The consensus is simple: NVIDIA rules AI. It just became the first company in history to cross a $5 trillion market cap, revenue is up 85% year over year, and the entire S&P 500 trades like a leveraged bet on Jensen Huang’s GPUs staying busy. A widely shared opinion piece on Hacker News argues that consensus is measuring the wrong race, and that the real king of AI is Apple. That take deserves a hard look, because the logic underneath it is stronger than the headline sounds.
The scoreboard everyone uses is the wrong one
Today we keep score two ways: who has the best frontier model, and who sells the chips to train it. By that measure NVIDIA is untouchable. But training is not where technology waves end. Mainframes became PCs. Timeshares became laptops. The cloud is bending toward the edge. According to Hacker News, the same thing is happening to AI right now, in public, faster than the market is willing to price in.
The question that follows isn’t “who has the best model?” It’s “when models become free, whose hardware runs them?”
The cadence is the real story
The author’s evidence is a release calendar, and it’s worth reading as a pattern rather than a list:
- February: Alibaba ships Qwen 3.5, a 397-billion-parameter model, weights public.
- April: DeepSeek drops V4 at 1.6 trillion parameters, MIT-licensed, hitting 80.6% on SWE-bench Verified.
- June: MiniMax, Moonshot, and Zhipu all ship near-frontier open models with million-token context windows.
- July 16: Moonshot announces Kimi K3, roughly 2.8 trillion parameters, reportedly third in the world behind only Anthropic and OpenAI, with weights promised by July 27.
Bloomberg framed this as China “closing the gap.” The sharper read: the distance between the best model money can buy and the best model you can download for free is now single-digit benchmark points, and it shrinks every few weeks. You can’t build a moat against a cadence. Price one release into your worldview and two more have already shipped.
Memory, not FLOPs, is the bottleneck
Here’s the part that flips the hardware story. Big mixture-of-experts models are memory-bound, not compute-bound. Kimi or GLM only fires a few dozen billion parameters per token, so the math is almost pedestrian. The catch is that the whole model, all trillion-plus parameters, has to sit in fast memory at once.
NVIDIA’s answer to memory is VRAM, and that’s where the consumer story cracks. The top workstation card, the RTX Pro 6000 Blackwell, tops out at 96GB. A 4-bit version of GLM-5.2 alone is around 467GB. One card physically cannot hold one modern open-weight frontier model.
Apple’s answer is unified memory, one shared pool at 819GB/s on the M3 Ultra. You can configure a Mac Studio with 512GB for about $9,500, drawing a couple hundred watts. The receipts, per Hacker News: DeepSeek V3 running 20+ tokens per second on a single box, a trillion-parameter Kimi model at about 15 tokens per second on two, GLM-5.2 fitting on one machine that’s quieter than a fridge and needs no internet connection.
What stands out, and what to watch
This is one person’s opinion, dated and self-aware about it, not a market report. The piece cuts off mid-argument on pricing, and “Apple is the undisputed king” is a stretch when Apple’s own frontier models and Siri story remain weak. But the core mechanic is real: open weights are commoditizing the frontier, and local inference is memory-bound in a way that favors Apple’s architecture over gaming-derived GPUs.
Practical takeaways if you build or buy AI:
- Stop assuming inference stays in someone else’s datacenter. Model a local or hybrid path now.
- Watch memory bandwidth and capacity, not just teraflops, when you spec hardware.
- Track the open-weight release cadence as a competitive signal. It’s compressing margins on closed APIs faster than most roadmaps assume.
- Treat “intelligence as a commodity” as a planning scenario, not a fringe bet.
The prediction may be early or overstated. The direction is not. More detail sits in the original Hacker News piece.