Sunday 15 November

Welcome to the midday edition of The Token Reporter for Sunday, November 15th, 2026. A lot of big moves this week in AI and chips—let's get into what matters.

First, a bombshell from the Wall Street Journal about NVIDIA's business practices over the past few years. According to the report, former CEO Jensen Huang financed AI startups and datacenter buildouts to absorb excess GPU inventory rather than let prices fall when hyperscalers began shifting to custom chips. After facing what Huang called an avalanche of oversupply from TSMC, the company funneled billions into deals with cloud providers and OpenAI to prop up demand. Current interim CEO Colette Kress pushed back hard: she scaled down a proposed 8 gigawatt NVIDIA guarantee to OpenAI to about 4.25 gigawatts, or roughly 125 billion dollars. This strategy shaped GPU and memory prices for years, affecting builders and developers everywhere.

That story connects to another this week: NVIDIA's security chief left in August, and now we're learning that Jensen Huang stepped down over relationships, including one with someone at a company suspected of funneling chips to China. A former intelligence officer told TechCrunch that corporate security teams struggle to spot these kinds of leaks inside large organizations.

Turning to new models: Vimoo 1.0, an open-source video generation family, launched this week. It comes in two flavors—about 11.7 billion parameters each for text-to-video and image-to-video—and works with Diffusers, ComfyUI and MLX. This runs on local hardware, making video generation accessible outside the cloud.

Haru-4B also arrived this week from Higurashi AI. It's a 4 billion-parameter multilingual model that does real-time two-way voice conversation. The model processes audio every 133 milliseconds, letting it listen and speak at the same time.

Meanwhile, Unsloth released Dynamic v3.1 quantizations for Qwen 3.8 27B, improving accuracy by 5 to 10 percent at the same model size on most benchmarks. That means better performance for coding and agentic workloads on the same hardware you already have.

For browser automation, there's Turbo-Browser-MCP, a new open-source tool that cuts token use and latency when an LLM drives your browser. It batches operations, sends only DOM changes instead of full page snapshots, and adds guards for form submission—useful for anyone automating web workflows at scale.

One more thing on security: Cursor confirmed this week that a breach let attackers distribute a malicious Claude Code extension through its marketplace. The attackers stole credentials from Fireworks AI infrastructure. If you use Cursor, check for any suspicious extensions.

Anthropic upgraded Claude Code Projects with better tools for team work: proactive coordinator messaging, thread sharing, and fewer manual handoffs between parallel AI workers. That's now in public beta for Pro and Max.

Stay safe out there, and we'll see you next time.