HOSTKEY

Dedicated server leasing and private clouds in the Netherlands, USA, Finland, Iceland and Turkey - premium WEB services provider. GPU servers with NVIDIA cards

[HOSTKEY @Hostkey]

LiteLLM or OmniRoute?

In our test, LiteLLM with two workers hit 130 req/s on standard responses. OmniRoute was slower there, but better suited to multi-provider subscriptions and agent tools.

Read →
https://t.co/zCwiavWzP9

[HOSTKEY @Hostkey]

Your GPU may not be the only inference bottleneck.

Supermicro, IBM and NVIDIA tested persistent KV cache for long-context LLMs — cutting Time-to-First-Token by up to 56× in their setup.

The idea: reuse context instead of recomputing it.

[HOSTKEY @Hostkey]

Talos Linux 1.14.2 is now available for automatic installation at HOSTKEY.

Built-in BGP, DoT/DoH, Btrfs, fewer reboot-required config changes — plus Kubernetes 1.37.1 and Linux 6.18.54.

Read more →
https://t.co/p8mTqw39gV

[HOSTKEY @Hostkey]

Talos Linux 1.14.2 is now available for automatic installation at HOSTKEY.

Built-in BGP, DoT/DoH, Btrfs, fewer reboot-required config changes — plus Kubernetes 1.37.1 and Linux 6.18.54.

Read more →
https://t.co/p8mTqw39gV

[HOSTKEY @Hostkey]

Buy a $7,000 RTX 4090 server or rent GPU capacity?

We compare ownership, AWS and specialized rentals — including depreciation, electricity, maintenance and utilization.

The break-even point changes fast once the GPU runs regularly.

Watch →
https://t.co/banWrP3Vyb

[HOSTKEY @Hostkey]

Anthropic’s planned capacity will use Broadcom XPUs and networking, with deliveries starting in 2027.

The takeaway: AI infrastructure is moving beyond individual GPUs toward integrated accelerator + network systems.

[HOSTKEY @Hostkey]

One local gateway for multiple AI providers.

OmniRoute is a free, open-source AI gateway with one OpenAI-compatible endpoint, provider routing and fallback — running on your own server.

Explore →
https://t.co/IkdGnKn2TX

[HOSTKEY @Hostkey]

Faster GPU or easier deployment?

Dell’s MLPerf results show AMD MI350P at roughly 40–50% of MI355X performance — but in an air-cooled PCIe server that is easier to fit into existing infrastructure.

[HOSTKEY @Hostkey]

One API for 100+ LLM providers.

HOSTKEY now offers LiteLLM — a free, open-source AI gateway you can run on your own server, with one OpenAI-compatible endpoint plus keys, budgets and routing.

Explore →
https://t.co/IkdGnKn2TX

[HOSTKEY @Hostkey]

AMD and Intel are unifying PyTorch CPU optimizations around oneDNN.

One software path, fewer vendor-specific stacks, and future AI Compute Extensions for faster matrix operations.

[HOSTKEY @Hostkey]

Manage HOSTKEY infrastructure from an MCP-compatible AI client.

Check server status, reboot, reinstall an OS, manage DNS or prepare a new server order — with explicit confirmation before the actual purchase.

Docs →
https://t.co/cEUqquQcDr

[HOSTKEY @Hostkey]

Same RTX 4090. €299/month as a GPU VPS, €449 as a dedicated server.

The difference is CPU, RAM, storage, multi-GPU options and hardware control — not the GPU itself.

Read →
https://t.co/3GZv4sq3kC

[HOSTKEY @Hostkey]

More CUDA cores. More VRAM. Full power. Better GPU?

Not always.

We test 4 common GPU myths across H100, A100 and workstation cards — and show where compute, bandwidth, VRAM and power actually matter.

Watch →
https://t.co/ka8YUwPOga

[HOSTKEY @Hostkey]

MinIO is out. RustFS is in.

HOSTKEY now offers RustFS as an S3-compatible storage option: built in Rust, Apache 2.0 licensed, actively developed, and designed as a lighter alternative to MinIO.

[HOSTKEY @Hostkey]

An AI factory in as little as 14 days instead of six months — at a prepared site.

NVIDIA DSX POD moves compute, networking, power and liquid cooling into pre-integrated modular units.

The result: less time from hardware delivery to production tokens.

[HOSTKEY @Hostkey]

Server won’t boot?

HOSTKEY Rescue mode lets you boot a temporary OS in memory, get root access and repair the system or recover data without reinstalling the server.

Rescue RHEL is available now. SystemRescue, GParted Live and a Windows option are next.

[HOSTKEY @Hostkey]

Which LLM actually fits on one GPU?

We tested DeepSeek, Gemma and Phi on a single RTX 4090 and compared VRAM use, context limits and inference speed.

The key lesson: “fits in memory” does not always mean “runs well.”

Read →
https://t.co/pLOd3FllLx

[HOSTKEY @Hostkey]

HOSTKEY S3 now spans five countries:

NL — Standard + Cold
FR — Standard + Cold
CH — Standard
PL — Standard
DE — Cold

Cold in Poland and the UK are next.

Docs →
https://t.co/ZvOniuyp5i

[HOSTKEY @Hostkey]

Dedicated server or cloud?

We compare real AWS and dedicated configs, including GPU, egress and bandwidth, to find the actual crossover point for always-on workloads.

Watch →
https://t.co/qdbLjHKlI8

[HOSTKEY @Hostkey]

The GPU is fast. Can memory keep up?

In a demo, SK hynix MARS shared and reused KV cache across servers and GPUs via CXL pooled memory.

Our takeaway: for long-context and agentic AI, memory architecture is becoming an inference bottleneck.