Research · Posts

Evenrock Signal — Week of May 3, 2026

2026-05-04·podcast digest·draft

Top insight of the week — worth being wrong about: Frontier AI model time-horizons are now doubling every ~4 months, not 6–7, and Anthropic's Opus 4.6 reportedly succeeds 50% of the time on tasks that take skilled humans 12 hours — nearly 2x the prior frontier (GPT-5 Codex at ~5h50m). If the 4-month doubling holds for even another 18 months, the question stops being "will AI replace junior analysts" and becomes "what does a fund look like when a single agent runs a week-long diligence loop unsupervised." METR's Joel Becker is appropriately humble on the absolute numbers (n≈3 human baselines per task), but emphatic that the rate has held across model generations and that data center commitments through 2027–28 are already baked in. Bet against the chart at your own risk.

"R&D spend on compute has risen exponentially at essentially the same rate as time horizon progress." — Joel Becker, Odd Lots: Understanding the Most Viral Chart in AI

The capex paradox: AI is printing money and incinerating cash flow simultaneously

Q1 2026 hyperscaler earnings put the AI capex cycle into sharp relief. The bull case: AWS reaccelerated to 28% (its fastest in 15 quarters at a $150B run-rate), Google Cloud grew 63% on $20B+, and Microsoft's RPO hit $625B (up 110%, ~45% from OpenAI). The bear case: Amazon's free cash flow fell 97%, Meta raised capex guidance another $10B on each end ($125–145B), and Amazon alone committed $200B in 2026 capex — the first company ever to cross that threshold.

  • Combined 2026 capex guidance across AMZN, GOOGL, MSFT, META is approaching $580B, with the four-firm total trending toward $700B annually. (Calacanis, All-In)
  • The MAG-4 trade at 17–25x forward earnings; compare Microsoft's 73x and Cisco's 200x at the dot-com peak. The mega-caps are not the bubble. (TBPN)
  • AMD trades at 130x, Palantir at 220x, and at least one AI-adjacent name at 900x. The bubble is real but localized. (TBPN)
  • Chamath: hyperscalers will "look like bulky industrial businesses in five years" as they lever up to fund capex that no longer fits inside operating cash flow.

"Cloud backlog nearly doubled, more than $460 billion. More than half of that backlog is to be recognized in the next 24 months." — TBPN, on Google Cloud

Compute is the moat — and the bottleneck

Two converging signals: AI demand is supply-constrained, not demand-constrained, and the constraint is shifting from training weights to inference capacity. This is good for hyperscalers, bad for pure-play labs, and creates an unexpected re-rating case for legacy silicon.

  • Chamath: less than half of announced gigawatt capacity is actually built; the rest is stuck in red tape. The losers are OpenAI and Anthropic, who must trade equity for compute. The winners are Oracle, Meta, Microsoft, Google.
  • Nomer Brown (cited on TBPN): "Model weights become less important as inference becomes more important — inference capacity becomes a competitive advantage."
  • Mark Lapidus upgraded INTC on the thesis that the CPU-to-GPU ratio could flip from 1:8 to 8:1 as workloads shift to agentic inference. Intel's Q1 data center revenue beat by 13% ($5.1B vs. $4.5B est.) — early evidence the thesis has legs. (TBPN)
  • Microsoft–OpenAI exclusivity is over: OpenAI now ships on AWS too. Azure loses a sales advantage; Microsoft's equity stake benefits from broader distribution. Net-net ambiguous; watch RPO trajectory.
  • Cursor reportedly has –23% gross margins while sitting in a rumored $60B acquisition conversation. The token-economics question for application-layer AI is no longer theoretical.

"By the end of the year, demand for an Opus 4.6 tier model could be $100 billion. It's 40 billion right now." — Patrick O'Shaughnessy (cited on TBPN)

Quality compounding is back in fashion (and the data backs it)

Two excellent investing conversations this week — Compound and Friends on 100-baggers, and Capital Allocators on GSP — converged on the same point: extreme concentration in long-tail outcomes makes selection of durable, high-quality businesses the only game that matters.

  • Only 4% of stocks from 1926–2022 produced 100% of net wealth creation; 84% of the top 50 wealth creators were high-quality by fundamentals. (Matt Ankrim, Compound and Friends)
  • 40% of Russell 3000 stocks suffer a 70%+ drawdown they never recover from. Avoiding losers ≈ finding winners. (Niraj Kislani)
  • Average holding period collapsed from 8 years (1960s) to 5.5 months today — structurally preventing investors from capturing the compounding curve. Buffett made 98–99% of his wealth after age 65. (Ankrim)
  • GSP's roll-up discipline: bolt-ons at 5–8x cash flow into platforms that re-rate to 12–16x; no leverage upfront until $15–20M EBITDA; tech stack adds 200–300 bps of organic growth that flows ~40% to the bottom line. They only invest in businesses ≥17 years old.
  • The 1999 e-commerce playbook didn't work: 15 of the top 20 US e-commerce sites today are old-line retailers that built websites, not pure-play internet companies. Worth holding in mind when projecting AI-native winners.

Geopolitical tail risk is back on the table

  • China ordered Meta to unwind its $2B Manus AI acquisition — an extraordinary extraterritorial intervention. Delian Asparouhov's read: any Chinese-founded AI company is permanently subject to PRC jurisdiction regardless of redomicile. Founder calculus changes immediately.
  • Ike Freiman on Odd Lots makes the case that a Taiwan disruption would be orders of magnitude larger than closing the Strait of Hormuz, with seven mega-caps representing ~40% of S&P market cap and all of them dependent on TSMC. Tracy Alloway's counter-thesis: as US domestic chip capacity comes online, the "silicon shield" weakens — potentially raising the probability of action.
  • China holds 10–20x Russia's pre-sanctions FX reserves, runs an existing capital-control regime, and has migrated its social contract from growth-based to ideology-based legitimacy. The sanctions deterrent is materially weaker than the 2022 playbook suggests.
  • A reminder: TSMC does not share US strategic objectives. They sell to whoever is allowed to buy.

Quick hits worth tracking

  • Retatrutide (LLY) phase 3: 37 lbs avg loss in 40 weeks vs. 6 on placebo, liver fat down 80%, A1C 7.9 → 6.0. Projected approval mid-2027. The GLP-3 cycle is real.
  • Cohere + Aleph Alpha merger. Two enterprise non-hyperscaler AI labs consolidating. Watch for a third.
  • Josh Kushner launched Thrive Eternal — permanent-capital holdco, first asset is the SF Giants. The Berkshire-for-trophy-assets trade is becoming crowded.
  • Private credit: John Sheehan on Odd Lots flagged that funds raised capital before sourcing deals, marks aren't real, and a redemption cycle forces sales of highest-quality assets first. 15% default scenarios from sell-side analysts are not absurd.
  • Vanguard VOO crossed $900B AUM — the first ETF ever — and is on track for $100B in net inflows for the third straight year. Indexing has not slowed despite everything.

Episodes worth your full attention

  1. Odd Lots — "Understanding the Most Viral Chart in Artificial Intelligence" (Joel Becker, Chris Painter, METR). The clearest, most calibrated explanation of where AI capability actually is. The 4-month doubling slide deserves a printout on your desk.
  2. All-In — "OpenAI Misses Targets, Codex vs Claude, Big Hyperscaler Beats". Chamath's compute-as-leverage-over-labs thesis and the financial-engineering trajectory of the hyperscalers are the most useful framing of the AI capex question this week.
  3. Odd Lots — "How Taiwan Became the World's Most Perilous Geopolitical Chokepoint" (Ike Freiman). The most rigorous tail-risk argument on the table for a portfolio that is, like it or not, ~40% AI-trade by S&P weight.