You Are Bidding Against a Datacenter

In the first piece I wrote about access being taken away from above. The government pulled two frontier models offline in a single month, and a handful of labs sit as the meter of intelligence with their hand on the tap. That is one way to lose access to the thing you depend on. Someone decides you no longer get it. But there is a second way, quieter and arguably harder to fight, and it has been happening in plain sight on every hardware retailer's product page for the last year. You do not get cut off. You get priced out. And once you start looking at what AI compute actually costs in the middle of 2026, you realize the on-ramp to owning your own intelligence is being pulled up behind the people who already made it on.

The Receipts

Let me just lay the numbers down, because they are genuinely hard to believe in aggregate. The RTX 5090 launched at a $1,999 MSRP and now sells north of $4,000, with credible leaks pointing at $5,000 before the year is out. The RTX Pro 6000 Blackwell, the workstation card I run, went from around $8,565 at launch to $13,250 in roughly a year, a 55% jump. Some vendors are even listing it for $16,000+ as I write this, a near doubling from launch pricing. NVIDIA's own DGX Spark, the unit my cluster is built on, climbed from $3,999 to $4,699 with not a single component changed, the company citing "memory supply constraints" and nothing else. Some third-party partners such as ASUS have raised the pricing of their equivalent 4TB Spark units to $5,999. Even Apple, which has spent decades shielding customers from exactly this kind of thing, just announced its first advance-warned price hike in modern company history. The Mac Studio with M3 Ultra went from $3,999 to $5,299 overnight, a $1,300 increase, and Tim Cook called the memory situation a "hundred-year flood" he had not seen the likes of in 40 years. None of these products got better. They just got more expensive to the people who use them.

The Part That Should Make You Angry

Here is the stat that reframes the whole thing. When Tom's Hardware broke down the RTX Pro 6000's 55% price increase, the analysis found that only about 8% of it came from higher manufacturing costs. The other 92% was pure supply and demand. That is the number to sit with. You are not paying more because the card is harder to make. You are paying more because someone else with infinitely deeper pockets wants the memory inside it, and the manufacturer has discovered exactly how much that lets them charge. The price of the hardware in front of you has quietly stopped being tied to its value to you and started being tied to its value to a datacenter. Those are very different numbers, and the gap between them is the gap you are now expected to cover.

Who You Are Actually Bidding Against

The mechanism underneath all of this is memory, and it is worth understanding because it explains why none of this is a temporary spike. The same handful of companies that make the memory in your GPU and your laptop, Samsung, SK Hynix, and Micron, also make the high-bandwidth memory that goes into AI accelerators. HBM carries margins of around 60% against roughly 40% for the commodity memory in consumer gear, so every wafer that can be turned into AI memory is being turned into AI memory. Contract DRAM prices rose 80 to 90% in a single quarter. Gartner is projecting a 130% rise across 2026. And the demand pulling those wafers away is not a fad you can wait out. The five largest cloud companies are spending roughly $725 billion on infrastructure this year alone, up nearly two thirds over last year, and McKinsey pencils the full build-out at $6.7 trillion by the end of the decade. You are not competing with other consumers for this supply. You are competing with trillion-dollar capital expenditure, and there is no version of that auction where you win on price.

And the Robots Haven't Even Shown Up Yet

If you are tempted to think this is a bubble that pops and frees up the supply, consider that the second wave of demand has barely started. Physical AI, the robots, is a brand new and hungrier buyer walking into the same constrained market. Tesla is halting production of the Model S and Model X to make room to build its Optimus humanoid at scale, and every one of those machines is an inference computer that needs memory and compute onboard. Every humanoid, every autonomous system, every edge device running a local model is another claim on the same wafers your gear depends on. The datacenter demand alone already broke the consumer market. The robotics demand is additive, and it is just getting started. Whatever you are imagining as the ceiling, the actual buyers in this market have more money and more reasons to spend it than you do.

The Two Squeezes

So now you can see the shape of it clearly. The first article was about the squeeze from above. They can revoke your access by directive, by policy, by a pricing change on a Friday afternoon, by simply deciding your use case is no longer welcome. This article is about the squeeze from below. They can price you out of ever building the alternative, by bidding away the raw materials of independent compute until owning your own intelligence becomes a luxury rather than a choice. Both pincers push in the same direction, and they push you toward the same place: total dependence on infrastructure you do not own and cannot influence. Notice that owning your own hardware is the single answer that defeats both. The model on my cluster cannot be revoked from above, and the cluster itself was bought before the worst of the squeeze from below. Sovereignty is the thing that sits outside both jaws.

What This Actually Means

Strip away the part numbers and here is the thing underneath. A line is being drawn right now between people who will own intelligence and people who will rent it, and almost nobody is objecting because the line is being drawn with price tags instead of policy. When a government revokes access, it makes the news, people get angry, letters get written. When a memory chip quietly triples in cost, there is no announcement and no villain to point at, just a slightly worse product page than last month, and you shrug and move on. But the outcome is the same and arguably more permanent. One group locks in compute now, at scale, and owns the ability to run frontier intelligence under its own roof on its own terms. Everyone else gets sorted into renting that intelligence by the token, from the same few providers, indefinitely, because the hardware that would have made them independent slipped quietly out of reach while they were waiting for a better price.

And understand what is actually being taken when you get priced out, because it is not just money. Intelligence-grade compute does not have a budget version that does the same job. Memory capacity is the hard wall between the models you can run yourself and the models you cannot, and when the high-memory hardware gets bid beyond what a person can pay, what leaves your reach is not a nicer card. It is the entire capability of running a serious model privately, at home, answering to no one. The squeeze does not make sovereignty more expensive. Past a certain point it makes sovereignty impossible for the ordinary person, and that is a different kind of loss than a high price. It is a door closing. The renting class can already have its access revoked from above, as the last piece showed. What this piece is about is that the door into the owning class is closing at the same time, in slow motion, while most people are not looking at it.

Why This Is the Window

I am not going to pretend my timing was perfect, but buying compute six months ago looks a lot smarter today than it felt at the time, and that tells you everything about the next year. With no new memory fabrication capacity coming online until 2028 at the earliest, and a second wave of robotics demand walking into the same line, the conventional instinct to wait for prices to fall is, this one time, almost certainly wrong. This is not a dip to buy. It is a floor that keeps rising, and every month you wait is a month the door drifts further shut. Which is why I keep landing on the same encouragement from the last piece: you do not need a cluster, you need to start. The gaming card already in your machine is a sovereign stack in miniature, and running a single model on it puts you on the right side of the line being drawn. The people building the future are spending trillions to own their intelligence outright, because they understand exactly what it is worth to not be dependent. The whole argument of this site is that you should own a piece of yours too, on whatever scale you can manage, while owning it is still a choice you get to make.