BlackRock just flipped positioning: US equities neutral, overweight short-term Treasuries. Meanwhile, a brutal arbitrage is playing out across the AI stack.
Silicon Data shows token prices down 52% from May peak to $0.98/million tokens. Same period: B300 GPU rental up 38%, B200 up 10%. 10-year Treasury hit 5.2%, highest since 2000. These aren't random—they're the same causal chain.
Model layer getting crushed. OpenRouter data: Chinese top models processed 76.12 trillion tokens in late Sept, US models only 14.2 trillion. Open weights commoditized inference, but compute demand exploded because volume growth outpaced per-token price collapse. Profit shifted from model providers to whoever owns the GPUs.
The GPU buying spree runs on credit. Microsoft, Google, Amazon, Oracle sitting on $2.34T in remaining performance obligations—roughly half from still-unprofitable OpenAI and Anthropic. Cloud providers are effectively vendor financing data centers for loss-making tenants.
Credit has a price: interest rates. BlackRock ended their Treasury underweight (held since Oct 2020 when 10Y was under 0.9%). They're de-risking AI infra exposure via "short-term risk management" but still long AI thesis, waiting for better entry.
PIMCO also extended duration to longest in years. Market priced in 160bps of additional hikes since Jan—arguably overdone. At current yields, Treasuries need another 75bps move up before 12-month total return goes negative.
CPI Wednesday (Oct 14): consensus 3.6% headline, 2.4% core. Whether 5.2% holds depends on this print.
The arbitrage: cheaper models, pricier cards. Money flows to scarce compute. Compute gets funded by credit. Credit costs are interest rates.
Silicon Data shows token prices down 52% from May peak to $0.98/million tokens. Same period: B300 GPU rental up 38%, B200 up 10%. 10-year Treasury hit 5.2%, highest since 2000. These aren't random—they're the same causal chain.
Model layer getting crushed. OpenRouter data: Chinese top models processed 76.12 trillion tokens in late Sept, US models only 14.2 trillion. Open weights commoditized inference, but compute demand exploded because volume growth outpaced per-token price collapse. Profit shifted from model providers to whoever owns the GPUs.
The GPU buying spree runs on credit. Microsoft, Google, Amazon, Oracle sitting on $2.34T in remaining performance obligations—roughly half from still-unprofitable OpenAI and Anthropic. Cloud providers are effectively vendor financing data centers for loss-making tenants.
Credit has a price: interest rates. BlackRock ended their Treasury underweight (held since Oct 2020 when 10Y was under 0.9%). They're de-risking AI infra exposure via "short-term risk management" but still long AI thesis, waiting for better entry.
PIMCO also extended duration to longest in years. Market priced in 160bps of additional hikes since Jan—arguably overdone. At current yields, Treasuries need another 75bps move up before 12-month total return goes negative.
CPI Wednesday (Oct 14): consensus 3.6% headline, 2.4% core. Whether 5.2% holds depends on this print.
The arbitrage: cheaper models, pricier cards. Money flows to scarce compute. Compute gets funded by credit. Credit costs are interest rates.