Leopold Aschenbrenner’s Situational Awareness: The Decade Ahead, June 2024, argues that AGI (artificial general intelligence) arrives around 2027 and that the constraint is industrial, not intellectual. In July 2026 the fund built on that thesis was liquidated after margin calls. That sequence invites a cheap article. This is not one.
Rules fixed before reading: only closed windows are scored, which leaves 27 forward claims. Roughly sixty dated 2027, 2028, 2030 or “the end of the decade,” including the headline AGI thesis, are not due. Retrospective claims are scored as sourcing accuracy, not prediction skill.
| Claims | Right | Partial | Wrong | Undetermined | |
|---|---|---|---|---|---|
| Forward claims, window closed | 27 | 13 | 3 | 7 | 4 |
| Retrospective sourcing accuracy | 43 | 33 | 4 | 1 | 5 |
| Arithmetic and internal consistency | 9 | 6 | 0 | 2 | 1 |
On the only part that tests prediction he is 13 right, 3 partial, 7 wrong, 4 undetermined, a respectable record for a two-year-old document of dated quantitative claims. He was right about the machines and wrong about what they would cost.
| Claim, June 2024 | What the record says | Verdict |
|---|---|---|
| Nvidia does over $200 billion of revenue in calendar 2025, against a sell-side consensus he named as $120 to $130 billion | Nvidia’s fiscal 2026 came in at $215.9 billion | Right; named the consensus, named his number, beat both |
| A 2026 cluster at roughly a million H100-equivalents and about a gigawatt | Colossus 2 came online in January 2026 at about a gigawatt | Right; a physical object forecast two years ahead |
| Aggregate AI investment of roughly $500 billion for 2026 | Hyperscaler AI infrastructure spend running near $725 billion | Wrong, and too small |
| OpenAI at a $10 billion run rate by late 2024 or early 2025 | Reached in June 2025 | About a year late |
| A big tech company at $100 billion of AI revenue by mid-2026 | Microsoft, the only one of the three he named that discloses a comparable figure, reported $37 billion | Wrong |
| The doubling rate assumed behind that call | Anthropic went from a run rate of roughly $1 billion of revenue to $47 billion in seventeen months | If anything too slow |
| Gas plant capital spending (capex) “under $1000 per kW”; combined cycle plants buildable “in about two years” | Gas turbine prices reported up roughly 300% in three years; current analyst estimates put GE Vernova’s heavy-duty turbine equipment alone near $790 per kW and its aeroderivative units near $1,800 | Neither survived contact |
The turbine appears once
He named the binding constraint early: “Probably the single biggest constraint on the supply-side will be power.” Then he priced it. The word “turbine” appears once in 165 pages, in a subordinate clause: “The harder part would be building enough generators / turbines; this wouldn’t be trivial, but it seems doable.” Both power misses live in those eleven words. That compression is fair in an essay about capability trajectories, not careless: his physics held, and his price failed because the price sits in a supply chain he never modelled. A forecast with a physical bottleneck is only as good as the model of that bottleneck.
Same object, two forms
On 1 August 2026 I submitted fourteen pre-registered forecasts with binding resolution criteria, and the slate froze at submission. Three price things his essay also discusses. In each pair one claim can be graded by a stranger who cares about neither of us.
| Object | His form | My form |
|---|---|---|
| Power | The biggest constraint on the supply side, undated, no threshold, closing on three modal verbs: it “can, must, and will be solved.” | The PJM Interconnection’s 2029/30 base residual auction clears at or above $325.00 per megawatt-day, at 72% probability, resolving on PJM’s own published results |
| Hyperscaler capex | 2024 dollar thresholds for four companies, three of four right; the miss was Meta at $39.23 billion against his “$40 billion plus,” a 2% shortfall that becomes 7% if finance lease principal is excluded, and which definition applied was never specified; implied growth 100% a year | The same four companies and the same line item, with the exclusions specified in writing before the fact; implied growth 20%, priced at 62% |
| Chips and China | Across 165 pages “rare earth” appears zero times, “chokepoint” zero times and “ASML” zero times; supply risk runs one way, from American capability outward | Two of the fourteen forecasts price the chokepoints running the other way |
What this does not show
- My slate has resolved zero of fourteen. No comparison of accuracy is computable, and I offer none.
- Resolvable is not the same as right. If my fourteen come back at ten wrong, I will have been wrong in public on schedule.
- An essayist and a pre-registered forecaster do different jobs. He wrote to change what people believe about the decade, and by that standard the essay worked.
- The fund is not evidence about the essay. A liquidation speaks to leverage, correlation and margin, not to whether AGI arrives in 2027. The sixty undue claims stay undue.
Method. Claims were extracted from the June 2024 PDF with page citations and classified as checkable now or not yet due before any scoring. Primary sources were used where they exist; trade press is flagged inline as secondary. The full extraction, including the sixty undue claims left unscored, is available on request.
| Quantity | Primary source |
|---|---|
| Revenue and capex | Securities and Exchange Commission (SEC) filings and company results |
| Capacity prices | PJM’s published auction releases |
| Electricity | US Energy Information Administration (EIA) series |
Educational research only. Not investment advice. This note concerns published claims and their resolution, and makes no assertion about any individual’s conduct or ability.
Educational research only. Not investment advice. This note concerns published claims and their resolution, and makes no assertion about any individual's conduct or ability.