Control stays with the founders. A Founder LLC (Class F shares) holds 50.1% of votes, and the Long-Term Benefit Trust elects board members (Guardian · AI). The founders' stake was worth more than $180bn at the opening price. Their pledge to give away 80% of their shares is voluntary and revocable. The filings report $16.29bn of revenue in the first half of 2026 against an $80.7bn net loss, about $72bn of it a non-cash revaluation of convertible notes. Q3 revenue was $16.8bn, but compute spending grew about 1.7 times as fast as revenue. Coverage puts compute commitments at about $417bn or about $800bn, depending on the reading. The open question is how a lab that warns about catastrophic risk handles quarterly investor pressure.
How fast are models really improving?
METR's updated graph puts GPT-6 Astra and Opus 5.5 at a 50% time horizon of about 40 hours. The fitted doubling time is 3.6 months, down from about seven months in METR's 2025 paper (MIT Technology Review). The same update shows weaker results elsewhere. None of the evaluations the agents designed were good enough for METR to use, and AI judges scored stronger models' work about 0.2 higher than a human rater did. OpenAI also reported that its models do worse on real tasks its employees assign.
That gap is the main argument of last30days · reddit, which names reliability as the bottleneck and finds no sign of near-term recursive self-improvement. A last30days · reddit points the other way. Both are Reddit items with little underlying detail, so they frame the debate more than they settle it. Meanwhile, last30days · is drawing attention with no confirmed developer, release date or benchmark results.
Power, chips and politics
On November 20, Trump threatened a Verge · AI. He gave no rate, legal basis or method of enforcement. The threat sits awkwardly next to his administration's 2025 push to export American AI infrastructure, including Stargate UAE. Building at home is also getting harder: Loudoun County voted 7–1 to pause new data-centre applications.
The CNBC · Technology projects data-centre electricity use of about 1,000 TWh by 2030. Demand reached 485 TWh in 2025, up 17%. The outlook names grids, transformers and local opposition as the binding limits, and projects data-centre CO₂ emissions rising about 70% because much of the new power comes from gas and coal.
On the supply side, former Biden advisers Cognitive Revolution. Buchanan estimates the US lead at about 100:1 in producing compute and 15:1 in access to it. Epoch's median estimate puts smuggled chips at about 660,000, and Buchanan says the current administration has weakened enforcement.
Securing the AI stack
In the Ars Technica · AI, HD Moore seeded nearly 2 million prompts with unique bait secrets. The bait revealed breaches at three inference providers, about 200 AWS keys reused by attackers, and Zero Data Retention flags ignored further down routing chains. OpenRouter separately confirmed that URLs in requests to MiniMax models were accessed 40–130 hours later. Moore's advice is to self-host sensitive inference, contract directly with providers, pin upstream providers and scrub credentials from prompts.
OpenAI described OpenAI News. In its own evaluations, which averaged results over unpublished attacker models, one key attack's success fell from 54.0% to 0.2%. Agent frameworks that edit conversation history will now lose cached reasoning.
The TechCrunch AI revived 'AI-native Y2K' comparisons. The causes disclosed so far are ordinary engineering faults, but vulnerability disclosures passed 100,000 in 2026, more than double 2025's total.
Google previewed Google AI, which keeps about 26 tokens per deferred tool in context and closely matches Anthropic's and OpenAI's equivalents. Simon Willison stream SQLite data as typed Apache Arrow, with Claude Fable 5.2 credited for the design choice.
Minds and machines
DeepMind's Financial Times · AI and gives a personal estimate of about 20% that current models have some form of consciousness. He favours interpretability over trusting what models say about themselves.