Microsoft and OpenAI CEO Sam Altman have a plan to bring nuclear fusion to the grid by 2028. We break down the strategy behind the bold bet, and what it has to do with ChatGPT.
Microsoft has signed a power purchase agreement with Helion Energy, a company backed by OpenAI and ChatGPT founder Sam Altman that specializes in nuclear fusion energy generation. The deal aims to connect the world's first commercial fusion generator to a power grid by 2028.
Firstly, the company plans to assemble its seventh prototype called Polaris in 2024 before it scales up the pilot reactor that will fulfill Microsoft’s power contract. The company’s sixth prototype, Trenta, was the first private fusion company to reach 100 million-degree plasma temperatures.
The plan is considered audacious due to the engineering challenges and scientific breakthroughs required to accomplish within a condensed time frame. However, the deal signals a critical need for additional public-private partnerships in the advancement of commercial-scale fusion power.
Nuclear fusion is a process where two light atomic nuclei (the center of an atom) combine to form a heavier nucleus, releasing a substantial amount of energy in the process. This process is responsible for the vast energy output observed in the sun and other stars.
Nuclear fusion works by bringing together light atomic nuclei under conditions of extreme temperature and pressure, similar to conditions present in the core of stars. These conditions overcome the electric repulsion between the positively charged nuclei, allowing them to fuse. When they combine, a heavier nucleus and one or more particles are produced, releasing a large amount of energy according to Einstein's equation E=mc^2.
Due to the high energy yield, nuclear fusion is being researched by companies like Helion Energy as a potential source of renewable energy on Earth, and as an alternative to nuclear fission.
Nuclear fission and nuclear fusion are two different types of nuclear reactions. While both processes release energy, they do so through fundamentally different mechanisms.
In nuclear fission, a neutron is slammed into another, larger atom, which causes the atom to split into two smaller nuclei. The result is the release of a large amount of energy.
On the other hand, nuclear fusion (which we just covered above) involves the combination of two light atomic nuclei. Instead of splitting, they combine; also releasing substantial energy.
Helion Energy is pursuing what’s called aneutronic fusion, a form of fusion that produces energy without generating neutrons (which are hard to contain and can damage reactors). The specific approach, called magneto-inertial fusion, is different from magnet-based tokamak or laser-based inertial confinement fusion. It uses tanks to heat up fuel into plasma on either end, shoots the plasma at each other at a million miles an hour, then uses magnets to compress it all.
By converting plasma energy into electricity directly, Helion's system could overcome the inefficiencies of tokamak or steam cycle fusion. However, the feasibility of maintaining control and stability in any kind of fusion reaction remains an area of deep skepticism.
To deliver a viable fusion generator, Helion will still have to overcome considerable hurdles, like achieving net-energy gain called "fusion ignition” (a nuclear fusion breakthrough accomplished by NIF in 2022, but not yet replicated) as well as producing commercial quantities of helium-3, a rare isotope necessary for Helion’s specific type of fusion.
In an interview with IEEE Spectrum, Helion’s founder and CEO, David Kirtley, said “The National Ignition Facility experiment last year proved key science in igniting a plasma for the first time… but in the process they threw away 99.9% of the input energy. We have proven our system can recover 95%, so we only lose about 5% of the energy that we put into the fuel. That means we have to do that much less fusion to reach net gain.”
For any company to create a commercial fusion pilot plant, three core engineering obstacles need to be overcome. First, the plant must produce more energy than it uses. Second, it needs to be cost-effective, avoiding the higher costs of large power units while producing reliable electricity. Third, the plant must maintain its power output long enough to prove its commercial viability.
Building this first fusion power plant will require substantial infrastructure. Helion is currently developing Polaris, its seventh-generation fusion device, which is expected to produce net electricity and prove the commercial viability of its approach when it comes online in 2024 (or shortly thereafter). A commercial facility is already in the works that will entail new construction, housing the reactor within a 30,000-square-foot building. Its exact location remains undisclosed.
The project, which will be licensed and regulated by the state Department of Health, also requires connection to the electrical grid to be managed by the clean energy company Constellation. Helion still requires design and construction approvals from the Nuclear Regulatory Commission (NRC) and local permits. Luckily, the NRC's recent decision to separate fusion regulation from fission could potentially speed up the timelines for license approvals.
Helion's lofty 2028 timeline reflects the market-driven behavior of companies in the field. There are now more than 50 fusion companies that have raised over $5B in private funding, plus an additional $117M in new government grants. Leading startups like Helion, CFS, and TAE have attracted the most substantial funding, with CFS and Helion alone raising a combined $2.3 billion in 2021.
Here's a (brief) summary of the other key players and their projects, with links to some of their latest breakthroughs:
Based on the number of startups entering the fusion race, we’re now at a potential handoff point from public investment to private. This is thanks to improved compute and machine learning, better magnets and materials, the diverse range of technical approaches that are finally becoming feasible, and dramatically increased capital availability for startups.
There is a strong incentive for fusion startups to set ambitious goals, as the influx of VC funding in the industry makes it hard to admit an inability to achieve fusion power. Former Rolls-Royce CEO, Warren East, recently joined the board of fusion energy startup Tokamak Energy, asserting that Tokamak Energy and other fusion startups will connect power to the grid by the 2030s. East emphasizes that engineering work, rather than scientific experimentation, will be the focus.
As Packy McCormick and Rahul Rana write for NotBoring, the odds of at least some companies hitting their targets in the early 2030s are considered promising, even despite the significant engineering challenges involved. Their success, however, will not only be measured by technical feats but also by their ability to produce cost-effective power that can compete with existing alternatives like solar and wind.
If Helion manages to become the first company to connect a fusion plant to the grid by 2028, then all bets are off. Despite this deal’s impressive nature, it could still turn out that governments ultimately pick the winner — just with an infusion of new capital unlocked by private market progress. This outcome would be akin to SpaceX’s impact on the space industry, which reinvigorated the space launch business through commercialization and private market incentives.
There could be an interesting correlation between the involvement of Sam Altman in this project and the timing of this announcement.
Both Microsoft and Altman have seen firsthand the massive power consumption driven by the demand for their popular large language model (LLM), ChatGPT. LLMs like ChatGPT and other generative AI products that make art and music require substantial computing power.
That’s because they use high-end graphical processing units (GPUs) from Nvidia and AMD, which are highly effective at resource allocation but consume significantly more power compared to typical CPUs. Therefore, it's been signified that generative AI produces more emissions than ordinary search engines due to the complexity of systems that use these energy-intensive neural network architectures.
As Manual G. Pascal writes for El Pais, training a natural language processing model creates emissions equivalent to five lifetime-running cars or 125 round-trip flights between Beijing and New York. According to Martin Bouchard, co-founder of digital infrastructure company QScale, a query based on generative AI could require 4-5x more computing power.
This power disparity poses a challenge for OpenAI, which is already limited by GPU availability. Since this lack of availability has already disrupted some of the company’s short-term plans, and the company’s power needs are on the rise, it's not so hard to believe that a moonshot bet on nuclear fusion could be part of a larger strategic pivot to lower the increasing energy requirements of GPU-driven server farms. Nvidia's new Grace Hopper "superchip" could help offset the availability issue, but it will increase the consumption issue as it demands substantial power; up to 1000W per chip.
Concerns have been raised regarding the environmental implications of this shift. Estimates predict that by 2040, the technology sector could be responsible for 14% of global emissions, and data center energy demand could increase 15x by 2030. These predictions might be outdated, as they both came out before the widespread use of generative AI.
While no concrete data has been released, it’s been estimated (somewhat cheekily) that to train GPT-3, OpenAI’s server farm could have released the same amount of CO2 emitted to drive by car to the moon and back. Projections suggest that OpenAI’s electricity consumption in January 2023 could match the annual consumption of 175,000 Danish households. As generative AI usage continues to spread, the equivalent electricity consumption could represent millions of people.
Regulatory measures are being prepared by the EU and the US to enforce energy efficiency and transparency in data centers. Efforts are also underway to manage and reduce the environmental impact of AI. Code Carbon, a tool developed by a team led by Turing Award recipient Yoshua Bengio, measures the carbon footprint of algorithm development and training. It encourages IT professionals to integrate it into their codes, helping them to make environmentally conscious programming decisions.
Much of AI’s carbon emissions come from the training phase, which is energy-intensive due to the massive number of examples needed to teach the algorithm. A study by OpenAI in 2018 warned that computing capacity for training the largest AI models doubles every 3-4 months. This is significantly faster than Moore's Law, which predicts the doubling of the number of transistors in an integrated circuit every two years. But reducing energy use during training, which requires sustained and intensive use, poses a more significant challenge.
The future may lie in reducing the complexity of the algorithms without compromising efficiency, leading to less polluting training processes. While some argue that shifting towards smaller, locally run LLMs could mitigate the power problem, OpenAI contends that despite their energy consumption, large models yield substantial benefits from increased scale. These large models offer generalization capabilities that smaller, specialized models cannot match (yet). Consequently, the demand for extensive GPU server farms will persist, even as AI progresses towards more specialized models.
The power usage of a data center varies widely based on its size, the number of servers it houses, and its efficiency measures. Small data centers, such as those within an organization, might consume power in the range of a few kilowatts (kW).
However, large-scale data centers operated by tech giants can consume significant power, often measured in megawatts (MW). On average, a typical data center can consume between 1 and 5 MW of power, but the US Department of Energy states that the largest data centers (those with 10,000+ devices) need 100MW or more. This energy is not only used for powering the servers but also for cooling systems, lighting, and other associated infrastructure.
Power consumption in itself isn’t always a carbon problem. The energy source and hardware used also contribute significantly to emissions. A study of the energy consumption of 95 AI models found that in regions where hydroelectric power is used, such as Quebec, carbon emissions are significantly less than areas relying on coal, gas, and other sources. Compare that then to China's data centers, which were predominantly powered by coal and produced at least 100 million tons of CO2 in 2018.
To its credit, China is implementing regulations to improve the power efficiency of its domestic data centers. This push for efficiency comes amid an expected 20% yearly growth in China's national compute fleet.
To offset further strain, Beijing has set a national Power Utilization Efficiency (PUE) target of 1.3 for most data centers, which is a metric that describes how much of the power used by data centers goes towards computing, storage, or networking equipment. The closer a PUE score is to 1.0, the higher the efficiency, which implies that less energy is used for non-computing functions like cooling and power distribution.
Recent regulations in China have also prompted the country to construct large-scale data centers in the cooler eastern region of the country, which generally has more abundant water resources than the arid north.
As a result of China’s regulations, there has been a significant increase in interest in liquid cooling technology across the country. Thermal management, including air conditioning, can account for up to 40% of a data center’s power consumption.
Liquid cooling reduces the need for these power-consuming thermal management systems by using liquid coolants to absorb and disperse the heat generated by computers. As Tobias Mann reported in The Register, analysts now expect liquid cooling technology to reach 19% of the data center thermal management market by 2026.
However, the use of liquid cooling in U.S. data centers could contribute to water scarcity, as many data centers are built in water-stressed areas like Arizona, which has 49 datacenters, and California, which has 239. As Shannon Osaka writes for The Washington Post, a single large data center can consume between 1 million to 5 million gallons of water per day due to water’s role in cooling data center equipment.
Interestingly, a study from April found that ChatGPT consumed approximately 185,000 gallons of water as its model was being trained; the researchers equate this to enough water to fill a traditional nuclear reactor cooling tower. Needless to say, the water component adds another environmental concern to the mix.
The short answer is, if nuclear fusion exists and is connected to the grid, it has the potential to solve any power consumption problem. The problem after the hard problem of creating a working fusion plant is then to get it connected to the grid.
Water consumption, however, is one area where Helion Energy’s approach to nuclear fusion could help (if it works). In general, nuclear fusion reactors, like the one Helion Energy is working on, tend to use water differently than traditional nuclear fission reactors. Traditional nuclear fission reactors use large quantities of water for cooling purposes as they split atoms. A typical fission reactor can consume between 270 and 670 gallons of water per MWh of electricity generated.
On the other hand, while fusion reactors still require cooling, water coolant has been avoided in conceptual fusion power plant design studies for over 25 years (at least in the U.S.) due to factors related to performance and safety. Therefore, it is plausible that a fusion reactor might require less water for cooling purposes, although the specific amount would depend on the details of the reactor design and the alternative cooling system used.
While certain fusion systems heat up water to generate steam that turns a turbine, Helion’s approach uses a pulsed fusion system to recover electricity directly, which eliminates the need for cooling towers, steam turbines, and their associated energy losses.
In addition to nuclear fusion, there’s one more moonshot technology that could help solve our growing datacenter demand problem. Quantum machine learning could provide a suitable alternative to the high-energy costs of other high-performance computers. That’s because quantum models express complex data structures using fewer parameters, potentially making them easier to train than classical AI models. This could allow for more efficient scaling of larger language models without a corresponding surge in resource requirements.
Google's quantum Sycamore processor is a potent example of this, as it can execute a quantum algorithm in seconds while consuming just 26 KW. While quantum technology remains in its infancy, the long-term goal is to achieve miniaturized, cost-efficient quantum devices. This shift would parallel the historical trajectory that saw room-sized classical computers shrink down to the size of today's smartphones.
It’s hard to say which technology is more likely to hit the market first, quantum computers or nuclear fusion. Many breakthroughs have supposedly sped up the roadmaps of IBM and Google’s efforts to build quantum computers, but the road to so-called “quantum supremacy” (when quantum tech surpasses classic computers) is still fuzzy.
Google has claimed it has reached the milestone (though its been disputed), and IBM just published a study in Nature that offered a demonstration of near-term quantum applications. IBM’s quantum computer was able to solve a complex problem that the best supercomputer approximation methods couldn’t handle. Instead, quantum creators are focusing on near-term "quantum advantage" that uses imperfect quantum computers available today to run short-term tests.
But it does seem like quantum tech is a lot farther along than fusion. IBM's advanced quantum roadmap has some meaningful milestones planned by 2026, and early adopters are already using partially working quantum computers for specific research tests to gain "quantum advantage" right now.
It may be that by 2028, when Helion Energy connects its pilot fusion reactor to the grid, we may already have a useful, (nearly) error-corrected quantum computer ready to help us solve the last remaining hard problems that arise. Although, don't be surprised if the company pushes back this ambitious timeline; the old joke with nuclear fusion is that it's always 30 years away.
Subscribe now for the latest trends, technology updates, and industry news.
