India's First Private Tokamak Is Live — and It's a Machine-Learning Problem

Pranos Fusion unveiled PRAGYA on September 3, 2026 — India's first privately developed tokamak, built in Bengaluru in roughly eight months. Co-founder and CEO Shaurya Kaushal showed the first close-up views of the machine, and the Indian tech feed did what it does with a homegrown hardware milestone.

The reflexive framing is "India joins the fusion race." That is true and not very useful. The more interesting framing is what Pranos says PRAGYA is for: a testbed for plasma control, high-temperature magnets, and diagnostics, planned to run about 3,000 plasma shots a year for 20 years.

Read that as a specification rather than a press line and it describes something specific — a machine whose primary output is 60,000 labelled experiments. Plasma control happens to be one of the very few physical-world domains where learned controllers have already beaten hand-tuned ones on real hardware. That is why this belongs on an AI blog.

TL;DR

What was actually built

PRAGYA is a low-aspect-ratio tokamak. Aspect ratio is major radius over minor radius: a conventional tokamak is a doughnut with an obvious hole, and a low-aspect-ratio machine squeezes that hole nearly shut — closer to a cored apple. The compact geometry can reach higher plasma pressure for a given magnetic field, which is why a number of private fusion efforts have picked it over the ITER-style route.

The numbers are modest by design:

A 0.1 Tesla field is small — orders of magnitude below what a power-producing machine needs. Nobody at Pranos is claiming otherwise. PRAGYA's job is to be cheap enough to run constantly, which is a materially different engineering goal from being big enough to break even, and arguably the smarter first move for a company with $6.8 million rather than $6.8 billion.

That funding round closed in March 2026 — about ₹63 crore, co-led by pi Ventures and Ankur Capital, with Industrial47 returning and angels including Groww co-founder Lalit Keshre, the founders of Razorpay, and Bhukhanwala Industries. The company is co-incubated at the Institute for Plasma Research and at JNCASR's Innovation and Development Centre, and collaborates with the Indian Institute of Science. For readers tracking India's deep-tech buildout, that pattern — small round, deep institutional partnerships, fast hardware — is the same one running through our India AI progress and top startups roundup.

Why this is a control-systems problem before it is a physics problem

Here is the part that makes PRAGYA relevant to people who will never touch a vacuum vessel.

A tokamak confines plasma with magnetic fields, and the plasma is violently unstable. Its shape and position have to be corrected on millisecond timescales by adjusting currents through many coils at once. Classically this is done with a stack of hand-designed controllers built on a linearised plasma model, tuned by physicists per machine and per configuration. It works, and it is enormously labour-intensive, and it does not transfer between machines.

In 2022, DeepMind and EPFL demonstrated reinforcement learning controlling the magnetic field of a real tokamak plasma, learning in simulation and transferring to hardware, sustaining plasma configurations that were difficult to hold with conventional control. It remains one of the cleanest published examples of deep RL beating engineered controllers on a real physical system with real safety constraints.

That result reframes what a shot-rate specification means:

• 3,000 shots/year is a data-generation rate.
• 20 years is 60,000 experiments on one machine with consistent instrumentation.
• A cheap, small machine is one you can afford to run into failure modes deliberately — which is where the informative data lives.

There is a reason the company's other named programme is JENGA, an integrated tokamak digital twin platform (alongside MAGGA for high-temperature superconducting magnets). A digital twin is the simulator half of exactly the sim-to-real loop that made the DeepMind result work. You train the controller in the twin, you transfer it to the machine, the machine's shots correct the twin, and you repeat. The physical tokamak and the simulator are two halves of one system, and Pranos is building both.

If you want the general version of that loop — learn in simulation, transfer to reality, use the real system's data to correct the simulation — it is the same structure as the RL post-training stacks we compared in open-source RL-as-a-service, with a plasma standing in for a language model and shot count standing in for rollouts.

The magnet programme is the other half

Pranos's second named programme, MAGGA, targets high-temperature superconducting (HTS) magnets — and it is worth understanding why a fusion company builds a magnet supply chain rather than buying one.

Confinement scales steeply with magnetic field strength, which is the entire reason HTS tape changed private fusion economics: a stronger field lets a much smaller machine reach conditions that previously required an ITER-sized one. The catch is that HTS tape is a constrained global supply with a short list of qualified manufacturers, and a company that cannot source it cannot build the next machine regardless of how good its physics is.

Building that capability domestically alongside the tokamak is a supply-chain bet as much as a technical one, and it is the same reasoning behind India's semiconductor and space-component pushes: the scarce input is not the design, it is the ability to make the part.

What this means for what you build or pay

Being direct, because it would be easy to write this post as pure national-interest cheerleading and leave the reader with nothing.

On electricity: nothing changes this decade. AI's power problem is a 2020s problem — grid interconnect queues, gas turbines with multi-year lead times, water for cooling, the constraints we worked through in the real environmental impact of data centres and California's data centre bills. The most aggressive private fusion roadmaps, Pranos's own 2030 net-energy ambition included, put demonstration in the early 2030s and grid contribution well after. Anyone selling fusion as an answer to your 2027 capacity plan is selling something.

On careers: this is where ML meets instrumentation. Plasma control, diagnostics, and digital twins need people who can do both reinforcement learning and real-time systems on physical hardware with hard safety limits. That intersection is thin globally and thinner in India, and a domestic machine running 3,000 shots a year is where it gets trained. If you have been looking for a way into applied RL that is not another chatbot, control problems on real instruments are the underrated door.

On method: the shot-rate insight generalises. Pranos's bet is that a small machine you can run constantly beats a large one you run rarely, because iteration count dominates. That is the same bet behind fast evals, small models, and short feedback loops in software — and it is usually right.

Honest limitations

• First plasma is a milestone, not a result. No confinement time, no plasma parameters, and no peer-reviewed performance data have been published. What exists so far is a working machine and photographs of it.
• We have not verified the "one of 16 tokamaks in the world" claim that circulated with the launch. Depending on whether you count operating, historical, or currently-fielded machines, the number varies a lot. Treat it as first-hand enthusiasm rather than a census.
• The 2030 net-energy target is an ambition, and fusion timelines have a well-documented history of slipping. Nothing about an eight-month 0.1 T build validates a 2030 date.
• The machine-learning angle here is ours, not an announcement. Pranos has stated that PRAGYA is a plasma-control testbed and that it builds a digital twin platform; it has not published details of learned controllers. The DeepMind/EPFL precedent is why the framing is reasonable, not evidence of what Pranos is doing.
• Fusion is adjacent to explainx.ai's usual beat. We are covering it for the control-systems and India deep-tech angles, not as energy-market analysis, and the electricity section above is deliberately deflationary.

Related on explainx.ai

• India AI progress 2026 — the top 15 startups
• Open-source RL-as-a-service — the post-training stacks compared
• The real environmental impact of data centres — water and electricity
• California's data centre and AI infrastructure bills
• Physical superintelligence — PSI's $58M seed
• Top AI instructors and trainers in India, 2026
• What are AI agents — the complete beginner's guide

Machine parameters, funding details, and institutional affiliations reflect Pranos Fusion's public materials and coverage as of September 3, 2026. PRAGYA's performance data has not been published; the reinforcement-learning framing in this post is our own analysis based on the stated purpose of the machine and prior published work on learned tokamak control, not a claim about Pranos's methods.