Confirmed in a primary source 2 Oct 2026 Hardware

Nvidia's desktop AI computer now comes with half the memory for a thousand dollars more than it first cost.

On October 2 Nvidia announced a 64 gigabyte version of DGX Spark, the small desktop machine it sells for running and tuning models locally, at a starting price of 4,999 dollars, on sale from October 23 through Acer, ASUS, Dell, Gigabyte, HP and MSI rather than from Nvidia directly. That is 1,000 dollars more than the 128 gigabyte original cost when it launched, for half the memory, and Nvidia's pitch is that two of the new machines cabled together pool to 128 gigabytes and run up to 1.7 times faster than one of the old ones.

The caveat that matters is the reason, which Nvidia does not give: memory prices have risen sharply this year, and the 128 gigabyte model is reported to now sell at 6,950 dollars, a figure that appears in trade reporting rather than in the Nvidia post we read.

Until now

Compute per dollar fell every year, and a machine for running models on a desk was expected to get cheaper and larger as it aged.

Now

The same machine costs 73.8 percent more than at launch on reported prices, the cheaper new model is the one with less memory, and the pitch is to buy two, because memory is now the expensive part of an artificial intelligence computer.

This is the smallest signal on the deck and the one a reader can check with a price tag. A year ago Nvidia sold a desktop machine with 128 gigabytes of unified memory for 3,999 dollars, after showing it at around 3,000. On October 2 it announced the same machine with 64 gigabytes for 4,999 dollars, available October 23 and only through six partner manufacturers, and the full version is reported to now cost 6,950. The chip is identical, the memory bandwidth is identical at 273 gigabytes per second, and the difference between the two products is how much memory is attached and how much it costs. Nvidia says nothing about why, and the reason is not in dispute anywhere else: memory contract prices have climbed all year and Micron's chief executive has told investors supply stays tight through 2028.

The engineering answer in the announcement is the part worth keeping. Rather than sell one machine with more memory, Nvidia tells developers to buy two of the smaller ones and cable them together over the networking already built into the box, where they pool to 128 gigabytes, hold models up to 200 billion parameters and run up to 1.7 times faster than a single large unit in Nvidia's own test. That is a sensible response to a shortage and it is also a quiet concession: when the scarce input is memory, the way to sell more computers is to stop putting memory in them. For anyone who has assumed that running models locally gets cheaper every year, the price list says otherwise. Compute per dollar fell every year, and a machine for running models on a desk was expected to get cheaper and larger as it aged; now the same machine costs 73.8 percent more than at launch on reported prices, the cheaper new model is the one with less memory, and the pitch is to buy two, because memory is now the expensive part of an artificial intelligence computer.

How it got here

  1. Jan 2025

    Nvidia shows the machine as Project Digits and points at a price around 3,000 dollars.

  2. 2025

    DGX Spark goes on sale with 128 gigabytes of unified memory at 3,999 dollars.

  3. Early 2026

    The 128 gigabyte price stands at 4,699 dollars, as reported in the trade press.

  4. Oct 2

    Nvidia announces a 64 gigabyte version at 4,999 dollars for October 23, sold only through six partner manufacturers, with two units clustering to 128 gigabytes; the 128 gigabyte model is reported at 6,950 dollars.

  5. Who has to react

    Every developer and small laboratory that planned to run models locally rather than rent them, whose budget just moved; the six partner manufacturers now carrying the retail relationship Nvidia used to hold directly; Micron, Samsung Electronics and SK hynix, whose memory pricing is the whole story and who are the beneficiaries; cloud providers, who get cheaper by comparison every time local hardware gets dearer; universities and hospitals that buy local machines for data they cannot send off site and have no cloud alternative; competitors selling desktop inference boxes with the same memory supply problem; and anyone who has been quoting falling cost per token as a law of nature.

The evidence

ObservedFriday October 2, 2026, established at Nvidia's own blog in a post titled NVIDIA DGX Spark 64GB Gives Developers More Ways to Build and Scale Local AI, which we read. What Nvidia publishes there: a 64 gigabyte configuration of DGX Spark starting at 4,999 dollars, available Friday October 23, 2026, exclusively through Acer, ASUS, Dell, Gigabyte, HP and MSI; the same GB10 Grace Blackwell superchip, the same DGX operating system and the same software stack as the 128 gigabyte model; support for models up to 100 billion parameters on a single unit; and two units connected over the built in ConnectX-7 networking pooling memory to 128 gigabytes, supporting models up to 200 billion parameters and delivering up to 1.7 times the performance of a single 128 gigabyte system in Nvidia's own testing on Qwen3.8 27B.

The whole observation, as recorded

Memory bandwidth is 273 gigabytes per second on both configurations and the 64 gigabyte unit also halves onboard storage, so it is the same machine with less and cheaper memory. Nvidia's product page confirms the two memory configurations and that the 64 gigabyte unit is sold only through partners. What Nvidia's post does not discuss at all is price history, and the following three figures come from trade reporting rather than from Nvidia: the 128 gigabyte DGX Spark was shown as Project Digits at around 3,000 dollars, launched at 3,999 dollars, stood at 4,699 dollars earlier in 2026, and is reported by The Register and by Hardware Busters on October 2 to now sell at 6,950 dollars. On those reported numbers the full machine is 73.8 percent above its launch price and 47.9 percent above its price earlier this year, and the new cheap model with half the memory is 25 percent above what the full machine cost at launch. The context the reporting supplies and Nvidia does not: memory contract prices have risen through 2026, and Micron's chief executive has said supply will stay tight through 2028. Nvidia gives no reason for any price, and does not mention memory costs.

Established

  • Nvidia announced on October 2, 2026 a 64 gigabyte DGX Spark starting at 4,999 dollars, available October 23, 2026 exclusively through Acer, ASUS, Dell, Gigabyte, HP and MSI.
  • The 64 gigabyte and 128 gigabyte models share the GB10 Grace Blackwell superchip, the DGX operating system, the software stack and 273 gigabytes per second of memory bandwidth.
  • Nvidia states that one 64 gigabyte unit supports models up to 100 billion parameters and that two clustered units pool to 128 gigabytes, support up to 200 billion parameters and reach up to 1.7 times the performance of a single 128 gigabyte system in its own Qwen3.8 27B testing.
  • Trade reporting on October 2, 2026 puts the 128 gigabyte model at 6,950 dollars against a 3,999 dollar launch price and 4,699 dollars earlier in 2026.

Not established

  • That Nvidia raised prices because of memory costs. Nvidia gives no reason for any price and does not mention memory supply.
  • The 6,950 dollar figure at a primary source. It is in trade reporting, not in the Nvidia post or product page we read.
  • That cloud inference prices will rise. Two frontier laboratories cut model prices on September 22 while this hardware got dearer.
  • That the clustering claim holds generally. The 1.7 times figure is Nvidia's own measurement on one model.
  • That local model running is in decline. Higher prices on one product line are not demand data, and Nvidia is adding a configuration rather than withdrawing one.

Why it weighs this much

Signal strength
CostLow
CommitmentModerate
ReversibilityEasy
PersistenceYears

Cost is low. Announcing a product configuration and a price costs Nvidia very little and this is a small line in a company whose data centre business is measured in tens of billions. Commitment is moderate: a published starting price, a dated availability of October 23 and six named partner manufacturers are commitments Nvidia would have to explain if they moved. Reversibility is easy, because prices are the most reversible thing a company publishes and a shortage that eases can unwind all of this in a quarter. Persistence is years rather than months, on one piece of outside evidence rather than on our own view: the memory supply constraint is described by the largest supplier's chief executive as lasting into 2028. We are scoring this deliberately modestly. It is a price list, not a policy, and it is the smallest action on the deck.

What to watch

  • strengthens A second vendor raises the published price of a shipping artificial intelligence workstation or accelerator by more than 20 percent, citing memory, or Nvidia raises DGX Spark prices again, by June 30, 2027.
  • weakens The 128 gigabyte DGX Spark returns to 4,699 dollars or below, or memory contract prices fall for two consecutive quarters, by June 30, 2027.

The words

DGX Spark
A small desktop computer Nvidia sells for running and tuning models on a desk instead of in a data centre.
Unified memory
Memory shared by the processor and the graphics parts of one chip, which sets the largest model the machine can hold.
LPDDR5X
A low power memory type used in laptops and phones, now in short supply because artificial intelligence hardware consumes so much of it.
Clustering
Cabling two machines together so that their memory and processors behave as one larger machine.

Your call

Which explanation do you believe?

A call is final. Other people's calls show after yours.

8 contributions. Thank you.

You have built
a record here.

Keep it: create an account to verify your contributions and put your name to the answers you choose. Your record stays yours.

Keep my record ↗
Optional: help show whose perspectives are missing
Kept with your answers under a random id, never with your name. Used for one thing: showing you, on your record, where you stand next to people like you, once at least 20 of them have answered.

Add your signature

Stand behind your judgment.

Create an account to attach your name to this answer. Your other answers show no name unless you sign them too.

Sign up to add your name ↗

Already have an account?