CxO · Weekly · September 14, 2026

the frontier ‘slowdown’ and emerging safety cartel

Musk, Altman, and Dario Amodei are publicly aligned on the need to "pace the frontier" and slow AI acceleration for safety reasons.[1] OpenAI has pushed its planned 2026 IPO off the table as "ill‑advised" in the current safety climate, while stating that alignment work is now its main bottleneck.[2]

Anthropic and OpenAI have embedded third‑party evaluators with permanent access to systems, positioning bodies like METR as de facto industry gatekeepers.[3]

In parallel, OpenAI is lobbying for mandatory national safety rules and Sanders is floating 20‑year prison terms and a corporate death penalty for superintelligence work, with 68% of U.S. voters backing some form of AI pause.[4]

Critics inside and outside the labs describe this as regulatory capture in safety clothing, noting that token consumption at OpenAI is up 124x and Anthropic researchers say RSI has "started to happen" even as they call for a slowdown.[5]

compute, power, and the capital land‑grab

Anthropic committed about $80B of compute in five days, including $45B with Nscale and $35B with Lambda, effectively pre‑buying a slice of the future GPU grid.[6]

BlackRock’s Larry Fink estimates the U.S. will need around $10T over the next decade to build out AI infrastructure, and AI revenue already ran at roughly $229B annualized by late 2023.[7]

Google is locking in energy and geography, signing a 22‑year deal for up to half the output of Finland’s Loviisa nuclear plant and pledging $15B of AI infra investment there.[8]

On the equity side, Cognition raised over $2B at a $48B valuation on a coding model business with run‑rate revenue jumping from $492M to nearly $900M in a few months, while Mistral raised €3B at a €21B valuation in Europe’s largest tech round.[9]

Nvidia, dubbed the "central bank of AI", is marketing $100 of return for every $1 it invests and is now under DOJ antitrust scrutiny just as skeptics compare AI infra exuberance to pre‑2008 circular financing.[10]

open‑weight squeeze on frontier pricing

Fine‑tuned open‑source models are reported to be up to 95% cheaper than frontier APIs while matching or beating them on targeted tasks.[11]

OpenBMB’s MiniCPM5‑2B tops open‑weight rankings for models under 4B parameters, and DeepSeek V4.1 Flash and Intern‑S2‑397B show that large, multimodal, open‑ish systems can reach frontier‑like benchmarks.[12]

Mistral’s record €3B raise is explicitly for "sovereign" open‑weight frontier AI, and Switzerland is piloting open‑source office stacks on thousands of government PCs as a hedge against Microsoft lock‑in.[13]

Garry Tan is publicly backing U.S. open‑weight labs for frontier distillation, framing them as a strategic counterweight to closed incumbents.[14]

Meanwhile, AI spending among the top 1% of businesses fell 10% in August, suggesting some large buyers are already pulling back from expensive frontier‑API experiments as cheaper options mature.[15]

china, IP leakage, and the split AI stack

China plans a fourfold increase in AI compute capacity by 2030 and already accounts for 13% of global datacenter power usage versus 43% for the U.S.[16] The NSA, FBI, and CISA jointly warned that Chinese AI companies are running industrial‑scale distillation of U.S. frontier models, effectively stripping proprietary behavior via API access.[17]

U.S. agencies have separately accused specific Chinese firms of "industrial‑scale" copying, and export‑control debates are intensifying.[18]

In parallel, Malaysia is weighing Huawei’s Ascend 910C chips for a $494M national AI build‑out despite long‑running U.S. security warnings about Huawei hardware, and Huawei is showcasing 7.2Tbps near‑package optics aimed squarely at AI infra.[19]

Commenters across political lines argue that any U.S. slowdown would simply hand the lead to China, while others see the arms‑race framing itself as the main driver of over‑investment and surveillance at home.[20]

ai surveillance and civil‑liberties blowback

Flock’s automated license‑plate readers are now tied to wrongful arrests, stalking cases, and over 1,200 scans in one Illinois town just to investigate loitering, and municipalities in places like Waukesha County are being sued over deployments.[21]

U.S. Republicans in some states are openly revolting against Flock’s systems, framing them as mass surveillance rather than crime‑fighting tools.[22]

Anthropic is reported to be building predictive surveillance to track anti‑AI activists "before crime" and its Claude model has been used by Iran to surveil thousands of people and analyze over 155,000 tweets.[23]

Palantir is deepening into public infrastructure, from integrating GP data into the NHS platform to a no‑bid contract with Britain’s grid operator, even as critics call its models a civil‑liberties risk.[24]

Across these cases, activists and many commentators describe pervasive AI surveillance as oppressive and under‑regulated, while vendors emphasize security and efficiency to justify sticky government contracts.[25]