Welcoming a buffet: Biting off more than we can chew?

Demand side: cheaper utility is displacing frontier models and default workflows, for three reasons.

- We must still understand and control our own processes. For high-risk AI, the EU AI Act requires that the people overseeing a system can understand its capacities and limitations, monitor its operation, and override or reverse its output.
Source: EU AI Act, Article 14 (European Commission)
https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-14

- Validating AI output is the bottleneck. In a randomized controlled trial, experienced developers using early-2025 AI tools took 19% longer to finish tasks than without them. Reviewing and correcting AI suggestions that were close but not quite right outweighed the gains.
Source: METR, July 2025
https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/

- Full automation is being pulled back. Gartner predicts over 40% of agentic AI projects will be canceled by the end of 2027, because of escalating costs, unclear business value or inadequate risk controls.
Source: Gartner press release, June 2025
https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027

Supply side: comparable capability is arriving in cheaper, smaller formats.

- Near-frontier performance now costs very little. GPT-3.5-level inference fell from $20 to $0.07 per million tokens between November 2022 and October 2024. The performance gap between top closed and open-weight models shrank from 8.0% to 1.7%.
Source: Stanford HAI, 2025 AI Index Report
https://hai.stanford.edu/ai-index/2025-ai-index-report

- Small, task-specific models are replacing general LLMs for business work. Gartner predicts that by 2027 their usage will be at least three times that of general-purpose LLMs, because general models lose accuracy on domain-specific tasks while smaller models are faster and use less compute.
Source: Gartner press release, April 2025
https://www.gartner.com/en/newsroom/press-releases/2025-04-09-gartner-predicts-by-2027-organizations-will-use-small-task-specific-ai-models-three-times-more-than-general-purpose-large-language-models

- Scaling is yielding less. Ilya Sutskever describes 2020-2025 as the "age of scaling" and argues that because pre-training data is finite, more resources no longer guarantee qualitative leaps.
Source: Dwarkesh Podcast interview, November 2025
https://www.dwarkesh.com/p/ilya-sutskever-2

- The payoff lies in downstream integration, not bigger models. MIT NANDA found 95% of organizations saw no measurable return from generative AI. It attributed the gap mainly to brittle workflows and poor fit with daily operations, not model quality.
Source: Fortune coverage of the MIT NANDA report, August 2025
https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo

Conclusion: Frontier scaling faces diminishing returns, while the value shifts to optimizing and integrating current models: cheaper, smaller, and under human oversight.

COMPUTE BUYERS (most to least exposed)

CRWV - CoreWeave
Segment: GPU cloud
Exposure: Highest
Main risk: Customer concentration, heavy leverage
Offset: None; pure AI capacity

ORCL - Oracle
Segment: Cloud / AI infrastructure
Exposure: Very high
Main risk: Debt-funded capex, OpenAI contract dependence
Offset: Database and software business

META - Meta
Segment: AI capex (internal use)
Exposure: Moderate
Main risk: Write-downs if AI payoff disappoints
Offset: Ad business cash flow

MSFT - Microsoft
Segment: Hyperscaler
Exposure: Moderate
Main risk: Azure AI growth tied to OpenAI
Offset: Diversified software, strong cash flow

GOOGL - Alphabet
Segment: Hyperscaler
Exposure: Lower
Main risk: Capex scale
Offset: Own TPUs and models, flexible capacity

AMZN - Amazon
Segment: Hyperscaler
Exposure: Lower
Main risk: Capex scale
Offset: Broad AWS base, own Trainium chips


COMPUTE SUPPLIERS (most to least exposed)

SMCI - Super Micro
Segment: AI servers
Exposure: Highest
Main risk: Thin margins, tracks accelerator shipments
Offset: None significant

NVDA - Nvidia
Segment: Training GPUs
Exposure: Highest
Main risk: Volume and margin compression
Offset: CUDA ecosystem, inference share

000660 - SK Hynix (Korea)
Segment: HBM memory
Exposure: High
Main risk: HBM demand tied to high-end accelerators
Offset: Conventional memory

MU - Micron
Segment: HBM memory
Exposure: High
Main risk: HBM exposure, memory cyclicality
Offset: Conventional memory

VRT - Vertiv
Segment: Data center power / cooling
Exposure: High
Main risk: Build-out slowdown
Offset: Long project backlogs delay impact

AMD - AMD
Segment: AI GPUs / CPUs
Exposure: High
Main risk: AI share-gain story stalls
Offset: CPUs, gaming, inference chips

MRVL - Marvell
Segment: Custom ASICs / optics
Exposure: Moderate-high
Main risk: Data-center concentration
Offset: Custom inference chips

ANET - Arista
Segment: AI networking
Exposure: Moderate
Main risk: Slower AI cluster builds
Offset: General cloud networking

AVGO - Broadcom
Segment: Custom ASICs / networking
Exposure: Moderate
Main risk: Hyperscaler capex slowdown
Offset: Custom inference chips, VMware software

005930 - Samsung (Korea)
Segment: Memory / foundry
Exposure: Moderate
Main risk: Trails in HBM
Offset: Phones, displays, commodity memory

TSM - TSMC
Segment: Foundry
Exposure: Moderate
Main risk: Advanced packaging overcapacity
Offset: Makes chips for nearly everyone

LRCX - Lam Research
Segment: Equipment (memory-weighted)
Exposure: Lower, lagged
Main risk: HBM capacity expansion slows
Offset: Broad memory and logic demand

ASML - ASML
Segment: Lithography equipment
Exposure: Lower, lagged
Main risk: Leading-edge expansion pauses
Offset: EUV monopoly pricing

KLAC - KLA
Segment: Process control equipment
Exposure: Lower
Main risk: Fab capex slowdown
Offset: Needed at every node

AMAT - Applied Materials
Segment: Equipment (broad)
Exposure: Lower
Main risk: Fab capex slowdown
Offset: Mature nodes; China the bigger driver

SNPS / CDNS - Synopsys / Cadence
Segment: Chip design software
Exposure: Low
Main risk: Fewer frontier designs
Offset: More custom and edge chip designs

ARM - Arm
Segment: CPU architecture / royalties
Exposure: Low / possible beneficiary
Main risk: Data-center CPU slowdown
Offset: Edge and on-device inference growth

QCOM - Qualcomm
Segment: Mobile and PC chips
Exposure: Low / possible beneficiary
Main risk: Minimal
Offset: Leading on-device AI chips

INTC - Intel
Segment: CPUs / foundry
Exposure: Least
Main risk: Own execution, not this thesis
Offset: Possible edge upside


Based on business mix and contracts as of mid-2026, not current valuations. Not financial advice.