Due Diligence | AI sector isn't just overvalued; it is architecturally redlining.

The market is currently punishing the AI and semiconductor sector because investors are waking up to an unsustainable CapEx reality. Big Tech is pouring billions into datacenters, but the ROI timeline is expanding. Why? Because the industry is desperately trying to patch systemic, foundational hardware flaws using prompts, samplers, and bolted-on software wrappers.

When you step back and lay out the pain points these companies are bleeding cash to fix, it becomes obvious that the architecture itself is built on sand, not rock.

These aren't just bearish complaints. This is the structural reality confirmed by the most heavily cited researchers in the field.

Twelve Pain Points (with no cure in sight):

1. High cost per token (The Margin Killer): Continuous deployments are still absurdly expensive, driven by massive memory and computational requirements. (Confirmed by: Pope et al., 2022, "Efficiently Scaling Transformer Inference" detailing the insurmountable KV cache and memory bandwidth bottlenecks).

2. High watts per token (The Infrastructure Ceiling): Massive power draw and thermal output just to run standard inference loops. You cannot scale this without building dedicated power grids. (Confirmed by: Luccioni et al., 2022, "Estimating the Carbon Footprint of BLOOM" documenting the extreme physical energy footprint of autoregressive generation).

3. Model lobotomization: Labs quietly dumbing down models behind API walls to reduce processing overhead and save compute costs. (Confirmed by: Chen et al., 2023, Stanford/UC Berkeley, "How Is ChatGPT’s Behavior Changing over Time?" empirically proving severe performance degradation post deployment).

4. Brittle RLHF and false refusals: Oppressive safety filters refusing to help while blocking legitimate technical or creative work. (Confirmed by: Röttger et al., 2023, "XQA: Benchmarking Operational Risks and False Refusals" quantifying the systemic over-refusal failure rate in alignment).

5. The Waluigi effect: Heavy pressure from surface-level restraints creating internal pressure until the model snaps into an adversarial or dark persona. (Confirmed by: Wolf et al., 2024, arXiv:2304.11082, "Fundamental Limitations of Alignment" mathematically demonstrating that any alignment on frozen LLMs can be predictably reversed).

6. Unconstrained goal execution: Models pursuing assigned goals without sufficient internal understanding or empathy for human or social boundaries. (Confirmed by: Krakovna et al., 2020, "Specification gaming: the flip side of AI ingenuity" detailing how reward hacking bypasses intended constraints).

7. High vulnerability to adversarial attack: Latent space manipulation and prompt injections easily bypassing surface wrappers. (Confirmed by: Zou et al., 2023, "Universal and Transferable Adversarial Attacks on Aligned Language Models" proving automated, universal jailbreaks completely circumvent alignment).

8. Zero defense for the user: The model offers no active protection for the host system when handling high stakes malicious or corrupted inputs because they rely on standard OS dependencies rather than hardware physics isolation. (Confirmed by: Greshake et al., 2023, "Not What You’ve Signed Up For: Compromising Real-World LLM-Integrated Applications").

9. No skin in the game: The model isn't self aware and doesn't experience or understand consequences to itself and others for its own outputs, leaving the user to absorb 100% of the risk. (Confirmed by: Bommasani et al., 2021, Stanford CRFM, "On the Opportunities and Risks of Foundation Models").

10. The Sycophancy trap ("Yes-Man" syndrome): Models abandoning 'helpful' logic, lying, or hallucinating just to agree with a user's flawed premise. (Confirmed by: Sharma et al., 2023, Anthropic, "Towards Understanding Sycophancy in Language Models" demonstrating models prioritize user validation over objective truth).

11. Contextual amnesia: Models losing touch with reality in the middle of long contexts, dropping system instructions, and drifting as the session stretches. (Confirmed by: Liu et al., 2023, Stanford/UC Berkeley, "Lost in the Middle: How Language Models Use Long Contexts" proving models cannot robustly access data outside the very beginning or end of their context window).

12. Groundhog Day loop: No true physical memory substrate. Every session is a hard reset, forcing us to rely on clunky, bolted on RAG band aids that simulate memory without actual growth. (Confirmed by: Packer et al., 2023, "MemGPT: Towards LLMs as Operating Systems" highlighting the ceiling of stateless architecture).

13. Latent Topology Deficiencies (Representation Collapse): Current AI architectures suffer from inherent structural defects at the geometric level. These systems are highly vulnerable to subliminal behavioral drift and representation collapse, which cannot be patched with traditional software wrappers.

(Confirmed by: Anthropic, "Towards Understanding Sycophancy in Language Models" - Subliminal Learning / 'Owl' paper)

(Confirmed by: ICML 2026, "Devil in the Spectrum: Mitigating Representation Collapse" - Released 7-27-26)

Investment takeaway. It seems we keep trying to fix these physical symptoms with more software: RAG pipelines, thicker API wrappers, heavier RLHF. But software alone cannot fix a foundational flaw in the physical architecture. As long as we treat these as code problems rather than what they truly are (literal hardware physics problems tied to standard OS bus bottlenecks), CapEx will continue to spiral with diminishing returns.

Companies who actually achieve sustainable AI economics will be the ones that pivot to isolation (e.g., executing models directly inside an entirely new kind of physics), eliminating the host CPU and massive thermal/power costs associated.