HPC News Bytes – 20260914

Doug Black (0:05)

Welcome to HPC News Bytes, a weekly show about important news in the world of supercomputing, AI, quantum computing, and other advanced technologies. Hi, everyone. Welcome to HPC News Bytes. I'm Doug Black, and with me is Shahin Khan. Shahin, if I could start off this week's conversation with an observation, it stems from an event I went to early last month, a pretty big data center conference in Northern Virginia. And the frequent comment we heard when people were comparing this year's conference to the previous year is that community issues related to data center, this is the growing resistance to data centers at the local, state, regional, and national levels, was the biggest change for the industry over the past 12 months. That resistance is only growing stronger. The latest development is that some of the most alarming AI warnings, with calls for AI controls and pauses, aren't just coming from AI doomers in academia, industry analysts and politicians, but from the industry itself.

As the Wall Street Journal stated in a recent headline, last week was the week everybody started talking about AI doomsday. We seem to have entered a new phase in the AI adventure we've been on for the last four years.

Okay, so getting into more specific industry news, the US Energy Information Administration, or EIA, expects electricity consumption to set records in both 2026 and 2027 with data centers a major contributor. But there's also a forecasting problem. Data center demand for new grid connections reportedly exceeds 700 gigawatts, more than 10 times current US data center requests.

As a result, some developers are applying for power at several possible sites before deciding where to build. Utilities are responding with deposits and other financial commitments that require developers to put real money behind a request. In some markets, projected demand has fallen substantially once speculative projects are screened out.

Shahin Khan (2:17)

It's hard to tell if the industry's warnings are motivated by the need to include all the downsides as they prepare to go public, or as a way of slowing down open source alternatives in a competitive move, or really only because of some new recognition that AI can cause harm, and crucially, whether the concern is shared globally.

But concern or not, like you were indicating, it doesn't look like anything is slowing down to the extent you can power it to begin with. The International Energy Agency projects data center electricity consumption through 2030, rising about 130% in the US, 170% in China, 70% in Europe, and 80% in Japan. China and the US alone account for nearly 80% of the projected increase. It is especially an acute problem because power availability is now a competitive advantage alongside accelerators and capital. Governments will have to wrestle with AI sovereignty also as you indicated, voter pushback, policy complexities and roadblocks, and infrastructure buildout starting with budgets, then power generation and distribution, securing the hardware, and developing indigenous AI models and applications. Many countries simply won't be able to build enough AI infrastructure, and just like many companies, we'll have to rely on access to cloud resources in other countries.

In this case, on-prem will mean domestic and cloud can mean offshore.

Doug Black (3:52)

The Tokamak is the most historically mature and heavily funded nuclear fusion design, using a donut-shaped magnetic field to trap plasma, though it suffers from engineering challenges like plasma instability. Private companies are racing to commercialize it using next-gen compact magnets, while simultaneously funding three major alternatives to bypass its flaws. One is called stellarators, which use twisted, ultra-complex external coils for a stable, continuous reaction. Another is inertial confinement, which uses lasers to implode fuel pellets, and magneto-inertial fusion, which uses colliding plasma rings for cheaper, smaller reactors.

Driven by AI datacenters, companies like Microsoft and Google have signed power purchase agreements with these private ventures, transforming fusion from an academic pursuit into a multi-billion dollar market race.

Researchers at the Princeton Plasma Physics Lab have developed PacMan, short for Prediction and Control Using Machine Learning. It's a modular AI control system for fusion experiments. They tested it in experiments on a large general atomics tokamak in San Diego. PacMan reads measurements from the plasma using AI models to predict its behavior and adjust the machine within safety limits. The loop runs about every 20 milliseconds. In one experiment, it predicted a potentially disruptive plasma instability about 200 milliseconds before it developed and changed the plasma conditions to avoid it.

Shahin Khan (5:41)

OpenAI, meanwhile, says an internal AI system has solved one of the Millennium Prize problems. These are seven of the most famous and difficult math problems, each with a million dollar prize for the first correct solution. It is organized by the Clay Mathematics Institute, a quiet philanthropic foundation that was founded in 1998, right next to Harvard University, before moving its operations to the University of Oxford in the UK in 2012, when a professor there became its president. The specific problem is known as the Navier-Stokes existence and smoothness problem. Referencing a set of equations, named after a French civil engineer and an Irish mathematician, which roughly speaking describe and predict how fluids move. Navier-Stokes are notoriously difficult to calculate. The problem asks whether these equations can break down mathematically, producing quantities that become infinite after starting with perfectly well-behaved conditions. So I see it fundamentally as a search for whatever pathological conditions you can devise to make the equations go haywire. And OpenAI says its AI has found such a condition and successfully proved the equations can in fact break down. We've talked about AI solving math problems a couple of times now, and I've been calling the impressive part of what AI does a perceptive shot in the dark, and then the logic to follow a good lead and synthesize the results. This sounds like the same type of thing. Now, nothing is suddenly changing in fluid dynamics R&D.