Hardware · Open Source AI

Apple Isn't Gouging You. AI Is.

· 6 min read

On Thursday, Apple raised prices on 14 products. The Mac Studio M3 Ultra jumped by $1,300. MacBook Pro added $300. Even the HomePod mini got a 30% hike. Tim Cook didn't blame tariffs, inflation, or supply chains. He blamed AI. And he's not wrong — but the story he told is only half of it. While everyone is staring at Apple's price tags, the open source AI community is losing the hardware it needs to exist.

What Apple actually did

On June 25, Apple raised prices across the board. MacBooks, iPads, Mac desktops, HomePods, Apple TV, and Vision Pro all got hit. The average increase was $246.67 per product line. The only products spared were iPhones, Apple Watches, AirPods, and accessories — likely because those have higher margins and Apple could absorb the component costs there.

The full damage:

Apple's stock dropped 5.5% on the news — its worst intraday decline in 15 months — which tells you even the market wasn't fully priced for this.

Tim Cook said the quiet part out loud

In an interview last week, Cook confirmed the company was preparing for price hikes and explained why in unusually direct terms.

"We're doing our best to mitigate the huge increases that are being passed to us, and we've been trying to shield our customers from the increases, but the situation has become unsustainable," Cook said. He specifically called out High Bandwidth Memory — the specialized DRAM used in AI accelerators — as the primary culprit. "There's less supply at a time when consumers want devices and the memory guys are passing along huge price increases."

In a statement to Reuters, Apple went further: "We have never seen a component price increase this much, this quickly. We have shielded our customers from these increases so far, but we have now reached a point where we need to begin raising prices."

This is not corporate spin. Apple sat on these cost increases for quarters before passing them through. When the company with the most ruthless supply chain operation on Earth says a component situation is unsustainable, pay attention.

The $100 billion reason your laptop costs more

The memory crisis is real and measurable. On the same day Apple announced its price hikes, Micron Technology reported Q3 2026 revenue of $41.5 billion — more than quadruple the $9.3 billion from a year earlier — and disclosed 16 multi-year supply agreements worth a minimum of $100 billion in contracted revenue. These are take-or-pay deals with pricing floors that guarantee margins above anything Micron has ever achieved.

CEO Sanjay Mehrotra told analysts the company "currently does not have line of sight as to when memory supply will be able to catch up with increasing demand." That is not a forecast. That is a structural admission.

Here is what is happening at the silicon level. Three companies — Samsung, SK Hynix, and Micron — control virtually all of the world's advanced memory production. AI accelerators like NVIDIA's H200 and B200 use HBM3E, a vertically stacked DRAM technology that delivers the memory bandwidth large models require. Manufacturing HBM consumes the same fabrication lines, the same advanced packaging capacity, and the same raw materials as the DDR5 and LPDDR5 memory that goes into every MacBook and iPad.

When hyperscalers sign $100 billion in take-or-pay contracts, they are not just buying memory. They are buying priority. Every wafer allocated to HBM for a Google TPU pod or an Amazon Trainium cluster is a wafer not available for the LPDDR5 in your next laptop. The AI industry is literally outbidding consumers for the same silicon.

This isn't just Apple. On the same day, Xbox raised its console price by $100 — its second increase in under a year — citing the same "components crisis." Microsoft's Surface Laptop 13.8-inch, now priced at $1,500, went up $300 for hardware that is two years old. Asus, Samsung, and Dell have all raised laptop prices this year. The New York Times Wirecutter's headline captured the mood: "RIP, Reasonable Laptop Prices."

The GPU market is already broken

The consumer GPU market is where the AI-driven hardware squeeze is most visible — and where the implications for open source AI are most alarming.

NVIDIA's RTX 5090 launched at a $1,999 MSRP. As of mid-2026, street prices have reached $3,000 to $5,000 — roughly 2 to 2.5 times MSRP. The discontinued RTX 4090 now costs more used than it did new. A prebuilt single-GPU 5090 system runs $5,000 to $8,000 complete. Dual-GPU builds start at $7,600. These are not enterprise workstation prices. These are consumer GPU prices.

The Render Network, a decentralized GPU compute platform, recorded negative GPU supply availability in Q2 2026 — meaning demand for processing power officially outstripped every node the network could throw at it. For the first time since 2018, the Render Network could not satisfy all incoming compute requests.

The shortage is so acute that memory contract prices rose sharply through the first half of 2026, with supply allocations "increasingly directed towards AI servers and enterprise storage, creating tighter availability for consumer electronics, industrial systems and embedded applications." Even companies with budget are being told "no."

The open source AI toll

This is where the story gets personal — and where I get worried.

Open source AI lives on consumer hardware. The Llama models, the Mistral releases, the DeepSeek open weights — they run on Macs with unified memory and on NVIDIA GeForce cards that fit under a desk. The entire open model ecosystem is built on the assumption that an independent developer can spend $2,000 to $4,000 and get a machine capable of running a 70B-parameter model locally with decent throughput.

That assumption is breaking.

A Mac Studio M3 Ultra — the machine a serious local inference builder would buy for its 192GB of unified memory — just went from $3,999 to $5,299. A MacBook Pro with enough RAM to run a 32B model locally now starts at $1,999. The RTX 5090, the enthusiast card that powered the local LLM renaissance of 2024-2025, is now a $3,000-5,000 purchase. The used RTX 4090 market, which was the budget entry point for local inference, has inverted: cards are appreciating, not depreciating.

The floor for running a capable open-source model locally has roughly doubled in 18 months. The ceiling — what you can do with a dual-GPU workstation — has gone from expensive to prohibitive.

This has structural consequences. When consumer hardware becomes too expensive for local inference, developers don't stop using AI. They route their prompts to an API. And when they route to an API, they route to Claude, or GPT, or Gemini — not to the open model they could have run themselves. Every $500 added to the cost of a MacBook Pro is a small nudge toward centralization. Multiplied across the developer ecosystem, it's a gravitational force.

The GPU allocation crisis is one facet of this. But GPU governance — who gets to use the GPUs that exist — is now the defining infrastructure challenge of 2026. Hyperscalers sign $100 billion take-or-pay contracts. Cloud providers reserve entire wafer allocations years in advance. Independent researchers and open source builders compete for whatever is left on the spot market. As one infrastructure engineer put it bluntly: GPU scarcity didn't disappear — it moved up the stack. The capacity exists. It's just not allocated to you.

The one thing that might save us

There is a counterintuitive bright side, and it's worth stating plainly: the hardware squeeze is forcing open source models to get dramatically more efficient.

When GPUs were abundant and cheap, model developers optimized for capability at any compute cost. The result was parameter bloat — 405B, 671B, trillion-parameter monsters that required data center hardware to run. The hardware crisis has inverted that incentive. The models that matter in 2026 are the ones that run on hardware people can actually access.

Z.ai's GLM-5.2 runs a 744B model at 239GB quantized — fitting on a single Mac Studio. DeepSeek's Mixture-of-Experts architecture activates only a fraction of its 671B parameters per query, slashing memory requirements. Unsloth's quantization tools have gotten so good that a 70B model can now run on a single 24GB consumer GPU at reasonable speeds. The efficiency breakthroughs of the past 12 months are not accidents — they are direct responses to hardware scarcity.

The question is whether efficiency can outrun price. If model quantization gets 30% better every year but hardware costs rise 40%, the math still pushes people toward APIs. Right now, it's a race — and the hardware side is winning.

What builders should do now

If you run models locally, the window for upgrading at reasonable prices is narrowing. The hardware you can buy today — even at inflated prices — may look cheap in six months. If you were planning a local inference build, do it now, not later. The component cost curve is pointing in the wrong direction, and nobody — not Micron's CEO, not Tim Cook, not Jensen Huang — claims to know when it reverses.

If you're building on open source models, prioritize efficiency. A 32B parameter model running at Q4 quantization on a MacBook Air is better than a 70B model you can't afford the hardware to run. The best model is the one you can actually use.

If you're routing to APIs by default, reconsider. The pricing pressure that hit consumer hardware is coming for inference tokens too. When memory is the constrained resource, every token you generate has a rising marginal cost. Lock in fixed-rate contracts where you can. Test self-hosted alternatives before you need them.

The AI revolution was supposed to democratize intelligence. Instead, its first physical consequence is making the hardware required for that democratization unaffordable. The companies building the largest models are not doing anything wrong — they are responding to demand. But the invisible cost of their success is quietly pricing independent builders out of the game.

Apple's price hikes are the warning light. What they're warning about isn't a bad quarter for consumer electronics. It's a structural realignment of the semiconductor industry around AI workloads — and the people who get squeezed out won't be Tim Cook's customers. They'll be the developers who can no longer afford a machine powerful enough to run their own models.