Howard Marks of Oaktree Makes a 180-Degree Turn on AI After Claude Tutorial. Here’s How He Suggests Investors Approach It
Howard Marks has built his career on skepticism, second-level thinking, and disciplined risk control. So when a hands-on tutorial with Anthropic’s Claude reportedly pushed him from guarded observer to pragmatic optimist on artificial intelligence, the shift wasn’t about hype. It was about evidence: tangible productivity gains, real-world workflows improved, and the unmistakable sense that AI is crossing from promise to practice.
What follows is an investor’s playbook—consistent with Marks’s long-standing philosophy—for approaching AI with clear eyes, a margin of safety, and an appreciation for cycles. It is a synthesis of Marks’s principles and current AI market dynamics, not a verbatim report of any single memo or interview.
What changed: from abstraction to application
– Utility over novelty: Hands-on exposure turns AI from a concept into a tool. When a system reliably drafts, analyzes, summarizes, reasons, and iterates with domain context, it ceases to be “tech for tech’s sake.”
– Productivity is the North Star: Investment returns ultimately follow cash flows. If AI raises output per knowledge worker, shortens cycles, and reduces waste, it can move income statements and valuations.
– Diffusion beats disruption theater: The bigger story may be broad-based adoption across incumbents, not only startups. That implies both opportunities and complexity across the stack.
The AI investing playbook, in Marks’s style
– Practice second-level thinking: Don’t ask “Is AI big?”; ask “Who captures the economics, when, and on what terms?” Separate narrative from net present value.
– Anchor to base rates: Most tech waves overshoot early, suffer a shakeout, then compound. Expect dispersion. Few capture outsized, durable profits.
– Seek bottlenecks and tollbooths: In periods of gold rush, scarcity earns rents. Watch compute, power, bandwidth, top talent, and high-quality proprietary data.
– Respect cycles and capital intensity: Semiconductors, memory, networking, and data centers are cyclical, with boom-bust capex. Good businesses at bad prices are still bad investments.
– Favor picks-and-shovels—prudently: Equipment, tooling, power, cooling, and foundational infrastructure can offer clearer cash-flow visibility than application-layer bets, but they’re not immune to supply responses.
– Focus on unit economics, not TAM: Track gross margins after compute, inference cost per user, payback periods, and pricing power. Beware “freemium” products subsidized by cheap capital.
– Build scenario-weighted models: Map bull/base/bear paths for compute demand, model efficiency gains, and regulatory risk. Assign probabilities and demand a margin of safety.
– Time diversification over market timing: Average into positions, especially in cyclical upstream names. Avoid pro-cyclical buying at peak optimism.
– Risk control before return: Size positions so you can be wrong and survive. Accept that dispersion will be high; avoid single-point failures.
– Use AI to improve the investment process: Let models accelerate research, parsing, and variant perception—but keep human judgment in the loop.
Where the opportunities may be
– Compute and networking: Accelerators, advanced packaging, high-bandwidth memory, optical interconnects, and Ethernet/InfiniBand fabrics. Opportunity is real but cyclical; watch supply responses and pricing power.
– Power and the grid: Utilities and independent power producers with credible data-center growth pipelines; transmission upgrades; backup/peaking capacity; next-gen cooling and energy efficiency. Regulatory timelines and fuel spreads matter.
– Data centers and real assets: Land near cheap, reliable power; high-density colocation; thermal management; modular builds; fiber backbones. Lease economics and tenant creditworthiness are key.
– Software infrastructure: Model orchestration, retrieval, vector databases, security, and compliance. Prefer vendors embedded into workflows with high switching costs.
– Application layer: Vertical software where AI can 10x outcomes—customer support, sales ops, coding assistance, design, legal review, drug discovery. Look for clear ROI, low churn, and net revenue expansion.
– Services and integration: Systems integrators and consulting with repeatable playbooks and industry depth. Revenue quality beats headcount growth.
– Credit and structured solutions: Financing for data centers, GPUs, and energy projects; asset-backed lending against hardware; convertibles for late-stage companies; secondaries in down rounds. Spreads can compensate for uncertainty if covenants are strong.
Red flags and failure modes
– Compute-cost gravity: If unit economics depend on perpetually falling compute costs, beware. Algorithmic efficiency gains can swing who captures value.
– Vendor concentration: Overreliance on one supplier or hyperscaler can turn into a margin handoff.
– Open-source pressure: Foundation models and tooling commoditize faster than expected, compressing pricing.
– Capital-cycle whiplash: Overbuilding in memory, networking, or data centers can crater margins.
– Regulatory and legal risks: Data rights, safety standards, and sector-specific rules (healthcare, finance, defense) can delay deployments or reshape economics.
– Vanity metrics: DAUs and demos without paying cohorts, expansion, or durable gross margins.
– Fragile moats: Products with weak switching costs, no proprietary data, or easy replication.
Portfolio construction and process
– Barbell exposure: Own a core of high-quality compounders with AI tailwinds (incumbent software, infrastructure leaders) balanced by selectively priced cyclicals and special situations. Keep a cash or short-duration buffer for volatility.
– Baskets over heroes: Where outcomes are power-law distributed, diversify across a curated basket instead of single-name heroics.
– Stage entries: Add on execution proof points—enterprise adoption, stable gross margins after compute, and expanding cohort profitability.
– Hedge thoughtfully: Consider pair trades (e.g., leaders vs. overhyped followers), tail hedges during capex frenzies, and duration matching for credit exposures.
– Monitor leading indicators: Power interconnect queues, HBM supply, model efficiency benchmarks, enterprise AI budgets, and regulatory milestones. Be ready to rebalance when they inflect.
What would make him change his mind again
– Persistent ROI shortfall: If productivity gains don’t translate into sustainable cash flows, reset expectations.
– Energy bottlenecks harden: Structural power scarcity or transmission constraints cap growth longer than modeled.
– Rapid commoditization: If open models and tooling undercut pricing faster than adoption scales, margins compress.
– Safety or legal events: High-profile failures triggering a regulatory re-think could slow diffusion materially.
Bottom line
A hands-on encounter can turn abstract excitement into practical conviction. Through Marks’s lens, the right approach to AI isn’t maximalist or dismissive—it’s pragmatic. Focus on cash flows and constraints. Prefer bottlenecks and real moats over narratives. Respect cycles. Use scenario analysis and insist on a margin of safety. And remember that, in a gold rush, discipline—not just shovels—separates durable winners from the crowd.
