Money in, revenue out — the full breakdown.

The read

A second consecutive week without a primary lab raise, and the interesting motion happened on the cost side. Anthropic committing to in-house silicon under a proven TPU-era architect is the highest-signal lab move of the week, and it turns Anthropic's forward cash needs upward without anyone yet putting a number on the delta. OpenAI opting to run a dated three-month Sol promotion while pushing ads into 31 EU countries and launching a teens product is the demand side of the same problem: it is trying to grow token intensity against a promotional rate rather than the reverse. Anthropic and OpenAI are visibly using different levers to answer the same question about durable unit economics, which is a healthy sign for the category and a slower sign for investors expecting a price war to settle in the next quarter.

The read

The hyperscaler row is where the residual-value logic became visible this week without a capex disclosure moving. NVIDIA's 8-K puts an OpenAI-affiliate lease inside the accelerator vendor's contingent-obligation footnote — a $105B cumulatively capped guaranty on approximately 4.25 IT-GW of initial commitment beginning 2028 subject to ready-for-service, with an additional ~3.8 GW optional. Marvell's warrant to Google indexes ~7% of the silicon vendor's equity to $500M-per-tranche Custom Products revenue on a $120B implied ceiling and no purchase obligation. Neither transaction moves reported hyperscaler capital in the quarter, but both change how the category's demand should be underwritten going forward. Product-side, Google's bundling of Antigravity into Gemini Enterprise on the same day Anthropic shipped Skills-and-computer-use GA converts the hyperscaler surface into the natural enterprise identity plane for agents, which will show up as an ARR conversation before it shows up as a capex one.

The read

The neocloud row moved on one instrument. Nebius pricing $5B of upsized converts across a 2030 and 2034 stack rolls the same duration story the category has been running: long paper against shorter contracts, with the residual-value assumption sitting inside the debt rather than being written down in the equity. The 0.50% and 4.50% split is the market rewarding Nebius for the growth story on the near tranche and demanding a step-up on the further one, which is a rational answer to a private-mark environment where GPU-collateralized paper is still learning to price. This confirms Prediction p84 from W33 in outline — a second public neocloud rolling tenor beyond stated contract durations — but the confirmation happens outside the deadline window and the exact filing form must be examined against the trigger before scoring. The important thing for this issue is that the category continues to raise durably without needing to price new equity, and the NVIDIA guaranty at PORTS-Pike will make the next round easier for whoever borrows against a site with that footnote.

The read

The on-prem row's motion was regulatory rather than transactional. PA EO 2026-05 makes GRID a binding stage for any DC above 25 MW in the largest US hyperscale market, removes those sites from Permit Fast Track, and conditions the state tax exemption on GRID compliance while requiring local approvals before DEP permits. That is the first state-level rule to explicitly rope hyperscale into a slower permitting lane, and it lands during a national permitting reset — the August 18 NSSTS action circulated in secondary coverage as removing data centers from the CET list, but the primary text does not support that reading and this publication is treating it as noise. Product-side, Cerebras's CS-4 SUPERNOVA is a genuine architectural claim (three WSE-3 Turbo processors, vendor-claimed 750 PFLOPS, first shipments end Q3 2026) that will start showing up in on-prem procurement conversations before the year is out — treat the number as a vendor harness result until independent measurement lands.