Every figure the video shows or says, with where it came from. Prices were checked on 13 October 2026 and move often; check the live listing before buying.
| Figure | What it is | Source |
|---|---|---|
| $5,499 | Cheapest M5 Ultra Mac Studio: 30-core CPU, 64-core GPU, 96GB memory, 1TB SSD | Apple store (“M5 Ultra … Buy from $5499”); configuration from Tom’s Hardware’s review |
| $12,299 | Full M5 Ultra (36-core CPU, 80-core GPU), 256GB, 4TB, as tested | Tom’s Hardware review (“Price as Configured $12,299.00”) |
| +$4,000 | Moving from 96GB to 256GB | Tom’s Hardware review (“Moving from 96GB of RAM to 256GB adds $4,000”) |
| $10,799 | The tested machine with a 1TB SSD instead of 4TB | Our arithmetic: $12,299 minus the review’s “$1,500 increase” for the 4TB SSD |
| 512GB, late October | Apple’s 512GB option for M5 Ultra | Apple store (“512GB memory option for M5 Ultra coming late October”) |
| $6,950 | NVIDIA DGX Spark, 128GB, 4TB, Founders Edition; listed Out of Stock | NVIDIA marketplace; the increase reported by ServeTheHome (3 Oct 2026, from $3,999 at launch) and The Register |
| $99 | NVIDIA “QSFP Cable 0.4m for DGX Spark”; listed Out of Stock | NVIDIA marketplace |
| $13,999 | Two Sparks and the cable | Our arithmetic: 2 x $6,950 + $99 |
| $3,200 | How much cheaper the 256GB, 1TB Mac is than the pair | Our arithmetic: $13,999 - $10,799 |
| two-unit bundle | “NVIDIA DGX Spark Bundle - 2 Units … 1 connecting cable … Contact Us” | NVIDIA marketplace |
| from $5,999 | Partner GB10 boxes with 128GB (ASUS Ascent GX10 1TB $5,999; MSI EdgeXpert $6,499.99; ASUS GX10 4TB $7,999), all Out of Stock | NVIDIA marketplace |
| $3,499.99 | GMKtec EVO-X2 (Ryzen AI Max+ 395), 128GB RAM, 1TB SSD; in stock (the 64GB version is $2,199.99) | GMKtec store |
| Figure | Source |
|---|---|
| DGX Spark: 128GB LPDDR5x, 273 GB/s | NVIDIA DGX Spark hardware guide (“273 GB/s bandwidth”); product page |
| 240 W external power supply; GB10 TDP 140 W | NVIDIA hardware guide (“240W external power supply (included)”, “GB10 SOC Thermal Design Power (TDP) is 140W”) |
| ConnectX-7, up to 200 Gb/s, two QSFP ports, pairs cluster directly | Tom’s Hardware DGX Spark review (“two QSFP ports for the onboard ConnectX 7 NIC running at up to 200 Gbps”) |
| Up to four Sparks clustered | NVIDIA product page (“connection of up to four NVIDIA DGX Spark systems”) |
| 29 dB operating, 13 dB idle (declared A-weighted sound pressure, operator position) | NVIDIA DGX Spark support |
| M5 Ultra: 1.2 TB/s memory bandwidth | Tom’s Hardware M5 Ultra review (“boosted the chip’s memory bandwidth to 1.2 TB/s”) |
| Mac Studio M5 Ultra: 480 W maximum continuous power | Tom’s Hardware M5 Ultra review, specifications table |
| Strix Halo (Ryzen AI Max): 256 GB/s | Framework (“the massive 256GB/s memory bandwidth that Ryzen AI Max delivers”) |
| Figure | Source |
|---|---|
| Two Sparks (vLLM, tensor parallel) read a 32K-token codebase in ~17 s; the M5 Ultra took ~50 s (DeepSeek V4 Flash) | Matched test of an M5 Ultra 256GB against two DGX Sparks |
| Writing: DeepSeek V4 Flash ~38 tok/s on both; Qwen 3.8 Flash 45 tok/s on the Mac, 38 on the Sparks | Same test |
| Cached follow-up at 16K context: ~3 s on the Mac, ~1.4 s on the Sparks | Same test |
| One Spark against the M5 Ultra on Qwen 3.8 27B (Q4_K_M): the Mac’s prompt processing is ahead, generation almost four times higher | Tom’s Hardware M5 Ultra review (“prompt processing speeds are even faster than Nvidia’s DGX Spark’s … almost four times higher than the DGX Spark across the board”) |
| The Spark’s higher raw compute makes it much faster than Strix Halo at prompt processing | Tom’s Hardware DGX Spark review, performance page |
| Reading a prompt (prefill) is compute-bound; writing tokens (decode) is memory-bandwidth-bound | Tom’s Hardware DGX Spark review; the matched test above |