The build-out goes looking for a lender

The read

The financing story grew a spine today. Jensen Huang expects the next phase of the build-out to be funded by Wall Street rather than corporate balance sheets, and two items read like the machinery for it being assembled. CME plans to list compute futures from 5 October, pending regulatory approval — a public forward curve is precisely what a lender needs before it can price compute-backed debt rather than guess at it. And Riot Platforms signed a 20-year, US$9.1bn agreement with Anthropic covering 191MW in Texas, which is a bitcoin miner turning itself into a landlord on contract terms a credit committee can actually read. Set against both, around US$315bn of hyperscaler infrastructure is built but not yet in service — a reservoir of depreciation that has not begun to run.

Which sharpens a question John has been putting for a while: why are the hyperscalers carrying the data-centre risk themselves when neoclouds exist precisely to carry it for them under long-term contract? If compute becomes hedgeable and third-party capital gets cheaper, that calculation changes for everyone. The constraint underneath is physical: data centres are projected to take about 75% of US power demand growth over the next decade on McKinsey's numbers — consultancy projections rather than measured outcomes, and worth holding loosely. And one small counterexample worth more than its size: a Taiwanese town that blocked a fab has petitioned to get it back after watching TSMC's capital go to Tainan and Kaohsiung instead. That one comes to us as an account rather than a sourced report, but siting fights evidently have a second act.

On the listed side, one name in every software category still clears a large premium — CrowdStrike, Cloudflare, Shopify — and Tom Tunguz's reading is that the market is paying for a credible AI path rather than for AI revenue anyone has actually booked. That makes the premium narrative-dependent, which is the thing to price. A rhyme rather than a connection, but worth noting the same day: OpenAI closed a roughly US$7bn secondary as another senior executive left. A secondary puts no capital into the company — it is liquidity for staff and early holders — but it sets a valuation mark ahead of any listing, and the staff deciding whether to sell into that mark are weighing much the same question the software premium poses: how durable the story turns out to be. Meanwhile Target appointed its first chief AI officer — an appointment Mark tied straight back to the financing thread, on the grounds that the debt markets reopen if AI revenue genuinely shows up at the buyer end.

The open-weights argument got its sharpest airing yet. The disagreement about local models turns out to be a disagreement about what counts as local: Sean Goedecke's economic case says batched hosted inference is some thirty times more efficient, and Mark rejects it from the other end — his own test of Meta's Muse Glimmer, a model rated 35, produced the most faithful recitation he has seen from anything. Benchmarks and usefulness continue to come apart. Meta itself rebooted its AI strategy again and is unwinding the Manus acquisition, and published a superintelligence manifesto whose account of history John disputed. Whether the near-frontier price war is being fought between the frontier labs rather than against the open-weight challengers is still an open question here, not a finding.

Three on agents, and one on supply. xAI is pitching Bots that sign in to your tools and come back with finished work — the vendor's own description, and notably the page does not say whose credentials they hold, which is the entire question for anyone who has to audit what happened afterwards. Google argues Go suits AI agents because it contradicts them quickly, which is Google advocating for Google's language, though the mechanism it describes stands on its own. A site claims frontier reasoning can be recovered from encrypted traces — worked examples now published, independently unverified. And Apple has reportedly lifted its display supply buffer from four weeks to six: second-hand, small in itself, and interesting only because Apple has spent twenty years treating inventory as a cost to be driven out.

Read all 19 signals →