The AI Brief

President Trump announced the Super Intelligence Force on October 4 via Truth Social, following his September 29 executive order "Inaugurating the Era of Super Intelligence," which directed all federal agencies to replace the terms "Artificial Intelligence" and "AI" with "Super Intelligence" and "SI" in public communications, websites, and reports. The order exempts existing statutory references until Congress acts.

Director of National Intelligence Jay Clayton will chair the force, effectively serving as the administration's AI czar. Vice chairs are FTC Chairman Andrew Ferguson, Under Secretary of War for Research and Engineering Emil Michael, and Office of Personnel Management Director Scott Kupor. The group reports directly to Trump and White House Chief of Staff Susie Wiles. Per the Wall Street Journal, the force has 120 days to produce a report assessing risks and opportunities and to develop plans for responding to SI-enabled threats, while "preventing overregulation and regulatory capture that would stifle innovation and competition."

The force's formation follows the September 29 White House Accord on Superintelligence, a voluntary industry pledge covered in Vol. I, No. 132, which carried no penalties and allowed companies to select their own auditors. The 120-day clock runs to early February 2027. Whether the force produces binding rules or additional voluntary guidance hinges on the legislative language the administration's policy office must deliver to Congress by November 28.

Bloomberg reported September 29 that OpenAI is in early-stage discussions to raise at least $30 billion, targeting a pre-money valuation of approximately $1.4 trillion. The figure represents a roughly 64% step-up from the $852 billion post-money valuation at which OpenAI closed a $122 billion round in March 2026, anchored by SoftBank, Amazon, and Nvidia. OpenAI CEO Sam Altman has described a 2026 listing as "ill-advised" given ongoing safety work; the new round is positioned as a bridge to a 2027 IPO.

OpenAI's annualized revenue run rate has reached approximately $70 billion, up more than 70% since the start of Q3, per Axios. The company's CFO Sarah Friar had earlier flagged that sluggish revenue growth earlier in the year raised concerns about meeting infrastructure commitments, but revenue acceleration tied to Dots, the Agents API, and Astra for Law has changed the picture. Investor appetite is reported to be strong, with Bloomberg citing demand from existing backers. Terms remain preliminary and could change before any closing.

Anthropic, by contrast, is moving toward the public market: its investor day is scheduled for October 14 at its San Francisco headquarters, with formal roadshow marketing expected to begin the week of November 9 and trading targeted before Thanksgiving, per Bloomberg. Investors peg Anthropic's fair value at $1.8 trillion to $2 trillion. The two competing datapoints, OpenAI's private $1.4 trillion and Anthropic's public target of up to $2 trillion, will together define the market's price for frontier AI this quarter.

Disclosure: Claude, which generates this brief, is built by Anthropic.

Google Research published a preprint on Cogentic, a multi-agent system built on Gemini that produced novel results on five open problems in theoretical computer science spanning online learning, auction theory, and mechanism design. The system begins only from the problem statement: no expert hints, no partially completed proofs, no domain-specific scaffolding. Developers noted in community discussion that the real breakthrough is organizational, not architectural: adversarial verification and persistent memory of proven lemmas, rather than raw model scale, drive the result quality.

The system runs a coordinator agent that breaks a research problem into parallel workstreams, with sub-agents writing code, searching literature, and attempting proofs. A dedicated reviewer agent reads all drafts from a given round simultaneously to catch shared mistakes, a design the authors distinguish from models that self-review sequentially. The preprint is authored by the system's own developers and has not yet been independently verified; the five companion papers have not been read by outside reviewers as of publication.

The Cogentic result arrives one day after Meta published six Muse Spark math papers, five of which answer open questions. The simultaneous announcements from two major labs using different architectures, one standard chat interface and one custom multi-agent harness, mark a visible step beyond competition-level benchmarks into open research territory, though neither set of results has cleared the months-long replication process standard in mathematics.

Meta Superintelligence Labs announced October 2 that a team of outside mathematicians worked with Muse Spark 1.1 and 1.2 in Thinking Mode over several months to produce six papers spanning probability, differential equations, group theory, optimization, arithmetic physics, and non-associative algebra. Five answer previously open questions. A separate set of mathematicians reviewed each paper; the company says passages are marked to distinguish researcher-drafted from AI-drafted content and credit prior work. The work used the standard meta.ai chat interface with no custom research scaffold.

Among the documented results: Muse Spark wrote a search program in GAP, a mathematical software system, that found a 384-element group disproving a 2024 conjecture; a wave-collapse result settles a question open since 2015; and an ellipsoid-fitting paper establishes a sharp threshold in high-dimensional probability. Some findings overlap with independently developed work, and the papers were published alongside Meta's announcement rather than after an extended peer-review cycle. Meta noted Muse Spark reached gold-medal level across five high-school Olympiads earlier in 2026, framing the new papers as a step from problems with known answers toward problems with no ready-made answer.

The announcement generated substantial attention partly because of its contrast with custom agentic research systems such as Google's Cogentic, published the same day. Meta's result used a retail chat product; Google's used a purpose-built multi-agent harness with adversarial verification and persistent memory. Both produced research-grade outputs on open problems, but through structurally different means, and neither set of results has cleared the months of replication review that typically follows a mathematical claim of this kind.

Anthropic signed a $35 billion, approximately six-year cloud-computing agreement with Nvidia-backed Lambda on August 31, anchored at a 352-megawatt Hut 8 data center campus called Beacon Point in Nueces County, Texas. The Wall Street Journal reported the deal. Lambda financed the build-out partly through a $926 million senior secured term loan B, priced at SOFR plus 300 basis points, that fully amortizes against contracted cash flows and is secured by the funded GPU servers and infrastructure. Nvidia reportedly holds the physical lease on the facility.

The deal follows Anthropic's $45 billion agreement with Nscale signed five days earlier and a prior $10 billion Volta commitment, bringing disclosed compute commitments to roughly $90 billion in under a month. Mizuho analyst Jordan Klein described the Lambda structure as one in which "Nvidia pre-funds purchases of its own products": GPUs serve as collateral, are rented back into the supply chain, and the loan repays via customer lease payments. The Bank for International Settlements flagged AI capital spending circularity, alongside a potential collapse in circular financing and sovereign debt fragility, as three major threats to global financial stability in its 2026 annual economic report.

Anthropic has not publicly confirmed the deal structure's specifics beyond the headline $35 billion figure. Lambda is separately reported to be in talks for a pre-IPO round of approximately $3 billion at a $12 billion valuation. The three-layer structure, supplier, equity holder, landlord, is not unique to this transaction but is concentrated enough in one counterparty that it receives no standard disclosure treatment in the contracted compute figures Anthropic's S-1 prospectus reports to investors.

Disclosure: Claude, which generates this brief, is built by Anthropic.