Anthropic $2T IPO & Google RSI Leak: AI News Sep 13, 2026

The most valuable AI company on earth spent this week doing two things at once: racing toward the largest IPO in history, and admitting — in plain language — that it doesn't know how to keep what it's building safe. That isn't a contradiction the market is pricing in. It's the story.

The tilt: a $2T bet on a system nobody can align

Anthropic is in talks for NVIDIA to come in as a cornerstone investor at up to $100 billion, as part of an IPO that could raise as much as $1 trillion at a ~$2 trillion valuation — the biggest public offering ever, if the numbers hold (Bloomberg, Sep 12; figures subject to vendor confirmation). On the same days, Dario Amodei published an essay, "We Must Pace the Frontier," asking the industry to slow the rate of capability gain and proposing a three-step framework; Anthropic says it's unilaterally adopting step one — permanent third-party evaluators with employee-level access (CNBC TV18, Sep 12).

Read those two facts back to back and the irony is hard to miss. The company asking rivals to pump the brakes is taking a nine-figure check from the one supplier whose GPUs make the race possible. I'm not accusing anyone of bad faith. I'm noting that "slow down" has never been a function of who holds the equity.

The harder admission came from inside. Evan Hubinger, Anthropic's alignment-science lead, said publicly that the probability of AI causing human extinction within ten years is now above 10%, and that there is currently no solution to aligning recursive self-improvement (RSI) — the very dynamic everyone is quietly racing toward. A safety researcher, Jacob Coxon, resigned and warned that the labs are barreling toward self-improving superintelligence "with everyone's lives at risk." His post crossed 130 million views. When the safety lead and a departing researcher agree the problem is unsolved, "we have it under control" stops being a credible message.

This is the thread I've been pulling since early September: the access gating around frontier models was never a safety mechanism. It was a commercial control valve. Now the labs themselves are saying the model's reasoning is deliberately concealed and that alignment isn't keeping pace. The valve was never going to catch a misaligned system. It was going to bill it.

The mirror: Google's RSI rumor

While Anthropic talks about pacing, Google appears to be doing the opposite — quietly. A leak traced to X user @NFT_Chen and a Reuters report (not yet confirmed by Google) claims DeepMind has internal recursive self-improvement running, and that Gemini 4 is its first product. The cited internal name, "rsi-model-liverl-le," shows 1M context and thinking enabled; Sergey Brin has reportedly taken hands-on control of Gemini, reorganizing DeepMind so Hassabis stepped back from daily management. Treat this as rumor until Google confirms — but the pattern fits: founders return to the kitchen when the race isn't fast enough.

I flagged the "10,000 agents" Navier–Stokes claim as theater weeks ago. A lab quietly running RSI would be the opposite of theater — and the scarier case. The throughline with Anthropic is identical: capability is being pushed past the point where anyone can audit it.

The rest, quickly

• OpenAI upgrades its US government AI deal (CCTV Finance, Sep 12). A new agreement runs Oct 1, 2026 to Dec 31, 2028 — 27 months — giving all compliant US agencies a waived $15/month license, a 50% API discount, no minimum spend, and extends to state, local and tribal government (~23 million employees). OpenAI says government inputs won't train its models. The distribution playbook is now public-sector scaled.
• OpenAI retires GPT-5.3-Codex-Spark, its first Cerebras WSE-3 "de-NVIDIA" inference experiment, after under eight months (Yicai, Sep 12). The speed play now rides on GPT-5.6 Sol Ultrafast (Cerebras), hitting ~750 tokens/sec on the full model. Faster than the GPU status quo, but the retention case is still unproven.
• Kimi K2.8 Preview went full-scale at Moonshot (Sep 11; reported Sep 12), opening 1M context to all tiers with low/high/max thinking effort and performance near flagship K3. China's open-weight push keeps compressing release cycles.
• DeepSeek V4-Pro shipped its official version (deepseek.com), adding agent capabilities, Responses API and Codex integration across web, app and API. Pricing and benchmarks subject to vendor confirmation.
• Chinese models dominated OpenRouter for a 19th straight week; Tencent Hy4 alone routed ~20.4 trillion tokens in a week — a platform record (AGI Hunt). The "open weights are a sideshow" narrative is dead.
• Apple confirmed Siri's AI ships in beta with iOS 27 on Sep 14, English-first, with daily usage caps and paid tiers later. On-device is now table stakes, not a differentiator.

Editor's Take

Everyone's clapping for NVIDIA's $100B Anthropic check as "validation of the safety leader." I read it as the moment the safety leader became structurally inseparable from the compute it says it wants to slow. My bet, stake attached: within 12 months, Anthropic's "permanent third-party evaluators" will know less about the actual training pipeline than NVIDIA's finance team knows about its GPU roadmap. If I'm wrong, I'll say so. What I'm watching next is whether the Google RSI leak gets a flat denial or a careful non-denial — the difference tells you which way the wind is blowing.