Fired OpenAI Scientist Published 165 Pages About Our Future and Bet 45 Billion on It
A former OpenAI scientist's 165-page forecast predicted AI data centers, power shortages, and superintelligence. We check his score against 2026 reality, and his fund's
Published November 2, 2026
In July 2024, a 22-year-old former OpenAI researcher, Leopold Aschenbrenner, published a 165-page document titled Situational Awareness. The paper forecast the scale of AI infrastructure, the constraints on power supply, the geopolitical contest with China, and the timeline for machines that could perform AI research. Aschenbrenner was fired from OpenAI in April 2024. By mid-2026, he had also built an investment fund around the paper's thesis. This article reviews his forecasts against reported developments through November 2026.
Background
- Aschenbrenner is German and the son of two doctors.
- At 15, he left to study in the United States at Columbia University.
- At 19, he graduated at the top of his class.
- He then joined OpenAI, where he worked on methods to keep future superintelligence under control.
In 2023, OpenAI suffered a security breach. Aschenbrenner wrote a memo to the board of directors. It stated that the company's security was weak and that it would not be able to stop Chinese intelligence services from obtaining its developments.
According to Aschenbrenner, human resources issued him a warning. In April 2024, OpenAI fired him. The company stated the reason was leaking information. Aschenbrenner says the company classified a routine draft as a leak after he shared it with three external researchers for feedback. He states that the real reason was the memo on espionage. OpenAI disputes this account.
The Forecast
The paper's opening line is: "the future can be seen in San Francisco." Aschenbrenner argues that a few hundred people working at leading AI companies see developments that the rest of the world does not.
Aschenbrenner does not state that artificial general intelligence (AGI) will certainly arrive by 2027. He describes this outcome as highly plausible. His paper presents it as a wager, not a guarantee.
His core predictions were:
- Compute race: Spending on compute clusters would grow from tens of billions of dollars to hundreds of billions, and a cluster costing one trillion dollars would appear by 2030.
- Electricity bottleneck: Power supply would become the main constraint on AI development.
- Espionage: China would steal American AI technology from laboratories.
- Revenue: By mid-2026, one major company would earn $100 billion per year from AI.
- Remote employees: By 2027, AI would function as a remote employee that works autonomously for weeks after connection to email and work chat.
- Automated research: AI would begin performing the work of AI researchers, after which progress would accelerate.
- Superintelligence: Superintelligence would arrive by the end of the decade.
Aschenbrenner also predicted a government-led "Project," comparable to the Manhattan Project, in which the state would control superintelligence development.
Scorecard: Forecast Versus Reported Developments
| # | Prediction | Reported development (as of November 2026) | Score |
|---|---|---|---|
| 1 | Compute spending grows to hundreds of billions, with a trillion-dollar cluster by 2030 | Google, Microsoft, Amazon, and Meta plan to spend about $725 billion this year, 77% more than last year | 1:0 for forecast |
| 2 | Electricity becomes the main bottleneck | Turbine manufacturers sell capacity through 2029 and 2030; GE Vernova has orders into 2031; banks estimate a shortfall of tens of gigawatts in US data centers by 2028 | 2:0 for forecast |
| 3 | China steals US AI technology | Former Google engineer Linwei Ding was convicted in January 2026 | 3:0 for forecast |
| 4 | One company earns $100 billion per year from AI by mid-2026 | Best available estimate is about $60 billion | 3:1 for forecast |
| 5 | AI functions as a remote employee by 2027 | Agents exist and perform many tasks, but no system can replace a human employee in this role | 3:2 for forecast |
| 6 | Strong AI arrives by 2027 | Two months remain before the deadline | Open |
| 7 | AI performs research tasks | OpenAI announced in September 2026 that its automated research intern was ready | 4:2 for forecast |
Compute and Spending
Aschenbrenner forecast a race for compute clusters. Reported capital expenditure by four major companies, Google, Microsoft, Amazon, and Meta, is about $725 billion for 2026. Those companies are also cutting staff.
Electricity
Gas turbine manufacturers are selling production slots for 2029 and 2030, and GE Vernova has slots reserved for 2031. Turbine prices have risen sharply. Banks estimate that US data centers will face a shortfall of tens of gigawatts by 2028, equal to several dozen nuclear reactors.
Espionage
In January 2026, a jury convicted former Google engineer Linwei Ding. He took secrets about the infrastructure used to train AI models and worked for Chinese companies. The US Department of Justice described the case as the first successful prosecution for AI theft for the benefit of China. In April 2026, the White House stated that China was stealing American AI technology "on an industrial scale."
Revenue
Aschenbrenner predicted that one company would earn 60 billion. The forecast did not meet this target.
Remote Employees
Aschenbrenner predicted that by 2027, AI would work as a remote employee that operates independently for weeks. AI agents already perform many tasks. A system that can fully replace a human employee in this role does not yet exist.
Automated Research
In October 2025, Sam Altman described OpenAI's internal goals during a broadcast:
- An automated research intern by September 2026.
- A fully automated AI researcher by March 2028.
In September 2026, OpenAI announced that its intern system was ready. The system works under human supervision and performs research tasks that would take experienced researchers several days. The claim comes from OpenAI. No independent verification has been found.
The Investment Fund
Aschenbrenner founded an investment fund named after his paper, Situational Awareness. Early backers included Patrick and John Collison, founders of Stripe, and investors Nat Friedman and Daniel Gross. According to press reports, the initial capital was about $225 million.
The fund invested in the infrastructure the paper identified: power plants, memory chips, graphics processing units (GPUs), and former Bitcoin miners with access to low-cost electricity.
Performance and Losses
- First half of 2026: The fund returned 439%, according to a letter to investors that the Financial Times reviewed.
- Early July 2026: Assets under management reached about 45 billion by summer 2026.
- Leverage: The fund used borrowed money from a broker to increase its positions. Leverage amplifies gains and losses in equal measure.
- Concentration: At the end of June, more than half of the fund's US stock holdings were in two memory producers, SanDisk and Micron.
- July decline: SanDisk lost nearly half its value, and Micron lost nearly a third. Power plant stocks also fell.
- Margin call: The broker demanded additional capital. The fund had no further funds to add.
- Liquidation: At the end of July, the fund sold most of its stock positions to Ken Griffin's fund at a large discount. Approximately $1 billion remained.
According to journalists' estimates, the loss ranks among the largest in dollar terms in hedge fund history.
Ashvin Bhalwani, a partner in the fund, reportedly married in the same period. His wife, Avital Balwit, is Chief of Staff to Dario Amodei, the head of Anthropic, which develops the Claude AI models.
In September 2026, the fund resumed buying the same categories of assets, chips and data centers, this time using options.
Sam Altman's Remarks on Memory
On October 29, 2026, Sam Altman appeared on John Collison's podcast. Altman said:
"From a global distribution standpoint, it's pretty terrible that everyone has a piece of the world's limited memory supply sitting in their computer."
Altman also stated that memory in personal computers is in use only about 5% of the time. He argued that this memory would be used more efficiently in data centers.
The fund's largest position, as described above, was in memory producers. The author of this article notes that consumer memory and computer prices have increased. In the author's observation, computers now cost at least 50% more than they did a year and a half earlier.
The Project and the Ending
Aschenbrenner's paper concludes with what he calls "The Project." He forecasts that by 2027 or 2028, governments will recognize superintelligence as a matter of national survival, comparable to the atomic bomb in the 1940s. In his scenario, the state takes control of the best researchers, the largest compute clusters, and development, under secrecy comparable to the Manhattan Project. He describes a race with China in which, in his words, "the free world must win."
In November 2025, President Trump signed the Genesis Mission executive order. It is a government program that connects the supercomputers of national laboratories with data and AI resources. The administration described it as the largest government science program since the Manhattan Project.
The program concerns scientific research. It is not yet the state-level superintelligence program described in the paper.
Parallel Forecast
Daniel Kokotajlo, another former OpenAI employee, left the company in April 2024 and also published a forecast of AI development. According to the source, Kokotajlo wrote his forecast so that it would not come true. In September 2026, he presented his forecast to the US Senate.
Summary
Aschenbrenner's forecasts on compute spending, electricity constraints, and espionage matched reported developments. His forecasts on AI revenue and remote AI employees have not been met. The automated research forecast is consistent with OpenAI's reported progress, although the company's claims remain unverified. The 2027 superintelligence forecast remains unresolved. The fund he built on the paper's thesis recorded a 439% gain in the first half of 2026, then suffered heavy losses in July after leveraged positions in memory and power-related stocks declined.
Introduction and the Forecast
Introduction to Aschenbrenner's 165-page essay, which predicted AI data-center scale, electricity shortages, and superintelligence by the end of the decade.
- Essay published in mid-2024 by Leopold Aschenbrenner, a 22-year-old former OpenAI employee
- Predicted AI computer clusters costing hundreds of billions of dollars and electricity as the main bottleneck
- Stargate project announced in January by OpenAI and partners: up to $500 billion for data centers over 4 years
- Predicted AI capable of AI research by 2027 and superintelligence by the end of the decade
- His investment fund was managing about 4.5 billion)
Who Aschenbrenner Is and Why He Was Fired
Background on Aschenbrenner: education, OpenAI role, and firing over a security memo about Chinese espionage.
- German, son of two doctors; left at 15 to study at Columbia University; graduated top of class at 19
- Joined OpenAI to work on keeping future superintelligence under control
- In 2023 OpenAI was hacked; he wrote a memo to the board warning that weak security could let Chinese intelligence steal developments
- Fired in April 2024 officially for leaking information; he says the leak was a regular draft shared with three external researchers
- OpenAI disputes his account; the host says the truth is unclear
Forecast Versus Reality
Point-by-point scoring of the forecast against 2026 reality, with the host's running score.
- Forecast titled 'Situational Awareness'; claims the future is visible in San Francisco
- Compute race: Google, Microsoft, Amazon, and Meta expected to spend about $725 billion this year, 77% more than last year (score 1:0)
- Electricity bottleneck: gas turbine makers booked into 2029-2030, GE Vernova into 2031; banks estimate data centers short by tens of gigawatts by 2028 (score 2:0)
- Espionage: former Google engineer Linwei Ding convicted in January 2026; White House accused China of AI theft in April (score 3:0)
- Missed: one big company earning 60 billion (score 3:1)
- Missed: AI as a remote employee by 2027; agents exist but cannot fully replace humans (score 3:2)
- Strong AI by 2027 left undecided
Automated Research Intern
OpenAI's internal goal of an automated research intern by September 2026 was reportedly met, counted as a point for the forecast.
- Sam Altman said in October 2025 that OpenAI's goals were an automated research intern by September 2026 and an automated AI researcher by March 2028
- In September 2026 OpenAI announced the intern was ready; claims are company-only, with no independent verification found
- Host counts this as a point, making the score 4:2 in favor of the forecast
The Fund Rise and Collapse
Aschenbrenner's fund, backed by the Collison brothers, Nat Friedman, and Daniel Gross, surged then collapsed due to leverage in July 2026.
- Fund named after his paper; initial capital about $225 million from Collison brothers, Nat Friedman, and Daniel Gross
- Invested in power plants, memory chips, graphics cards, and former Bitcoin miners
- H1 2026 return reportedly 439%; fund managed about $4.5 billion by early July per press estimates
- Heavy leverage: more than half of US stock holdings in SanDisk and Micron as of end of June
- July: SanDisk lost nearly half its value, Micron nearly a third; broker margin calls forced sale of most positions to Ken Griffin's fund at a large discount
- About $1 billion left, reportedly one of the largest dollar losses in hedge fund history
Sam Altman, RAM, and Fund Options
Sam Altman's remarks on memory supply, the RAM price increase, and the link between Bhalwani's marriage and Anthropic.
- Bhalwani's wife Avital Balwit is Chief of Staff to Dario Amodei, head of Anthropic (maker of Claude)
- On October 29 Sam Altman, on John Collison's podcast, said it is terrible that everyone has a piece of the world's limited memory supply in their computer
- Altman argued that RAM in personal computers is used only about 5% of the time and belongs in data centers
- Host claims RAM prices rose so that computers are at least 50% more expensive than a year and a half ago
Conclusion and the Project
The host's view that the forecast is largely accurate on hardware and security, but early on revenue and remote employees; the 'Project' national-security scenario is emerging.
- Final score given as 4 to 2 in favor of the forecast, with strong AI by 2027 still open
- Accurate: hardware, electricity, and espionage; early: company money and remote employees
- Aschenbrenner's 'The Project' predicts superintelligence becoming a national-security matter by 2027 or 2028, like the Manhattan Project
- In November 2025 Trump signed the Genesis Mission decree, a government program unifying national laboratory supercomputers, described as the largest since the Manhattan Project
- Daniel Kokotajlo, another former OpenAI employee, described the scenario as the most dangerous and presented his forecast to the US Senate in September
