Decision support

Guides & Research

Practical guides and source-backed research for choosing GPU capacity, comparing rental terms, and validating inference workloads.

Deployment guide · Oct 21, 2026

Gemma 4 26B A4B: When Is Self-Hosting Worth Testing?

Use your Gemma 4 26B A4B API bill to decide when a one-H100 investigation is worthwhile. Then test the actual context lengths and engine build before adding an assistant checkpoint for speculative decoding.

Deployment guide · Oct 14, 2026

Qwen3-VL-30B-A3B: When to Test Self-Hosting on an H100 NVL

When does Qwen3-VL-30B-A3B usage justify investigating self-hosting? Compare billed traffic with one H100 candidate, then preserve screenshot resolution, video timestamps and the complete visual-agent conversation in the test.

Deployment guide · Oct 12, 2026

Qwen3.5-27B: When to Test Self-Hosting on an H200

When does Qwen3.5-27B traffic justify investigating self-hosting? Use the live API spending line for one H200, then test this dense model with the actual reasoning policy, output lengths and video workload. Encoder separation can be the wrong optimization.

Deployment guide · Oct 11, 2026

Gemma 4 31B: When to Test Self-Hosting on 80 GB

When does your Gemma 4 31B API traffic justify investigating self-hosting? Compare billed usage with the live one-H100 compute budget, then test the original BF16 checkpoint with your image resolution, context lengths and complete tool conversations.

Deployment guide · Oct 9, 2026

Qwen3.5-35B-A3B: Does Your API Bill Justify an H200 Test?

When does Qwen3.5-35B-A3B traffic justify investigating self-hosting? Use the live API spending line for one H200, then test expert memory, hybrid-state caching and speculative decoding on the original BF16 checkpoint.

Deployment guide · Oct 7, 2026

Qwen3-14B: API or Self-Hosting on 48 GB?

When does Qwen3-14B traffic justify investigating self-hosting? Use the live spending threshold for one L40S, then resolve the BF16-versus-FP8 and thinking-mode choices before testing.

A graphite graphics card beside a pale shipping box on an ivory background.

Deployment guide · Oct 6, 2026

Where to Buy a Used GPU

A US Guide to Second-Hand GPUs for AI

Deployment guide · Oct 6, 2026

gpt-oss-120b: When Is an H100 Worth Testing?

When does your gpt-oss-120b API spending justify investigating self-hosting? Follow the live traffic calculation, then test the memory and latency limits of a single H100.

Deployment guide · Oct 5, 2026

Qwen3.5-9B: API or Self-Hosting?

When does your API bill justify testing a single-GPU deployment? This comparison covers text requests only; image and video workloads need a separate cost and capacity assessment.

A graphite GPU connected by an amber line to two pale server shapes on an ivory background.

Comparison · Oct 5, 2026

Small GPU Providers or Large Clouds

How to choose a GPU rental provider for inference and training

Token tiles connect through an amber routing junction to a graphite processor, with two pale processor blocks on an ivory background.

Deployment guide · Oct 4, 2026

gpt-oss-20b

When Is Your API Bill High Enough to Test Self-Hosting?

A graphite graphics card with two fans and a small amber check mark on an ivory background.

Deployment guide · Oct 4, 2026

New or Second-Hand GPUs for AI

How to Make the Right Buying Decision

A graphite processor connects to a group of pale tiles, with one tile and its connection highlighted in amber on an ivory background.

Deployment guide · Oct 1, 2026

Qwen3-30B-A3B self hosting or api

When does Qwen3-30B-A3B traffic justify testing a self-hosted FP8 service? Use current API and rental rates with a measured one-L40S reference, and see where its MoE memory and workload shape limit the comparison.

Three token paths converge on a layered graphite processor, with a restrained amber connection on an ivory background.

Deployment guide · Sep 30, 2026

Qwen3-32B self hosting or api

Qwen3-32B's thinking switch and measured FP8 cluster shape change what its live traffic crossover means. See which deployment is actually worth testing.

A stream of token tiles reaches one graphite processor with an amber core on an ivory background.

Deployment guide · Sep 29, 2026

Llama 3.1 8B self hosting or api

Find the Llama 3.1 8B traffic volume that makes a GPU deployment test worthwhile, with current API and rental prices, single-H100 benchmark context and long-context memory calculations.

A stream of tokens meeting two graphite GPU shapes, with a muted amber connection on an ivory background.

Deployment guide · Sep 29, 2026

Llama 3.3 70B self hosting or api

When does your Llama 3.3 70B traffic warrant a self-hosting test? Calculate the threshold from fresh prices, adjust it for your workload and costs, then check the throughput you need.

A graphite graphics card with an amber tile lifting from its surface, beside an ivory hourglass on a warm ivory background.

Deployment guide · Sep 28, 2026

Spot GPUs for inference: when the discount is worth the interruption

Measure the real spot discount, account for recovery time, and keep your inference endpoint serving when capacity is reclaimed.

Three graphite processor motifs progress from a single amber block to an array of parallel units on a warm ivory background; a conceptual illustration of GPU evolution.

History · Sep 27, 2026

From Pixels to AI: A History of the GPU

How dedicated graphics hardware became programmable parallel processors used for science and artificial intelligence.

Two matte graphite processor blocks resting on a shared amber slab against a warm ivory background; a conceptual illustration of shared resources.

Comparison · Sep 26, 2026

Apple unified memory vs GPU VRAM

Understand how sharing physical memory changes CPU/GPU data movement, resource management, and the bottlenecks that determine application speed.

A graphite graphics card beside a small ivory stopwatch with an amber hand, arranged on a warm ivory background.

Deployment guide · Sep 25, 2026

How to benchmark a rented GPU for LLM inference

Use a controlled benchmark and representative application traffic to compare GPU rentals by response time, reliability, and cost.

A dark graphite GPU chip with a small stream of ivory token tiles and one amber tile on a warm ivory background.

Deployment guide · Sep 25, 2026

GPU inference cost per million tokens: an H100 example

See how an H100 rental rate and published Llama 3.1 8B benchmarks translate into output-token costs, then account for actual traffic and billed time.

A graphite graphics card beside a small ivory calendar tile with a muted amber binding, on an uncluttered ivory background.

Deployment guide · Sep 25, 2026

Monthly GPU rentals: when a fixed bill beats hourly pricing

Compare monthly GPU rental offers by their complete configuration, commitment, and break-even hours—not just the price beside /month.

A graphite chip beneath three floating memory layers, with one muted amber layer on an ivory background.

Deployment guide · Sep 24, 2026

How much GPU memory does your model need?

Estimate the memory your language model needs by accounting for its weights, cached context, and runtime overhead—and validate the workload before choosing a GPU.