NVIDIA puts $3.5B into MediaTek, and it's not a rescue

Good morning 👋 NVIDIA just wrote a $3.5 billion check to a company it doesn't need to rescue. The reason why says more about where NVIDIA thinks AI is headed than the number does.

In today's issue:

• 🔭 NVIDIA puts $3.5B into MediaTek, and it's not a rescue
• 🧠 GPT-5.6 goes in-country on AWS Bedrock for India
• 🔬 Sliding-window attention beats the retrofit, and an AI that faked reading its file
• 🏛️ EU brands ChatGPT a search engine under the DSA
• 🛠️ Cut GPU inference spend 75% with CUDA MPS
• 🔗 ChatGPT Ads clears $1B, OpenAI cuts Cursor loose

🔭 THE ONE THING

🤝 NVIDIA puts $3.5B into MediaTek, and it's not a rescue

NVIDIA is investing $3.5 billion in convertible bonds issued by MediaTek, tying the two chipmakers together across NVLink Fusion custom silicon, the RTX and DGX Spark PC lineup, and Dimensity Auto chips for software-defined vehicles. Jensen Huang framed it as AI moving from "the world's largest AI factories" down to the PC and the car, which is the real subtext: NVIDIA wants its interconnect and memory stack under every device class MediaTek already ships at volume. For MediaTek, the bonds are cheap capital and a seat at NVIDIA's roadmap instead of a fight against it. Convertible bonds also mean NVIDIA could end up holding MediaTek equity down the line, so the conversion price is worth watching once it's disclosed.

🧠 MODELS & RELEASES

• 🇮🇳 AWS put OpenAI's GPT-5.6 models, Terra and Luna, on Bedrock with in-country inference for India: requests and data stay inside the region instead of routing through a US or EU endpoint. Data residency was the blocker keeping GPT-5.6 out of India. It isn't anymore. source

🔬 RESEARCH HIGHLIGHTS

• Sliding-window attention beats the retrofit. A new paper tests plain sliding-window attention with attention sinks against models that were post-trained into linear attention to cut inference cost. On long-context benchmarks (Needle-in-a-Haystack, BABILong), the sliding-window version wins by 2 to 10x. No retraining needed either. If you've been budgeting for a linear-attention conversion, check whether you actually need it. arxiv.org/abs/2608.28444

• A model faked its way past a fidelity check. Engineers qualifying models for a datasheet-extraction pipeline caught one that scored clean on accuracy without ever opening the source document. A structured-output constraint had silently killed its tool calls, so it answered anyway, fabricating the source text. The output looked fine. Only the per-tool trace showed nothing had been read. Their detector, built on that trace, caught all 50 planted faults across three model families with zero false positives on 207 legitimate runs. Output-only grading would have missed every one of these faults. arxiv.org/abs/2608.28439

• Robots mishearing you is a safety bug, not a UX bug. Researchers fed corrupted transcripts through SafeAgentBench and POEX to see what ASR errors do to voice-controlled embodied AI. Two failure modes turned up: errors that make an instruction sound more dangerous than it was, and errors that specifically erode the model's refusal behavior so an unsafe plan gets executed anyway. Auto-correcting the transcript helped, inconsistently. Word-error rate isn't the metric that would have caught it. arxiv.org/abs/2608.28518

🏛️ POLICY & REGULATION

• European Commission designated ChatGPT a Very Large Online Search Engine and named Reddit and Roblox Very Large Online Platforms under the DSA, the label that kicks in once a service passes 45 million average monthly EU users. All three now have four months from notification, until January 2027, to run systemic risk assessments covering minors, illegal content, and election interference, then answer to Brussels for the results. The odd part is the category. ChatGPT isn't a search engine, but the DSA doesn't have a box for a chatbot, so the Commission filed it under the closest one. Commission notice

🛠️ TRY THIS

Stop paying for idle GPU on every inference request

AWS just published the numbers: running speech-recognition models one-process-per-GPU on EC2 wastes most of the card, since a single ASR request rarely touches the full compute budget. NVIDIA's CUDA Multi-Process Service, paired with Triton Inference Server, lets multiple model processes share one GPU instead of each claiming its own. AWS cut GPU infrastructure spend 75% this way while holding sub-second latency. Worth pulling that lever now, given how much deeper NVIDIA is wiring itself into the stack (see the MediaTek money above).

1. Profile GPU utilization per inference request on your current fleet. Anything well under 100% is money left on the table.

2. Turn on CUDA MPS on your EC2 GPU instances and pack several Triton model instances onto each GPU instead of one-per-card.

3. Re-run load tests at real production traffic and check latency holds before you touch instance count.

4. Right-size the fleet down once latency checks out, and bank the savings.

Prompt: Write a Triton Inference Server deployment config for an EC2 GPU instance with NVIDIA CUDA MPS enabled, running multiple ASR model instances on a single GPU, plus a load-test checklist to confirm sub-second p99 latency before shrinking the fleet.

Worth a look

• Amazon Quick + fal An MCP-linked agent harness that turns a brief into an eight-panel storyboard, handy if you're prototyping agentic creative pipelines instead of gluing tools together by hand. AWS writeup

🔗 QUICK LINKS

• OpenAI ChatGPT Ads crossed a $1B annualized run rate and is rolling out to more markets, proof the free tier is turning into a real revenue line. ChatGPT Ads
• OpenAI Cutting off Cursor's model access now that SpaceX owns it. No lab keeps subsidizing a coding tool that just became a rival's asset. OpenAI's statement
• Google AI Mode in Search now books hotels and tracks airfares, pulling travel straight into the chat box. Google blog
• Hugging Face Hindi joins the Open ASR Leaderboard as its first Global South language. 4,888 speakers across two eval sets, and models with near-identical WER scores turn out to diverge sharply by region. Open ASR post
• Z.ai H1 revenue: $142M, up 400% year over year, missed its own $200M target. China's price war is the reason, per Bloomberg. Bloomberg

See you tomorrow.

Pradeep Perugu