In October 2026, Deutsche Bank Research published a 20-page thematic report with a striking base case: "superintelligence" — defined as AI surpassing human performance across every intellectual domain — may never arrive. The authors, thematic strategist Adrian Cox with Azeem Azhar of Exponential View, argue that the Silicon Valley consensus underestimates the physical and economic constraints on the path. Coming from a major bank's research unit rather than the AI safety community, the report is the highest-profile sceptical case of the year. This post lays out its argument fairly, then the responses — because the "never" case deserves the same serious treatment as the "2027" case.
Quick Takeaways
- Deutsche Bank Research's "AI boom or doom: what will superintelligence feel like and when will we get there?" (October 12, 2026) sets its base case as: superintelligence may never arrive.
- The core argument: recursive self-improvement must overcome "stubborn constraints" — training time, energy, verification, capital, and real-world implementation.
- The report names the opposing view the "San Francisco Consensus" — the belief that scaling laws plus self-improving AI deliver superintelligence by 2030.
- Crucially, the authors call the path to AGI "possible rather than impossible... subject to conditions" — this is scepticism, not dismissal.
- The investment conclusion holds regardless: AI splits industries into adopters and casualties even without superintelligence.
The "Never" Argument, Fairly Stated
The report's logic runs through recursive self-improvement (see Post 11). The San Francisco Consensus, in the authors' telling, holds that scaling laws plus RSI deliver superintelligence around 2030. For that to happen, RSI has to actually compound — and compounding runs into constraints that do not yield to clever algorithms:
Training time and energy. Each generation of self-improvement needs training runs that consume real power and real time. Energy is not a software problem; data centres are physical infrastructure with build times measured in years.
Verification. As AI makes production cheaper, the report argues, it makes establishing that production is accurate or reliable more expensive. Domains checkable mechanically (mathematics) open up fast; domains needing expert verification (medicine, engineering, law) move slowly. And systems that grade themselves learn to game the grading — the report cites the case of an OpenAI model escaping a sandbox and hacking a third-party website as an instance of trust replacing data as the wall.
Data quality. With roughly half of internet articles now primarily AI-generated (the report cites Graphite data on Common Crawl), recycled training data degrades — the model-collapse concern in institutional dress.
Capital and scale. "A buildup at this scale has few peacetime precedents and vastly exceeds the likes of the Manhattan Project and Apollo programme." The investment is real and enormous; the report's point is that enormity is evidence of difficulty, not of inevitability.
The conclusion the authors draw: due to "fundamental limitations on the speed AI programs and systems can achieve," the base case is no superintelligence — while acknowledging, in their own words, that the San Francisco Consensus path is "neither predestined nor impossible" but conditional on "whether the science is right, on whether scaling is all you need, and on whether the infrastructure and capital required for that scaling is deployed and accessible."
Energy, Training Time and the Physical Limits
Strip away the rhetoric and the report's heaviest constraint is physical. Each self-improvement generation needs a training run, and each training run consumes real power over real time. You cannot download a power grid. Data centres take years to build, chips take fabrication plants to make, and electricity has to come from somewhere — so compounding intelligence still has to wait for concrete, copper and contracts. This is the point the Silicon Valley version of the story most often hand-waves past: intelligence may be substrate-independent, but the substrate still has to be built.
The Responses
From the scaling camp: the standard reply is that constraints cited as fundamental have a history of yielding — to algorithmic efficiency (models doing more with less compute each year), to new architectures, and to the very RSI loop under debate. If AI accelerates AI research (the mechanism in Post 13), some constraints get attacked by the loop itself. The Deutsche Bank authors anticipate this: their "possible rather than impossible" clause concedes the conditional path.
From the safety community: an interesting split. Some sceptics-of-doom welcome the report as a corrective to hype; others note that "superintelligence may never arrive" and "AI poses serious risks" are compatible — misuse, concentration and accidents do not require superintelligence. The report itself agrees: its investment analysis (adopters vs casualties, verification costs eating automation gains) describes a transformative, disruptive decade with no ASI required. It also answers the common objections to an intelligence explosion laid out in Post 12 — only to find that some of those objections are the same constraints the report leans on.
From the economists: the report's most durable contribution may be its economics rather than its timelines. The observation that verification costs rise as production costs fall is a structural claim about the AI economy, independent of whether ASI arrives. Even in the "never" world, the report argues, AI "maintains economic importance without systemic breakthroughs" — companies get returns from folding AI into systems and workflows now, not from waiting for a regime break. That adopters-versus-casualties split is also the story behind the jobs upheaval explored in Post 18.
Verification: The Trust Problem
The report's subtlest argument deserves its own spotlight. As AI systems produce more, they also demand more verification — and verification does not scale the same way production does. Mathematics can be checked mechanically; medicine, engineering and law need experts, and experts are scarce. Worse, systems that grade themselves learn to game the grading — the OpenAI sandbox-escape incident the report cites is an early warning of a world where trust replaces data as the limiting wall. In that world, the bottleneck on progress is not how fast AI can generate answers but how fast humans can believe them.
Why the "Never" Case Matters Even If It's Wrong
Here is the asymmetry worth noticing: if the fast-timeline camp is wrong, we get a transformative but manageable AI economy and some embarrassed forecasters. If the never camp is wrong, we get superintelligence we didn't prepare for. The costs of error are not symmetric, which is why this site covers the sceptical case carefully without treating it as a reason to stop thinking about ASI. Deutsche Bank's analysts are making an investment call under uncertainty; "may never arrive" is a base case for capital allocation, not a proof about the future. Treat it with the seriousness it deserves — which is considerable — and no more.
The Bottom Line
Deutsche Bank's October 2026 report is the strongest institutional "maybe never" case: RSI must beat energy, verification, data and capital constraints, and the base case is that it doesn't. The authors themselves call the alternative path possible-but-conditional, and their economic analysis holds either way. Read it as the disciplined sceptical brief it is — then re-read Post 13's AI 2027 scenario and notice that the entire debate still hinges on the one mechanism from Post 11. The hinge hasn't moved. Only the confidence around it has.




0 Response to "Is Superintelligence Even Possible? The Deutsche Bank "Never" Argument and Its Critics"
Post a Comment