Local AI Machine Prices: Strix Halo About $27 per GB, GB10 $46.87-$65.98

AMD Strix Halo local-AI boxes, built on the Ryzen AI Max+ 395, sell memory for about $27 per GB after price rises of 72% to 75%, against $46.87 to $65.98 for NVIDIA GB10 boxes listed on Nov. 11.

The Framework Desktop ($26.95 per GB, out of stock) and the GMKtec EVO-X2 ($27.34, in stock) lead on memory per dollar. Tokens per dollar tell another story. In StorageReview's side-by-side test of the 120B-class gpt-oss-120b model on the vLLM engine, the DGX Spark, NVIDIA's own GB10 box, served 3.2 times the Ryzen AI Halo's throughput across 64 requests: 100.86 tokens per second per $1,000 against 55.51, on the reported $6,950. Cheap memory is not cheap tokens.

Local AI Price Rises Since 2025: GMKtec EVO-X2 and Other Moves by Date

Notebookcheck reported the 128 GB GMKtec EVO-X2 at $1,999 in 2025, storage unstated. GMKtec's US store listed the 128 GB, 1 TB model at $3,499.99 on Nov. 11, 2026 (shown against a $3,999.99 reference price): 75.1% more ($3,499.99 ÷ $1,999). No source says when between those dates the price moved, or why, as our EVO-X2 review noted.

On Jan. 13, 2026, Framework added $460 to the 128 GB Desktop, taking it to $2,459, Notebookcheck reported; Tom's Hardware quoted CEO Nirav Patel blaming rising LPDDR5X memory prices from its suppliers. Its store showed $3,449 on Nov. 3 and Nov. 11, out of stock both times: 72.5% above the $1,999 that the $460 rise implies.

NVIDIA's own record covers one step. A staff post on its developer forums, dated Feb. 25, 2026, raised the DGX Spark Founders Edition from $3,999 to $4,699 worldwide, citing memory supply constraints. Implicator reported on Oct. 4 that the 128 GB, 4 TB model had moved to $6,950 as NVIDIA launched a 64 GB version from $4,999; NVIDIA's Oct. 2 blog post states only the 64 GB price. NVIDIA's marketplace showed $6,950 as of Oct. 31. If that price holds, it is 73.8% above $3,999, as our DGX Spark review tracks.

Dell.com listed the 128 GB Dell Pro Precision with GB10 at $6,593 on Oct. 30, $8,103 on Nov. 2, as our Dell review recorded, and $7,901 on Nov. 11. CDW's public list price for Gigabyte's AI TOP ATOM with 4 TB went from $7,312 on Nov. 2 to $8,445.99 on Nov. 11, up 15.5% in nine days. Exxact's DGX Station build rose from $98,450 to $104,830 between Nov. 4 and 11. None gave a reason.

Dollars per GB of Model Memory Across the Catalog

The figure is listed US price divided by memory, and what counts as memory differs by tier. Prices are from maker or named retailer pages on Nov. 11 unless dated, and include storage from none (Framework's DIY kit) to 4 TB, which the figure does not adjust for.

Strix Halo, 128 GB of unified LPDDR5X shared by CPU and GPU: Framework Desktop $3,449 ÷ 128 = $26.95 (out of stock); GMKtec EVO-X2 $3,499.99 ÷ 128 = $27.34 (in stock); AMD Ryzen AI Halo $3,999 ÷ 128 = $31.24, a price StorageReview reported on July 6. Not all of it reaches the GPU: HP caps graphics at 96 GB on its Z2 Mini G1a.

Ryzen AI Max+ PRO 495 boxes, 192 GB unified: Bosgame M5 MAX $6,199 ÷ 192 = $32.29, a pre-order price Notebookcheck reported on Nov. 3, no ship date; GMKtec EVO-X5 Pro $6,799 ÷ 192 = $35.41 (in stock); Framework Desktop 495 $6,799 ÷ 192 = $35.41, as of Nov. 3, Batch 3 due in December; AMD Ryzen AI Halo Linux edition $7,299.99 ÷ 192 = $38.02, per Tom's Hardware on Oct. 15. GMKtec says up to 160 GB can go to the GPU, and Notebookcheck reported the same for Bosgame; counted on 160 GB, GMKtec's $6,799 ÷ 160 = $42.49 and Bosgame's $6,199 ÷ 160 = $38.74.

Apple Mac Studio M5 Ultra, 512 GB unified: $22,299 ÷ 512 = $43.55, orderable on Apple's store since Oct. 28, MacRumors reported.

GB10 boxes, 128 GB of coherent unified LPDDR5X, each ÷ 128: ASUS Ascent GX10 $5,999 = $46.87 (out of stock); DGX Spark $6,950 = $54.30 (reported, Oct. 31); Dell Pro Precision with GB10 $7,901 = $61.73; Gigabyte AI TOP ATOM 4 TB at CDW $8,445.99 = $65.98. The HP ZGX Nano 4 TB is not counted in the $46.87 to $65.98 range, which covers Nov. 11 listings: Notebookcheck reported it on Oct. 5 at $4,759 on Amazon ($37.18) or $7,399 at HP ($57.80). Two 128 GB RTX Spark boxes list in the same range: Microsoft's Surface RTX Spark Dev Box $5,999.99 = $46.87 (pre-order) and MSI's EdgeMesa N AI+ $7,999 = $62.49 (out of stock).

Workstation card and GB300 towers such as the DGX Station: the RTX PRO 6000 Blackwell's 96 GB of GDDR7 is card memory only, with no host PC in its $16,000 price: $166.67 per GB, out of stock on NVIDIA's store as of Aug. 10. A GB300 tower's 748 GB is 252 GB of HBM3e plus 496 GB of LPDDR5X. Towers ÷ 748: Supermicro Super AI Station $92,887.47 = $124.18 (TechRadar, Sept. 1); MSI XpertStation WS300 $99,999 = $133.69 (Newegg, Nov. 4); Exxact $104,830 = $140.15; ASUS ExpertCenter Pro ET900N G3 $111,999.99 = $149.73 (Central Computers); HP ZGX Fury $132,000 = $176.47; Dell Pro Precision with GB300 $160,195 = $214.16. On the HBM3e alone, the towers priced on Nov. 11 cost $415.99 (Exxact) to $635.69 (Dell) per GB.

The Acer Veriton GN100, HP Z2 Mini G1a and Lenovo ThinkStation PGX (coming soon, Tom's Hardware reported on Nov. 10) stay out for lack of a current US price.

Tokens per Second per $1,000: Only Matched Runs Count

Tokens per second per $1,000 is computed only where model, engine, workload and batch match. The GB10 set is StorageReview's separate reviews of gpt-oss-120b in vLLM online serving at batch 1; the write-ups give equal input and output lengths but not the length or the weight precision, so treat the order as indicative. The HP ZGX Nano's 69.6 tokens per second at $4,759 gives 14.62, or 9.41 at $7,399 (both reported Oct. 5); the ASUS GX10's 70 at $5,999 gives 11.67 (out of stock); the Dell's 68.57 at $7,901 gives 8.68. Gigabyte's 60.71, measured on one ATOM, gives 10.12 at $5,999.99 (1 TB, Newegg, Nov. 9) and 7.19 at CDW's $8,445.99 (4 TB). Speeds sit within 60.71 to 70, so price sets the order, as our GB10 comparison found.

StorageReview's July 6 review ran the Ryzen AI Halo and the DGX Spark side by side: gpt-oss-120b in vLLM, 256 input and 256 output tokens, batch 64. On the reported $6,950, the Spark's 701 tokens per second gives 100.86; on AMD's July $3,999, the Halo's 222 gives 55.51. Neither is a Nov. 11 price; in July the Spark listed at $4,699. Our Ryzen AI Halo review covers the gap.

For GB300 towers, StorageReview ran gpt-oss-120b in MXFP4 on vLLM, 512/512 chat workload, 128 concurrent streams, in separate reviews on Aug. 24 and Sept. 23. The MSI WS300's 9,294 tokens per second at $99,999 (a Newegg quote-request listing) gives 92.94; the ASUS ET900N G3's 9,280 at $111,999.99 gives 82.86. The RTX PRO 6000's 3,064 on the same workload would give 191.5 per $1,000 of card price alone, but it is not ranked: StorageReview measured the 600 W Server Edition, not the priced Workstation Edition; our RTX PRO 6000 review has the detail.

No matched set exists for the Mac Studio, measured by Tom's Hardware on GLM-5.3-Flash in MLX, or for single-user Strix Halo, where ServeTheHome's roughly 45 tokens per second on the Framework Desktop and StorageReview's close to 40 (HP Z2 Mini) came from different setups. The Acer GN100 is left out too.

What Fits: Model Size Limits and Memory Bandwidth

To write each token, a model reads its active weights from memory again, so for one user the limit is how fast memory delivers data, not how fast the chip calculates. Capacity decides only which models fit. Framework's 192 GB Desktop reaches 273 GB/s against 256 GB/s for its 128 GB model, 50% more room for 6.6% more bandwidth. Serving many requests changes the result: StorageReview measured the ZGX Nano at 69.6 tokens per second at batch 1 and 722.9 in total across 64 requests, because one read of the weights serves the whole batch.

By the makers' statements, a 64 GB GB10 runs models up to 100B parameters (NVIDIA) and a 128 GB GB10 up to 200B (NVIDIA, ASUS), at 273 GB/s. A 128 GB Strix Halo runs Qwen3 235B by GMKtec's claim at 256 GB/s. A 192 GB GMKtec or Framework box on the Ryzen AI Max+ PRO 495 runs 300B-class models at 4-bit by Framework's claim, or 320B by GMKtec's, beside a 160 GB GPU cap, at 273 GB/s; Bosgame states no bandwidth. Framework aside, the makers give no quantization; by our arithmetic these ceilings fit only near 4 bits per weight, counting the weights alone, before cache and the operating system. The RTX PRO 6000 holds 96 GB at 1,792 GB/s; the Mac Studio, 512 GB at 1.2 TB/s, with no Apple parameter claim. A GB300 tower pairs its HBM3e at 7.1 TB/s with LPDDR5X at 396 GB/s, and NVIDIA says it runs models up to 1 trillion parameters.

For engineers, dollars per GB says whether a model fits, not how fast a team is served: an in-stock Strix Halo box costs about 42% less per GB than the cheapest GB10 box listed on Nov. 11 ($27.34 against $46.87), yet the DGX Spark served 3.2 times the Ryzen AI Halo's batch-64 throughput. For investors, prices on identical GB10 hardware moving by double digits within days suggest they now track memory supply more than the chip.

What to Watch

NVIDIA reports third-quarter fiscal 2027 results on Nov. 17, its first call since the reported move to $6,950, and holds GTC Washington, D.C. from Nov. 30 to Dec. 3. AMD, which told investors on Nov. 3 it is planning for PC market unit shipments to fall by a double-digit percentage in 2027, appears at UBS's technology conference on Dec. 1 and Barclays' on Dec. 9, its investor calendar shows.

Digital Trends says the first 512 GB Mac Studio orders arrive on Dec. 3, the date our Mac Studio M5 Ultra review recorded. Expected in December: Framework's Desktop 495 Batch 3 and, from partners, the DGX Station for Windows. GMKtec's EVO-X2 price move remains undated, and NVIDIA has published no statement confirming the DGX Spark's move to $6,950.