dailyai.stream

dailyai.stream

your daily ink of AI news

TUE, NOV 17 · 4 STORIES

Tuesday, November 17

BUSINESS & POLICY

Data centers over 12M TPP need US licenses to serve China

A Commerce Department interim rule requires a license to export, reexport, or transfer AI chips and servers to a data center anywhere in the world if it may provide compute to companies in China or other restricted countries. It covers any data center or network that can supply over 12 million TPP. TechCrunch puts that at roughly the max compute of about 750 Nvidia H100 or H200 chips. The targets are companies in the D:5 embargo group, which includes China, Russia, Iran, North Korea, and 18 others. Indirect access through brokers counts too, and operators must monitor customers' ownership and activities. Malaysia, Singapore, Indonesia, Thailand, and Brazil get new chip licensing requirements, which Commerce says is because they present a higher-than-normal risk of diversion.

Why it matters for builders

Existing operators covered by the rule have 90 days after it takes effect to get licensed. Expect GPU clouds outside the US to add ownership checks and ask harder questions at signup. If your team has ties to a D:5 country and rents compute overseas, line up a backup provider now. Check your providers in the five named countries first.

Read on TechCrunch →
OPEN SOURCE

Hugging Face tokens now run open-weight models via OpenRouter, no markup

Hugging Face users can now call open-weight models with their Hugging Face token through the same API, and the best provider is picked for each request. Partner OpenRouter says it serves more than 800 trillion tokens a month to over 10 million users, across 600+ models from 80+ providers. Hugging Face adds no markup on provider pricing, and PRO users get $2 of credits a month. The partners say open-weight models average $0.28 per million input tokens and $1.62 per million output, against $1.27 and $6.36 for closed models. That makes them about four times cheaper, though the figure compares averages, not equivalent models.

Scatter chart of Artificial Analysis Intelligence Index against cost per task on a log scale, with open-weight models such as DeepSeek V4.1 Flash 1017 and GLM-5.3 clustered at lower cost and lower scores than proprietary models.
Image: Hugging Face
Why it matters for builders

If you pass your own provider API key for these language models, that no longer works. Usage is now billed to your Hugging Face account or organization, so audit any code that sends provider keys. You do get routing control in exchange. Add a suffix to the model id, like zai-org/GLM-5.3:fastest or zai-org/GLM-5.3:cheapest, to choose speed or cost per call.

Read on Hugging Face →
MODELS

Tencent's Hy4 Flash lands on Vercel AI Gateway at 50% off

Tencent's Hy4 Flash is now on Vercel AI Gateway. Input, cached input and output tokens are 50% off as an introductory discount through November 29, 2026. Vercel lists these features: reasoning with thinking preserved across turns, tool calling, structured outputs, prompt caching and image input. The context window is 1M tokens. You can turn reasoning off for workloads that don't need it, and Vercel says that makes the model a fit for high-volume apps. The model ID is tencent/hy4-flash, and it works with the AI SDK's streamText.

Why it matters for builders

You have under two weeks of half-price tokens to test it. A 1M-token window with reasoning you can switch off makes it worth trying on bulk extraction or classification jobs that need structured output. Vercel's post gives no per-token prices and no benchmarks. Check the gateway's model listing and run your own evals before you send production traffic to it.

Read on vercel.com →
CODING & AGENTS

Vercel AI Gateway can now route to your own model endpoints

Vercel's AI Gateway now supports Private Inference, which lets you connect your own model endpoints to the gateway. You can switch between private models and catalog models, and the gateway translates each request into that endpoint's format. Private models stay scoped to your team. Credentials are encrypted, and Zero Data Retention is optional. Requests to private models show up in the gateway logs next to catalog model requests. Setting it up through the CLI requires Vercel CLI 65.16.0 or later.

The Private Inference page in the Vercel AI Gateway dashboard, listing connected providers such as a self-hosted vLLM endpoint with their private model IDs and API formats.
Image: Vercel
Why it matters for builders

If you serve your own model, it can now sit behind the same gateway as the hosted ones, and both show up in one log view. The practical win is failover. You can set a catalog model as the fallback when your private endpoint fails. Configure that before you send production traffic to your own endpoint, not after its first outage.

Read on vercel.com →

That's today's ink. Come back tomorrow for more.