Skip to main content
AISupercycle
Projects

Thesis sourcebook

The bull and bear cases

A balanced reading list for the AI trade. Each note preserves the author’s argument, maps it to the thesis, and makes the evidence or falsifier worth monitoring explicit.

Bull views
2
Bear views
2

Constructive evidence

Bull views

The strongest case for sustained AI infrastructure demand, alongside the conditions that still have to hold.

X POSTBULL view

SemiAnalysis X thread on Kimi K3 and linear attention

SemiAnalysis · SemiAnalysis

Published Jul 17, 2026 · submitted Jul 18, 2026

gpumemorynetworkingThesis MIXED

TL;DR

Reviewed: Full eight-post X thread

SemiAnalysis argues that Kimi K3 is not bearish for AI infrastructure despite using less KV-cache capacity. The thread's case is that K3's 2.8-trillion-plus parameter footprint favors rack-scale accelerators, distributing its weights raises interconnect demand, and limited HBM headroom pushes cache storage into DDR5 and NVMe. The bullish conclusion still depends on efficiency expanding total AI usage enough to offset lower cache intensity.

Summary points

  • — K3's parameter count requires a large scale-up domain to hold its weights; SemiAnalysis presents that as a favorable workload for NVL72-class systems.
  • — The thread says KDA can reduce networking needed for KV-cache transfers, while WideEP weight distribution can increase traffic across GPUs.
  • — WideEP spreads 896 experts across accelerators. SemiAnalysis argues rack-scale copper fabrics are better suited to that pattern than lower-bandwidth multi-node systems.
  • — Because the weights reportedly occupy more than 1.5 TB of HBM, the thread expects some KV-cache storage to spill into CPU DDR5 and NVMe even at relatively low concurrency.
  • — The author says optimal K3 inference needs a rack with a scale-up domain of at least 64 chips, reinforcing the rack-scale rather than single-server read-through.
  • — Investment read-through: constructive for GPU systems and networking, with demand shifting across HBM, DDR5, and NVMe rather than disappearing. It remains mixed evidence until real deployments validate utilization and aggregate infrastructure spending.
  • — Watch: production node counts, network throughput, the HBM-to-DDR5/NVMe memory mix, accelerator utilization, and tokens served. The thesis weakens if per-workload savings outpace usage growth.

Editorially published Jul 18, 2026.

↗
PDFBULL view

Situational Awareness: The Decade Ahead

Leopold Aschenbrenner · Situational Awareness

Published Jun 6, 2024 · submitted Jul 18, 2026

power coolingsuppliersadvanced packagingmemorygpunetworkingThesis CONTEXT

TL;DR

Reviewed: Full PDF; infrastructure and supply-constraint sections emphasized

Aschenbrenner's report is a scenario map for a multi-year AI infrastructure buildout: power, data-center construction, HBM, advanced packaging, and networking can become binding constraints. For the equity thesis, those projections are useful as a checklist—not evidence that the buildout or its returns have already materialized.

Summary points

  • — The report frames the AI buildout as a full-stack capital cycle rather than a GPU-only story.
  • — Power delivery, sites, and data-center construction can set the physical ceiling on how quickly compute is deployed.
  • — HBM, advanced packaging, and networking translate the high-level compute thesis into measurable supply-chain bottlenecks.
  • — Its forecasts are author scenarios from 2024. They should be tested against realized capital spending, power procurement, project starts, lead times, and capacity additions.
  • — Investment read-through: infrastructure scarcity can support suppliers across several layers, but utilization and monetization must justify the capital intensity before constraints ease.
  • — Watch: capex, grid interconnections, HBM and packaging lead times, deployment utilization, and financing. The scenario weakens if efficiency or weak demand reduces infrastructure needs faster than new workloads expand.

Editorially published Jul 18, 2026.

↗

Challenge evidence

Bear views

The best arguments against the cycle, translated into observable failure modes rather than dismissed as generic skepticism.

SUBSTACK POSTBEAR view

Peak Cheap: The AI Boom Isn't 2000, It's 2008

Groundbreaker · Groundbreaker

Published Jun 7, 2026 · submitted Jul 19, 2026

power coolingsemiconductorsgpuneocloudsThesis CHALLENGES

TL;DR

Reviewed: Full publicly accessible article; no paywalled content accessed

Groundbreaker argues that the AI buildout resembles a credit and earnings bubble more than a dot-com valuation bubble. The bear case rests on three linked claims: reported profits are flattered by long depreciation lives and financed demand, short-lived GPU collateral is supporting longer-lived debt, and end-user AI revenue may arrive too slowly to justify the buildout. The article is most useful as a financing and return-on-capital stress test, not as proof that a credit unwind has begun.

Summary points

  • — The author distinguishes a price bubble from an earnings bubble, arguing that under-depreciation, non-cash markups, and supplier-funded demand can make headline valuation multiples look safer than the underlying economics.
  • — The article maps AI financing through neocloud debt, special-purpose vehicles, long leases, project finance, and GPU-backed loans, then compares the distribution of that exposure with pre-2008 originate-to-distribute structures.
  • — Its central asset-liability concern is duration mismatch: rapidly depreciating accelerators may support loans and leases that outlast their economic usefulness.
  • — The proposed unwind is reflexive: weaker AI returns reduce marginal spending, lower utilization and rental rates, weaken collateral and debt-service coverage, and force additional capacity into the market.
  • — The author explicitly identifies falsifiers: strong end-user AI revenue, durable older-GPU rental values, better-capitalized debt structures, or hyperscaler balance sheets absorbing losses without contagion.
  • — Investment read-through: the bear case challenges neoclouds and the financing-dependent parts of the GPU and data-center chain first; it does not by itself establish that every supplier or committed project is impaired.
  • — Watch: end-user AI revenue and margins, accelerator utilization and rental prices, asset useful-life assumptions, debt-service coverage, loan-to-cost, hyperscaler guarantees, project cancellations, and credit spreads.

Editorially published Jul 19, 2026.

↗
SUBSTACK POSTBEAR viewPreview only

The Heretic’s Guide to AI’s Stars Part III: Tracepalooza & the Bezzle

Michael Burry · Cassandra Unchained

Published May 22, 2026 · submitted Jul 19, 2026

semiconductorsmemorygpuneocloudsThesis CHALLENGES

TL;DR

Reviewed: Public preview through the paid subscription boundary; no paywalled content accessed

In the public preview, Michael Burry argues that Nvidia's demand is unusually concentrated, amplified by a temporary training and benchmarking phase, and transmitted into bespoke supply commitments and data-center financing. He points to Microsoft deployment constraints and Nvidia receivables as signs that purchases may be running ahead of deployed capacity. Because the remainder is paywalled, this note records the preview's testable claims and does not present itself as a review of the complete article.

Summary points

  • — Burry's accessible argument centers on customer concentration: he says Nvidia is more exposed than prior networking leaders to a slowdown or merely slower growth at one large buyer.
  • — He compares the current setup with older DRAM cycles, where concentrated customers and a single demand theme increased the amplitude of the cycle; he discloses a short view on the Philadelphia Semiconductor Index.
  • — The preview argues that Nvidia's custom TSMC commitments are less fungible than the components Cisco once sourced, raising write-down risk if demand slows.
  • — Burry highlights Microsoft capacity constraints and construction-in-progress accounting, then questions why one large customer's Nvidia receivables grew faster than Nvidia's total revenue and receivables in his analysis.
  • — Investment read-through: this is a concentration and deployment-timing challenge to the semiconductor bull case, not evidence that aggregate AI demand has already reversed.
  • — Watch: customer revenue and receivables concentration, forward purchase commitments, hyperscaler capex and cancellations, installed-versus-purchased accelerators, capacity utilization, custom-supply write-downs, and the breadth of semiconductor demand.

Editorially published Jul 19, 2026.

↗

Editorial contract

From link to thesis evidence

This is a thesis tracker, not a social feed. A source earns a published note only when the interpretation is attributable, falsifiable, and useful for a specific layer of the AI stack.

  1. 01

    Preserve

    Store the canonical source and provenance without reproducing the full work.

  2. 02

    Categorize

    Map the source to relevant sectors and five-layer thesis facets.

  3. 03

    Analyze

    Separate the author's argument from our interpretation and testable read-through.

  4. 04

    Publish

    Release only after an administrator approves both the source and the exact note.