AGI by 2030 Is Constrained by More Than Models—It May Run Out of Electricity
Can AGI arrive by 2030 if compute runs into electricity, energy infrastructure, and the cost of scale?
Assess the path to AGI not only through model progress, but also through available power, compute, infrastructure lead times, and the cost of scale.
What to watch for
Key takeaways
The boundary of the “Today on the ToTheMoon podcast” case is defined by this point: behind the forecast is scaling logic — give the model far more compute and the ability to learn from its own results — yet AGI remains a concept with no single measurement, and an exact date easily becomes a myth.
The boundary of the “OpenAI is no longer the leader of AI? Why are people leaving?” case is defined by this point: John Schulman's move from OpenAI to Anthropic matters as more than a personnel story: when a co-founder and key reinforcement-learning researcher changes teams, the industry stops treating OpenAI's leadership as unconditional.
The “What will AI be capable of by 2030?” topic becomes clearer once this point is included: by strong AI the labs mean a system that improves itself and does almost all tasks better than almost all people: that takes roughly a hundred times more compute than GPT-4 — about two generations of data-center construction.
In the context of “Why 2030?,” this criterion applies: the arithmetic is simple: GPT-4 trained on 25,000 H100s, the new clusters run 100,000 — six to eight times the compute, and with a year-long cycle instead of a hundred days, 15–20 times; the next clusters, including Musk's build in Tennessee, cover the rest of the way to 2030.
The discussion of “Is 2030 realistic or not? What has to happen for it?” yields a practical test: the hosts call AGI a "mythical story" — it may arrive earlier or not arrive in our lifetime; the one-year horizon is more reliable: chatbots and search improved dramatically over the year, a change visible without any abstractions.
The “Strong AI needs twice the electricity that exists today” issue should be assessed with one constraint in mind: Trump told Musk the country needs twice its entire electricity output "just for AI": even if the number is inflated, it shows the scale — the U.S. will have to build plants and grids, and the model race becomes a race for gigawatts.
The boundary of the “When will the problem of the AI hallucinations be solved?” case is defined by this point: ChatGPT can mix up university rankings or confidently state false information, and as long as the system needs constant verification it is hard to call it universal intelligence — no matter how many parameters or how much energy stand behind the answer.
The working conclusion from “Is China's energy sector ahead of the whole planet?” is that while the U.S. has yet to build the plants and grids for AI, China is already expanding its energy base aggressively — a head start in the race for gigawatts that model quality alone cannot offset.
In the context of “The U.S. antitrust commission vs. Google,” this criterion applies: the antitrust case against Google and the EU's questions for Apple, OpenAI, and xAI show that technological scale inevitably turns political: the rules governing how the largest players may use their power shape the market no less than the products do.
The “Pixel is a really good phone” scene leads to a working conclusion: Pixel may be an excellent phone and Amazon may put AI into commerce, but the future market will not be decided by products alone: it will be decided by energy, people, access to data, and the rules for the biggest players.
What this episode is about
Researchers identify 2030 as a possible point for the arrival of powerful AI, but the next scale requires tens of times more compute and a new energy base. At the same time, OpenAI is losing key people, Google faces antitrust pressure, and hallucinations remain an unresolved everyday problem.
The year 2030 sounds convincing because researchers and leaders of the largest companies keep repeating it. Behind the forecast is a simple scaling logic: give a model far more compute and the ability to learn from its own results, and it may advance to a new level of tasks.
But AGI still has no single accepted measurement, so an exact date can easily become a myth.
Even if the algorithmic part works, the physical system must withstand the growth. The next generation of AI is discussed in terms of energy volumes comparable with a huge share of a country's entire electricity generation. The United States will need to build power plants and grids; China is already expanding its energy base aggressively. The model race is becoming a race for gigawatts.
Talent provides another signal. John Schulman's move from OpenAI to Anthropic matters as more than another corporate personnel story. A co-founder and key reinforcement-learning researcher changes teams, which means OpenAI's leadership is no longer treated as unconditional inside the industry.
At the same time, ordinary users still encounter hallucinations on simple tasks. ChatGPT can mix up university rankings or confidently provide false information. As long as the system requires constant verification, it is difficult to call it universal intelligence, regardless of how many parameters or how much energy stand behind the answer.
The antitrust case against Google and the European Union's questions for Apple, OpenAI, and xAI show that technological scale inevitably becomes political. Pixel may be an excellent phone and Amazon may integrate AI into commerce, but the future market will not be determined by products alone. It will be determined by energy, people, access to data, and the rules governing how the largest players are allowed to use their power.
AGI is constrained by more than algorithms. Energy, talent, data, and the rules governing access to infrastructure may determine the pace before the next model breakthrough does.
Episode transcript
The episode is in Russian; below is an English reading guide to the transcript (the full EN transcript is a machine translation). Voice matching applied to 45 segments: 27 identified, 3 mixed, 14 probable, and 1 unresolved.
Read transcript on a separate page
Loading…