Summary: Axios reported on March 23, 2026 that Helion Energy was in advanced talks to sell OpenAI a guaranteed 12.5% of future output, potentially reaching 5 gigawatts by 2030 and 50 gigawatts by 2035. Sam Altman resigned as Helion board chair the same day and reportedly recused himself, though his equity stake remains a conflict concern. Helion has no commercial deliveries yet, so the deal is an early-stage bet on unproven fusion capacity.
Helion Energy, the fusion startup backed by OpenAI’s own chief executive, is in advanced talks to sell electricity to OpenAI. That is the report Axios published on March 23, 2026. Under the agreement taking shape, OpenAI would be assured 12.5% of Helion’s future production — roughly 5 gigawatts by 2030, growing to 50 gigawatts by 2035. On the very day the report appeared, Sam Altman stepped down from Helion’s board chairmanship.
This would not be one more data-center power contract. Commercial fusion has never before appeared in the AI industry’s decade-long procurement scripts; this would be a first. For scale, 50 gigawatts approaches half of the United States’ existing nuclear fleet and surpasses the total installed generating capacity of most countries. The talks remain early-stage, yet the figures carry the message by themselves: chips have stopped being the binding constraint on AI growth — electricity now is — and frontier companies will commit to power that does not yet exist.
What Would Be Bought: a 12.5% Share and a Ten-Year Buildout
Axios lays out the structure as a guaranteed share: of everything Helion produces, 12.5% would flow to OpenAI. TechCrunch translated that share into 5 GW by 2030 and 50 GW by 2035. Context puts the size in view: Helion’s existing customers number two — Microsoft, which signed a power purchase agreement in 2023 and expects deliveries to begin in 2028, and the steelmaker Nucor. Both commitments are demonstration-scale, in the hundred-megawatt class. What is being discussed with OpenAI stands two orders of magnitude above them.
The guaranteed-share structure is the part that deserves a second look. OpenAI would not be purchasing a fixed volume of electricity; it would hold a fixed slice of the output of every reactor Helion ever builds. In effect, AI’s demand for power becomes a long-term offtake contract for the fusion industry, with capacity risk and demand risk lashed together.
A Textbook Conflict, Handled in Two Layers
The entanglement is personal and early: Altman was among Helion’s first backers and remains one of its most prominent. The $425 million raise in early 2025 counted Altman himself, Mithril, Lightspeed, and SoftBank among its investors. A company buying 50 GW of power from a firm its own CEO put money into is a conflict of interest straight from the textbook.
The response came in two layers. First, Altman resigned the Helion board chairmanship; Helion CEO David Kirtley told TechCrunch the step “enables Helion and OpenAI to partner on future opportunities.” Second, reportedly, he recused himself from the negotiations from start to finish. The playbook is not new — a year earlier, he left the chairmanship of Oklo, a fission startup, in similarly framed terms ahead of AI-company partnerships. The critics’ objection stands: giving up a chairmanship leaves the equity stake exactly where it was. What has been removed is the institutional overlap, and only that.
Fusion’s Appeal, and the Reason for the Timing
Helion’s engineering route is a rarity among fusion startups. Where the conventional design heats a working fluid into steam to drive a turbine, Helion applies magnetic fields to convert the kinetic energy of charged fusion products directly into electricity. In principle, that yields smaller and cheaper plants — and it accounts for the modular 50 MW rating of each reactor.
The timetable tells its own story: in February 2026, Helion’s Polaris prototype reached 150 million degrees Celsius, closing in on the roughly 200 million degrees the company believes commercial operation requires. Even so, not a single commercial kilowatt-hour has been delivered yet. Meanwhile, the race between models keeps pushing compute and power demand toward their limits — a trajectory we flagged in our opening outlook for 2026 — and every lab is signing contracts for next-decade electricity well ahead of need.
Running the Numbers: About a Thousand Machines
Count at 50 MW per reactor: reaching 5 GW means roughly 100 machines built and grid-connected before 2030, and 50 GW means about 1,000. Measured against reality, Helion fields zero commercial units today, and the fusion industry has never sustained a production cadence even in the double digits, never mind four digits.
Even so, the arrangement pays off for both parties. Helion obtains an OpenAI offtake it can present to capital markets and supply chains as evidence that demand exists. OpenAI holds what is essentially a low-cost option: fusion arriving on schedule means clean baseload power no rival can bid away; a slipped timeline costs it mostly the priority, not the capital.
Three Takeaways for AI Infrastructure Teams
Three observations. First, procuring power is becoming a core function inside AI companies: gas and renewables projects, geothermal PPAs, now an equity share of fusion output — frontier firms reach across every generation technology. Second, schedule risk stays with you: entrusting a 2030 capacity plan to any single pre-commercial technology is a wager, and a diversified portfolio remains the only robust answer. Third, conflict governance will face much sharper scrutiny: once a CEO’s personal investments overlap company procurement, the resignation is where the work starts, not where it ends — transparent pricing and an independent audit complete it.