The Investor

Meta paid more for each point of capability

Meta's own lineup is the clearest sign that frontier leadership is getting expensive to hold. AI Breakfast reports that moving from Muse Spark 1.3 to Muse Wonder raised Meta's score on the Artificial Analysis Intelligence Index from 48 to 52. Cost per benchmark task rose from $1.60 to $4.52 over the same step. That is 4 index points for roughly 2.8x the cost. Meta's top model, Wonder Full, scores 56 on that independent index, against 58 for Claude Opus 5.5. Meta's headline lead of 57.5 vs 46.3 on SWE-Marathon comes from a benchmark Meta runs itself.

Wonder costs $4 per million input tokens and $20 per million output tokens. That is cheaper than Anthropic's lineup and more expensive than GPT-6 Astra and Gemini 4 Argon. Four suppliers now sit inside a narrow band of price and quality. When buyers can swap between near-equivalent suppliers, none of them keeps pricing power for long.

The floor is falling faster than the ceiling is rising

Prices at the bottom of the market are collapsing. Turing Post reports that Claude Haiku 5.5 runs about 75% below Haiku 4.5, and that Sonnet 5.5's price for reading cached input has been halved. Microsoft-Decision-1, a model that returns calibrated probabilities instead of text, costs $0.042 per million input tokens and charges nothing for output. Microsoft built it by further training Alibaba's open Qwen3.5-9B. That tells you where the commodity base layer now sits.

Google also released EmbeddingGemma 2, a 740M-parameter model that turns text, images, audio and video into one shared search format. It ships under the permissive Apache 2.0 license and is small enough to run on a phone.

Put the top and the bottom together and the usage data makes sense. Buyers send most workloads to 'good enough' models and save premium models for narrow tasks. Meta's lead use case for Wonder is its consumer agent, Muse, which is the workload least likely to need the best model. The Information's Martin Peers argues that the main effect may land on Meta's own costs. That is a columnist's opinion, not a disclosed figure.

Where the sources disagree

The Information itself flags the main weakness in the usage data. OpenRouter skews toward developers and price-sensitive buyers, and it measures share of usage, not spend. Top-10 share can fall while dollars spent on premium models still grow. All four labs aimed their launches at coding, agent and professional work, which is where we would expect the premium tier to hold up. We take the shift as real, but we read it as dispersion. Premium models keep the high-value niches while most of the volume moves to cheaper prices.

A leaderboard lead now lasts weeks, and the companies that benefit are the ones whose margins widen every time a model gets cheaper.

What this means for your book

The winners are companies that switch between models, own a customer workflow, or hold proprietary data, because cheaper inference raises their gross margin. The losers are model-layer bets valued on benchmark rank, and application companies tied to a single premium model. Both groups will report healthy usage for now, so the difference shows up in COGS sensitivity, not top-line growth. For new deals, the test is whether a company can show pricing power, switching costs or proprietary data beyond its place on a leaderboard.

The capital stack under the chips

The largest single financing in this reporting sits one step away from the data centers themselves. TLDR Fintech reports that Bank of America, Citi and Morgan Stanley are syndicating a record $60B package to fund Anthropic's lease of Google-designed chips. It splits into $42B of senior loans backed by Broadcom and $18B of riskier junior debt, so the junior tranche is 30% of the stack.

Bridge's enlarged loan would be roughly 2.1x the $3.16B Bain paid in 2023 to take its parent, Chindata (now Wintrix), private, per The Information. A purchase price and a loan are different measures, but they point the same way. Lenders are now writing cheques that dwarf the original equity.

The debt is chasing real scarcity. Bloomberg reports that Lumentum's optical components for cloud computing are 'completely sold out' through early 2029. It also reports that Cboe's former Chicago headquarters will hold 33 MW of AI capacity, at rents well above the downtown office average. Techpresso reports that the IMF credits AI infrastructure spending with helping push global growth to about 3% a year. Major tech companies plan $660-690B of AI spending this year.

The revenue line underneath is contested

Debt like this is ultimately repaid from AI lab revenue, and the headline figure for the largest lab now comes in two versions. TLDR Fintech reports that OpenAI has told investors its annualized revenue is approaching $50B. That is below the roughly $70B figure that had circulated in comparisons with Anthropic, and the report puts the gap down to different revenue definitions. Separately, Bloomberg, citing people familiar with the matter, reports that OpenAI expects to reach or exceed $70B in annualized revenue by end-2026, driven largely by its enterprise business.

The two figures can partly be reconciled, since one describes revenue now and the other a year-end target. But neither report says which accounting choices drive the difference: gross vs. net, current run-rate vs. recognized revenue, or contracted vs. usage-based. Both figures are reported, not confirmed. The ambiguity matters because it sits underneath financings sized on expected demand.

One hyperscaler is already cutting its supplier bill

The Information reports that Microsoft plans 'multigigawatt deployments' of its in-house Maia 300 chip from 2027. Microsoft claims 40% more output per unit of power than leading Nvidia systems. Techpresso reports that analysts expect Maia to cover only about 10% of Microsoft's AI chip spending by early 2027, and Microsoft will not rent the chip to outside customers. We read Maia as a cost lever rather than an immediate threat to Nvidia. Still, it shows the richest buyers working to lower what it costs them to serve AI, which is the same pressure squeezing the model layer.

The smart move

Value infrastructure on megawatts that are actually powered up and on committed grid-connection dates, not on announced capacity. The scarce inputs justify a premium. The tenant revenue behind them deserves a discount until its definition is clear. Bridge's syndication, targeted for November, will show whether lenders price tenant quality or simply price the AI theme.

One asset class, two markets

PitchBook's Kyle Walters told Fortune why the Q3 decline compounds. Fewer exits mean fewer distributions to LPs, who then have less to reinvest and back fewer managers. Fundraising 'has definitely gotten harder over the last 18 to 24 months,' he says, and managers outside the top tier are 'fighting over scraps.' The top tier is having a different year. KKR hit its $300B fundraising goal in 30 months against a three-year plan.

Fortune's own figures disagree with each other. The subhead puts the drop in deal value at 14.1%, while the body has deal value at -13.6% and exit value at -14.1%. We would confirm with PitchBook before either number goes into an LP letter.

Venture has the same shape. Term Sheet listed five rounds totalling $1.8175B, and about 98.5% of it went to three: TypeSafe AI's $870M, Oratomic's $475M Series B and Oxide Computer's $445M Series D. Oxide is the instructive case. TLDR Hardware reports it was profitable at the operating level before raising, and the money funds manufacturing capacity, inventory and a growing order backlog. Its valuation was not disclosed. Below the megarounds, Kapta Space's $24M Series A was priced normally.

The rate test lands mid-week

The 10-year Treasury yield of 5.244%, up 108.1 basis points this year, is Friday's close, as Morning Brew points out. Bond markets were shut Monday for the holiday, so Tuesday is the first session to price a weekend of Houthi strikes on Riyadh. Stocks did not wait (the Nasdaq is up 17.74% this year). CME FedWatch still shows the market expecting no change at the late-October Fed meeting. A hold is already priced, so it moves nothing. The one outcome left that could move prices is a hike.

Walters argues PE has 'penciled a lot of this in,' and that a hike gives a Q4 that looks like Q3: weaker than 2025, stronger than the 2022-23 slump. We think deal volumes already reflect that and the debt load of portfolio companies at these yields does not, though Walters may be right that the industry has absorbed more than it shows. Diesel at $6.28 a gallon, 68% above its pre-war level, adds margin pressure for logistics- and food-exposed holdings.

Fewer exits mean fewer distributions, and fewer distributions mean fewer managers get funded.

Which exits still clear

Sponsor-to-sponsor sales still work for specialized, mission-critical businesses: AEA Investors and BCIMC bought Rensa Filtration from Audax on undisclosed terms. The second of the two open routes is the take-private, which is clearing the 2021 IPO class at deep discounts. The Information, citing the WSJ, reports Bumble is in talks to go private at about $1B, against a $9.5B market cap after its IPO and a $43 IPO price. The buyer is a new vehicle led by ex-Sequoia partner Ravi Gupta and ex-Meta CRO John Hegeman. It is unclear whether $1B is the equity value or the enterprise value. If it is the latter, shareholders receive $1B minus whatever debt Bumble carries.

Bank earnings come next. JPMorgan, Goldman, Citi and Wells Fargo report Tuesday, with BofA and Morgan Stanley on Wednesday. Their IPO and M&A fee backlogs will show whether the exit window is opening.