AI News Roundup – October 20, 2026

TL;DR

• Reuters: ByteDance, Alibaba and Tencent have ordered more than 3 million Nvidia RTX PRO 5500 cards worth over $35 billion, but Nvidia has told partners it can’t guarantee delivery, and some may get only 5% to 10% of what they asked for.
• OpenAI says its latest models can now migrate COBOL code while preserving behavior: GPT-6.1 Sol scores 86.6% at low effort (up from 5.4%) and 97.9% at higher effort.
• Google open-sourced TorchTPU and Amazon released TorchNeuron, so PyTorch code can run natively on TPUs and Trainium with minimal or no rewrites.
• DeepSeek open-sourced DeepInfer, the inference engine it says powers its production services, under Apache 2.0.
• Perplexity Computer can now run fully local on a Mac, with both the model and the agent on-device.
• Mistral launched Voxtral 3 speech models and a voice agent SDK, with transcription from $0.0035 per minute.
• OpenAI acquired Orchestra Bio to build drug-development agents alongside its Rosalind life-sciences tools.
• Anthropic says Claude designed 17 lab-confirmed protein binders against scorpion neurotoxins, out of 211 designs tested.

---

Why can’t Nvidia fill China’s $35 billion chip order?

Short answer: According to a Reuters exclusive published October 20, citing people familiar with the matter, ByteDance, Alibaba and Tencent have ordered more than 3 million of Nvidia’s new RTX PRO 5500 graphics cards, with a combined sticker price of over $35 billion. The RTX PRO 5500 is a workstation chip, less powerful than Nvidia’s data-center processors but export-compliant, and among the most powerful Nvidia chips Chinese buyers can still get.

Nvidia has told customers and partners it cannot guarantee it can fill the orders because demand from Chinese tech companies far exceeds planned production. One source said partners might receive only 5% to 10% of the quantities requested. Another said Nvidia indicated it could make no more than 3 million RTX PRO 5500 cards globally in 2027, while total orders from China have passed 6 million. The buyers want the cards mainly for AI inference, which means running trained models rather than training new ones. Nvidia’s Blackwell-based B300 can’t be sold to Chinese buyers under US export controls, and H200 exports are reviewed case by case.

Why it matters: Chinese giants ordered roughly twice what Nvidia plans to build. Tight chip supply keeps pressure on the price of AI compute worldwide and pushes buyers toward alternatives, including Huawei, Google TPUs and Amazon Trainium (more on that below). Compare what models cost to run on our Models page.

Sources: Reuters via U.S. News · Reuters via Moneycontrol · The Economic Times

---

Can AI finally migrate COBOL code?

Short answer: In an October 19 post, OpenAI said its latest models are much better at low-resource programming languages such as COBOL, Fortran, OCaml and Visual Basic, which have far less training data than Python or JavaScript. OpenAI notes COBOL supports more than 40% of online banking systems and 95% of ATM transactions.

OpenAI built two new evaluations. Its COBOL migration test asks a model to move COBOL programs to C++, Java, Kotlin, Python or Rust and requires byte-for-byte matching output on hidden tests. GPT-6.1 Sol scored 86.6% at low reasoning effort, versus 5.4% for GPT-5.6 Sol, at 39% lower API cost. At higher effort it reached 97.9%, nearly saturating the eval. On MultiLangBench, where a model builds a SQLite-compatible database engine from scratch in 20 languages, GPT-6 Astra averaged 87.4%, up from 26.8%, and scored above 83% in 19 of the 20 languages. COBOL is still the hardest, at 72.1%. OpenAI is releasing the MultiLangBench test suite on GitHub.

Why it matters: Banks, governments and insurers run critical systems on code that few developers can still maintain. Behavior-preserving migration at this accuracy could make modernization faster and cheaper, though OpenAI says production use still needs expert review and testing. See OpenAI Codex, ChatGPT and the Models page.

Sources: OpenAI: Advancing coding in low-resource programming languages · OpenAI on X · Thurrott

---

How are Google and Amazon taking on Nvidia’s software lead?

Short answer: At PyTorch Conference 2026 on October 20, Google made TorchTPU, its native PyTorch stack for TPUs, open source on GitHub. Co-developed with Meta’s PyTorch team, it supports eager mode and torch.compile and works with vLLM, SGLang, Ray, TorchTitan and Hugging Face. In Transformers v5, deploying to a TPU is as simple as device_map="tpu". Google also shipped llm-d 1.0 for distributed inference, brought LiteRT into the PyTorch ecosystem, and previewed a TPU agent suite that helps move models over.

Amazon released TorchNeuron as open source under Apache 2.0 the same day. It runs existing PyTorch code natively on Trainium2 and Trainium3 with a single configuration change and no rewrite. Amazon is also contributing changes to PyTorch itself to remove code that ties the framework to specific chips, and Hugging Face has added native Trainium support to its tools.

Why it matters: By some accounts about 85% of AI apps use PyTorch, and most default to Nvidia GPUs. If switching hardware becomes a one-line change, you can shop for cheaper compute without rewriting your app.

Sources: Google Cloud: TorchTPU is now open source · PyTorch: Introducing TorchTPU · Amazon: TorchNeuron · TechCrunch

---

What did DeepSeek open-source today?

Short answer: On October 20, DeepSeek released DeepInfer, which its GitHub README calls “the high-performance inference framework that powers DeepSeek’s large-scale production services.” It’s released under the Apache 2.0 license and supports DeepSeek-V4.1-Flash on NVIDIA SM100 GPUs and Huawei Ascend 950 chips.

The release covers offline inference with DeepSeek’s optimized kernels (including FlashMLA, DeepGEMM and DeepEP), expert parallelism for its mixture-of-experts models, draft-token decoding, asynchronous scheduling, deterministic sampling and structured outputs. DeepSeek says a complete inference service would still need pieces that aren’t included: a KV cache system, prefill/decode disaggregation and load balancing across replicas.

Why it matters: Teams that self-host DeepSeek models can now study and run the same engine DeepSeek uses in production, a rare look inside a frontier lab’s serving stack. It isn’t yet a drop-in serving system. See DeepSeek and the Models page.

Sources: GitHub: deepseek-ai/DeepInfer · DeepSeek on X

---

Can Perplexity Computer really run fully offline?

Short answer: Yes, on the right Mac. Perplexity said on October 20 that Perplexity Computer can now run fully local in its Mac app: “The model and the agent both run on your Mac.” With permission, it works with local files and Mac apps such as Notes, Messages, Mail and Numbers, and it can run tasks without an internet connection. Files reach the cloud only if you approve it, for example when a task needs current information or browser use.

The local mode launches with Qwen 3.6 and PPLX Pro (a post-trained version of Qwen, per The Verge) as model options. It’s available to Pro, Max and Enterprise subscribers, and Perplexity recommends an Apple silicon Mac with at least 36GB of unified memory.

Why it matters: An agent that never sends your files to the cloud makes AI usable for sensitive work like legal, finance or HR documents, as long as your Mac has enough memory. See our Perplexity review.

Sources: Perplexity on X · Perplexity blog · 9to5Mac · The Verge

---

What is Mistral’s Voxtral 3?

Short answer: Mistral launched Voxtral 3 on October 20: new speech-to-text and text-to-speech models plus a voice agent SDK, built on LiveKit, that listens, reasons, calls tools and speaks in one real-time pipeline. Mistral says each turn completes in under one second and the agent stops talking when a caller cuts in. It can answer and place phone calls over SIP and run in browsers and apps over WebRTC.

Voxtral Small Transcribe 3 handles 50+ languages with speaker identification (diarization), and Mistral says it beats Whisper, Scribe v2 and Gemini 3.5 Transcribe. A realtime version runs at sub-500ms latency. Voxtral Speech 3 covers 35+ languages and 50+ dialects, with voices that adapt emotion and pacing. Prices start at $0.0035 per minute for batch transcription, $0.008 per minute for realtime, and $13 per million characters for speech. It’s available through Mistral’s European, US and global API endpoints or for self-deployment.

Why it matters: Building a phone or support voice agent usually means stitching together several vendors. Mistral now sells the whole stack at low per-minute prices, with a self-hosting option for data control. See Mistral Vibe and compare voice tools like ElevenLabs.

Sources: Mistral: Voxtral 3 · Mistral on X · Mistral docs: Voice agents

---

Why did OpenAI buy Orchestra Bio?

Short answer: OpenAI announced on October 20 that it has acquired Orchestra Bio, a platform that helps biopharma teams coordinate the complex process of drug development. Orchestra’s team and technology will join Rosalind, OpenAI’s models and tools for life-sciences research, including GPT-Rosalind and Rosalind Workbench.

On Orchestra, scientists and program leaders hand off the monitoring and tracking of studies to AI agents. OpenAI’s example: if a manufacturing batch fails, the agents identify which activities and timelines are affected, gather the relevant information and compare ways to respond. OpenAI says developing a new medicine takes roughly 10 to 15 years on average from target discovery to US approval. Deal terms weren’t disclosed.

Why it matters: OpenAI is moving from answering science questions to running the operational side of drug programs, a sign that labs see agents as project managers, not just chatbots. See ChatGPT.

Sources: OpenAI: OpenAI acquires Orchestra Bio · OpenAI on X

---

How did Claude design proteins against scorpion venom?

Short answer: In an October 20 Science blog post, Anthropic said an internal research model working inside Claude Science designed minibinders, small proteins that attach tightly to a target, against scorpion neurotoxins. Scorpions sting more than 1.2 million people a year and cause more than 3,250 deaths, mostly children in places with limited medical care.

Working mostly autonomously for 64 hours, from one prompt and four follow-ups, Claude used 12 design tools to generate 318,484 designs and picked 180 for testing; the team added 31 more with Claude’s help. Lab partners Adaptyv Bio and Twist Bioscience tested all 211, and 17 met Anthropic’s criteria as binders, each against a single toxin. The two strongest bound the Mexican scorpion toxin Cn2 tightly and likely attach at the same site as an antibody known to neutralize it. Anthropic is clear that this is not an antivenom: no neutralization or animal tests have been done yet. It’s releasing all 211 designs and their data.

Why it matters: AI agents are starting to run lab-validated science on neglected diseases that get little investment. It’s an early step, but it shows where tools like Claude are heading beyond coding.

Sources: Anthropic: How Claude designed protein binders against scorpion neurotoxins · Anthropic on X

---

What should builders watch next?

1. Chip supply: whether Nvidia’s RTX PRO 5500 allocations really land at a fraction of China’s orders, and how much demand shifts to Huawei, TPUs and Trainium.
2. Hardware choice: real-world benchmarks of TorchTPU and TorchNeuron against Nvidia GPUs on cost per token.
3. Legacy code: whether banks and governments start COBOL migration pilots with OpenAI’s latest models, and what independent tests on the MultiLangBench suite show.
4. Private and voice agents: local agents like Perplexity Computer and full-stack voice agents like Voxtral 3 are getting cheaper and easier to deploy.
5. AI in science: follow-up tests on Claude’s scorpion-toxin binders and what OpenAI builds with Orchestra Bio.

For tool deep-dives, browse the AI tools directory or the Models page. This roundup summarizes public announcements and reporting; it is not hands-on product testing.

Published by Andrew for AI Tools Optimizer.