Leopold Aschenbrenner, a former OpenAI superalignment team member and Columbia valedictorian, argues that AI progress is on a trajectory toward superintelligence by 207–2028, driven by exponential growth in compute clusters, and that this will trigger a geopolitical crisis comparable to the Manhattan Project era — with the US, China, and authoritarian states racing to build trillion-dollar data centers, steal AI secrets, and automate military and economic power.
The trillion-dollar cluster and unhobbling
AI development is fundamentally an industrial process: each new model generation requires building massive new compute clusters, power plants, and eventually semiconductor fabs — not just writing better code.
Training compute for the largest AI systems has grown by roughly half an order of magnitude (0.5 OOMs) per year for nearly a decade.
GPT-4 (2022) used 25,000 A100s ($500M cluster, ~10 MW). By 2024, clusters are ~100 MW with 100,000 H100 equivalents costing billions. By 2026, a gigawatt-scale cluster (Hoover Dam output) costing tens of billions with a million H100 equivalents. By 2028, 10 GW (more than most US states), 10 million H100 equivalents, hundreds of billions of dollars. By 2030, a trillion-dollar cluster at 100 GW — over 20% of US electricity production.
Companies like OpenAI and Microsoft are already planning $100B+ clusters; AMD forecasts a $400B AI accelerator market by 2027.
The economic justification: if AI can automate cognitive labor at scale, $100B+ annual revenue is plausible — e.g., selling a $100/month AI add-on to a third of Microsoft’s 300M Office subscribers yields $100B.
Unobbling: Current models like GPT-4 are “hobbled” — they’re smart but limited to chatbot-style interactions. The key unlock is making them into agents that can do long-horizon tasks, use computers, and work autonomously like remote employees.
By 2025–2026, models will surpass most college graduates in capability. By 2027–2028, they’ll match the smartest experts and function as “drop-in remote workers” — attending Zoom calls, using Slack, writing and iterating on code, running tests, and completing projects with minimal human oversight.
Intermediate models could have been integrated into businesses, but it would require painful workflow changes (“schlep”). Overpowered agentic models make adoption trivial — you just don’t need the human worker anymore.
Test-time compute overhang: GPT-4 can “think” for a few hundred tokens (equivalent to ~3 minutes of human thought). If models could think coherently for millions of tokens (months of working time), they’d gain enormous problem-solving ability — roughly equivalent to a model 3.5 OOMs larger.
Unlocking this requires learning “System 2” thinking tokens: error correction (“I made a mistake, let me reconsider”), planning (“here’s my plan of attack”), self-critique (“let me review my draft”).
This is the “unhobbling path” to agents — as opposed to the “scaling path” that just improves reliability (more “nines”).
Pre-training vs. self-play: Pre-training on internet text gives models rich world representations, but it’s sample-inefficient (like passively listening to a lecture). Real learning requires active engagement — trying problems, failing, discussing, and distilling insights.
Models are entering a regime where they can learn from self-play, synthetic data, and RL — like a student who has learned enough basics to start teaching themselves.
This is analogous to the difference between GPT-2 (preschooler) and GPT-4 (smart high schooler); scaling alone will produce another such jump by 2027–2028.
AI 2028: The return of history
The intelligence explosion: Once AI can automate AI research itself, progress could accelerate dramatically — 100 million automated AI researchers running on inference clusters could compress a decade of ML progress into a year.
This then cascades into robotics, biology, materials science, and other fields — potentially compressing a century of technological progress into less than a decade.
A lead of even a few years could be as decisive as the US technological advantage in the first Gulf War (100:1 kill ratio from 20–30 years of lead in sensors, GPS, stealth, precision weapons).
Superintelligence applied to military technology could undermine nuclear deterrence entirely — e.g., millions of mosquito-sized drones finding and destroying nuclear submarines and mobile launchers.
Historical parallels: The post–Cold War period of peace and liberal democratic dominance is historically abnormal. The norm is intense great-power competition — World War II saw 50% of US GDP go to war production, borrowing over 60% of GDP. The Seven Years’ War killed 20–30% of Prussia; the Thirty Years’ War killed up to 50% of parts of Germany.
People in the US have grown complacent; most don’t yet “feel” the trajectory. But exponential trends will become undeniable — like COVID in February 2020, when most of the world didn’t yet grasp what was coming.
Authoritarian implications of superintelligence: A regime with superintelligence could achieve perfect surveillance, perfect lie detection, and perfectly loyal security forces — eliminating dissent, coups, and reformers like Gorbachev. Truth could be permanently locked in by the party, with no pluralistic evolution of ideas.
Espionage and American AI superiority
Why clusters must be in the US (or allied democracies): Building AGI clusters in authoritarian states like the UAE creates irreversible security risks.
They could steal the model weights (a literal copy of the AGI, like stealing the atomic bomb design) or seize the compute outright.
Even a 25% compute share gives authoritarian states enormous leverage — 33 million superintelligent AI researchers could design novel WMDs. A 3:1 compute ratio is dangerously close.
Middle Eastern states have financial capital but lack leading AI labs, talent, and hardware (they’re export-controlled from receiving Nvidia chips). Their “seat at the AGI table” is being bought with money, not earned with capability.
Two paths to powering US clusters:
Natural gas: The US has ample natural gas; production has nearly doubled in a decade. A 10 GW cluster is a few percent of US natural gas output. 100 GW is doable with continued expansion. This conflicts with climate commitments made by Microsoft, Amazon, etc.
Green energy megaprojects: Solar, batteries, small modular reactors (SMRs), geothermal — but requires massive deregulation: FERC reform, NEPA exemptions, streamlined permitting, rights-of-way for transmission lines. Currently, hooking a solar installation to the grid can take years due to state-level regulations.
Ideally both paths are pursued; at least one is necessary.
The “they’ll go to China anyway” argument: Some claim that if the US doesn’t work with the UAE, they’ll partner with China instead. Leopold is skeptical:
The UAE can’t translate money into AI progress on its own — it lacks labs, talent, and hardware.
There are reports that OpenAI leadership once planned to fund AGI by starting a bidding war between the US, China, and Russia — effectively selling AGI to authoritarian governments.
Benefit-sharing (offering last-gen models for civilian use) is reasonable, but giving authoritarian states a seat at the AGI development table is a profound strategic error.
Secrecy and algorithmic lead: The US has a significant lead in algorithmic progress (~0.5 OOMs/year). If secrets are protected, this compounds — a few years of lead could mean a 10–100x effective compute advantage.
Weights theft: Stealing model weights is trivially easy — an employee at a US lab reportedly copied critical AI code to Apple Notes and exported it as a PDF, bypassing monitoring. Google has the best security (enterprise-grade); other labs have startup-level security.
Algorithmic secrets: The fundamental approaches (next-token pre-training, scaling laws, transformers, MoE) were all published openly until recently, which is why China can build decent models from Llama and other open-source work. If the next paradigm (e.g., self-play RL to get past the data wall) is kept secret, China could be stuck — like Nazi Germany going down the heavy water path instead of graphite for nuclear reactors.
Tacit knowledge: Large-scale engineering for training runs involves hard-won tacit knowledge, but China can likely figure this out. The critical thing is protecting the ideas — the next paradigm — not just the engineering details.
Why a 1–2 year lead matters enormously: At the current pace, three years ago models couldn’t solve competition-level math problems; now they can. With a billion superintelligent researchers accelerating R&D, a year of lead could mean the difference between human-level and vastly superhuman AI — and decades of technological advantage.
Geopolitical implications
The stakes: What’s at stake is not just cool products but whether liberal democracy survives, whether the CCP survives, and what the world order looks like for the next century.
The CCP will eventually recognize superintelligence as decisive for national power and mount an all-out espionage and industrial effort — billions of dollars, thousands of people, full Ministry of State Security involvement.
China has enormous latent industrial capacity (they added as much power in the last decade as the entire US grid) and are already producing 7-nanometer chips despite export controls.
The danger of a tight race: If the US and China are neck-and-neck (e.g., 3-month lead), the situation is incredibly dangerous — both sides rush, throw caution to the wind, and destabilizing new WMDs emerge every few weeks, making deterrence volatile.
A comfortable lead (6+ months to 2 years) gives the US “wiggle room” to dedicate compute to alignment, slow down if needed, and avoid catastrophic mistakes.
Why most people aren’t talking about this: Being “in the trenches” of AI development gives a myopic view — researchers focus on the next model, the next benchmark, the next data problem. Zooming out just a few years reveals the exponential trajectory. Most people outside SF don’t yet “feel” it, just as most of the world didn’t grasp COVID until March 2020.
Once AGI becomes undeniable, societal reaction will be radical and fast — like Congress spending over 10% of GDP on COVID within weeks.
