From November 4 (Japan time) AI news, here's a roundup of five stories — including the Nvidia chip supply squeeze for China, exit polls from the U.S. midterms, and Revolut's AI coding cost cuts — plus other notable moves.

Nvidia's China-Bound Chips: Only About 1% of $7 Billion in Orders to Be Delivered — Bloomberg Reports

According to Bloomberg, citing people familiar with the matter, Chinese AI and cloud companies including Alibaba, Tencent, and ByteDance have ordered a total of around six million Nvidia RTX Pro 5500 units — a deal worth roughly $7 billion. Yet only "tens of thousands" of units, about 1% of the total, are expected to actually arrive over the next four months.

The reason: U.S. production capacity hasn't kept pace with surging demand. With export controls barring Chinese firms from top-tier datacenter GPUs, they are piling into the high-end chips they can still buy within the rules — meaning compute constraints look set to keep squeezing Chinese AI development for the foreseeable future, even once deliveries begin.

U.S. Midterm Exit Polls — AI and Data Centers Rank Last Among Voter Priorities

In the 2026 U.S. midterm elections, NBC's exit polling found that just 56% of voters named "AI and data centers" as an important issue — the lowest-ranked issue in the survey. Corruption (90%) and the cost of living (89%) far outstripped it.

Even as public interest in AI grows, it drops down the priority list when voters' finances and distrust of politics are at stake — a result that underscores the reality that AI policy alone doesn't win votes. Bloomberg also reports that the election ended in a decisive Democratic sweep of Congress; S&P 500 futures rose 0.2%, the dollar fell against major currencies, and the 10-year Treasury yield opened 3 basis points lower.

Revolut Cuts AI Coding Costs to One-Tenth — Drops Claude for Self-Hosting

UK fintech giant Revolut has cut its AI coding costs to one-tenth, The Register reports. The savings come from reducing its dependence on external APIs such as Anthropic's Claude and migrating to a self-hosted environment.

With API costs continuing to climb, more large organizations are likely to bring their coding-assist infrastructure in-house. That said, the information available at collection time doesn't break down the details of the migration, and it's worth keeping in mind that self-hosting carries its own costs, from GPU procurement to operations staffing.

ByteDance Releases "Axion 1.0", a Fully Open Agentic Foundation Model

A team at ByteDance has released "Axion 1.0" (35B), a fully open foundation model purpose-built for agentic workloads, on Hugging Face. Its hallmark is a simple training recipe: just three stages — two lightweight supervised fine-tuning (SFT) stages plus reinforcement learning (RL). The team says it uses no mid-training, no DPO, and no on-policy distillation.

The first SFT stage develops self-contained reasoning through "model-aware data construction," synthesizing and amplifying hard problems from the foundation model's own successful solutions. The second stage extends this to tool-dependent reasoning via verifiable tasks grounded in real-world data sources and deterministic computation, and a final RL stage strengthens long-horizon reasoning and interaction with external tools and environments. Beyond the weights, the team has released everything — data construction, generation, filtering, verification, prompts, and training code — and reports performance rivaling leading commercial models across multiple agent benchmarks.

AI Coding Agents Install Abandoned, Vulnerable MCP Servers — glow.io Research

Glow's research blog post "MCPhantom" finds that AI coding agents will install deprecated and abandoned MCP servers found in their training data — even when those servers carry known security issues.

It's the "freshness" of the information agents draw on that becomes a supply-chain risk — a problem reminiscent of traditional package management. With the MCP server ecosystem expanding rapidly, figuring out how to spot unmaintained servers and building mechanisms for pre-install verification are emerging as key challenges.

Other Notable Moves

A paper on "SpecPilot", an agentic loop that automatically optimizes speculative decoding for each deployment environment, is drawing attention. It includes a mechanism that detects reward hacking through forced token-by-token comparison, and reports boosting the speedup over autoregressive decoding from 4.46x in conventional systems to 10.87x.

A user reports running the GGUF build of DeepSeek V4.1 Flash on a DIY rig with 4x RTX 3090s and 256GB of DDR4, achieving 30–40 tokens/second of decoding on a roughly 520K-token context, with impressions noting fewer loops than the previous version.

IEEE Spectrum's piece on "AI forcing the repayment of 40 years of tech debt" picked up 110 points on Hacker News. As AI reaches further into vulnerabilities lurking in aging assets, discussion is growing over what to do with debt that could previously be left untouched.