripheral capex and recurring software/service spend tied to that endpoint. That structural shift favors controller/bridge semiconductor vendors and small-form-factor OEMs who capture the last-mile connectivity and cooling engineering — a multi-quarter revenue reallocation rather than an immediate bump to traditional GPU ASPs.

For GPU vendors, the economic impact diverges: the addressable base for inference and local experimentation expands, improving future software lock‑in for the vendor whose software stack performs best on mixed-host environments. However, per-unit ASP pressure can emerge if third‑party integrators bundle commodity GPUs into turnkey accelerators, compressing OEM margins and changing go‑to‑market leverage between silicon vendors and enclosure integrators.

Key risks and reversal mechanics are governance and performance ceilings. Platform-level controls, driver certification changes, or limited PCIe/TB4 bandwidth ceilings could cap usable GPU performance and therefore the willingness of enterprises to substitute cloud spend; conversely, fast, stable performance and a credible software stack could accelerate endpoint adoption inside 6–18 months and materially re-route some cloud spend to hybrid local endpoints.

Key Decisions for Investors

• Long AAPL via a 3–9 month call‑spread (buy 1 ATM call, sell 1 +15% OTM call) sized to lose <3% of portfolio — thesis: small but visible uplift in accessory/Pro Mac demand and higher attach rates for services over the next 2–4 quarters. Target 40–80% return if Mac units or ASPs move up 3–6% vs consensus; cut if Apple issues restrictive platform controls.
• Go asymmetric long NVDA vs short AMD for 6–12 months (long NVDA 6–9 month ATM calls, sell/short AMD 6–9 month near‑term puts or reduce AMD gross exposure) — rationale: broader endpoint deployment favors the vendor with dominant AI software ecosystem; reward if NVDA captures incremental dev/inference volume, with downside protection from selling premium in AMD. Size to a 2:1 notional long NVDA : short AMD exposure; stop‑loss at 20% adverse move on the pair.
• Tactical short or avoid pure GPU integrator small caps that lack software IP for 3–12 months — these businesses can be disintermediated by OEMs or large silicon partners if margins compress. Use this as a watchlist trade: initiate only after earnings that show inventory buildup or ASP pressure, target 30–50% downside over 12 months if execution falters.