Research report 2026-05-11 – 2026-07-03

The AI Capital Cycle: Bubble or Supercycle?

6 min read 7 episodes 2 podcasts 27% cited

Updated 2 weeks, 5 days ago

Photo by Google DeepMind on Pexels

Overview

Hyperscalers are on pace to spend $600–700 billion on AI infrastructure in a single year, yet enterprise customers cannot demonstrate ROI and CFOs are pulling back. Scott Galloway draws a direct parallel to the 1999 dot-com crash — B2C stalled first, then B2B, and now infrastructure is next — and predicts a basket of leading AI infrastructure stocks including NVIDIA, CoreWeave, and Vertiv will fall 20–40% within 12 months. A dissenting supercycle case is anchored by radical cost-reduction bets: space-based data centers cooled by the void and powered by solar could deliver AI inference tokens at 50–200% lower cost than ground-based compute, while a $920 million-per-month Google–SpaceX compute deal signals that at least some hyperscalers are making multi-year infrastructure commitments regardless.

Most represented

Top voices

  • Pivot 9
  • My First Million 6
  • Donald Trump 91
  • Elon Musk 64
  • Sam Altman 17
  • Bill Maher 12
  • Nancy Pelosi 12

The arguments

Supercycle — Radical Cost Curves Change the Math

Space-based GPU clusters, cooled by the void and solar-powered, could cut AI inference costs 50–200% versus ground-based compute, unlocking demand at a scale that justifies the current build-out.

Bubble — Capex Vastly Outpacing Demand

Scott Galloway argues the $600–700B annual infrastructure spend mirrors the dot-com infrastructure overbuild, with enterprise ROI absent and CFOs visibly retreating. He predicts AI infra stocks fall 20–40% within 12 months.

Permitting Constraints Make Alternative Infrastructure Inevitable

Local government opposition to terrestrial data centers — not just economics — is pushing innovators toward unconventional infrastructure (space, aesthetically integrated hubs), suggesting the build-out will find new channels rather than simply deflate.

Government Backstop Will Mask the Bust

Galloway predicts the biggest bailout since the banks will arrive disguised as a 'growth opportunity' — a government backstop for AI companies whose infrastructure commitments have outrun actual demand.

Smart Money Is Already Positioned Long

Leopold Aschenbrenner, a former OpenAI researcher who published the influential 'Situational Awareness' white paper, launched a hedge fund at 22 that has grown to ~$5 billion — betting the AI infrastructure thesis holds.

Mixed verdict

Selective Survival — Disciplined Operators Will Break Even

Anthropic projects it will break even by 2030 despite massive current losses, contrasting sharply with OpenAI's open-ended trillion-dollar CapEx commitments — suggesting discipline, not scale, determines survivors.

The Capex Mountain: $600–700 Billion in a Single Year

The largest cloud companies are collectively on pace to spend $600–700 billion on AI infrastructure in 2025 alone. That figure — concentrated in data centers, chips, and networking — dwarfs any prior technology capital cycle and is the central fact around which the bubble-versus-supercycle debate orbits. For context, the entire US electrical grid runs at roughly 1 terawatt; the AI CEO Award proposed by Elon Musk requires delivering 100 terawatts of compute per year from non-Earth data centers — 100 times that capacity.

The scale of commitment is not in dispute. What divides analysts is whether demand will grow into it. CFOs at enterprise customers are already pulling back because they cannot demonstrate ROI on current AI deployments, and the gap between hyperscaler supply promises and enterprise willingness to pay is widening in real time.

Galloway's 1999 Playbook: Infrastructure Is the Last Shoe to Drop

Scott Galloway has mapped the current AI cycle directly onto the dot-com crash sequence: consumer applications stall first, then enterprise B2B, and finally the infrastructure layer collapses under the weight of stranded assets. His prediction is specific — a basket of AI infrastructure stocks including NVIDIA, Astera Labs, Marvell, Vertiv, Supermicro, CoreWeave, and the hyperscalers themselves will fall 20–40% within 12 months.

Galloway's most provocative extension of this thesis is that the bust will not be allowed to look like a bust. He predicts the largest government backstop since the 2008 bank bailouts will arrive rebranded as a 'growth opportunity' — a quasi-fiscal intervention to protect companies whose infrastructure commitments have structurally outrun demand. The political economy of this scenario matters: if the government absorbs the downside, the capital misallocation signal is suppressed and the cycle extends.

The Bull Case: Cost Curves, Space, and Locked-In Demand

The supercycle argument rests less on current demand and more on anticipated cost collapse. Shaan Puri of My First Million articulates the most radical version: building a data center in space is now arguably easier than securing local government permits for a terrestrial one, and solar-powered, vacuum-cooled GPU clusters in orbit could deliver AI inference tokens at 50–200% lower cost than ground-based equivalents. If that cost curve materialises, it unlocks demand tiers — consumer, developing-world, real-time inference at scale — that current pricing leaves inaccessible.

Concrete deal flow supports the multi-year commitment thesis. Google agreed to pay SpaceX $920 million per month over three years for computing power, including access to at least 110,000 NVIDIA chips. That is not speculative — it is a signed contract representing roughly $33 billion in committed compute spend from a single counterparty. Deals of this structure suggest at least some hyperscalers have internalised the supercycle thesis operationally, whatever their public messaging about ROI discipline.

Terrestrial Friction: Permitting, Aesthetics, and Community Resistance

A structural constraint the purely financial analysis misses is local opposition to data center construction. Communities are pushing back against the aesthetic and environmental footprint of industrial-scale compute facilities, creating a permitting bottleneck that independent of capital availability limits how fast terrestrial infrastructure can be built. This friction simultaneously validates the space-based data center thesis and raises the floor for returns on approved ground-based sites.

One proposed response — architecturally beautiful, community-integrated hubs — signals that the industry is beginning to internalise NIMBYism as a real cost, not a minor friction. The implication for the capital cycle is that supply will be more constrained than raw capex commitments suggest, which could actually prevent the worst overbuilding scenarios Galloway fears.

Survival Discipline: Anthropic vs. OpenAI as a Case Study

Within the AI application layer, the contrast between Anthropic and OpenAI illustrates a fork in capital strategy. Anthropic projects breaking even by 2030 despite current losses — a defined glide path with a visible endpoint. OpenAI, by contrast, has made open-ended trillion-dollar CapEx commitments with no comparable break-even horizon. If Galloway's correction thesis is correct, disciplined operators with defined runways survive while those betting on indefinite capital availability face existential risk.

Leopold Aschenbrenner represents an interesting data point at the intersection of insider knowledge and capital markets. The former OpenAI researcher who authored the 'Situational Awareness' white paper launched a hedge fund at 22, grew it to approximately $5 billion, and found the overlapping consensus on AI infrastructure worth a multi-billion-dollar long bet. The fact that someone with direct access to OpenAI's internal trajectory chose to build a large fund around the AI thesis — rather than hedge against it — carries at least some evidential weight against the pure bubble narrative.

Open questions

Angles still unanswered — threads worth pursuing.

  1. If the US government does backstop AI infrastructure companies as Galloway predicts, what is the mechanism — loan guarantees, strategic reserve purchases of compute, or direct equity stakes — and how does that reshape the private capital cycle going forward?

    The form of any government intervention would determine whether the misallocation is corrected or perpetuated, and no podcaster has yet mapped the specific policy instruments.

  2. At what inference cost-per-token does space-based compute actually unlock the demand tiers needed to justify the build-out, and what is the realistic timeline for SpaceX to reach that price point?

    The entire supercycle bull case pivots on a cost curve that remains speculative; quantifying the threshold would transform the debate from qualitative to testable.

  3. Is the dot-com analogy structurally valid given that hyperscalers — unlike 1999 telcos — are also the primary customers of the infrastructure they are building, creating a vertical integration that changes the default-and-abandon dynamic?

    Galloway's 1999 parallel assumes infrastructure builders and demand aggregators are separate entities; the current concentration may mean the bust, if it comes, takes an entirely different form.

  4. How much of the $600–700 billion annual AI capex is contractually committed versus discretionary, and what does the cancellation-clause landscape look like for deals like the Google–SpaceX $920M/month contract?

    The difference between committed and discretionary spend is the difference between a slow-motion restructuring and a sharp crash, yet no public breakdown of this figure appears to exist.

  5. Can the 'aesthetic data center' model — community-integrated, architecturally designed facilities — achieve the energy density and cooling efficiency required for frontier AI training workloads, or is it only viable for inference?

    If permitting friction is genuinely structural, the viability of community-friendly designs at training scale determines whether terrestrial supply can grow at all.

References

  1. Pivot Trump's Crypto Windfall, Dems' Anti-Establishment Wave, and the Supreme Court’s Big Week — Scott Galloway “Hyperscalers spending $600–700B on AI infra in 2025: The largest cloud companies are on pace to spend $600–700 billion on AI infrastructure in a single year, yet enterprise customers are struggling t…”
  2. Pivot Trump's Crypto Windfall, Dems' Anti-Establishment Wave, and the Supreme Court’s Big Week — Scott Galloway “AI infra stocks predicted down 20–40% in 12 months: Scott Galloway predicted a basket of leading AI infrastructure stocks — NVIDIA, Astera Labs, Marvel, Vertiv, Supermicro, CoreWeave, plus hyperscale…”
  3. Pivot Trump's Crypto Windfall, Dems' Anti-Establishment Wave, and the Supreme Court’s Big Week — Scott Galloway “In the dot-com crash, it went B2C first, then B2B, then infrastructure. Scott Galloway sees the same pattern now: AI application spending is stalling, demand is not as exponential as promised, and in…”
  1. Pivot Trump's AI Stake, SpaceX's IPO Froth, and Apple's Siri Overhaul — Scott Galloway “Scott Galloway predicts the biggest bailout since the banks will come disguised as a 'growth opportunity' — a government backstop for AI companies whose infrastructure commitments outpace their actua…”
  2. My First Million The most simplified breakdown of the SpaceX IPO on the internet — Shaan Puri “Building a data center in space is easier than getting local government approval to build one on land. Elon's thesis: solar-powered, space-cooled GPU clusters that deliver AI tokens at a fraction of …”
  3. My First Million The most simplified breakdown of the SpaceX IPO on the internet — Shaan Puri “SpaceX is three businesses stapled together: a launch monopoly, a satellite internet service (Starlink), and an AI play. You're not buying a rocket company — you're buying the infrastructure layer of…”
  4. Pivot Trump's AI Stake, SpaceX's IPO Froth, and Apple's Siri Overhaul — Kara Swisher “Google–SpaceX compute deal: Google agreed to pay SpaceX $920 million a month over 3 years for computing power including access to at least 110,000 NVIDIA chips.”
  5. My First Million I put 80% of my money in the S&P — Shaan Puri “As the AI race demands massive new power and computing infrastructure, local communities are pushing back against ugly data centers. Building architecturally beautiful, community-centric hubs could s…”
  6. Pivot Comcast Splits, OpenAI Weighs IPO Delay, and Buttigieg Targeted — Scott Galloway “Anthropic projected to break even by 2030: Despite massive losses, Anthropic projects it will break even by 2030, contrasting with OpenAI's open-ended trillion-dollar CapEx commitments.”
  7. My First Million 3 weird businesses doing $10M, $20M, $30M — Shaan Puri “Leopold Aschenbrenner, a former OpenAI researcher who published the 'Situational Awareness' AI white paper, launched a hedge fund at 22 and has grown it to approximately $5 billion. He found the over…”

No sections match your search.

End of report.

Create a new report to explore a different angle.

New report All reports