How to Stop the AI Oligarchy Before It's Too Late: Why "Good AGI" Demands We Break Big Tech's Grip on the Grid in 2026
The AI safety movement has a blind spot the size of a power grid.
For a decade, the smartest people in the field have asked how to align artificial intelligence with human values. It is the wrong question — or rather, a dangerously incomplete one. A perfectly aligned AGI owned by three companies and two sovereign funds is not a safe AGI. It is a loaded weapon pointed at democracy, held by whoever signs the checks.
On April 14, 2026, that is the trajectory we are on. And the window to change it is measured in quarters, not decades.
The facts of concentration are no longer in dispute. A handful of hyperscalers ration the GPUs. Nvidia sells 80–95% of them. Frontier labs raise sovereign-scale capital — OpenAI's $122 billion round in March, Anthropic's $380 billion post-money valuation in February — and lock in multi-gigawatt power contracts that foreclose the grid to everyone else. When Anthropic's Claude Mythos model proved capable of autonomously surfacing zero-day exploits, the company shared it with roughly 40–50 "trusted" firms through Project Glasswing. No democratic body decided who qualified. That is the template: safety redefined as selective access for incumbents.
This is the part most analyses stop at. But cataloging the problem is not a plan. Here is what actually has to happen if we want an AGI transition that does not end in oligarchy.
First, stop treating alignment as the whole game. Technical safety research — making models truthful, controllable, non-deceptive — is necessary. It is also politically naive when pursued in isolation. A model aligned to its principal is only as good as the principal. If the principals are boards, investors, and national security establishments operating behind NDA walls, "aligned AGI" means AGI aligned to them. The alignment community needs a structural turn: the question is not only can we make this model safe, but who gets to deploy it, withhold it, and audit it. Those are governance questions, and they are being answered by default every day we don't answer them deliberately.
Second, treat compute as infrastructure, not product. Electricity, water, and spectrum are all regulated as essential facilities because a functioning society cannot let private actors ration them at will. AI compute has crossed that threshold. When a single inference burns up to 1,000 times the energy of a web search, when U.S. data-center demand is projected at 70–120 gigawatts by the early 2030s, when 40% of planned capacity may be blocked by power shortages — this is not a consumer market. Competition authorities should mandate transparent GPU allocation, nondiscriminatory cloud access for public-interest research, and an end to the preferential treatment hyperscalers extend to firms they've invested in. India's IndiaAI common-compute grid, pooling tens of thousands of GPUs for smaller players, is the prototype. It needs regional and multilateral equivalents, funded at the scale of highway systems.
Third, build the data and model commons now, while it is still legal and possible. Open-weight ecosystems — Llama, Mistral, DeepSeek, Qwen — are imperfect, sometimes released by actors with their own agendas, and not automatically democratic. But they are the only part of the stack currently resisting full enclosure. Governments, universities, and foundations should be funding governed open models with interpretability and misuse safeguards as first-class features, and data commons with transparent governance and real representation from the communities whose information trains them. Every month without this, the default hardens.
Fourth, rewrite the safety-as-secrecy script. Glasswing-style coalitions will keep appearing, and some variant of restricted access for genuinely dangerous capabilities is defensible. What is not defensible is letting the frontier labs alone decide who is "trusted." Independent, internationally chartered bodies should govern access to the most dangerous capabilities — with public audit rights, whistleblower protections, and mandatory vulnerability disclosure to the broader ecosystem, not just coalition members. Safety that flows only to incumbents is not safety. It is a moat with a compliance department.
Fifth, name the politics. The current AI trajectory requires "the unraveling of core social, political, and economic fabrics" to function at scale. Entry-level developer jobs are already down roughly 20%. The productivity gains are flowing to capital, not labor. Presenting this as inevitable — as the natural physics of a technology rather than the chosen strategy of a few hundred executives — is itself a political act. Unions, antitrust movements, climate coalitions, and democracy advocates need to stop treating AI as a separate issue. It is the substrate on which all their other fights will be won or lost.
None of this requires halting AI development. The false choice between "accelerate" and "pause" has consumed too much oxygen. The real choice is between an AGI transition governed by a handful of capital constellations and one governed by plural, contestable, accountable institutions. The first produces bad AGI even if every model is technically aligned. The second produces something worth the name.
The grid is being built right now. The PPAs are being signed right now. The coalitions are being formed right now. Every week of delay in public investment, antitrust action, and international governance is a week in which the private stack hardens into permanence.
Good AGI is not a model property. It is an institutional achievement. We are not going to stumble into it by hoping founders stay benevolent or alignment researchers stay brilliant. We get it by building the counter-infrastructure — public compute, open models, data commons, democratic oversight — before the window closes.
It is closing. But it is not closed.
That distinction is the whole fight.
not investment advice
Sources: Overview page (with links to full PDF and sections) https://hai.stanford.edu/ai-index/2025-ai-index-report
CNBC: round closing, $122B raised, valuation range, investor mix https://www.cnbc.com/2026/03/31/openai-funding-round-ipo.html
Official: $30B Series G at $380B post‑money https://www.anthropic.com/news/anthropic-raises-30-billion-series-g-funding-380-billion-post-money-valuation