Sol at half price, and whether GPUs work as collateral
The board
▼ AMKR 10.5% ▼ ASX 8.8% ▼ MRVL 7.8% ▼ MU 7.0% ▼ GEV 6.9% ▼ VRT 6.8%
- Markets — Tuesday's session went the other way, and hard: the memory, packaging and power complex sold off together — Amkor −10.5%, ASE −8.8%, Marvell −7.8%, Micron −7.0%, GE Vernova −6.9%, Vertiv −6.8% and ARM −6.7%. Nothing in the slice we track finished higher.
- Open weights — Qwen3.8-27B keeps compounding: 665,513 downloads in thirty days, against 415,000 yesterday and 268,000 the day before.
- Models — Opus 5 still holds both ends of the model board: top intelligence at 63.1, best value at $10/M blended.
The read
The price war stopped being a trend line and became a price tag. GPT-5.6 Sol is listed at half price on OpenRouter — 50% off a list of US$2.50 and US$15 per million input and output tokens, on a model released only in July — while Anthropic told subscribers it is extending a 50% increase to weekly Claude limits. Mark's read on seeing it: shots fired. Alongside that, Anthropic's annualised run rate was reported at US$65bn for July, against the US$47bn it claimed in May. Both are happening at once; we would not assume either explains the other.
The Information's figures for OpenRouter give the layer Stripe is reported to be buying an actual shape: roughly US$40m of annualised serving costs at 28.5% of revenue, about US$100m of annualised gross profit on a 70% gross margin, implying annualised revenue near US$140m. Set against a reported price above US$7bn, the denominator is the thing to keep in view — whatever is being bought, it is not this year's revenue.
Open weights kept compounding. Qwen3.8-27B passed half a million downloads within three days of release, doubling in twenty-four hours, with Tunguz scoring it 52 on Artificial Analysis. What that does to frontier revenue is genuinely unresolved: one Reddit user's claim of saving US$650 in API costs in a single evening is a self-report we could not verify, and no disclosed number from any provider settles the question either way. Two useful things sit alongside it — shoehorn fits a model to the memory you actually have, solving a per-tensor precision assignment against what is left after inference, rather than picking a quantisation and hoping; and Tomasz Tunguz argues the right measure is time to answer, not tokens per second, which reorders which models look fast.
The physical layer kept asserting itself. Micron and SK hynix are adding memory capacity, but almost none of it lands before 2028 — which is the capex test we set yesterday, answered sooner and less encouragingly than expected. Power allocation is becoming industrial policy on two continents: PJM has proposed buying more power for data centres, and Australian taxpayers turn up in the Tomago arrangements. And an argument that begins as arithmetic and ends somewhere more useful — data centres are measured in megawatts, which gives no prize for efficiency. Mark's objection is that we are measuring energy input rather than compute output; John's proposal is to measure in FLOPS.
On the financing question, whether GPUs work as loan collateral is now load-bearing — Benedict Evans reads Nvidia's new joint venture as an effort to make GPUs fungible, tradable assets that can secure lending, so that when a borrower fails someone else picks up the hardware. The dot-com analogy keeps returning, and it is worth being precise about where it breaks: the honest version of the problem is that nobody knows where supply and demand come into balance. Meanwhile TrendForce expects Chinese accelerators to take 90% of China's domestic market as export controls and Beijing mandates push AMD and Nvidia out — a forecast, and one carried through two outlets.
Three from the research side. Microsoft 365 Copilot disclosed the parameter that defeated its own guardrail: researchers building a zero-click exfiltration exploit found the assistant refused the request, then answered a question about how the protection worked. MIT researchers report attribution decay — the more data a diffusion model is trained on, the harder it becomes to attribute any output to any training example, which lands squarely in the middle of the licensing argument. And an AI proof of Sendov's conjecture, followed by a digestion at a sixth the size: Lech Mazur resolved the remaining cases with an AI tool and verified it in Lean, and Terence Tao's compression of it strengthened the result.
Two on the work itself. AI accelerates zero to one and does nothing for what comes after — an argument about work theatre in large organisations that cuts against most productivity claims made for these tools. And a new developer platform pitched at humans and agents together, which is at least an honest statement of who the customer now is.