The Investor

Washington's fiscal arithmetic explains one leg of this, and there are four. UK 30-year gilts sit at a 1998 high, France's 10-year is near 2008 levels, and Treasury's expanded buyback program bought a brief reprieve before yields resumed climbing, per Morning Brew's reporting. That makes this investors charging more to hold long paper everywhere at once, not a US headline, and the policy tools already in use are losing potency, which argues for the move running further rather than mean-reverting.

Japan is the leg with direct portfolio consequences. A government expanding its budget while cutting taxes as its 10-year breaks 3%, with Treasury Secretary Bessent publicly expecting the Bank of Japan to hike to defend the yen, is the textbook yen carry-trade unwind setup. Carry funds global risk assets at the margin, and it surfaces in crossover-fund behaviour and secondary pricing well before it reaches any primary round. The lag between yields and marks works as exposure.

The one asset class that improves as rates rise

Google contracted 400MW from Fervo Energy, the largest geothermal offtake on record, deliveries beginning in 2028 out of the $2B Cape Station project in southwestern Utah, grid connection slated by end-2026. Fervo re-rated 28.4% in a single session after falling more than 50% since its May IPO. Underwrite the mechanism rather than the megawatts: an offtake turns speculative development equity into contracted, inflation-resilient cash flow, and contracted cash flow is worth relatively more as the discount rate rises.

A hyperscaler PPA prices the electron where a funding round prices the story, and only the first of those improves when the hurdle rate moves against you.

Google also holds an option to add 600MW by 2030, on top of a pre-existing ~115MW agreement, which is uncompensated optionality moving from developer equity to the offtaker as the price of bankability. Price that option in energy term sheets rather than concede it. Scarcity sits in siteable geology and interconnection queue position; the drilling stack is adapted oil-field fracking technology.

Where the sources combine

The same tape carries the downside comps. The Information's morning report has GoPro selling 90% of itself for $285M, or $1.14 a share against a $93 high in 2014, after flagging going-concern doubt. That is the honest clearing price for a category-leading consumer hardware brand with global distribution, and it governs every hardware-first "AI device" mark in the book. Dell printed a $47B quarter with revenue up 58% and net income up 255%, having run 237% year to date, and fell 6.94% into that print. Momentum names carry no valuation cushion while the hurdle rate rises, and neither do their private comparables.

Sequencing is the take: public comps have absorbed the discount-rate move, private marks have not, and the gap closes at the next financing process rather than the next quarter-end. Probably wrong on timing, but dollars chasing momentum-priced private rounds are dollars not funding contracted cash flow.

Begin with the number that sorts everyone into camps. OpenAI's GPT-5.6 Sol scores 61 on the Artificial Analysis intelligence index at $0.95 a task against Claude Fable 5.1's 66 at $3.76, which is roughly 92% of the capability for a quarter of the money, or about 3.6x better intelligence-per-dollar for anyone who prefers the ratio dressed up. The frontier ceiling has compressed into a five-point band across four labs (66, 63, 62, 61) with more than 4x of price spread sitting inside it. Any term sheet still live this week on the premise that the company's edge is picking the smartest model is priced against a distinction the index no longer supports.

The underwriting error lives in the cost arithmetic rather than the scores. Anthropic's cache-read cut saved roughly $1.40 per task, and per-task tokens went from about 83M to 140M, so the net bill rose anyway. Models are now trained to think longer. Thinking longer bills at $50 per million output tokens.

The disagreement is the finding

All three are measuring honestly, which is the uncomfortable part. The variable is cache-hit ratio: cache-heavy workloads capture the 75% input cut, output-heavy workloads eat the 1.7x token inflation. So a market-level question becomes a company-level one (or rather, a diligence-level one), because the same model migration is a margin windfall for one portfolio company and a margin event for the one next to it. The difference is engineering, not the vendor's price list.

Diligence hygiene gets worse from there. Artificial Analysis ran its evaluation with Anthropic's default server-side fallback, meaning roughly 4% of output tokens in the headline benchmark came from different weights. The unit being priced is a routing policy, and the artifact that produced the benchmark can differ from the one a portfolio company actually deploys in production.

Where the money went instead

Two pricing moves inside one cycle suggest the model layer will not hold a price floor. Alibaba's Qwen3.8-Max-0902 debuted #1 on Code Arena WebDev at 1691, ahead of Claude Opus 5 Max, at $2 input and $6 output per million against Anthropic's $10 and $50. Meta shipped Muse Spark 1.3 to third place globally with a discount of more than 90% for customers who opt into training on their data, which is a competitor monetizing customer code instead of tokens, priced below any pure-play API vendor's sustainable margin.

That is the context for the day's hardest valuation print: Stripe paid approximately $7.5B for OpenRouter, per Fortune's Term Sheet. There are three readings and they are not mutually exclusive. The generous one: no lab holds a capability lead for more than weeks, cost outcomes are workload-dependent, so the layer that lets a buyer avoid single-model commitment is the durable position, and a payments company paid strategic-scarcity pricing for genuine scarcity. The counter-position is equally investable and has a short fuse, since a payments owner of the router is a real conflict for developers and model vendors, and that objection sells best in the months before Stripe integrates and neutralizes it. The third is probably wrong, but worth saying: the price may just be what Stripe would pay not to be on the other side of the router trade. First reading, held loosely, fuse noted.

Worth being precise about what disappears here. The Azure-containing segment ran a 41% operating margin against 58% for the Microsoft 365 segment, and that spread was the only public proxy anyone had for what AI capex and depreciation do to cloud profitability. Microsoft's stated rationale (the two units share AI chips and pooled compute capacity) is technically defensible and financially convenient at the same time, which is the most common combination in the history of segment reporting, and Satya Nadella described the result as "increased transparency to investors." Revenue clarity improves; margin clarity simply stops existing. Even granting that the accounting genuinely no longer separates, the number every AI infrastructure committee argues about stops printing.

What replaces it comes from the private side, or rather from the one private filing that will have to answer questions Microsoft has now stopped answering. Nscale is telling prospective investors it holds roughly $103 billion of total contracted revenue after a single $45 billion Anthropic compute deal, with an IPO possibly landing this month, per The Information's reporting. That prospectus becomes the reference disclosure for AI compute unit economics: contract structure, take-or-pay terms, delivery schedules, capacity cost per megawatt. Either it prints with the comps intact, or it prints with take-or-pay terms softer than the headline implies, or it slips and the reference disclosure never gets written. The middle case is the one worth modelling, and this may well be wrong, but a successful print reads as an exit window for private neocloud exposure rather than a permanent revaluation of the category.

Three customers carry Broadcom's guide

Nothing in the demand data shows a crack, and the bull case deserves that concession stated plainly. Broadcom printed $29.6B, up 86%, in the quarter ending August 2, with Hock Tan guiding AI revenue to $115B in 2027 and potentially $230B in 2028. Snowflake beat product revenue guidance by more than $70M and the stock moved over 20%. KKR is deploying $10B into AI infrastructure with two dedicated executives on it, which is $10B and two executives not deployed anywhere else.

Concentration is a separate matter, and it takes the same shape in every one of those lines. Broadcom's entire guide rests on custom accelerator work for three customers, namely Google, Meta and OpenAI. Nscale's book rests substantially on one. Anthropic's commitment rests on a pre-IPO counterparty delivering capacity that does not yet exist.

In AI infrastructure, revenue quality now runs inversely to headline backlog size. The bigger the contracted number, the fewer names it traces back to.

The policy leg that changes a different book

The Justice Department filed a brief in the New York Times' copyright case arguing that training AI models on copyrighted material does not violate copyright law, on the theory that rules making US AI development harder "threaten national security and give a competitive advantage to foreign adversaries." The brief settles nothing legally but signals where federal thinking sits, and in doing so it converts content licensing from a pricing-power business into a goodwill business. Any position whose revenue model is being the clearinghouse between labs and rights-holders now needs a base case in which the labs owe nothing.

The constructive half is narrower and more durable: proprietary data that is locked behind a contract or buried inside a customer's workflow where scrapers never reach just gained value relative to data that is merely copyrighted. That distinction has not moved through private marks yet.