Local AI hardware prices rose by something close to double across 2026, and the reason is the same AI buildout the hardware is sold as an escape from.
The pitch for a local AI box has always been one sentence. Buy it once, run models on it, stop paying a subscription. That pitch rests entirely on a break-even calculation, and both sides of that calculation moved this year. The capital cost went up hard. The cost of renting the alternative held flat or came down.
We have covered the principles of this decision before, in hardware for AI agents, which lays out when local inference genuinely beats an API, and in the M6 Mac mini piece, which covers why memory bandwidth rather than TOPS decides how fast a model runs, and more broadly in local AI models. We also covered the memory market itself in August, in server memory prices, which is about what the same shortage does to an ordinary refresh budget. This piece does not repeat any of them. It is the price history, the mechanism behind it, and the arithmetic redone at today’s numbers.
The short version: a 128GB desktop inference box that cost roughly $2,000 eighteen months ago costs roughly $3,500 to $4,700 today, vendors have said on the record that memory is why, the API side got no more expensive and in one documented case got cheaper, and the first sign of relief on the hardware side appeared about a week ago in one memory format on one product line. If you were about to buy, the case is weaker than it was a year ago but not dead, and the threshold where it flips has moved up.
What local AI hardware prices did in 2026
Dated, vendor by vendor.
Framework Desktop, 128GB. Announced 25 February 2025 at $1,999 for the Ryzen AI Max+ 395 with 128GB. Raised to $2,459 on 12 January 2026. Listed today at $3,449, out of stock in every configuration. That is a 73 percent increase on the same machine.
NVIDIA DGX Spark. Went on sale 15 October 2025 at $3,999, with 128GB of unified LPDDR5x and 273 GB/s of memory bandwidth on NVIDIA’s own spec table. NVIDIA raised the Founders Edition to $4,699 effective the week of 23 February 2026, an 18 percent increase with no hardware change.
Minisforum MS-S1 Max. Launched at $2,299. The 128GB with 2TB configuration is listed on Minisforum’s own store at $3,799, marked down from a $4,749 regular price. Roughly 65 percent above launch.
Beelink GTR9 Pro. 128GB with 2TB at $4,349, reduced from $4,699, listed as pre-sale with orders shipping within 35 days.
AMD Ryzen AI Halo. AMD’s own branded developer desktop, launched June 2026 at $3,999 with 128GB of LPDDR5x. AMD’s product page positions it directly against NVIDIA, and its own footnote reads: "Retail price for NVIDIA DGX Spark is $4699, Retail price for AMD Ryzen AI Halo is $3999." When a chip vendor’s headline selling point is being $700 cheaper than the competition, the category has a price problem.
Apple. Apple is the cleanest signal in the set, because of something that happened on 6 March 2026 with no announcement at all. Apple removed the 512GB memory option from the M3 Ultra Mac Studio and raised the 96GB to 256GB upgrade from $1,600 to $2,000. Same machine, same silicon, no new product. A $400 increase on a memory upgrade and the deletion of the top capacity tier is a pure memory-price event with nothing else mixed in. It was a silent configurator change spotted by the press rather than a statement, so attribute it accordingly.
The newer Apple machines announced 25 August 2026 are not clean evidence in the same way, because they are new silicon with more bandwidth. For the record: Mac Studio with M5 Max starts at $2,499 with up to 128GB and up to 614 GB/s, M5 Ultra starts at $5,499 with up to 512GB and 1.2 TB/s, and Mac mini with M6 starts at $899 at up to 170 GB/s but configures only to 32GB. General availability is 22 September, though Apple says the M5 Ultra with maximum memory arrives in late October. If you are comparing capacity per dollar against a Strix Halo box, the 32GB ceiling on the M6 mini matters more than its price.
The mechanism, in the vendors’ own words
We laid out the supply-chain mechanism in August and will not re-derive it. What is new since then is that two vendors have put the cause on the record in their own words, which is rarer than it sounds.
NVIDIA, announcing the DGX Spark increase: "The price adjustment reflects industry wide memory supply constraints."
Framework has been more detailed, because it publishes a running blog on memory pricing and has updated it roughly monthly since December 2025. CEO Nirav Patel, on 12 January: "The memory outlook as we enter 2026 continues to get worse. From what we learned in meetings throughout the week at CES with suppliers, distributors, and partners, it’s clear that this is going to be a challenging year and possibly even years for consumers."
The same post explains where the memory went, and this is the sentence worth carrying into a client conversation: "A single rack of NVIDIA’s GB300 solution uses 20TB of HBM3E and 17TB of LPDDR5X."
And where the supply is going: "Higher-margin server-focused memory like HBM and the server markets for DDR5 are prioritized over PC."
So the loop closes on itself. The memory in a desktop AI box and the memory in an AI training rack come off the same production lines, the racks pay more, and a company selling you a way to run AI without renting datacenter capacity is bidding against that datacenter for parts. A reader who takes nothing else from this piece should take that.
The August piece reported TrendForce putting the Q1 2026 rise in conventional DRAM contract prices at 93 to 98 percent quarter over quarter. Here is what the rate did after that, and it is the part nobody has picked up. TrendForce’s 31 March release projected 58 to 63 percent for Q2. Its 3 July release forecast 13 to 18 percent for Q3.
Read that series in order: roughly 95, then 60, then 15. Prices kept climbing every quarter, and the rate of climb fell by more than half each time. The cumulative level got much worse through the year. The rate has been improving since spring. Both are true, and coverage that quotes only the first number is describing March.
Two cautions on those figures. The Q2 and Q3 numbers are forecasts published before their quarters, not measured outcomes, and several write-ups have repeated the Q2 projection as a result. And TrendForce is independent of the memory makers but sells the research these releases promote.
The other side of the calculation
While this was happening, hosted inference did not get more expensive.
The documented case is Anthropic’s. Claude Sonnet 5 lists at $2 per million input tokens and $10 per million output. Sonnet 4.5 and 4.6 are $3 and $15. And Anthropic’s pricing page states: "The $2/$10 per million input/output token pricing for Claude Sonnet 5, announced at launch as introductory pricing through August 31, 2026, is now the standard price. The previously scheduled increase to $3/$15 per million input/output tokens on September 1, 2026 will not occur."
An increase was on the calendar for two weeks ago and was called off. That is the only scheduled API price rise in this window we could find, and it did not happen.
The rest of the current card, for the arithmetic below: Claude Haiku 4.5 at $1 and $5, Claude Opus 5 at $5 and $25. Cache reads across the range are a tenth of base input, so $0.10 per million on Haiku. The Batch API is half price on both input and output. OpenAI’s GPT-5.6 ships in tiers, with Luna at $0.20 and $1.20, Terra at $2 and $12, and Sol at $5 and $30, cached input a tenth of input on each.
Those two discounts matter more than the headline rates for the work most agencies and publishers actually do. Retrieval over a fixed corpus is mostly cache reads. Overnight classification, transcription backlogs and bulk tagging are batch jobs by nature. A workload that is 80 percent cached input and runs on the Batch API is not paying anything close to the list price.
The break-even, redone
Take the Framework Desktop 128GB at $3,449, which is about as cheap as a 128GB box gets right now, and compare it against Haiku 4.5 at $1 and $5, which is roughly the right class of comparison for a good open-weight model in the 20B to 30B range. Comparing a local model against Opus is the error that makes every one of these calculations come out wrong.
| Monthly volume | Annual API cost | Years to pay off $3,449 |
|---|---|---|
| 20M in / 4M out | $480 | 7.2 |
| 50M in / 10M out | $1,200 | 2.9 |
| 100M in / 20M out | $2,400 | 1.4 |
| 200M in / 40M out | $4,800 | 0.7 |
Add power. A box drawing 120 to 240 watts through a business day runs roughly $50 to $120 a year at typical US commercial rates. Then add the line everybody omits, which is staff time. If keeping the thing patched, serving and monitored costs one day of a systems person per quarter, that exceeds the entire API bill in the first row of the table.
Now run the same middle row at the old price. At $1,999, a 50M by 10M month paid the box off in about 1.7 years. At $3,449 it takes 2.9. The threshold did not vanish. It moved up by more than a year, on identical usage.
And if the workload is cacheable or batchable, the API column drops by half or more and the payback period roughly doubles again.
A note on performance so nobody sizes a box from marketing numbers. LMSYS measured a DGX Spark running gpt-oss-20B in MXFP4 at 49.7 tokens per second of decode in October 2025, then reported roughly 70 tokens per second on SGLang in a follow-up three weeks later after working with NVIDIA on quantization and Triton. Both are real; the second supersedes the first, which is why any figure in this category needs a date attached. The same tests put Llama 3.1 70B at FP8 at 2.7 tokens per second of decode, which is unusable for anything interactive, and showed Llama 3.1 8B going from 20.5 tokens per second at batch 1 to 368 at batch 32. That last pair is the most under-reported number in the category. These boxes are mediocre at serving one person and respectable at serving twenty, and almost every review tests them the first way.
For anything at frontier scale, the arithmetic is not close, and we went through it separately in what self-hosting Kimi K3 actually costs. Nothing here changes that conclusion.
What changed a week ago
On 8 September, Framework posted its first good news in nine months. It secured "a limited quantity of Micron 32GB and 64GB LPCAMM2 memory modules at lower cost," began "rolling out price reductions across both already shipped and a subset of currently pending pre-orders," and wrote: "While these new costs remain above our original pre-order launch baseline, they are lower than the last cost update."
Read that precisely before treating it as a turn. It covers LPCAMM2 modules, which are the removable memory used in laptops. The desktop AI boxes in this piece use soldered LPDDR5x, and the Framework Desktop 128GB is still listed at $3,449. So the relief so far is in one memory format on one product line, and it has not reached the machines this article is about.
Taken with the deceleration in the DRAM forecasts, the honest reading is that the worst of the rate of increase is behind us and the level is not coming back down soon. Anyone who told a client in January to wait for prices to normalize has been wrong for eight months.
What this does and does not change
It does not change the cases where local was never about money.
If the data cannot leave your environment for contractual reasons, price is not the deciding variable and never was. If the network is genuinely restricted or air-gapped, there is no hosted option at any price. If the workload is steady and heavy enough to keep the hardware busy, the table above still crosses over, just later than it used to. If you are doing high-iteration experimentation where per-call billing discourages trying things, a paid-for box changes behavior in a way the spreadsheet does not capture.
What local AI hardware prices did change is the marginal case, which is most of them. An organization spending $40 or $100 a month on inference and thinking about a box because the subscription feels like a leak is now looking at a three to seven year payback before power and labor, on hardware whose useful life for AI work is probably shorter than that. That was a closer call a year ago.
Three practical points for anyone making this decision this quarter.
Price the comparison against the cheap tier of a hosted model, not the expensive one, and apply caching and batch discounts to the hosted side before you compare. Most local-versus-cloud calculations you will see quoted skip both and are wrong by a factor of several.
Size for concurrency, not for one user. The batch-1 to batch-32 numbers say the economics change completely once more than a couple of people are using the box, and that is the configuration nobody benchmarks.
If you are buying anyway, buy the memory you need now rather than planning to upgrade. Every machine in this piece has soldered or non-upgradeable memory, and the one vendor with a configurable tier deleted its largest option and raised the price of the next one down without saying anything.
Frequently Asked Questions
Why did local AI hardware prices rise so much in 2026?
Memory. Both NVIDIA and Framework have said so on the record. The LPDDR5x and DDR5 that go into a desktop inference box come off the same lines as the memory in AI training racks, and the racks pay more. Framework’s figure is that one NVIDIA GB300 rack uses 20TB of HBM3E and 17TB of LPDDR5X.
How much did the Framework Desktop actually go up?
The 128GB configuration was announced at $1,999 in February 2025, went to $2,459 on 12 January 2026, and is listed at $3,449 today. That is 73 percent. All configurations are currently out of stock.
Did NVIDIA raise the DGX Spark price?
Yes. From $3,999 at its October 2025 launch to $4,699 effective the week of 23 February 2026, with no hardware change. NVIDIA’s stated reason was industry wide memory supply constraints.
Are API prices going up too?
Not in the same window. The one scheduled increase we could find was Anthropic’s plan to move Claude Sonnet 5 from $2/$10 to $3/$15 on 1 September 2026, and Anthropic canceled it and made $2/$10 the standard price.
What is the break-even on a $3,449 box?
Against Claude Haiku 4.5 at $1 and $5 per million tokens, roughly 2.9 years at 50 million input and 10 million output tokens a month, about 7 years at a fifth of that volume, and under a year above 200 million input tokens a month. Add $50 to $120 a year for power and whatever your staff time is worth.
Is it cheaper to just keep paying for an API?
For most small organizations, yes, on cost alone. Below roughly 50 million input tokens a month the hardware does not pay for itself within a sensible replacement cycle. Above that it starts to, and the heavier and steadier the load the better it looks.
Does prompt caching change the math?
Substantially. Cache reads are a tenth of base input price, and the Batch API is half price on both input and output. A retrieval workload over a fixed corpus is mostly cache reads, which can cut the hosted side by most of its cost and push the payback period out by years.
Are memory prices coming back down?
Partly, and not yet for these machines. Framework reported lower costs on some LPCAMM2 laptop modules on 8 September and passed reductions through, while noting costs remain above its original baseline. Desktop boxes use soldered LPDDR5x and have not seen it. TrendForce’s quarterly series runs roughly 95 percent actual for Q1, 58 to 63 percent projected for Q2 and 13 to 18 percent forecast for Q3, which is deceleration rather than reversal.
Which box is the best value right now?
On published prices, AMD’s own Ryzen AI Halo at $3,999 and the Framework Desktop at $3,449 are the two 128GB machines with the clearest pricing, and Minisforum lists the MS-S1 Max at $3,799. Availability is the real constraint: Framework is out of stock across the line and Beelink is taking pre-sale orders with a 35 day ship window.
How fast do these actually run models?
LMSYS measured a DGX Spark at 49.7 tokens per second of decode on gpt-oss-20B in October 2025 and roughly 70 tokens per second in a November follow-up after software fixes. A dense 70B model at FP8 ran at 2.7 tokens per second, which is not usable interactively. Sparse mixture-of-experts models are what make this class of hardware viable.
Should I wait for prices to fall?
Waiting has been the wrong call for eight months running. The forecasts point to slower increases rather than reductions, and supply is being allocated to server memory first. If the purchase is justified by data residency or network restrictions, those reasons have not changed. If it is justified only by cost, run the table above at your real volume first.
Does any of this apply to running frontier-scale open models?
No. The machines here top out around 128GB of unified memory, which suits models in the 20B to 120B sparse range. Anything at the scale of the largest open-weight releases needs a multi-GPU server node, and that comparison runs differently.
