Observer published its 2026 AI Power Index on July 15, ranking 90 people by a single criterion: who controlled the flow of capital in artificial intelligence over the past twelve months. The list itself is interesting. Its composition is more interesting.

The List Is an Infrastructure Map

The top ten includes the names you’d expect: Elon Musk (#1, after rolling xAI, X, and SpaceX into a $1.75 trillion Nasdaq-listed conglomerate), Sundar Pichai (#2, with Alphabet’s 2026 capex guidance lifted to $180-190 billion), Dario Amodei (#3, with Anthropic projecting $47 billion annualized revenue and reportedly pursuing an IPO that could push valuation above $1 trillion from its last-known $965 billion mark). Jensen Huang, Sam Altman, Mark Zuckerberg round out the model lab tier.

But the list doesn’t stop at lab founders and chip executives. Prince Mohammed bin Salman (#7) and Sheikh Tahnoon bin Zayed Al Nahyan (#11) represent sovereign wealth capital that is reshaping where AI infrastructure physically lives. Stephen Schwarzman (#8) of Blackstone anchors the private equity capital that funds data center construction. Patrick Collison (#12) of Stripe makes the list for building agentic commerce infrastructure: payment and transaction systems designed for AI agents to autonomously execute financial workflows, per Observer.

Harry Sideris (#37) runs Duke Energy’s AI datacenter division. Bob Blue (#47) runs Dominion Energy, which powers Virginia’s data center corridor. Their presence on an AI power list would have been unthinkable two years ago. It signals that the binding constraint on AI deployment has shifted from model capability to physical infrastructure: power, cooling, and permitting.

The Model Scaling Plateau

Observer’s framing question, “Who controls the capital flow in AI?” implicitly asks what that capital is buying. Two years ago, the answer was straightforward: training runs. Bigger models, more GPUs, higher benchmark scores. That market still exists, but the returns have diminished.

The evidence is visible in the July 2026 model release cycle. xAI’s Grok 4.5, OpenAI’s GPT-5.6, and Meta’s Muse Spark 1.1 all launched within days of each other. All three marketed themselves on cost efficiency and agentic reasoning performance rather than raw intelligence gains. Grok 4.5 uses a 295-billion-parameter MoE architecture that activates only 21 billion parameters per token, delivering what xAI calls “Opus-level” agentic performance at $2 per million input tokens, according to BigGo Finance reporting. OpenAI replaced its flagship/mini naming convention with scenario-based tiers (Sol, Terra, Luna) where cost follows the task, not the model.

When every frontier lab optimizes for the same metric (cost per agentic task), capability convergence is close. The differentiation moves downstream: who builds the infrastructure that runs agents in production?

Where Capital Actually Moved in H1 2026

The Observer index aligns with where venture capital actually flowed in the first half of 2026. June 2026 US venture capital totaled $19.27 billion across 429 companies, with AI-focused startups capturing 59.8% of all capital, according to AlleyWatch data NCT previously reported. The largest rounds went to inference infrastructure (Baseten, $1.5B Series F) and custom inference chips (Groq, $650M), not model training.

In the agent platform category: Emergent closed a $130 million Series C at $1.5 billion valuation after reaching $5 million ARR and 5 million users within seven months of public launch. Nous Research, maker of the open-source Hermes AI agent, is finalizing a $75 million round at $1.5 billion valuation. Lyzr used its own agent, SivaClaw, to autonomously manage a $100 million Series B fundraise at $500 million valuation.

In governance and sovereignty: Valarian raised $50 million for sovereign AI control planes. Citrix shipped its NetScaler MCP Gateway for centralized agent authentication. Ant Group open-sourced SingGuard for agent guardrailing. Anthropic is hiring 32 specialists in catastrophic misuse prevention. Each of these represents a capital allocation decision that favors agent runtime infrastructure over model training.

The Collison Signal

Patrick Collison’s inclusion at #12 may be the most telling data point on the list. Stripe processes payment infrastructure. Its expansion into agentic commerce means building systems where AI agents autonomously initiate, authorize, and complete financial transactions without human confirmation for each step.

That’s a different product category from “API access to a language model.” It requires transaction security, identity verification, fraud detection, and regulatory compliance designed for non-human actors. If Stripe’s agentic commerce infrastructure becomes standard, it creates a new trust layer between agents and the financial system, and whoever controls that layer captures margin on every agent-initiated transaction.

Observer’s framing is right: the question that matters in H2 2026 is not which model leads the benchmarks. The benchmarks have converged. The question is who controls the infrastructure that agents run on, the governance layers that constrain them, and the commerce systems that let them spend money. Capital has already answered.