Hardware Just Entered Its Inflation Era, and the Real Power Play Is Liquid Labor

AI outbid the consumer for memory. Next it outbids everyone for power. Here's the four-year arc, why centralized compute stays cheaper than the edge for now, and the layer where the value actually pools.

Apple raised prices across its Mac and iPad lineup this week and blamed an unusual culprit: AI data centers. The MacBook Air is up $200, the base MacBook Pro up $300, iPads up $150–$200. The official line was “unprecedented demand for memory and storage.”

It would be easy to file that under corporate excuse-making. It isn't. For the first time in roughly four decades, consumer computing is getting more expensive and less capable year over year, and the cause is structural rather than seasonal. What follows is the mechanism, the four-year trajectory, and three second-order consequences most people aren't pricing in, ending with the one I think actually matters: liquid labor.

01Memflation

The world's memory is made by three companies, Samsung, SK Hynix, and Micron, and they can only point their fabs at one thing at a time. Right now hyperscalers are buying high-bandwidth memory (HBM) and server DRAM at essentially any price to feed AI training and inference. Those parts carry 60–70% gross margins; consumer DRAM and NAND do not. So the manufacturers are doing the rational thing and converting capacity toward AI.

The numbers, all fact-checked against Gartner, TrendForce, IDC, and vendor earnings calls: DRAM contract prices rose ~95% in Q1 2026 with another ~60% in Q2; NAND is climbing 70–75% quarter-over-quarter; Gartner's full-year estimate is DRAM +125% and NAND +234%. Memory has gone from ~16% of a laptop's bill of materials to ~35%. At HP, that shift happened in a single quarter. Data centers may consume close to 70% of high-end memory produced this year.

Apple didn't raise prices because it wanted to. It raised them because memory is the one component it can't engineer around fast enough.

02The four-year arc

2026, the shock. Pure pass-through. Dell, HP, Lenovo, Acer, and ASUS have already warned of 15–20% increases; IDC sees average PC prices up as much as 8% (Gartner models a steeper ~17%). Watch for quiet de-speccing: base configs capped at 8GB, some prebuilts shipping without RAM, phones holding storage tiers flat instead of doubling them.

2027, the normalization. Prices don't snap back, because supply physically can't respond. New fabs take ~3 years and HBM is contracted years out. Late 2027 is now the optimistic end of the relief window; several analysts have pushed consumer relief into 2028–2029. This is the year today's shocking numbers quietly become the baseline.

2028, the bifurcation. New capacity arrives but it's earmarked for AI. The market splits: premium tiers absorb the cost because their buyers don't blink; mid- and low-end tiers de-spec and push you toward cloud and subscriptions. “Buy the device cheap, rent the capability” becomes a dominant model.

2029, the new equilibrium. Memory is brutally cyclical, so a glut probably does eventually arrive and prices dip. But the floor resets permanently higher. Compute-per-dollar doesn't reverse, it pauses, for years. That pause is the inflation era.

03Cheaper how? Centralized still wins the bill

Here's the question a couple of AI-crypto VCs put to me this week, betting on decentralized compute: what happens when the giants outbid retail on memory and components? Doesn't that hand the edge its opening?

Short run, no, but you have to be precise about what “cheaper” means, because the obvious read is backwards.

On raw $/GPU-hour, the edge already wins, Akash, io.net, and Render undercut AWS on-demand by 50–86%. So “centralized is cheaper” is false if you're reading a price list. It's true on the axis that decides production workloads: fully-loaded cost of trust plus guaranteed access to scarce hardware. Decentralized nodes run at 40–70% utilization, carry no enterprise SLAs, and, critically, can't get HBM-class silicon during a sold-out cycle. The bottleneck doesn't make centralized cheap until it ends; centralized is cheap-to-trust because of its privileged allocation while the bottleneck lasts. Access, not price, is the binding constraint.

Which is also why I'd temper the decentralized-compute thesis without dismissing it. Inference is now ~70% of GPU demand, and inference, bursty, geographically distributed, fault-tolerant, is the edge's natural home. Decentralized networks currently serve the long tail and price-sensitive batch work. That's a real, growing business. It is not, yet, where frontier-grade reliable capacity lives.

04The surprise winners: miners become AI landlords

The same boom draining your RAM is the best thing that ever happened to bitcoin miners, because the real ceiling on AI isn't chips, it's power. Miners spent 2017–2024 quietly locking up the cheapest electricity in the country: long-term PPAs, substations, grid interconnects, land. In a buildout where “megawatts available now” is the number-one constraint, that portfolio became one of the most valuable asset classes around.

The pivot is already ~$70B in announced deals: IREN's $9.7B Microsoft arrangement, TeraWulf's ~$12.8B, Hut 8's multi-billion Google-backed lease, Applied Digital's CoreWeave commitments, Riot's AMD deal (25MW initially, expanding toward 150–200MW). CoinShares projects ~70% of listed-miner revenue will come from AI/HPC by year-end, up from ~30%.

One caveat worth stating plainly: the “80–90% margin landlord” story is a bull-case snapshot. These leases lean on a handful of counterparties, Microsoft, CoreWeave, Nvidia, and the neoclouds are themselves leveraged on the same GPU thesis. A capex wobble would hit landlord and tenant at once. Durable, but not bulletproof.

05The casualty nobody mentions: local AI

Here's where I'd correct a lazy take, including a version of my own. People say “open-source AI stalls.” That conflates two different layers. Open weights are thriving: Qwen, DeepSeek, and friends trail the frontier by single-digit points at a tenth to a thirtieth of the cost. What stalls is open weights running locally, on hardware you own.

Local inference is gated almost entirely by RAM and VRAM, the exact components being rationed. The enthusiast who'd have bought 64GB to run models at home in 2025 faces a much steeper bill in 2027; the entry laptop that might have democratized local inference ships with 8GB instead. So the open-weights movement keeps sprinting, but it runs on rented cloud, not on your desk.

That's the reflexive part. The hyperscalers consuming the memory also operate the clouds. De-spec the edge and demand routes back to them. Cheap, abundant consumer memory was the open movement's structural tailwind. For the next few years it becomes a headwind, and local AI decouples from the “just add RAM” curve, advancing instead on quantization and smaller architectures.

06The through-line

AI outbid the consumer for memory, and it's outbidding everyone for power. Your laptop gets pricier; your favorite miner becomes a data-center landlord.

It all ladders up to one prize, liquid labor, work itself backed by compute rather than headcount and rented by the minute, and whoever controls power, then compute, then the scarce memory that gates both, gets to price it.

The skeptic's fair rebuttal: this is a cyclical spike dressed as an epoch, and memory relief in 2027–2028 will date the whole frame. Maybe. But the prices are the symptom; the durable disease is allocation power, who gets the scarce inputs first. That outlives any single cycle. The era isn't really about the price of RAM. It's about who controls the inputs to intelligence, and therefore the price of work.

A note on the post that kicked this off: there's no “MacBook Neo,” so treat that specific image as embellished, but every underlying trend here was fact-checked against primary sources, and two adversarial reviews were run for recency and accuracy. “Liquid labor” is a deliberate framing, not an established term of art; I've defined it rather than assumed it.