Fusion was 'always 50 years away.' AI's power appetite may have just changed that

The 50-megawatt fusion reactor is not a grid infrastructure project. It is a power purchase agreement, sized for exactly one hyperscale AI data center, and the first one was signed: Microsoft agreed in May 2023 to buy electricity from Helion starting in 2028, with Constellation acting as power marketer. CNBC and Reuters independently confirmed the same terms, and Orrick's transactional summary confirmed the structure. That off-take commitment: a named buyer, a defined unit of deployment, a 50 MW or larger target, is the hinge that prior fusion cycles never had.

The reason the unit of sale is suddenly 50 MW rather than 500 MW or 5 GW sits in a single piece of physics hardware, the high-temperature superconducting (HTS) magnet. In 2021, MIT's plasma science group demonstrated an HTS magnet with field strength sufficient to confine burning plasma at a fraction of the size previously required, reducing reactor scale to 1/40 of ITER. That technology, per 36kr, underwrote Commonwealth Fusion Systems (CFS), which in 2021 closed the largest private fusion round on record at roughly USD 1.8 billion. The 36kr anchor piece frames this as the moment private fusion stopped asking for a national-laboratory budget and started asking for a venture round. Once a reactor can fit on a factory floor and serve a single industrial customer, the off-take math starts to look more like a long-duration software contract than a 30-year utility bond.

AI is what turned that math into a bidding war. Goldman Sachs, cited by 36kr, projects global data-center electricity demand will roughly 2.75x from 2023 to 2030, growing from about 55 GW today to roughly 150 GW. The buyers that need that power do not want renewable-plus-storage intermittency or nuclear fission's decade-long licensing tail. They want 24/7 carbon-free baseload, and they are willing to sign multi-year contracts to lock it in before the price goes up. A hyperscaler buying 50 MW of fusion at a known price, for delivery starting 2028, is a hedge against being the AI lab that has models but no megawatts.

That hedge is now being priced in. According to 36kr/Yingke, Chinese fusion startups raised roughly RMB 3 billion in the first third of 2026 alone, with several rounds still closing. The National Energy Administration has stood up a RMB 20 billion fusion industry fund, and Shanghai's future-industries allocation has been expanded to RMB 15 billion. Add it up and Beijing-aligned capital has committed roughly RMB 35 billion to fusion across separate commitments in early-to-mid 2026, before counting any dollar-denominated rounds from US-based players like Helion, CFS, or TAE.

The froth is also showing up in the line items. 36kr profiles Nova Fusion (诺瓦聚变), founded in March 2025, which reportedly raised about RMB 1.2 billion in its first year and is now approaching a RMB 10 billion valuation. Founder Guo Houyang frames the company's 50 MW small-fusion reactor as sized for a single AI supercomputing center, which is exactly the unit Microsoft pre-bought from Helion. The symmetry is the point: the customer pitch is now "we sell to one data center, not one country."

That symmetry is also what is making some investors nervous. The same 36kr piece carries the bubble signals. VCs who rejected fusion deals at internal investment committees a year ago now watch their peers mark up portfolio companies. One CFO, quoted anonymously by 36kr, says a RMB 10 billion valuation was a three-year target and arrived in six months. A founder described in the article declined to be interviewed; another CEO, also quoted anonymously by 36kr, admits that companies whose "cold-bench discipline" was unfashionable two years ago are now the ones getting term sheets. And 36kr reports that at least one recent round included terms so punitive that one limited partner, quoted by 36kr, called them "丧权辱国的条款," provisions that, even on a successful exit, would not return capital to the early funds.

The skepticism is not just posture. Fusion's long-running joke, that it is "always 50 years away," earned its punchline for a reason. Helion's own newsroom language on the Microsoft PPA is careful: 50 MW or greater commercial operation in 2028, after a one-year ramp from first electricity. Orrick's deal alert matches that framing. Three years out from a target is not delivery, and HTS-magnet confinement in a commercial-grade compact device has not yet produced net energy gain. The 36kr piece itself uses the "always 50 years away" line, signaling that even the bullish Chinese coverage knows it is running ahead of the physics.

The political layer is what makes the timeline unforgiving. China's 15th Five-Year Plan recommendation, per 36kr, issued in November 2025, lists fusion alongside quantum and embodied AI as a future industry, which is the bureaucratic precondition for state fund flow and pilot-project siting. Fusion has been folded into the AI-power stack the same way grid transmission and advanced packaging have been, as critical infrastructure for compute. Whoever locks in compact fusion baseload first may set the ceiling for everyone else's AI scale, because the off-take contract is what makes the next reactor factory financeable.

What to watch next: any Helion 2027 commissioning update, the next Chinese state-fund fusion allocation after the RMB 35 billion already deployed, and whether any non-Chinese, non-US sovereign fund, from the Gulf states, Korea, or Japan, writes a fusion off-take of its own. The capital has decided that the bottleneck is real. The physics still has to deliver.