Topics Pioneers Insight Method Author
Back

AI is Heading for a Massive Oversupply: Top Hedge Fund Manager’s Prediction | David Orr

·The Monetary Matters Network — Hand-picked

Rendering the long image…

Topics Trillion-Dollar AI Capex and Bubble Risks SaaS Valuation Reset Under AI Disruption Macro Asset Allocation in High-Interest Regimes Commercial Shifts in Mobile Recommendation Algorithms
Pioneers David Orr

Summary

  • Orr sold all his AI stocks in June because he expects computing supply to outrun usable demand. The risk is temporarily “too much intelligence”: capex extends through 2028 while physical applications take years to scale. If oversupply arrives in 12–18 months, June could have been the peak; if it takes four years, AI stocks could rise another 100% first.
  • Google, Amazon and Microsoft remain Orr’s preferred AI exposure, although he has stepped aside. He describes projects earning 35% returns on capital, with roughly half the capacity already contracted, making “their downside risk pretty small” unless software customers fail rapidly in large numbers. A decade of continuing scarcity would make that investment exceptionally attractive.
  • Replacing expensive human labor does not guarantee lasting software profits or enormous computing demand. Orr expects routine accounting to become almost free within five years; competitors replicate the tools and consumers capture the savings. Companies with durable moats could become a “printing press,” while competitive businesses surrender their improved margins.
  • Scientific breakthroughs could overturn the oversupply thesis, but Orr cannot establish that demand yet. Expanding drug experiments 100-fold requires laboratories, subjects and fresh data, potentially leaving intelligence waiting on the physical world. Pure intellectual breakthroughs might instead send computing prices soaring: “I really don’t know.”
  • Orr sees taxes and capital issuance as clearer market risks than political restrictions on data centers. A 30–50% probability of higher corporate taxes around 2030 could justify the S&P 500’s roughly 4% decline, perhaps 6–7%. An Anthropic IPO would add another liquidity demand: “another blow.”
  • AI threatens incumbent moats beyond software, including search advertising and potentially social media. Orr doubts ordinary searches can sustain Google-style “walls of advertising” once competing models become good enough. Meta’s increasingly entertainment-driven platforms might become commodities within five years, although he explicitly presents that as a pessimistic scenario.
  • Orr’s response is to limit exposure rather than make a large directional short. His fund’s beta has been slightly negative over 30 days, but his target is approximately zero. AI has a real bullish case, and he expects to be wrong “at least 40% of the time”: “I’m just going to stay out of this.”

Deep dive

1. Crowded positioning makes the sell-off hard to diagnose

  • Orr thinks AI shares had been vulnerable since June: too many investors were stuck at high prices, with no new margin buyers. “The market was just waiting to fall.” He describes companies operating with 6–10 times leverage; a 5% Nikkei decline alongside rising Japanese small caps looked like forced liquidation rather than coherent fundamental repricing.
  • Wiethe asks how that stress travels through Orr’s ten portfolio managers. The other seven managers are small enough that even a mistake would not materially affect the fund, Orr says, and the worst result seemed to be about 8%. Mike and Rodrigo made good money. Fund-level beta has been slightly negative over 30 days, although Orr questions whether to trust the figure and aims to remain near zero.
  • The comparison with 2022 has limits: Orr then shorted “really bad” companies, while AI is “a fundamentally real thing” with a genuine bullish scenario. He is not buying the drawdown yet, although he thinks he will probably buy a semiconductor ETF for a short-term trade.

2. Oversupply timing matters more than AI’s long-run promise

  • Orr was bullish through much of the rally before selling all his AI stocks in June. “You can never have too much intelligence,” but other bottlenecks can temporarily leave excess compute. Oversupply in a year or a year and a half could validate June’s peak; a four-year delay could allow another 100% rise in AI shares.
  • His evidence is colossal capex figures through 2028 against demand concentrated in software companies still learning integration. Eventually customers may say, “We don’t really need this,” and resell capacity. He cites small operators whose quoted computing prices fell from $12 an hour in September to about $4. That is potentially important but “very ambiguous”: small-provider capacity is not directly comparable with major contracts.
  • Wiethe notes IGV falling alongside SOXX and SMH after crowded long-hardware, short-software trades. Orr dismisses the supposed fundamental correlation: “These narratives don’t really make any sense.” His separate, multiyear concern is that software without strong moats becomes vulnerable regardless of hardware spending.
  • His preferred exposure is hyperscalers Google, Amazon and Microsoft, although he exited anticipating selling pressure. He describes projects with 35% returns on capital and roughly half the capacity contracted. Downside looks limited unless software customers go bankrupt rapidly and in large numbers; continued intelligence scarcity for ten years would make the returns extraordinary, potentially outweighing weakness elsewhere in their businesses.

3. Taxes and liquidity matter more than the Texas catalyst

  • Orr considers higher corporate taxes a plausible explanation for broader market weakness. If Democrats do well in ’28, tax changes could perhaps arrive around 2030. At a 30–50% probability, a roughly 4% S&P 500 decline seems reasonable: “Maybe it should be 6% or 7%, I don’t know.”
  • A hypothetical government block on 95% of new projects would benefit existing capacity owners. But Orr doubts the Texas election news changes much: power plants were already the bottleneck, and projects can move elsewhere. Beyond taxes, he generally sees limited political consequences, though stronger antitrust enforcement could hurt stocks.
  • Wiethe raises a possible Anthropic IPO before year-end. Orr treats its arrival as conditional, adding that G7 debt issuance, especially in the bond market, is already absorbing liquidity. Another enormous offering would provide a further reason for pessimism.
  • Google is Orr’s likeliest antitrust target, but AI’s threat to monopoly profits already featured in his 2023 letter. Maps and YouTube offer strengths; search faces competing models capable of ordinary restaurant recommendations without “walls of advertising.” He expects eventual decline, though not cable television’s roughly 10% annual contraction.

4. Eliminating work may require surprisingly little compute

  • Orr’s premise is that many digital products do not require extraordinary intelligence. Windows 7 was already “really good”; subsequent versions illustrate how useful functionality can plateau. Similarly, he expects routine office work, including accounting, to become nearly free within five years rather than support limitless computing expenditure.
  • Medical research could consume enormous resources, but each promising computational direction still requires experiments and fresh data. Increasing experiments 100-fold means building laboratories over many years. “No one knows for sure how this will work,” Orr emphasizes; his expected bottleneck is physical execution rather than intelligence.
  • Wiethe pushes on basic science, including solid-state batteries: could AI revive breakthroughs achieved largely through thought? Orr allows that sufficiently superior intelligence might do so, sending computing prices sharply higher. Wiethe also raises cancer-treatment narratives supporting a $1.5 trillion Anthropic valuation. Orr hopes the applications work but says, “I really don’t know.”
  • Wiethe’s Progressive Insurance example tests who captures labor savings. Orr’s base case is that most office jobs disappear. A business protected by scale or network effects could cut employment and become a “printing press”; a competitive business loses its bureaucratic barrier and passes savings onward. Neither outcome, he argues, necessarily consumes enough compute to justify hyperscaler capex.

5. A free terminal illustrates software’s threatened pricing

  • Wiethe asks whether vertical AI can charge for outcomes rather than usage. Orr allows that it may look attractive “for a few years,” but expects increasingly capable AI to reproduce straightforward solutions within five years. His fund has already built a risk system and terminal: “Maybe it’s not as good as its professional counterparts, but I like it better.”
  • Orr envisages users connecting an IBKR account and receiving automated portfolio-risk analysis, perhaps requiring virtually no setup within a year or two. He expects to offer the product free, possibly with users paying their own token costs, and gain marketing exposure through its logo. Wiethe questions whether Orr’s initiative generalizes; Orr says there may be a dozen others who build similar tools and give them away.
  • Bloomberg’s historical data-cleaning effort looks vulnerable to AI processing old annual reports; Orr bets he will trust AI-generated numbers more within a few years. Against incumbent prices of $25,000–$35,000 per user annually, he imagines a dozen people offering free alternatives. Wiethe preserves a distinction: institutional IB chat, broker access and recorded transactions retain value beyond interchangeable data.

6. Better abundance does not settle the labor or platform debate

  • Orr hopes displaced talent builds better physical environments rather than optimizes advertising. His concrete example is a partner’s hotel beside a national park, designed by architects to fit the landscape, but delayed by higher interest rates. The aspiration is more thoughtfully designed places to live.
  • Wiethe challenges the transition: standardized housing deliberately minimizes costs, displaced workers could depress wages, and professional status does not automatically transfer into craftsmanship. Orr still imagines millions finding enjoyable work producing things people value, at decent incomes relative to their alternatives. Wiethe calls it a “beautiful vision” while remaining unconvinced about the path after widespread job losses.
  • On Meta, Wiethe distinguishes voluntarily engaging with social content from frustrating search results. Orr worries that platforms are becoming “less social every year”: “Now it’s just media.” Without the old social dependence, platforms might become commodities in five years. He explicitly acknowledges the bullish alternative—continued usage and more engaging AI-enhanced content.
  • Wiethe argues that creator-audience network effects and endless scrolling can be serious barriers, while also suggesting niche services could draw users away. Orr says social media has a network effect, but sees Spotify’s recommendation algorithm as better than TikTok’s or Instagram’s, which push addictive content users may not even like. He speculates that AI might learn a person’s preferences and make better recommendations.