I still see people quote the GeForce RTX 5090's $1,999 MSRP like it is a price you can pay. On a good week, street prices land somewhere in the $3,500 to $4,300 range. Refurbished listings have floated around $4,360. The premium bins clear five thousand without blushing. Late 2025 had a brief mood of "normalization", supply catching desire, scalping cooling off, maybe a path back toward list. That mood did not survive contact with AI data centers vacuuming up DRAM, HBM, and GDDR7. VRAM's share of the bill of materials keeps climbing in every teardown-adjacent write-up I trust. Samsung has been hiking DRAM. Nvidia has been reported as considering further increases. Micron's CEO has talked about shortage lasting into 2027. If you were waiting for 2026 to make high-end GPUs boring again, the wait is still waiting.
The 5090 receipt does not match the slide
MSRP is a press-cycle number. The number that hits your card is whatever inventory, AI demand, and memory pricing allow that afternoon. I have watched 5090 listings behave less like a refresh cycle and more like a commodity desk, wide spreads, weird regional gaps, refurbished stock that costs more than last year's fantasy of a "reasonable" street price. When used-ish silicon asks for $4,360, the market is telling you new supply is not loose.
Gamers feel robbed. Creators feel stuck. Small labs feel priced into cloud credits they did not want to depend on. I am in the last camp more often than I admit. Renting an A-class or H-class slice by the hour can look rational next to a scalp on a single 5090, until you add up a year of inference and realize you bought a car in cloud invoices. Neither path is "normalized." Both are coping strategies.
Prebuilt PCs sometimes pencil out cheaper than a bare card plus the rest of a build, which is an ugly sentence to type in a year when we were supposed to be past shortage logic. System integrators with allocation can bundle in ways the DIY queue cannot. If your goal is a machine that exists in your apartment in August, not a purity test about picking every SKU, the prebuilt math is worth running before you refresh the Newegg tab for the twelfth time.
I am deliberately not inventing a shelf of exact board partners and clock bins beyond what the reporting supports. The story is the spread between $1,999 and the mid-four-thousands, not a catalog. Anyone promising you a secret stable of MSRP cards in bulk is selling a narrative, not inventory you can purchase-order. Screenshot culture makes every temporary dip look like a trend. A single in-stock listing at a less insane number is not a market clearing.
Power and the rest of the BOM still exist, which gets forgotten when the GPU is the villain. A tower that can feed a 5090-class card needs a PSU, cooling, and a case that is not a toaster. Those parts did not get cheaper because the GPU got mythical. If your upgrade plan is "card only" on a five-year-old chassis, price the silent extras before you celebrate finding a listing.
Memory is eating the BOM
AI data centers did not invent GPU demand, but they changed who wins allocation. HBM for training clusters, GDDR7 for the boards people actually put in towers, DRAM across the rest of the build, the same suppliers keep hearing from hyperscalers first. Coverage of VRAM as a soaring fraction of GPU cost matches what you would expect when memory is the constrained input. The silicon die is not free, but it is not the only villain on the invoice anymore. When memory is the scarce input, every consumer product that needs fast DRAM starts bidding against a training cluster's purchase order.
Samsung's DRAM hikes and chatter about Nvidia considering further price increases feed each other. If memory costs rise, board prices follow or margins compress. Nvidia has rarely volunteered to be the charity in that squeeze. Micron's CEO talking about shortage into 2027 is the cold water on any plan that assumed a soft 2026 landing. I treat executive shortage commentary as directional, not as a delivery date for your local Micro Center. Directional is already enough to stop pretending Q4 will casually fix a DIY build budget. Companies do not warn about multi-year tightness to make hobbyists feel cozy.
AMD's next-gen showing up late in some reports, late 2027 or even 2028, removes a pressure valve people were counting on. Competitive supply is how prices remember manners. If the alternative flagship calendar slips, Nvidia's street market keeps the pricing power even when MSRP slides look polite. That is not a fandom argument. It is what happens in hardware when the second source is late and the first source has AI buyers on speed dial. Waiting for a savior SKU is a strategy only if your workload can wait on the same calendar.
Late-2025 normalization hopes were not crazy at the time. Channel inventory looked healthier in places. Crypto-mining echoes had faded. Then training clusters kept eating boards and memory lines, and the consumer channel discovered it was still the residual claimant. I got burned emotionally by that whiplash once already this decade. I am not writing a hopeful price target into this piece. Hope is not a hedge against HBM allocation.
How I would buy compute in this market
Decide whether you need a 5090-class card or whether you need FLOPs. Those are different shopping trips. If you need local VRAM for models that refuse to fit elsewhere, you are in the scalp economy, set a hard ceiling, consider a prebuilt with allocation, or wait knowing waiting has a calendar cost too. If you need training or batch inference, price cloud and colo against twelve months of ownership, including power, before you congratulate yourself for "avoiding" a $4,000 GPU. Avoiding a purchase order is not the same as avoiding spend.
Used and refurbished can be rational when the premium on new is nonsense, but $4,360 refurbished is not a bargain narrative. It is evidence. Inspect return policies like you mean it. For teams, standardize on a cloud SKU for burst work and reserve local GPUs for the latency-sensitive or air-gapped jobs. Spreading emotional energy across every rumor of a restock helps Discord, not your roadmap. I would rather approve one boring cloud budget line than three heroic eBay threads that each almost worked.
Creators who edit locally and hobbyists who just want frames in a game are stuck in the same market with different patience levels. Frame-rate dreams do not get a separate DRAM fab. If a 40-series or last-gen card still clears your titles or your timeline exports, keeping it another year is not failure, it is reading the tape. Bragging rights about owning a 5090 at five thousand dollars are expensive jokes.
I still want a world where MSRP means something closer to the median transaction. We are not in that world in August 2026. We are in a world where AI capex sets the weather, memory vendors hike into a shortage they publicly expect to linger, and AMD's next answer might not arrive until the problem has had another birthday. Buy the machine you can justify on real workloads. Stop quoting $1,999 as if the universe owes you that number. The universe is busy selling HBM to someone with a bigger PO.
