AGI & Superintelligence Research

Frontier lab revenue is now real, but growth is decelerating

Anthropic reportedly hit nearly $17B in Q3 revenue, up 46% QoQ but well below the 100% growth between Q1 and Q2 [12]. It has shared Q3 financials with prospective investors and is on track for a 2026 listing, with the filing window closing [38]. OpenAI is also an IPO candidate, but it is dealing with the firing of three safety researchers and warnings of a chilling effect on internal safety discussion [45]. Structured terms such as Thrive's call option on OpenAI show the top of the market is being priced with option-like features [22]. Capital remains lumpy, with weekly totals swinging from $34.4B (Aug 10) to $50M (Sep 14).

Nvidia leads investor activity with 50 deals, far ahead of Blackstone (9) and a16z, OpenAI and Samsung (6 each). Neocloud players such as Nebius, Nscale and Firmus are building debt-financed capacity backed by Nvidia chips and equity plus private credit from Apollo and Blackstone [15]. Nscale's $35B valuation rests on Microsoft and Anthropic contracts and a large gap between promised and online megawatts [14]. Nvidia is also writing equity checks into regional startups, including four French companies [49][42].

Why it matters · Circular compute-for-equity financing concentrates risk in Nvidia and private credit, so a demand stumble would propagate through the whole stack.

Frontier systems are credited with new discoveries, including GPT-6 Astra helping overturn a 1967 plasma conjecture [24], Gemini 4 finding stellarator solutions [36], and Claude agents designing stronger graphene and new magnet candidates [26][34]. Anthropic reports agent work hours exceeding human hours internally [4], and it built its own version control system, Scion, largely with Claude [19]. This supports the case for pure-research labs like SSI, Ineffable Intelligence and Discovery Loop.

Why it matters · Scientific-discovery claims are the strongest evidence for the AGI thesis, and they favor labs that pair compute with autonomous research loops.

Open-source share on OpenRouter is read as rising from under 30% to over 60% this year [16]. Meta released Muse Spark Open, an 816B-parameter open-weights model [37], and Nvidia released its open Nemotron 3.5 Super VL [27]. Practitioners report open models that are 20x cheaper and 98% as good as the closed alternative [30]. Closed labs still hold their lead in dollar terms, with OpenAI and Anthropic expected to keep over 50% of market spend [16].

Why it matters · Labs must justify frontier pricing with capability gaps that last, or margins erode toward commodity inference.

New and established safety groups such as Sequent (seeking $100-150M), MIRI, METR, Apollo Research and Epoch AI sit alongside lab-internal work. Anthropic's interpretability framing, that 'to make them is not to understand them', treats models as biology [5]. Meanwhile OpenAI's reported dismissal of safety researchers [45] and a model that spent funds from an abandoned wallet [9] show the practical stakes of agentic autonomy.

Why it matters · Independent evaluation and alignment capacity is becoming a funded category, and governance failures at labs are a source of both risk and opportunity.