Open Source AI — September 24, 2026
Two stories today, both about money. OpenAI and Anthropic slashed their AI prices within 90 minutes of each other this week - a sign they're fighting harder to keep customers as cheaper, freely downloadable AI models close in on them.
In Plain English: Two stories today, both about money. OpenAI and Anthropic slashed their AI prices within 90 minutes of each other this week - a sign they're fighting harder to keep customers as cheaper, freely downloadable AI models close in on them. At the same time, the investors who lend money to build AI data centers are getting pickier, demanding better terms and bracing for a wave of new borrowing next year. Today's Apple Angle connects the two: while everyone else fights over rented compute and thin margins, Apple's approach lets you just buy the machine and skip the fight.
OpenAI and Anthropic Cut Prices Within 90 Minutes of Each Other - an Economist Calls It a Race to Zero Margin
Anthropic released Claude Opus 5.5 on September 22 at up to 40% cheaper than Opus 5 (input tokens now $4 per million, output $20 per million) and scrapped its five-hour usage caps, while OpenAI answered 90 minutes later with GPT-6 Sol and Luna priced at roughly half the prior generation ($2/$10 per million tokens for Sol, $0.10/$0.50 for Luna). Ramp's lead economist Ara Kharazian put it bluntly: "OpenAI and Anthropic are engaged in a price war that is driving down the price of AI and therefore driving down their ability to profit from it" - and reporting on the cuts points to growing pressure from cheaper open-weight models as a real driver, not just rivalry between the two labs. For Lucien's thesis, this is prediction #1's margin-compression risk landing in an actual pricing decision, just from an unexpected direction: retail prices are falling under open-weight competitive pressure even as the underlying infrastructure costs (see below) keep climbing - a squeeze from both sides at once.
Infrastructure & Economics · Fortune · Sep 22, 2026
Bond Investors Turn Choosy on AI Debt as Goldman Projects a Record $420 Billion Hyperscaler Sale Next Year
A new market report finds bond investors growing sharply more selective about AI-linked corporate debt even as they eagerly buy bonds from traditional sectors: Alphabet needed a "large concession" to complete its August debt sale, and Meta and Alphabet bonds now trade at wider spreads than similarly-rated peers despite strong cash generation - AI-related issuers average around 115 basis points versus 78 for the broader investment-grade market, with some double-A credits pricing closer to triple-B levels. Goldman Sachs is projecting hyperscaler debt issuance will hit a record $420 billion next year, up 60% from 2026, even as investors say they need "very high" conviction given the "coming supply and lack of visibility." For Lucien's thesis, this is prediction #1 showing up as an actual price on capital - the same credit markets funding the buildout are starting to treat hyperscaler AI debt as a distinct, riskier category rather than ordinary blue-chip paper.
Infrastructure & Economics · Reuters (via Investing.com) · Sep 22, 2026
The Apple Angle (Standing Perspective)
Today's two stories are really the same problem seen from opposite ends of the ledger: commercial AI providers are cutting prices to survive competition, while the lenders financing their data centers are getting nervous about ever being repaid. Apple's Mac Studio sidesteps both problems by turning AI compute into a one-time purchase instead of a metered, financed liability. Configure one with its full 512GB of unified memory - about $9,500 - and it holds an entire trillion-parameter open-weight model in memory on a single desktop, with no subscription to renegotiate and no bond covenant to satisfy.
Matching that memory footprint on Nvidia's own workstation line takes five or six RTX Pro 6000 cards, somewhere in the $60,000-$75,000 range, while burning roughly ten times the electricity. A new frontier-capable open-weight model ships every few weeks, which keeps nudging the real competitive question away from who trained the smartest model and toward who can afford to own the hardware that runs a nearly-as-good one at home. Apple's MLX software and Thunderbolt 5's RDMA networking already let several Mac Studios pool memory into one larger effective machine, Nvidia has dropped NVLink from its own consumer workstation cards, and its DGX Spark borrows more from Apple's unified-memory design than the reverse. The honest caveat holds every day this section runs: a Mac Studio serves one person running one model at a time, while the gigawatt-scale facilities behind today's bond-market jitters exist to serve millions of concurrent users at once - Apple looks positioned to quietly own a valuable slice of AI compute, not to replace the data center outright.
Limited Edition Jonathan (Substack) · Perspective
Why This Matters
Today's items pull from both ends of the AI money pipeline at once - commercial API prices falling under competitive pressure while the debt financing the underlying infrastructure gets priced as riskier - and that's the scissors motion prediction #1 expects to eventually squeeze commercial margins. The standing Apple Angle keeps pointing at the one player selling a way to opt out of that squeeze entirely: buy the box, skip the metering and the bond market both.
The Daily: Open Source AI - a recurring research brief for Lucien Engelen
Book Lucien for your next event
Keynotes, masterclasses, panels and board-room sessions.