Fusion Data Centers: the 2028 Microsoft Deadline and What the Fusion Race Can Deliver

Fusion data centers are no longer a slide — the deadline is contractual: Microsoft has purchased 50 megawatts of fusion electricity from Helion’s Orion plant, targeted for 2028, and this week alone the United States government committed up to $175 million in CHIPS funding to Helion’s semiconductor work (with a fresh $50 million private round pushing the company past $3.2 billion raised) — while rival Commonwealth Fusion Systems signed its own $85 million letter of intent and GaN specialist Vertical Semiconductor won $73 million to build power electronics for AI racks. The AI power race that already drove our nuclear scoreboard and the copper analysis now has a third frontier lane — and this guide measures it honestly: how fusion actually produces electricity, why the Microsoft PPA matters even though Orion delivers into the grid (not next to a campus), what the verified milestones in the US and China actually are, and the timeline table that separates company targets from engineering reality.

The discipline throughout: every figure is dated and attributed, company targets are labeled as targets, and the physics is in plain language — because the fusion story’s history is exactly a story of mistaking demonstrations for deliveries.

Key Takeaway

• ⚡ The one date that matters for fusion data centers: Microsoft’s 2028 PPA for 50MW from Helion’s Orion — the first grid-connected fusion data center power purchase by a tech company; Orion is under construction in Washington state and the deadline is a company target, not a certainty.
• 🇺🇸 This week’s money moved the needle: up to $175M (CHIPS, 3-year) to Helion for high-voltage power semiconductors + $50M private; up to $85M to Commonwealth Fusion for plasma-heating and rad-hard electronics; up to $73M to Vertical Semiconductor for vertical GaN power devices aimed at AI data-center racks.
• 🔥 The physics in one line: fuse hydrogen isotopes at over 100 million degrees; the energy released becomes electricity — in Helion’s design without a steam turbine, via direct energy conversion whose efficiency is decided by power semiconductors (“cut the loss in half — that’s twice the power output,” per the CEO).
• 🇨🇳 China runs the state-scale program: EAST’s plasma-density record (1.3-1.65× past a decades-old theoretical limit, published in Science Advances) and a national fusion company seeded at $2.1B with ~$3B/year estimated spending — the strategic-race lane alongside the AI-compute race.
• 🕒 The honest timeline: Microsoft PPA 2028 (target); first commercial fusion plants 2030-2035 (analyst range); grid-scale pilots early-to-mid 2030s; ITER’s deuterium-tritium operations 2034 — fusion is a 2030s power source for AI, arriving after the gigawatt buildouts this decade need.

What a fusion data center actually means (term first, headline second)

A fusion data center is shorthand for a fusion-powered AI campus — and the live deal structure matters: Microsoft does not buy a reactor, it buys 50 megawatts of fusion ELECTRICITY delivered to the grid by 2028 under the 2023-announced power purchase agreement, which the company then draws for its workloads. The distinction is structural: fusion competes as a power-market entrant (like the nuclear PPAs in our nuclear data centers scoreboard), and every campus that runs on it still needs the grid, the copper and the cooling this series tracks. No fusion data center with an on-site plant exists; nobody credible is promising one this decade. The honest definition the market uses: a fusion data center is an AI campus whose clean-firm power layer comes from a fusion plant with a signed delivery date.

Why the fusion data center earns a lane at all: it is the only firm, carbon-free power source whose fuel supply (deuterium/tritium isotopes) is effectively unlimited and whose waste profile is minimal — the theoretical endgame for the gigawatt-scale AI power problem. The practical question the rest of this guide answers: when and in what form.

How Helion’s design gets electricity without a turbine

Most fusion concepts heat plasma and capture heat indirectly — boil water, drive a steam turbine, generate electricity. Helion’s field-reversed configuration (FRC) pulses plasma to over 100 million degrees in an hourglass-shaped chamber, compresses it with magnetic fields until fusion occurs, and captures energy DIRECTLY as the expanding plasma pushes back against the magnetic field — an induction pulse, not a steam cycle. The commercial hinge of that design is electrical: how much of the recovered energy survives the power electronics. “If we can cut that loss in half, that’s actually twice the power output of the fusion system,” Helion’s CEO told reporters this week — which is precisely why the CHIPS award funds high-voltage semiconductor manufacturing (an automated module, packaging and test line in Washington State, plus research with GlobalFoundries in Vermont) rather than plasma physics alone.

That semiconductor logic connects the two frontier lanes the chip-industry press tracked this week: Helion’s $175M award for high-voltage power devices, Commonwealth Fusion’s $85M letter of intent (semiconductors for radio-frequency plasma heating, power electronics, diagnostics and radiation-hardened electronics), and — the AI tie in the award list itself — Vertical Semiconductor’s $73M for vertical GaN power transistors explicitly aimed at power density in AI data-center racks. The same power-electronics engineering discipline runs through fusion plants and AI campuses; the money is now saying so out loud.

The verified fusion scoreboard: US companies and China, dated

The honest reading of the table: the US private lane races to sell power (Helion’s PPA deadline, CFS’s net-energy demo); China’s state lane races to own the science baseline (EAST records, national companies); and the semiconductor programs this week are the quiet tell that BOTH lanes’ bottlenecks have shifted from plasma to power electronics — the same components AI racks compete for.

The timeline table: what happens when (and with what confidence)

The fusion data center fit follows the timeline: AI’s 2026-2030 gigawatt buildouts ride nuclear restarts, grid expansion and the efficiency ladder (our power pipeline and copper analyses); fusion arrives as the clean-firm layer of the 2030s — one reason the PPA now (2023) and delivery later (2028) structure matters to every buyer pricing long-dated power. The risk the timeline carries is the field’s own history: demonstration-to-delivery gaps have swallowed fusion budgets before, which is why the contract structure (buy the electricity, not the reactor) is the honest hedge.

WorldNgayon Insight: the semiconductors under the fusion race

What changed: the week’s awards moved the fusion data center’s bottleneck into plain sight — power semiconductors — and put government money behind hardware the AI buildout ALSO needs (the vertical-GaN award names AI racks explicitly). Why now: Helion crossed the D-T demonstration line in February, the Microsoft deadline moved from talking point to construction schedule, and the CHIPS window opened for high-voltage devices. Why it lasts beyond the news cycle: the dated table above runs to 2034+ with checkable rungs — Orion’s delivery, SPARC’s net-energy demo, EAST-class records, ITER’s D-T operations — every one a named update trigger. Who is affected: AI campus planners pricing the 2030s power layer, the semiconductor industry (new high-voltage capacity), fusion’s investor class, and energy ministries watching China’s state lane. What to do next: treat fusion PPAs the way this series treats nuclear PPAs — read the delivery date, the counterparty and the fallback. Risks, opportunities, costs, skills: the risk is the 2028 target slipping with technology risk unresolved; the opportunity is clean-firm supply for the AI century plus a US high-voltage semiconductor base; the cost is decade-scale and capital-heavy; the skills gap runs through power electronics, plasma engineering and high-voltage manufacturing. Connection to work and money: the fusion lane is the newest spoke on the sovereign AI cornerstone — nations that own the power stack will own the compute stack.

For Filipino readers: the energy lesson and the honest distance

The Philippines will not host a fusion data center this decade, and pretending otherwise would be the mistake this guide exists to avoid. What Filipino readers can take from the race is planning arithmetic: the country’s data-center corridor (tracked in our SEA infrastructure scoreboard) runs on grid expansion, gas and renewables this decade; the fusion question arrives — if it arrives — as one more firm-power option when regional utilities price the 2030s. The OFW lane is more concrete: high-voltage power electronics, plasma engineering and semiconductor manufacturing are hiring worldwide (the awards above are contracts with real payrolls), and Filipino engineers already sit inside the global semiconductor and energy workforces that these programs staff. The household-level takeaway is the same one our energy-cost coverage has carried: firm clean power is the binding constraint of the AI century, and every country that solves it — by reactor, by fusion, by transmission — prices its own AI future accordingly.

What is a fusion data center?

An AI data campus whose firm, clean power comes from a fusion power plant under contract — today structured as grid-delivered power purchases (Microsoft bought 50MW from Helion’s Orion for 2028), not on-site reactors. The term describes the power source, not the cooling or chips; the campus still needs everything the rest of the infrastructure series covers.

When will fusion power AI data centers?

The honest layers: Microsoft’s 50MW PPA target is 2028 (a company target, medium confidence); first commercial fusion plants are projected 2030-2035 with grid pilots in the early-to-mid 2030s — fusion becomes a meaningful AI-power layer in the 2030s, after this decade’s buildouts rely on nuclear restarts, gas and renewables.

How does Helion’s fusion design make electricity?

It pulses plasma to over 100 million degrees in a field-reversed configuration, compresses it magnetically until fusion occurs, and captures the expanding plasma’s energy directly as electricity — no steam turbine. The design’s commercial hinge is power-electronics efficiency: halving the conversion loss doubles net output, which is why the company’s CHIPS award funds high-voltage semiconductor manufacturing.

Why did the US award $175 million to Helion?

The three-year CHIPS award funds research, development and manufacturing of high-voltage power semiconductors for fusion systems — including an automated module, packaging and test line in Washington State, the first US line for that semiconductor class, plus research with GlobalFoundries. The devices also serve broader industry (manufacturing, wastewater treatment, medicine, defense), and the award signals federal interest in the power-electronics layer fusion and AI data centers share.

Is fusion energy actually close to commercial reality?

More than a decade ago in demonstration, still short in delivery: measurable D-T fusion reactions are demonstrated (Helion, February 2026), plasma-density records are crossed (China’s EAST), and net-energy demos are the current frontier (CFS’s SPARC) — but analyst confidence puts first commercial plants at 2030-2035, and the field’s history of demonstration-to-delivery gaps justifies the caution.

What is China doing in fusion?

Running the state-scale program: the EAST “artificial sun” tokamak published a stable-plasma result at 1.3-1.65 times the Greenwald density limit (Science Advances, January 2026) — a decades-old theoretical barrier crossed — while the national fusion company CFEC launched with $2.1 billion (2025) and the country’s annual fusion spending is estimated around $3 billion, the strategic-scale competitor to America’s startup-led lane.

Do fusion data centers help the grid like nuclear plants do?

In ambition, yes — firm, carbon-free baseload-class power with minimal fuel and waste constraints; in practice, not yet: no fusion data center operates commercially at grid scale today. The live comparison runs through the PPAs — nuclear restarts deliver megawatts this decade (our nuclear data centers scoreboard), fusion targets the late 2020s-2030s window — so the grid impact depends on which deadlines hold.

What is the Microsoft fusion deal?

A power purchase agreement announced in 2023: Microsoft buys 50 megawatts of electricity from Helion’s first commercial plant, Orion, targeted to begin delivering in 2028 — the first fusion PPA signed by a tech company. The deal made Orion’s schedule a market event and made the 2028 date the fusion industry’s most-watched deadline.

Financial Disclaimer

This article is editorial analysis of publicly announced contracts, awards and research milestones, not energy, investment or procurement advice, and no return or outcome is promised. Targets (2028 delivery, 2030-2035 commercialization) are company or analyst projections with stated confidence levels. Fusion, semiconductor and energy-industry positions involve long-duration technology and execution risk; verify current terms with the named companies and agencies before any financial decision.

Sources: TechCrunch — Helion’s $175M CHIPS award, the $50M round, and the Orion-2028 Microsoft delivery target (with CEO quotes); SemiEngineering — the week’s CHIPS awards table (Helion $175M, CFS $85M, Vertical Semiconductor $73M GaN for AI racks, Sibelco $56M). Additional context cited without links: the mofa fusion-2026 compilation (EAST’s Greenwald-limit result in Science Advances; the Helion D-T demonstration of February 2026; CFEC funding; the milestone confidence table); Livescience’s Petal coverage for the parallel trans-Atlantic infrastructure lane. Company delivery dates are targets; investor figures are as announced.