AI Notebook] The G20 and AI Panels

EJ's note

I spent time this week going through the full G20 Innovation Ministerial sessions from North Carolina — Jensen Huang (Nvidia), Sam Altman (OpenAI), Tom Brown (Anthropic), Elon Musk (virtual), Treasury Secretary Scott Bessent, and Commerce Secretary Lutnick alongside Palantir's Alex Karp. Rarely do you get the frontier labs, the chip supplier, and the two most powerful economic policymakers in the U.S. on the same stage inside a single week, all saying broadly the same thing.

Here's my bottom line first, then what they actually said and my take, and — for those with time — a detailed summary of each session at the end.

Enjoy!

I. Bottom line

1. Policymakers and frontier labs are now fully aligned on a single framing — AI is infrastructure, not software. The comparison every speaker reached for was electricity. The stated worst outcome, for a country or a firm, is non-adoption.

2. The binding constraint has shifted from models to the physical world: power, grid, permitting, and construction labour. Musk was the most pointed — he expects a meaningful power shortfall as early as next year, and argues it is already handing China an edge. Brown called data-centre permitting and power the single biggest bottleneck. When the bottleneck is physical, the opportunity migrates to those who finance and build physical assets. That is our business.

3. For alternatives, the alpha pockets are (i) the AI infrastructure complex — power, grid, data centres, chips, cooling, fibre; (ii) the application layer — vertical AI that owns domain data and workflows; (iii) embodied AI and robotics as the second derivative; and (iv) the financing structures that intermediate all of it — project finance, private credit, yieldcos, sale-leasebacks.

4. Every speaker was talking their book. Nvidia sells the picks and shovels, the labs sell tokens, Treasury and Commerce are selling a growth narrative. The direction of travel is credible; the specific numbers deserve your usual underwriting scepticism.

II. What the panels shared — and my take

The consensus message

Strip out the individual styles and you get a remarkably coherent thesis. Huang: AI is national infrastructure, and every country must build local capacity across an energy-to-applications stack. Altman: adoption is "non-negotiable," and the goal should be abundant, cheap intelligence rather than concentrated scarcity. Brown: the build-out is larger than the 19th-century railroad boom, capabilities are still scaling predictably, and AI's real economic mechanism is multiplying effective intellectual labour. Musk: AI could lift the global economy 20–30%, digital work gets largely automated within a couple of years, and humanoid robotics is the next leg. Bessent and Lutnick: Washington intends to win this — deregulation, full expensing, energy expansion, and a claimed rise in the U.S. share of global compute from roughly 60% to 80% by 2028.

My take

The physical bottleneck is the investable story: When the labs themselves say the constraint is no longer intelligence but electrons and permits, the value shifts toward what alternatives firms actually do: underwrite, finance, and build long-duration physical assets. Data-centre development, generation and transmission, storage, behind-the-meter and firm power, advanced packaging, specialised EPC and cooling — this is a private-markets menu, and much of it will be financed through private credit and project structures rather than public equity.

The railroad analogy cuts both ways: Brown's comparison is apt, but worth remembering that the railroad boom was transformational for society and ruinous for many of its investors. Capex supercycles produce overbuild, stranded assets, and brutal shakeouts alongside the winners. Discipline on counterparty quality, contract tenor, and residual-value assumptions will separate the infrastructure funds that compound from those that become cautionary tales. I would also treat single-tenant, single-technology exposure with care — the pace of chip and cooling evolution is itself a residual-value risk.

In the application layer, moats are data and workflow, not models: Altman and Brown both conceded, in effect, that the labs supply the engine and someone else supplies the context. With last year's frontier models already roughly 20x cheaper, inference is deflating fast and thin "wrapper" businesses will see margins compress. For PE and VC, the filter is simple: does the company own proprietary domain data, embedded workflows, and distribution — and does it deploy AI as a co-worker with tools and permissions rather than a chatbot bolted on the side?

III. Session-by-session Summaries

Jensen Huang (Nvidia) — AI as national infrastructure

Huang frames AI as the next universal infrastructure layer alongside water, roads, electricity and the internet, and argues every country must build local capacity or be left behind. He describes a five-layer stack running from energy through chips and data centres to models and applications, notes the U.S. is invested across all of it, and suggests other countries can specialise in the layers that suit them. He references U.S. AI infrastructure investment approaching the trillion-dollar range, positions Nvidia's GPU architecture as fungible, durable infrastructure that runs every model and modality, and points to embodied AI — robotics, self-driving, autonomous labs, advanced manufacturing — as the next wave. On risk, his line is that safety matters but the single worst outcome is non-adoption; regulate real harms, not hypothetical ones. He expects AGI-level capability within a couple of years but insists firms will still need to supply context, purpose and domain tooling — jobs transform rather than vanish.

Sam Altman (OpenAI) — adoption is non-negotiable

In conversation with Secretary Lutnick, Altman traces OpenAI's arc from 2015–16 experiments to GPT‑4 as the internal inflection point, and previews a substantial capability jump in the next model. His core argument: AI adoption is non-negotiable for growth, comparable to electrification, and countries that block or over-constrain it will fall behind. He predicts the largest entrepreneurship and small-business boom in history as AI substitutes for teams and capital, stresses that compute is the primary bottleneck amid exploding token demand, and frames the policy goal as abundant, cheap intelligence — the alternative being concentration and worsening inequality. Model providers supply the engine; users and firms supply the context and use cases. Near-term risks he flags are cybersecurity and biosecurity, where a failure could trigger backlash and slow adoption.

Tom Brown (Anthropic) — scaling intellectual labour

Brown's three messages: capabilities are rising rapidly and consistently; demand is growing by multiples each year; and the resulting global infrastructure build-out exceeds the 19th-century railroad boom. He traces Anthropic's founding to the 2019 scaling-laws work, which turned AGI from "decades away" into "under ten years." Frontier models have, in his telling, reached once-in-a-generation-genius level in several mathematics subfields, with similar leaps in key sciences expected within roughly twelve months. The binding constraint is data-centre permitting and power, and he calls for "red carpet" policies on land, labour and grid access. His economic mechanism: AI multiplies effective smart labour — orders of magnitude more oncology-researcher-equivalents means multiplicatively faster progress on cancer, not incrementally faster. The biggest gains come from treating models as co-workers with full context, tools and permissions, and he argues the application ecosystem is a larger opportunity than the base-model layer, with last year's best models already about 20x cheaper.

Elon Musk (virtual) — the power warning

Musk's headline claim is that AI will probably increase the global economy by 20–30% — on the order of $20–30 trillion a year. His sharpest warning is near-term: a significant power shortfall arriving as soon as next year, not in some distant future, which he argues is already advantaging China. He calls for urgent data-centre build-out, criticises heavy-handed regulation — singling out the EU's "default illegal" posture — and argues new technologies should be default legal. He expects AI to handle almost anything digital by the end of next year, with humanoid robots delivering the next dramatic productivity leg beyond that.

Scott Bessent (U.S. Treasury) — the growth agenda

With Larry Kudlow at the G20 finance track in Asheville, Bessent argues the G20 should return to its core mission of growth — his framing being that a world "awash in debt" can only grow its way out, not austerity its way out. He credits regulatory, tax and energy certainty (including 100% expensing) for a capex boom across machinery, factories and data centres, claims the U.S. share of global compute could rise from roughly 60% in 2025 to 80% by 2028, and emphasises onshoring of critical nodes — semiconductors, critical minerals, drug precursors — under the banner that maximum efficiency is not maximum security. The session also covers a hardline sanctions and enforcement posture on Iran, including secondary-sanctions warnings, which keeps geopolitical risk premia in energy and shipping firmly on the table.

Lutnick & Karp (Commerce + Palantir) — the industrial stack

This session frames AI as an end-to-end industrial stack — compute, data, data management, model control, training — tied explicitly to national competitiveness and jobs. The signal for investors is policy certainty: faster permitting, incentives, and export-control clarity reduce long-duration policy risk on U.S. infrastructure and chips. Karp's presence underscores the government, defence and enterprise-operations layer — long contracts, high switching costs, and data platforms that let AI agents operate securely at scale — a distinct sleeve worth watching for public-private co-investment opportunities.

If you have time...