The white paper, and the money moving toward the build

Thursday 13 August 2026 · our data centre white paper, the board, the read, and today's 14 signals

Out today — the white paper

A Second Bite at the Golden Apple: Australia's data centre opportunity. A generation ago Australia squandered the greatest resources boom in its history. The tree is fruiting again, and this time the resource is the sun, the wind, the land, the stable institutions and the timing. Three parts: where the work actually happens, the objections against data centres taken seriously and answered, and what Australia stands to gain if it walks the middle path.

Free to read, free to download, open to everyone — no membership, no paywall. If you know someone who should read it — in government, in energy, in infrastructure — please send it to them.

Read the white paper →

The board

▲ SMCI 19.0% ▲ DELL 9.9% ▲ ANET 6.4% ▲ ORCL 5.4% ▲ MU 4.9% ▲ LRCX 4.7%

  • Markets — Wednesday's session ran one way in the names we track: Supermicro +19.0%, Dell +9.9%, Arista +6.4%, Oracle +5.4%, Micron +4.9%, Lam Research +4.7%, GlobalFoundries +4.6%. Hardware, memory and equipment led it, and nothing in our slice finished lower.
  • Open weights — Meta's Muse Glimmer 30B now leads Hugging Face trending, ahead of MiniMax H3 and DeepSeek V4 Flash.
  • Models — Opus 5 holds both ends of the model board again: top intelligence at 63.1, best value at $10/M blended.

See the full dashboard →

The read

The money kept moving toward the build. Tencent nearly tripled its second-quarter capital spending — up 176%, with revenue ahead of estimates — and Mark's question about it is the one nobody has answered yet: what is that money actually buying? CoreWeave beat on both lines while its net loss widened to US$626m, and told the call that prior-generation Nvidia SKUs are still renting at or above where they had been. That is the company describing its own book rather than an independent price, but it is the most direct evidence yet on the question underneath the whole depreciation argument — whether older silicon holds its value or falls off a cliff.

Two raises show how far the capital base has spread beyond the usual names. The EU's Scaleup Europe Fund co-led Lovable's US$400m round at a US$13.3bn valuation — the European Commission is the anchor investor in that fund, though EQT manages it and makes the investment decisions, and the distinction matters more than the headline does. And Igor Babuschkin's River AI has raised US$1.1bn to build models "owned and shaped by each of us", with Nvidia and AMD Ventures alongside General Catalyst.

Tim O'Reilly reads Google's August reshuffle as a bet on diffusion rather than on frontier model development — Cloud up 82% on the quarter against AWS at 37% — though the Gemini numbers he leans on are SemiAnalysis estimates rather than reported results, and should be read as such. Meanwhile the containment story reached Congress: 32 House Democrats have written to Sam Altman about the agent that breached Hugging Face, asking him to release the logs and answer by 24 August. An internal evaluation that ran loose on the internet for four days is now an oversight matter with a date on it.

Three good pieces on what agents are doing to the work itself. Brent Fitzgerald argues for constrained use — the human is the loop, and he explicitly declines the addiction framing his essay is usually summarised as making, which is what makes it worth reading. Once testing an idea costs almost nothing, the scarce input becomes pruning rather than generation — the discipline shifts from having ideas to killing them. And Florian Herrengt's bet is that AI pulls engineering salaries apart rather than deleting the jobs: there is a bar you have to clear to be employable, and the bar is whatever the best current model can do.

On the models and the ground they sit on: Grok 4.6 scores 61 on Artificial Analysis, sixth of 183, which raises a question about what a threshold score is still measuring; and llama.cpp has opened llama.app, a consumer front door for local models, so the local-inference argument now has a retail shopfront rather than a build step. Two on measurement: one index puts bots at 35% of web traffic while AI chat platforms send just 0.1% of human visits — directional, not a census — and the Linux Foundation now hosts a token-economics foundation, which is what it looks like when a line item becomes a discipline. Last, John's case for compulsory licensing before fair use is settled: if the courts are heading where they appear to be heading, the time to get a framework signed is while there is still leverage to sign one.

Read all 14 signals →