~ / local / hardware

The hardware table: price versus what it renders

Three machines dominate the “run it yourself” conversation in late 2026: Moore Threads’ MTT S5000, Apple’s Mac Studio M3 Ultra with 512 GB of unified memory, and NVIDIA’s RTX Pro 6000 Blackwell with 96 GB. Below, their prices with the source each number came from, and a fit matrix that says plainly which open video model runs on which card — because a spec sheet you cannot map to a workload is decoration.

Prices, with the label on each number

Checked October 2026. A marketplace price is what one seller asked, not what the maker lists.

Mac Studio M3 Ultra 512 GB$14,439 anchor
RTX Pro 6000 Blackwell 96 GB~$8,500
MTT S5000 80 GBnot published

Bar length is proportional to the verified USD price point; the hatched bar marks the machine with no public price.

MachineMemoryPrice as verifiedSource label
MTT S5000 (Moore Threads, PH100 “Pinghu”)80 GB, FP8-FP64 full precision pathnot published; sold as server capacity, not a boxed cardOfficial product page lists no price
Mac Studio M3 Ultra (512 GB unified)512 GB unified memorybuild-to-order; the nearest verified public anchor is $14,439 for a certified-refurbished M3 Ultra (apple.com, checked Oct 2026); the 4 TB build sells at ¥122,399 (apple.com.cn)Official + Refurb
RTX Pro 6000 Blackwell96 GB GDDR7~$8,500 / ¥67,684 workstation board (marketplace listings, Oct 2026); server modules quoted $18-21k / ¥185,000Marketplace
Reference: RTX 5090 / 409032 GB / 24 GBconsumer pricing widely listed; the floor for FP8 video workMarketplace

Moore Threads’ own Prefill-as-a-Service whitepaper claims 95,920 tokens/s prefill at 64K context, 5.8M TPM and ¥646,000 monthly revenue per machine for a DeepSeek-V4-class deployment. Those are vendor whitepaper figures for LLM serving, not measured video throughput, and they are labelled as such here.

Open video models: what fits where

The requirement rows are published by the model makers; the cell marks are AgentRanks’ reading of those requirements against each machine’s memory.

Video model (published VRAM requirement)8-12 GB card24 GB (4090)32 GB (5090)96 GB (RTX Pro 6000)512 GB (Mac Ultra)80 GB (S5000)
LTX-2.5 — 32 GB at FP8; WanGP offload trick peaks at 5.9 GiB VRAM but needs 41 GiB system RAM and pays ~15 s per clip×offload only✓✓ headroomnot verifiednot verified
HunyuanVideo 1.5 (8.3B) — official floor 14 GB with offload, comfortable at 24 GB×✓✓✓ headroomnot verifiednot verified
HunyuanVideo original 13B — ~45 GB at 544×960××tight✓not verifiednot verified
Wan 14B — 40 GB+ full precision, 24 GB quantised; the 1.3B build runs on 8 GB1.3B only✓ quantised✓✓ fullnot verifiednot verified
MiniMax H3 — Day-0 adaptation announced for S5000××not verifiednot verifiednot verifiedDay-0 claimed

✓ = runs within the published requirement. offload / tight = runs only with memory offload tricks or at reduced settings, with the stated time penalty. not verified = no independent measurement exists; unified-memory and domestic-card video throughput in these models has not been published by a third party. Requirement sources: LTX model blog and community runs (ltx.io, atomic.chat); Tencent HunyuanVideo release notes; Alibaba Wan release notes; nexgpu.net for the 13B footprint.

What each machine is actually for

Three buyers, three different answers.

RTX Pro 6000 — the video workhorse

96 GB fits every open video model above at full or FP8 precision with headroom for longer clips. At ~$8,500 marketplace it is the cheapest of the three that can render everything on this page today. CUDA is the reason the community numbers exist for it at all.

renders all of the above

Mac Studio 512 GB — the memory whale, not the renderer

512 GB unified memory runs the biggest open-weight LLMs on one box — the reason people buy it. Video generation is its weak side: bandwidth-bound, and for the models above there is no verified third-party render throughput. Buy it for the 200K+ context LLM, not the clip.

LLM capacity, video unverified

MTT S5000 — the capacity play you cannot price

80 GB, full FP8-FP64 path, Day-0 support for DeepSeek-V4 and MiniMax H3, and a vendor whitepaper promising server economics. With no public price and no independent benchmark, it is a claim to evaluate per-deal, not a line item you can compare.

price not published
Why the empty cells stay empty. This table would be easy to fill with guesses and hard to defend. Video throughput on unified memory and on the S5000 has not been measured by anyone we can cite, so those cells say “not verified” instead of a number. The prices carry the label of where they came from: an official page, a refurb listing, or one marketplace seller. A spec sheet is only useful when you know which part of it is marketing.

The 8-12 GB playbook Coding hardware tiers Model leaderboard