The thesis
Leopold Aschenbrenner's Situational Awareness: The Decade Ahead(2024) argues that AGI by 2027 is “strikingly plausible,” driven by steady scaling of training compute, algorithmic efficiency, and the unhobbling of models into agents. The implications cascade through chips, fabs, packaging, memory, networking, data centers, power, cloud, models, applications — and ultimately national security. These five cards summarize the thesis and flag the public companies most directly impacted by each pillar.
AGI by 2027 is plausible
Capability gains have been driven by predictable scaling. If trends continue, models could be doing serious cognitive work by the late 2020s.
Aschenbrenner argues that the most striking thing about AI progress is how steady it has been when measured along the right axes — training compute, data, algorithmic efficiency, and 'unhobbling' from chatbot to agent. Extrapolating those curves another four to five years lands at systems that can plausibly do the work of a junior research scientist or engineer.
This is not a guarantee. It is a forecast that says: if the same inputs keep stacking, you get something qualitatively new. Investors should treat AGI-by-2027 as a probability-weighted scenario, not a base case — but a high enough probability to drive significant infrastructure spend in the meantime.
Capability keeps compounding; AI revenue becomes a meaningful share of GDP; whoever supplies compute, models, and AI-enabled workflows captures durable cash flow.
Scaling laws bend (data exhaustion, RL plateaus, eval saturation); capex outruns monetization; valuations correct violently as the multi-trillion-dollar buildout meets a cooler revenue ramp.
Compute is the bottleneck and the accelerant
Each ~10x of training compute has unlocked a qualitative capability step. The next 10x requires unprecedented capital, power, and silicon.
The single most leveraged input to AI capability has been training compute. The cluster-scale needed for the next generation of models is large enough to put real strain on the global supply of advanced packaging, HBM, leading-edge wafers, power, and skilled labor.
The corollary: most of the AI dollars in the next several years flow into infrastructure. NVIDIA, TSMC, ASML, SK Hynix, the hyperscalers, the data center REITs, the grid contractors, and the gas-turbine OEMs all benefit, in different ways and at different points in the cycle.
Compute demand stays vertical through 2027; HBM and CoWoS remain sold-out; pricing power persists for bottleneck holders.
Inference becomes the dominant workload; specialized cheaper silicon eats general-purpose accelerator margin; recipes get more efficient (DeepSeek-style); supply catches up and pricing normalizes.
Infrastructure bottlenecks are real and binding
Power transformers, EUV scanners, HBM packaging, and grid interconnects are all running multi-year backlogs.
The AI buildout is not bottlenecked at the same place every quarter. In 2023 it was wafer supply; in 2024 it was CoWoS packaging and HBM; in 2025 it is increasingly power, transformers, and grid interconnect queues; by 2026 it may be skilled construction labor and substation siting.
For investors, the implication is that the highest-quality returns will come from owning the supply where the bottleneck has just moved — not where the bottleneck used to be. That's a moving research target, which is what makes this dashboard worth maintaining.
Bottlenecks persist long enough for pricing power to translate into durable cash flow at infrastructure suppliers.
Supply ramps faster than demand once new HBM lines and turbine slots energize; the cycle inflects sharply downward for the most leveraged names.
AI is becoming national-security infrastructure
Export controls, sovereign AI deals, and chip diplomacy are reshaping where compute is built and who gets to use it.
Once you take AGI-by-2027 seriously, AI compute starts to look less like commercial software and more like aerospace or nuclear. Governments behave accordingly: US export controls on advanced chips, sovereign AI cloud agreements with Gulf states and Europe, and the politicization of model weights are all consequences.
The investment implication is that geography matters more than it used to. Taiwan concentration risk is the largest single tail in the supply chain. US-based fabs, OSATs, and power infrastructure get a strategic premium. Sovereign cloud spend creates secondary winners in countries that build their own AI capacity.
US/allied infrastructure spend accelerates as a national-security priority; bottleneck holders within friendly jurisdictions earn structural premiums.
Geopolitical accident (Taiwan, Middle East, export-control retaliation) triggers a supply shock; some of the most-owned AI names dislocate hard.
Investment implications: layered, not monolithic
AI is not one trade. It is a stack of trades with different margins, capex profiles, durations, and risks.
Treating AI as a single theme is the single most expensive mistake an investor can make right now. The accelerator layer has different economics from the foundry layer, which differ from cloud, which differ from frontier models, which differ from enterprise apps, which differ from disrupted incumbents.
This dashboard exists to help investors see all of those layers at once: who is in each, how the market sizes look, where the value is being captured today, and which companies look strongest on fundamentals, technicals, and AI strategic position when judged within their own segment. Composite scores are the starting point of a research conversation, not the end of one.
Disciplined investors who pick by layer outperform the index-weighted AI trade.
AI exuberance reverses; index-level damage is concentrated in the most owned names; segment-aware allocation provides only partial protection.