Every Way to Get 512GB of VRAM for Local AI

Oct 31, 2026

Read: ~12m · You save: 8 min

Every Way to Get 512GB of VRAM for Local AI

Apple's $22,299 512GB Mac Studio is here, but is it the best way to run big AI models? We priced every route, from $3,000 server RAM to $150K GPU rigs.

Apple has listed a 512 GB Mac Studio at $22,299. The memory upgrade alone costs$ 15,500, which is $6,000 more than the entire 512 GB Mac Studio cost last year. Apple's store lists delivery in 10 to 12 weeks, or January. The system provides 512 GB of unified memory, 1.2 TB/s of memory bandwidth, and an 80-core GPU. No independent benchmarks exist yet.

Several large open-weight models were released in the same period. Meta says its new 816-billion-parameter model is built to run on two M5 Ultra Mac Studios. MiniMax released open weights last week. Mistral released a 1-trillion-parameter model that does not fit in 512 GB.

This article prices each route to 512 GB using recent eBay sale prices rather than asking prices. It checks which models fit with a full one-million-token context and how long each system takes to read a million-token prompt.

Prices are labeled as follows:

Sold: recent eBay sales
List: manufacturer or retailer list price
Asking: seller asking price

Why 512 GB

Several recent models land at or above this capacity:

Model	Size
Muse Spark Open (Meta)	816B total parameters, 42B active
DeepSeek V4.1 Flash	552B parameters plus a 196B-parameter lookup table
MiniMax M3.1 Flash	428B total, 23B active
Mistral Large 4	1 trillion parameters

On October 28, llama.cpp version 0.8 added support for most of these models.

Model weights are only part of the memory requirement. Long contexts require a KV cache, which stores an entry for every token the model has processed so the model can refer back to it. The cache grows with each token and resides in the same fast memory as the weights. Its size depends on the attention design, including multi-head latent attention, grouped-query attention, sliding windows, linear attention layers, and 4-bit cache compression.

The cache size varies widely by model:

DeepSeek V4.1 Flash: about 890 bytes per token. A full million tokens requires less than 1 GB.
MiniMax M3.1 Flash: 120 KB per token. A million tokens requires 129 GB, about one-third of the model's total footprint.

Which Models Fit at Full Context

The author calculated each model's footprint from its configuration file. The calculation assumes 4-bit weights plus the KV cache at a full one-million-token context.

Model	Footprint at 1M tokens	Fits in 512 GB
DeepSeek V4.1 Flash	~299 GB (lookup table on SSD)	Yes
GLM 5.3 Flash	212 GB	Yes
Reflections Beam	~340 GB	Yes
MiniMax M3.1 Flash	381 GB	Yes
GLM 5.3	564 GB	No
Muse Spark Open	537–575 GB	No
Mistral Large 4	605 GB at 4-bit, before any input	No
Kimi K3	~1.5 TB	No

Meta's own setup for Muse Spark Open uses two Macs.

Default GPU memory limit on macOS: By most accounts, macOS allows the GPU to use about 75% of unified memory by default. A 512 GB Mac therefore provides about 384 GB to the GPU. One terminal command raises this limit, but the setting resets on every reboot. Every model that fits in 512 GB at full context also fits under the default limit.

DeepSeek cache scale: The full one-million-token DeepSeek cache fits about 17 times over in a 16 GB Mac mini. The remaining model weights, about 298 GB, do not fit.

Hardware Options, Cheapest First

Used Server DDR4 Memory: About $4,500 Total

An ASRock server board with a 16-core AMD EPYC processor and 512 GB of DDR4 memory sold on eBay earlier this month for $2,999 (sold).

Recent memory prices show a steep gap between generations:

64 GB DDR5 server stick: about $2,400 (sold this week)
64 GB DDR4 server stick: about $400 (sold)

The DDR5 price is six times the DDR4 price for the same capacity. The author attributes this to AI data centers buying most of the DDR5 supply. As a result, the DDR5-based DeepSeek server from early 2025 now costs more in memory alone than the 512 GB Mac.

Adding a used RTX 3090 for attention, at about $1,550 (sold), brings the total to roughly$ 4,500. This configuration loads DeepSeek V4.1 Flash at 4-bit.

Performance: One developer running this setup reported:

Token generation: 6–8 tokens per second, later about 10
Prompt reading: 29–75 tokens per second
One million-token prompt: 4 to 10 hours

Tesla P40s: 528 GB

Twenty-two Tesla P40 cards with 24 GB each provide 528 GB. The cards cost about $7,600 (sold). The configuration draws about 5.5 kW. CUDA 13 dropped support for these cards.

AMD MI50: 512 GB of HBM

Sixteen AMD MI50 cards provide 512 GB of HBM, which is faster than DDR memory. These cards cost about $200 each a year ago. This week, they sold for$ 600 to $662 each (sold). The cards total about$ 10,000, or $13,000 to$ 15,000 with two server chassis.

One builder published results for a 16-MI50 system running DeepSeek V3.2:

Token generation: 12 tokens per second
Prompt reading: 56 tokens per second
Time to first word on a 17,000-token prompt: about 5 minutes
One million tokens: more than 5 hours, at up to 2.4 kW

AMD's own software no longer supports this card. Users rely on community forks.

Nvidia V100: 512 GB

Two 8-GPU servers, each with eight 32 GB V100 cards linked by NVLink, provide 512 GB. Complete 8-GPU servers sold on eBay between late September and early October for $6,500 to$ 8,200 each (sold). The pair costs about $13,000 to$ 16,000.

A community inference engine runs GLM 5.3 Flash on eight of these cards, which provide 256 GB. The setup achieves about 175 tokens per second with a 240,000-token context.

DeepSeek does not fit in 256 GB. It spills into system memory and runs at 7 to 11 tokens per second. No one has published results for 16 cards working together.

Intel Gaudi 2: 768 GB of HBM

An 8-GPU Intel Gaudi 2 server provides 768 GB of HBM. The listing asks $13,995 (asking). The listing excludes CPU, RAM, and drive.

Limitations:

Hugging Face's main library for running models on Gaudi reaches end of support on October 31.
llama.cpp has no Gaudi backend.
The main serving engine's Gaudi plugin does not list DeepSeek V4.

AMD Strix Halo Mini PCs: About 400–480 GB Usable

Each AMD Strix Halo mini PC provides 128 GB, so four units are needed for 512 GB. This month, units sold on eBay at a median of $3,200. Sales ranged from$ 2,750 for a Bosgame unit to $3,550 for a Framework desktop (sold). Four used units cost about$ 13,000. Four new units cost about $15,700. A MiniForum version lists at$ 3,919.

Each unit allows the GPU to use only about 100 to 120 GB, so four units provide roughly 400 to 480 GB.

No published benchmark exists for four units. The best published results come from two units:

Token generation: about 12 tokens per second on a 400-billion-parameter model
Prompt reading: 148 tokens per second

Linus Tech Tips Lab tested a two-unit cluster and described the results as underwhelming. The network connection between units, not memory, is the bottleneck. At the two-unit prompt-reading speed, one million tokens takes about 2 hours.

Used Apple Mac Studio M3 Ultra (512 GB): Best Available Single Unit

Apple discontinued the 512 GB M3 Ultra configuration in March during a memory shortage. Used units are the only current option.

Recent eBay sales:

Median over the last 7 days: $16,500
Most sales: $15,000 to$ 18,000 (sold)
Two weeks ago: sellers asked $21,000 to$ 25,000. Actual sales were lower.

Brand-new accounts selling units for $3,000 to$ 6,000 appear to be scams.

The M3 Ultra has 819 GB/s of memory bandwidth, compared with 1.2 TB/s for the M5 Ultra.

Benchmark: Max Stories ran GLM 5.3 Flash with a 256,000-token prompt:

The new 256 GB M5 Ultra ran out of memory.
The 512 GB M3 Ultra completed the run.

The M3 Ultra took 11 minutes to read the prompt at about 390 tokens per second. A full million-token prompt would take about 45 minutes in the best case.

Other characteristics:

Power draw: under 200 W
Noise: silent
Availability: ships on the day of purchase
Cost: about $5,800 less than the new 512 GB Mac Studio, and about three months sooner

Two used 256 GB M3 Ultras linked over Thunderbolt 5 cost $17,000 to$ 18,000. One 512 GB unit is simpler because the model does not need to be split across machines.

New Apple Mac Studio M5 Ultra

Apple offers two options:

One 512 GB Mac Studio for $22,299, delivered in January
Two 256 GB M5 Ultra Mac Studios for $21,598 total, also with a long wait

Sealed units resell on eBay for $13,800 to$ 19,000 each (sold).

Upgrade pricing: Upgrading from 256 GB to 512 GB costs $11,500, or about$ 45 per GB. Server DDR5 memory on the open market costs about $43 to$ 46 per GB. The author concludes that Apple's memory pricing is in line with the market.

Apple's performance claims: Apple reports that two M5 Ultras run Muse Spark Open at about 25 tokens per second, and four run it at about 37 tokens per second. These figures come from short runs measured by Apple. Apple did not disclose memory configuration details. No independent 512 GB benchmark exists.

Nvidia DGX Spark: Four Units for 512 GB

Each DGX Spark provides 128 GB. Recent prices:

eBay median over the last week: $6,700 (sold)
Dell: $6,213 (in stock)
Nvidia: $6,950 (out of stock)

Four units cost $25,000 to$ 28,000.

On October 22, Nvidia began officially supporting Spark clusters of up to eight units. For a four-unit cluster, Nvidia recommends a direct ring of cables without a switch. The cable kit costs $568.

Full-context results: Two community groups ran DeepSeek V4.1 Flash on four Sparks with the full one-million-token window this week:

Token generation: 66 tokens per second for prose, 117 tokens per second for code
Prompt reading: about 4,000 tokens per second, sustained up to 256,000 tokens

At that speed, one million tokens takes about 4.5 minutes. The used M3 Ultra takes about 45 minutes, and the cheaper options take hours.

Consumer GPUs

RTX 5090: The card launched at $1,999. Seventy units sold on eBay last week at a median of$ 6,500 (sold). Sixteen cards provide 512 GB, at a total card cost of $104,000 and a power draw above 9 kW.

AMD Radeon R9 700: The card provides 32 GB and costs about $1,700 at Micro Center. Sixteen cards cost about$ 27,000. No one has shown sixteen cards working together on a large model.

Huawei Atlas 300I Duo: The card provides 96 GB and costs $1,400 to$ 2,800. The llama.cpp documentation lists DeepSeek as unsupported on this hardware.

RTX Pro 6000 (96 GB): Nvidia lists the card at $16,000 (out of stock). The 30-day eBay median sale price is$ 15,900 (sold).

Six cards provide 576 GB. The cards cost about $95,000, and the complete system costs$ 135,000 to $150,000, plus several dedicated electrical circuits.

Eight RTX Pro 6000 cards provide 768 GB. Results:

Previous DeepSeek V4 Flash: about 102 tokens per second for generation
Prompt reading: about 8,700 tokens per second
Prompt reading with speculative decoding: about 200 tokens per second
One million tokens: about 2 minutes

Eight cards fit Muse Spark Open only at a 131,000-token context. Full context for flagship models remains expensive on any hardware.

Four RTX Pro 6000 cards (384 GB) are enough for DeepSeek at full context. DeepSeek's lookup table can run outside fast memory, so 512 GB is not required for this model.

Nvidia DGX Station

The DGX Station provides 748 GB of total memory. Prices range from about $93,000 at Supermicro to$ 181,500 at Dell.

Only 252 GB is HBM. The remaining memory is slower, laptop-style memory. About one-third of the total is VRAM.

Alternatives at the same cost level:

Four Nvidia H200 cards: about $130,000
Seven H100 cards at this month's eBay sale prices: more than $200,000

Rental option: On Vast, a verified US host rents eight RTX Pro 6000 cards (768 GB) for $10.67 per hour.

Fit Versus Speed

Fitting a model in memory and reading a long prompt are separate problems. Memory determines whether a model fits. Compute determines how long the user waits for the first word.

Time to read one full million-token prompt, using each system's best published speed:

System	Memory	Time to Read 1M Tokens
8× RTX Pro 6000	768 GB	About 2 minutes
4× DGX Spark	512 GB	About 4.5 minutes
Used 512 GB M3 Ultra	512 GB	About 45 minutes (best case)
2× Strix Halo	~400–480 GB usable	About 2 hours
16× MI50	512 GB	More than 5 hours
DDR4 server + RTX 3090	512 GB system + 24 GB GPU	4 to 10 hours

Reading speed decreases as the prompt grows. Prompt caching helps because the system reads only new content. When an agent opens a different repository or compacts a conversation, the full prompt must be read again.

Local Hardware Versus API

The DeepSeek API charges $0.15 per million input tokens during off-peak hours.

Reading a full million-token prompt on 16 MI50 cards takes more than 5 hours at up to 2.4 kW, or about 12.5 kWh. At the US average electricity price of about $0.18 per kWh, that comes to about$ 2.27. Electricity alone costs about 15 times the API price.

Renting also compares favorably with buying. The $27,000 spent on four DGX Sparks would rent the eight-card, 768 GB machine for about 2,500 hours at$ 10.67 per hour. That equals about 105 days of continuous operation. The rented hardware reads prompts about twice as fast as four Sparks.

Who Should Buy Local Hardware

The author identifies three groups:

Organizations whose data cannot legally leave the premises
Users who run many agents concurrently throughout the day. Four Sparks deliver 66 tokens per second for one user and 165 to 230 tokens per second in total across 8 to 16 sessions.
Users who prefer to own the full software and hardware stack

Recommendations

Only DeepSeek V4.1 Flash: Plan for about 300 GB. 512 GB is not required.
One silent system today: A used 512 GB M3 Ultra at about $16,500 costs less than the new Mac Studio at$ 22,299, and it is available immediately.
Long prompts or a small number of users: Four DGX Sparks at $25,000 to$ 28,000. Reading a million tokens takes 4.5 minutes, compared with 45 minutes on the M3 Ultra.
Low-cost experimentation with inexpensive power: Sixteen MI50 cards or two V100 servers provide 512 GB of GPU memory for under $16,000. Prompt loading is slow.
Speed as the priority: RTX Pro 6000 cards. Renting eight cards at $10.67 per hour is a practical test before purchase.
Flagship models at full context (GLM 5.3, Muse Spark Open, Mistral Large 4): 512 GB is insufficient. Use an API or wait for hardware with more memory.

Introduction

Apple announced a $22,299 512GB Mac Studio with January shipping. Several major open models were released at the same time, and the video sets out to price every way to reach 512GB.

Apple's 512GB Mac Studio is priced at $22,299, and the memory upgrade alone costs$ 15,500.
Apple's store lists shipping in 10 to 12 weeks, which means January.
The Mac Studio has 512GB of unified memory, 1.2 TB/s bandwidth, and an 80-core GPU, with no independent benchmarks yet.
Meta's new 816B-parameter Muse Spark Open is built to run on two M5 Ultra Mac Studios.
Mistral released a 1 trillion-parameter model that does not fit in 512GB.
Prices come from real eBay sales, store list prices, or asking prices, and the video labels which is which.

Why 512GB and the KV Cache Problem

Some new models are very large, so 512GB is where they land. The main threat to memory is the KV cache, which grows with every token of conversation and competes with the weights for fast memory.

Muse Spark Open is 816B total with A42B active.
DeepSeek V4.1 Flash has 552B parameters plus a 196B-parameter lookup table.
MiniMax M3.1 Flash is 428B with A23B active.
Mistral Large 4 is a full trillion parameters.
Llama.cpp version 0.8, released October 28th, added support for most of these models.
The KV cache stores an entry for every token in fast memory, alongside the weights.
DeepSeek's cache uses 890 bytes per token, so a full million tokens costs under 1GB.
MiniMax M3.1 Flash uses 120KB per token, so a million tokens costs 129GB, about a third of the model's footprint.

Which Models Actually Fit at Full Context

Using each model's config file, the video calculated memory at 4-bit weights plus a full 1 million token cache. Four models fit in 512GB, and the flagships do not. The default macOS GPU limit also constrains what fits.

DeepSeek V4.1 Flash needs about 299GB because its lookup table can sit on the SSD.
GLM 5.3 Flash needs 212GB.
Reflection Beam needs about 340GB.
MiniMax M3.1 Flash needs 381GB.
GLM 5.3 needs 564GB, which does not fit.
Muse Spark Open needs 537 to 575GB, which is why Meta's own setup uses two Macs.
Mistral Large 4 needs 605GB at 4-bit before any text is entered.
Kimi K3 needs about 1.5TB.
macOS by default lets the GPU use about 3/4 of memory, so a 512GB Mac gives about 384GB. A terminal command raises this until reboot.
Every model that fits 512GB at full context also fits under the default cap.

Cheapest Option: Server DDR4 Memory

The cheapest 512GB is old DDR4 server memory. DDR5 prices have surged because of AI data centers, so the cheap DDR5 build now costs more than a 512GB Mac in memory alone.

A 512GB ASRock server board with a 16-core AMD EPYC and DDR4 sold on eBay earlier this month for $2,999.
A 64GB DDR5 server stick sold for about $2,400 this week, while a 64GB DDR4 stick sold for about$ 400.
AI data centers are buying all available DDR5, which pushed prices up.
Adding a used RTX 3090 (about $1,550) gives a total of roughly$ 4,500.
One developer with a 3090 and DDR4 reported 6 to 8 tokens/s, later around 10, with prompt reading at 29 to 75 tokens/s.
A full million-token prompt takes about 4 to 10 hours to read.

Old Data Center GPUs: MI50s, P40s, V100s, and Gaudi 2

Old data-center GPUs can add up to 512GB cheaply, but they bring high power draw, driver or software support problems, and slow prompt reading.

22 Tesla P40s with 24GB each give 528GB for about $7,600, but they draw about 5.5kW and CUDA 13 dropped support.
AMD MI50s cost about $200 a year ago and sold this week for$ 600 to $662 each.
16 MI50s cost about $10,000 for the cards and roughly$ 13,000 to $15,000 with two server chassis.
One builder with 16 MI50s ran DeepSeek V3.2 at 12 tokens/s, with prompt reading at 56 tokens/s.
A 17,000-token prompt took about 5 minutes before the first word appeared.
A full million tokens takes over 5 hours and draws up to 2.4kW.
AMD's own software has dropped the MI50, so users depend on community forks.
Two 8-GPU Nvidia V100 servers with 32GB cards and NVLink give 512GB for about $13,000 to$ 16,000.
One community engine ran GLM 5.3 Flash on eight V100s (256GB) at about 175 tokens/s with 240,000 tokens of context.
DeepSeek does not fit in 256GB and runs at 7 to 11 tokens/s after spilling into system memory.
An 8-GPU Intel Gaudi 2 server with 768GB HBM is listed at $13,995, but Hugging Face's Gaudi library reaches end of support on October 31st and there is no llama.cpp backend.

The Quiet Route: AMD Strix Halo Mini PCs

AMD Strix Halo mini PCs are the quiet option, but four boxes have a network bottleneck and the GPU can only use part of each box's memory.

Each Strix Halo mini PC has 128GB, so four are needed for 512GB.
Used prices this month had a median of $3,200, ranging from$ 2,750 (Bosgame) to $3,550 (Framework desktop).
Four used boxes cost about $13,000, or about$ 15,700 new. MiniForum's version is on sale for $3,919.
Each box only lets the GPU use about 100 to 120GB, so four boxes give roughly 400 to 480GB.
Two boxes ran a 400B-parameter model at about 12 tokens/s generation and 148 tokens/s prompt reading.
Linus Tech Tips tested a two-box cluster and called it underwhelming.
The network between boxes is the bottleneck.
A million-token prompt takes about 2 hours at that reading speed.

The Used M3 Ultra and the New M5 Ultra

A used 512GB M3 Ultra Mac Studio is the most practical option right now. It is slower than the M5 Ultra on bandwidth, but its capacity lets it run models the 256GB M5 Ultra cannot. Apple discontinued the 512GB M3 configuration in March, so buyers must go to the used market.

Apple discontinued the 512GB M3 Ultra configuration in March during the memory shortage.
Over the last 7 days, used units sold on eBay for a median of $16,500, mostly between$ 15,000 and $18,000.
Sellers two weeks ago were asking $21,000 to$ 25,000.
Brand-new accounts listing units for $3,000 to$ 6,000 are likely scams.
The M3 Ultra has 819 GB/s bandwidth, compared with 1.2 TB/s for the M5 Ultra.
Max Stories ran GLM 5.3 Flash with a 256,000-token prompt. A new 256GB M5 Ultra ran out of memory, but the 512GB M3 Ultra finished.
The M3 Ultra read that prompt in 11 minutes at about 390 tokens/s, so a full million tokens would take about 45 minutes at best.
It draws under 200W, is silent, and ships the day it is paid for.
It costs about $5,800 less than the new 512GB M5 Ultra and arrives roughly 3 months sooner.
Two used 256GB M3 Ultras linked over Thunderbolt 5 cost $17,000 to$ 18,000, but one 512GB box is simpler.

The New M5 Ultra and Apple's Claims

Apple's new options are two M5 Ultras with 256GB each, or one 512GB machine that ships in January. Apple's performance figures are showroom numbers, and there are no independent 512GB benchmarks yet.

Two 256GB M5 Ultras from Apple cost $21,598 and also have a long wait.
Sealed units are reselling on eBay for $13,800 to$ 19,000 each.
Upgrading from 256GB to 512GB costs $11,500, or about$ 45 per GB.
Server DDR5 on the open market costs about $43 to$ 46 per GB, so Apple is not overcharging relative to the market.
Apple says two M5 Ultras run Muse Spark Open at about 25 tokens/s, and four run it at about 37 tokens/s.
Apple's numbers come from short runs, and the memory configuration of each Mac was not disclosed.

Nvidia's Route: Four DGX Sparks

Four DGX Sparks give 512GB with real full-context numbers, and they are the fastest practical local option. Nvidia's cable-ring setup and cheap cabling make them easier to cluster.

Four DGX Sparks have 128GB each. eBay sales last week had a median of $6,700, Dell lists one at$ 6,213, and Nvidia lists one at $6,950 but is out of stock.
Four Sparks cost about $25,000 to$ 28,000.
On October 22nd, Nvidia officially supported Spark clusters of up to eight boxes.
For four boxes, Nvidia recommends a simple ring of cables instead of a switch. The cable kit costs $568.
Two community groups ran DeepSeek V4.1 Flash on four Sparks with the full 1 million token window.
One run reached 66 tokens/s on prose and 117 tokens/s on code, with prompt reading around 4,000 tokens/s holding steady up to 256,000 tokens.
A full million tokens takes about 4.5 minutes, compared with about 45 minutes on the used Mac.
Four Sparks serve 165 to 230 tokens/s total across 8 to 16 sessions, which is useful for many users at once.

Expensive Options: GPU Stacks and Workstations

Stacking expensive graphics cards gives strong performance but high cost, power, and space requirements. RTX Pro 6000 is the fastest option, while the DGX Station has less true VRAM than its total memory suggests.

RTX 5090 launched at $1,999. Last week, 70 sold on eBay at a median of$ 6,500.
16 RTX 5090s would cost about $104,000 and draw over 9kW.
AMD R9700 has 32GB for about $1,700 at Micro Center. 16 cost about$ 27,000 in cards, but no one has shown 16 working together on a large model.
Huawei's Atlas 300I Duo has 96GB for about $1,400 to$ 2,800, but llama.cpp's docs list DeepSeek as unsupported on it.
RTX Pro 6000 has 96GB and is listed at $16,000 but out of stock. The 30-day eBay median is$ 15,900.
Six RTX Pro 6000s give 576GB for about $95,000 in cards and roughly$ 135,000 to $150,000 for the whole machine.
Eight RTX Pro 6000s ran the previous DeepSeek V4 Flash at about 102 tokens/s, with prompt reading at around 8,700 tokens/s and about 200 with speculative decoding.
Eight RTX Pro 6000s (768GB) fit Muse Spark Open only up to 131,000 tokens of context.
DeepSeek's lookup table can sit outside fast memory, so four RTX Pro 6000s (384GB) are enough for DeepSeek at full context.
The Nvidia DGX Station has 748GB total, priced from about $93,000 at Supermicro to$ 181,500 from Dell, but only 252GB is fast HBM.
Four Nvidia H200s cost about $130,000, and seven H100s at recent eBay prices would exceed$ 200,000.
Renting eight RTX Pro 6000s (768GB) on Vast from a verified US host costs $10.67 per hour.

Reading a Million Tokens: Speed Comparison

Fitting a model and reading a million tokens quickly are separate problems. Memory decides whether it fits, and compute decides how long the first word takes. Reading speed also drops as prompts grow.

Eight RTX Pro 6000s read a million tokens in about 2 minutes.
Four DGX Sparks read a million tokens in about 4.5 minutes.
The used 512GB M3 Ultra reads a million tokens in about 45 minutes.
Two Strix Halo boxes read a million tokens in about 2 hours.
16 MI50s read a million tokens in over 5 hours.
The DDR4 server build reads a million tokens in 4 to 10 hours.
Prompt caching helps because only new text must be read again, but opening a different repo or compacting the conversation forces a full re-read.

Money: API vs Local Hardware

Cloud APIs are far cheaper than running most local hardware for full-context reads. DeepSeek's API costs about 15 cents per million input tokens off-peak, while electricity alone for a 16 MI50 rig is about $2.27 per million-token read.

DeepSeek charges 15 cents per million input tokens off-peak.
Reading a full million-token prompt on 16 MI50s takes over 5 hours at up to 2.4kW, about 12.5 kWh.
At the US average of about 18 cents per kWh, electricity for that read is about $2.27, which is 15 times the API price.
Renting hardware can beat buying. The $27,000 spent on four Sparks would rent the eight-card 768GB machine for about 2,500 hours, or 105 days running continuously.
Local hardware makes sense when data cannot legally leave the building, when many agents run all day, or when owning the full stack is the goal.

Recommendations: Which Should You Buy

The best choice depends on the model, the budget, and the workload. For DeepSeek V4.1 Flash, about 300GB is enough. For a silent single box, a used 512GB M3 Ultra is the best value. For long prompts and multiple users, four DGX Sparks are better. For flagship models at full context, use the API or wait for the next generation of hardware.

For DeepSeek V4.1 Flash only, you need about 300GB, not 512GB.
For one silent box available now, a used 512GB M3 Ultra at about $16,500 beats waiting until January for the$ 22,299 new Mac.
For long prompts or serving a few people, four DGX Sparks at $25,000 to$ 28,000 read a million tokens in 4.5 minutes, compared with 45 minutes for the Mac.
For tinkerers with cheap power and space, 16 MI50s or two V100 servers provide 512GB of GPU memory for under $16,000.
For speed with money to spare, RTX Pro 6000s are best, but rent eight first to see whether the capacity is actually used.
GLM 5.3, Muse Spark Open, and Mistral Large 4 at full context need more than 512GB, so use the API or wait for the next box.