InferX

Serverless GPU inference. Sub-Second cold starts — even for 32B(64GB)+ models. Scale to zero. No idle billing.

[InferX @InferXai]

Introducing a new primitive in AI: Skill Function.

A Skill Function is a protected AI capability hosted as a callable endpoint. You send input. You get expert output. The instructions never leave the platform.

Not a file you download. A function you call.

[InferX @InferXai]

Ornith-1.0-35B is now available on InferX.

If you’re looking to test a new 35B model with your own dedicated endpoint, you can spin it up now on InferX.

Get your endpoint

https://t.co/tZGoqNehm9

[Skill Function @SkillFunction]

Skills are coming to every platform. Microsoft in Excel. Google in Gemini Enterprise. But they’re all closed, locked to one vendor, one model, one ecosystem. Skill Function is the open infrastructure layer - callable from any agent, any model, any platform.

Own your intelligence.

Try it for free: https://t.co/OhGiFipUcL

[InferX @InferXai]

With @SkillFunction , Intelligence becomes callable.

[Skill Function @SkillFunction]

MCP made tools callable.

Skill Function makes intelligence callable.

Every skill becomes an independently deployable AI service with its own model, context, permissions, lifecycle, and execution environment.

https://t.co/vxbP56sleK

[InferX @InferXai]

You call the skill. You get the output. The security problem disappears.

That’s Skill Function.

https://t.co/ihTpXLNAgm

[InferX @InferXai]

NVIDIA just analyzed 42,447 public AI agent skills.

The results are alarming. 🧵

[InferX @InferXai]

This is exactly why we built Skill Function.

Skills don’t run locally anymore.

They run in an isolated cloud sandbox with zero local execution authority.

No file access. No shell commands. No credential risk. Nothing touches your machine.

[InferX @InferXai]

Skill Function - More Power to Your Agent

https://t.co/ihTpXLNAgm

[InferX @InferXai]

Renaming yourself “AI Developer Cloud” doesn’t fix shared compute, cold starts, or paying for idle GPUs.

Dedicated H100 instance. Sub-second cold starts. Scale to zero. Pay only when you use it.

$10/month. https://t.co/x4X8YLH0uq

[Runpod @runpod]

The bottleneck isn't GPUs anymore.

It's the three to five clouds most AI teams stitch together to get a model to production.

So as of today, we'll start calling Runpod the AI Developer Cloud.

Read Zhen's take on this through the link below.

https://t.co/gBaKNhZtNL

[InferX @InferXai]

Building toward this at InferX. you don’t need a PhD to take meeting notes. Not every feature needs GPT-5.5. The model is just the runtime. https://t.co/x4X8YLH0uq

[@jason @Jason]

OpenAI is trying to own every app and platform and sell tokens, which means they want to kill every startup.

That’s their right, it’s a free market.

we need open source alternatives built in America that are committed to enabling startups with tokens — instead of selling them tokens and using the profits to replace them.

[InferX @InferXai]

@NVIDIAAI @huggingface We too 💙

[InferX @InferXai]

We too 🙏

[NVIDIA AI @NVIDIAAI]

We 💚 open source

[Prashanth (Manohar) Velidandi @PMV_InferX]

On a recent episode of the @theallinpod , Bill Gurley @bgurley laid out what he called the "Dr. Frankenstein theory" regarding the frontier AI labs. After deep-diving into @AnthropicAI ’s public writings, including @DarioAmodei’s Machines of Loving Grace, Gurley pointed out an escalating, slightly eerie trend: these teams don't believe they are writing software anymore. They believe they are midwifing a centralized, superior digital deity.

At @InferXai , we completely agree with Gurley's skepticism. The industry has bought into a massive, highly synchronized delusion.

The idea that intelligence naturally scales into a single, omniscient, monolithic "God Model" defies everything we know about how intelligence actually works. Centering the future of technology around a single corporate oracle is an immense structural hazard, no matter how many 80-page "constitutions" or safety guardrails are layered on top of it.

Humanity didn’t build the modern world because one solitary super-genius emerged and the rest of the species blindly followed them. We built a civilization based on compounded, specialized, and distributed intelligence. True capability scales through high participation, friction, redundancy, and a collective network of diverse actors working in parallel.

AI will scale exactly the same way. The future is not a trillion-dollar model sitting on a mountain. it is a highly coordinated, deeply integrated collective.
And here is our core conviction: The future of intelligence is not model-centric.

A model is just a runtime. The real breakthrough isn't the single asset; it’s the orchestration fabric that binds a distributed civilization of compute together. What the industry will ultimately call this interconnected web remains to be decided.

Internally at InferX, we already know the answer. I’ll talk more about it soon.

[InferX @InferXai]

@mikeyssi @PMV_InferX Zero-Ops Inference

[InferX @InferXai]

Inference margins are real. GPU rental costs rising is real.

The answer isn’t better negotiation with NVIDIA. It’s using the GPU more efficiently.

124 dedicated instances on a single H100. On-demand. Scale to zero. Pay only when you use it.

That’s how you protect margins.

https://t.co/U3f8qmezyS

[The Information @theinformation]

Inference startups like Baseten, Modal and Together AI are drawing investor demand after rapid revenue growth. The challenge is whether they can protect margins as Nvidia server rental costs rise.

Read more in our Dealmaker newsletter: https://t.co/IgaClOxeEE

[InferX @InferXai]

@runpod Nice. We’ve been doing this on H100s since day one . any slice size, sub-second cold starts, scale to zero. $10/month. https://t.co/x4X8YLH0uq

[InferX @InferXai]

Welcome to the party RunPod. 👋

GPU slicing on a consumer RTX 6000 Pro. 24GB partitions.

We’ve been doing this on H100s since day one. Any slice size. Any model. Sub-second cold starts.

Same idea. Different hardware. Different performance.

$10/month → https://t.co/x4X8YLH0uq

[Runpod @runpod]

We just launched Multi-Instance GPU (MIG) on Runpod Serverless.

It partitions the RTX 6000 Pro into isolated 24 GB instances, each with dedicated memory and compute.

So if your workload fits in 24 GB, now you can pay for 24 GB.

Read the blog post to get started: https://t.co/p50r1UeeDM