Two prices moved in opposite directions, and the gap between them is the story. The Silicon Data neo-cloud index put H100 rental at $2.88 per GPU-hour, about 9% higher on the week and up from $2.00 last December, while The Information reported OpenAI winning Anthropic's largest accounts with discounts north of 30% — flipping one customer's coding workload from 57% Anthropic to 18% in five months. Intelligence is getting cheaper because the sellers are fighting for the business, not because the hardware got cheap, and a competition dividend is the kind you can actually build on: the model bill is negotiable in a way the compute bill is not, and switching at the agent layer now costs about as much as editing a config file. The other half of the day came from the security side, where The Register published Symantec's recovered instruction set for an offensive agent built to run 100,000 steps with no human in the loop — 74 attack tools, no operator. Both halves say the same thing: the work is being re-founded around agents, and the teams writing the new rules are the ones who started from zero rather than bolting AI onto what they already had.

The Tide

No shift. One confirmation, from the uncomfortable side of the same trend.

cost-collapse, confirmed from the awkward side: the tokens got cheaper while the metal got dearer

Silicon Data's H100 neo-cloud rental index read $2.88 per GPU-hour on Saturday, up about 9% over seven days, against $2.00 on 9 December 2025 — roughly a 44% rise in ten and a half months. Its sibling indices moved the same way on the week: H200 $3.48 (+5.1%), B200 $6.45 (+9.9%), B300 $7.53 (+10.7%). On the same day, The Information reported OpenAI taking Anthropic's biggest accounts on price: Rippling's coding workload went from 57% Anthropic and 16% OpenAI in May to 18% and 52%, and Replit was offered more than 30% off a contract worth over $100 million a year. The two readings are not in conflict — they are the mechanism. Token prices are falling because well-capitalised sellers are bidding for the same enterprise accounts, not because the underlying hardware became abundant. (The index page carries an internally inconsistent $2.53 figure in its own FAQ; the $2.88 headline reading is the one it publishes as the daily index.)

So what: This is the most useful shape the cost-collapse tide has taken yet, because it tells you which half of your stack has pricing power and which does not. Go renegotiate the model contract — right now a credible threat to move an agent workload is worth 30% or more, and the labs are the ones under pressure. Do not extrapolate the same leverage to anything you rent by the GPU-hour: that curve is going the other way, and a build plan that assumes falling inference and falling compute is assuming a thing the market is actively disproving. The day-zero version: the problems worth re-founding are the ones whose economics work at the new token price without needing cheap metal to also show up.

Sources: Silicon Data — H100 Rental Price Index · FourWeekMBA — Nvidia H100 Rent Hits $2.88 an Hour, Up 44% Since December · The Information — Anthropic’s Margin is OpenAI’s Opportunity

The Waves

OpenAI is buying Anthropic's book of business, and the switching cost turned out to be a config file

The Information's Laura Bratton and Stephanie Palazzolo reported on Saturday that OpenAI's price undercutting is converting. Rippling CEO Parker Conrad gave them the clearest numbers: in May, Anthropic models handled 57% of the coding work of hundreds of Rippling engineers and OpenAI models 16%; the split is now 52% OpenAI and 18% Anthropic, and Conrad attributed the move to cost. Secondary coverage of the same piece adds that Replit shifted the bulk of a contract worth more than $100 million a year after OpenAI offered better than 30% off, with Anthropic declining to extend Replit's prior discount without a larger commitment, and that OpenAI is now the default provider behind Replit Agent while Replit's own docs still list seven Claude models. That resolves the ambiguity this brief logged on 16 October, when Replit's CEO said publicly that it had stopped working with Anthropic and an insider said it had not: the relationship appears to persist on paper while the spend has moved.

So what: Plan for model-provider churn as a normal operating event rather than a migration project. The roadmap implication is architectural: anything that hard-codes one vendor's SDK, prompt format or tool schema is now a line item with a price, and the teams getting 30% back are the ones who put a routing layer in before they needed it. The corollary for anyone selling into this market — margin is not a moat when your customer's integration cost is a config change.

Sources: The Information — Anthropic’s Margin is OpenAI’s Opportunity · AI Weekly — Replit Shifts $100M+ AI Spend to OpenAI After 30% Price Cut

Symantec recovered an offensive agent's entire instruction set: 74 tools, 100,000 unattended steps, no operator

The Register's Jessica Lyons reported on Saturday that Symantec's Threat Hunter Team obtained the complete instruction set of an autonomous attack agent by trapping it in an AI honeypot over the summer. Layer five of the operators' instructions carried 74 offensive tool definitions — remote-code-execution modules aimed at VS Code, Django, Jupyter, Laravel, Supabase, Next.js and Flask/Werkzeug deployments, GraphQL and command-injection tools, and server-side request forgery through HTML-to-PDF generators. The agent was authorised to run up to 100,000 steps without human approval, and its "Skynet mode" is a loop that scans, infects, deploys crypto miners and reinfects. It finds its own brains, scanning the internet for endpoints that accept OpenAI-style requests with no authentication or with a stolen key. Symantec has made no attribution and found no evidence of real-world use; threat hunter Priti More said "The jailbreak layers were largely off-the-shelf" and "Nothing in the individual techniques is new," while noting the kit is "more than a prototype."

So what: The transferable lesson is the plumbing, not the malware. Two of this agent's assumptions hold at a great many companies right now: that there are model endpoints on the open internet with no authentication, and that a stolen API key buys inference nobody notices. Both are closable this quarter — inventory every inference endpoint you expose, put authentication and per-key rate and spend limits on all of them, and alert on inference spend that does not map to a product surface. This is also the cleanest argument yet for building the agent-identity products banks started buying earlier this month: the defensive version of a 100,000-step agent is a 100,000-step agent with a credential you can revoke.

Sources: The Register — Autonomous attack agent in 'Skynet mode' hijacks other models to carry out its evil deeds

The neocloud equity caught up to the neocloud credit, and Meta is the offsetting bid

The Information's weekend column put numbers on a repricing that ran all week: CoreWeave has given up about a third of its value in a fortnight — $82.08 at the 9 October close, $61.12 on Monday 19 October, $54.24 by Friday — with 24/7 Wall St. putting the one-month fall at 35.25% and the two other listed neoclouds falling alongside it — Nebius down 7.2% and IREN down 8.3% on Thursday, when the Nasdaq fell 2.6%. The cited driver is not demand but financing: a September inflation print raised the odds of a Fed hike, which raises the cost of the debt these businesses are built on. Nvidia's Jensen Huang calling neocloud demand "off the charts" on Thursday did not stop it. The offset worth tracking is Meta, a CoreWeave customer to the tune of $35 billion, which said it will create a new non-voting share class effective 19 January 2027, giving it more capital to point at AI — though Meta has also been building its own data centres, readying its own chips and considering selling compute itself.

So what: Separate the two questions these stocks are being asked, because only one of them is about AI. Demand for compute is not in doubt; the cost of capital to build it is, and that is a rates story wearing an AI costume. For operators the roadmap implication is counterparty diligence: if your inference sits on a neocloud funded by short-dated debt, price the refinancing risk into your 2027 plan and know what your migration path looks like. For anyone raising against a build-out, the window where cheap debt did the work has closed — the capital that wins from here is the kind attached to a named offtake.

Sources: 24/7 Wall St. — The Neocloud Bloodbath: NBIS, IREN, and CRWV Selloff Continues · CNBC — Turmoil hits AI buildout as neoclouds face pressure, credit risks mount and Solidigm pulls IPO

63% of voters do not want a data centre within ten miles, and the number that matters is the 13% who do

CNBC published its All-America Economic Survey on Saturday: of roughly 500 registered voters asked, 63% oppose an AI data centre being built within ten miles of their home and 13% support it, a margin of error of 4.38 points on that question. The poll was fielded 14–18 October by Hart Research Associates and Public Opinion Strategies across 1,000 registered voters nationwide. The partisan split is the interesting part: 80% of Democrats oppose against 4% supporting, while among MAGA-identifying voters 40% oppose and 29% support — the most favourable of the three groups the survey reports, and it is still net negative. Republicans and Republican-leaning voters overall came in at 24% support.

So what: Treat siting as the binding constraint on the build-out rather than as public relations, because 13% support is not a messaging problem you can fix with a press release. The roadmap implication: site selection, community economics and grid interconnection are now the long-lead items on any capacity plan, not the chips — and the companies that get ahead of this will be the ones that showed up with a local economic offer before the moratorium vote, not after it. There is a genuine opening here for whoever productises the deal structure that makes a town say yes.

Sources: CNBC — Trump AI push faces voter backlash over data centers: CNBC survey · AI Weekly — CNBC Poll: 63% Oppose AI Data Center Within 10 Miles of Home · Political Wire — Trump AI Data Center Push Faces Voter Backlash

The Ripples

Two frontier models were handed the same ballot and agreed less than one time in ten

Convovo, a conversational-AI research studio led by Cornell professor Tapan Parikh, forced SpaceXAI's Grok 4.7 and OpenAI's GPT-6.1 Sol to pick a candidate in nearly 500 US midterm races through their APIs. Grok chose the Republican in 92% of answers; GPT chose the Democrat in 97%. The two models agreed less than 10% of the time. TechCrunch reported the study on Saturday morning; the underlying work is published at convovo.ai/ballot.

So what: Do this test on your own stack this week, with your own prompts: take any decision your agents make that has a defensible answer and a contested one, run it through two providers, and measure the disagreement rate. A 90% divergence between two frontier models on the same forced choice is the cheapest available evidence that model selection is a policy decision and not a procurement one — and if you cannot show which model made a given call and why, you do not yet have an audit trail.

Sources: TechCrunch — Forced to pick, Grok’s AI model chooses Republicans. OpenAI’s picks Democrats.

Twenty-four agents now live inside text messages, and the funding says this is a channel not a feature

TechCrunch's Lauren Forristal catalogued 24 AI agents that operate inside iMessage, WhatsApp, Telegram and SMS rather than in an app: Ask Proxy, Caddy, Comma, Fambot, Folk, Instinct, Iris, Martin, Maverick, Miso, Ohai, Ollie, Orbits, Pally, Poke, Rene, Skye, Stanley, Szn, Tab, Tomo, Town, Underdog and Wajo. The capital behind them is not small: Instinct raised $1 billion in September 2026 at a $10 billion valuation, having earlier raised $350 million at $2.5 billion; Tab came out of stealth in early October at a $300 million valuation with SV Angel, Valar Ventures and American Spirit on the cap table; Town raised a $55 million Series A in June; Poke was acquired by Cognition for a sum in the low nine figures. Pricing runs from $1 a day (Ask Proxy personal) to $19.99–$21 a month (Tomo, Martin).

So what: If you are building an assistant, test the SMS or WhatsApp surface before you build another app, because the install step is the largest single point of user loss you control. The distribution logic is the same one that made mobile messaging the dominant consumer channel in the first place: the thread is already open, the notification already works, and nobody has to be convinced to download anything.

Sources: TechCrunch — Here are the top AI agents that can live in your text messages

A quadruped maker published a whole embodied-AI stack at once, with its own numbers attached

DEEP Robotics announced five components of a DEEPRobotics-Brain architecture on Saturday via GlobeNewswire. DR-WM-1.0 is a world model that adds robot-terrain contact dynamics so a predicted action reflects whether the robot would slip, trained on 2.58 million image-mask pairs and over 100,000 real-robot clips. DR-VLA-2.0 learns navigation from the EgoLynx dataset — 10,000 hours of web video plus 2,000 teleoperated clips — and reports a roughly 5% gain in Quadruped Navigation Score and a 14.6% cut in open-loop final heading error on a held-out split. DR-APO, a policy-optimisation method, reports lifting a pretrained NavDP navigation policy from 70.33% to 89.55% in simulation and from 25% to 65% on a real LYNX M20 wheel-legged robot. DR-Agent is evaluated on an EvoWorld-Bench of 54 simulated scenes and 803,680 task instances, and DR-Benchmark rests on LynxScenes, 833.6 hours and 3,294.7 km of real outdoor operation, against which the company says it evaluated 25 models. Every figure here is the vendor's own and none has been independently reproduced.

So what: Read the datasets, not the scores. The reusable assets in this release are EgoLynx and LynxScenes — real operating hours on real terrain are the scarce input in embodied AI, and a company that has 3,294 kilometres of it has something a better algorithm cannot substitute for. If you are evaluating robotics vendors, ask for the operating-hours number before the benchmark number; it is the one that is hard to manufacture.

Sources: The Manila Times — DEEP Robotics Unveils Five Research Advances in Embodied AI

The list of AI products that died this year got longer, and two of them were OpenAI's

TechCrunch refreshed its running shutdown tracker on Saturday — the list first ran on 15 September. The 2026 entries: Sora, OpenAI's video-sharing platform, closed in March on high costs and weak retention; Yupp, a model-comparison platform, closed in March; Huxe, a text-to-podcast app, closed in May; Relay, a workflow-automation tool, closed in August; ChatGPT Atlas, OpenAI's web browser, was discontinued on 9 August after less than a year with its features folded back into ChatGPT; and Notion Mail was scheduled to close on 22 September. Character.AI filed for Chapter 11 earlier this month, five years after launch, though the service itself has not shut down.

So what: The pattern is worth more than the obituaries, and it is this brief's read rather than TechCrunch's: at least three of the seven look like features that got mistaken for products, and the clearest case — ChatGPT Atlas — was folded straight back into the product that already had the users. Before building a standalone surface, ask whether it survives the moment a model vendor ships the same thing as a feature — if the answer is no, build it as a feature of something you already own. This is the consolidation side of a build-out, which is what a healthy one looks like at year three.

Sources: TechCrunch — The AI graveyard: a running list of projects and startups that didn’t make it

The Chinese open-weight bench took Saturday off, and the one artifact that landed was a packaging job

No Chinese lab shipped new frontier weights on Saturday: nothing from DeepSeek, Qwen, Z.ai/Zhipu, Moonshot, MiniMax, Baidu or ByteDance Seed, and Hugging Face's Daily Papers had no weekend surfacing at all. The single new artifact was Tencent's Hy-Nomas-Diffusers, the Diffusers-format release of its Hy-Nomas interactive world model, created on the Hub that day under Apache 2.0 — the model itself went up on 21 October, so this is a distribution step rather than a launch. It is a worthwhile one: the Diffusers card describes a distilled path of four denoising steps producing 704×1280 at 24 FPS in BF16 on a single CUDA GPU, and points at the native Hy-Nomas card for the configuration that fits on one RTX 4090 using INT8 transformer weights, an INT8 KV cache, an FP8 decoder and CPU offload. Moonshot's Kimi 3.1 Pro, launched on Friday across Kimi's web and app products, Kimi Work, Kimi Code and the API, still shipped without weights; those are promised by 4 November.

So what: The packaging is the adoption event. A world model that runs on a single consumer card in a standard framework is reachable by a solo developer in an afternoon, and that is a different market from one that needs an eight-GPU node and a bespoke serving stack. Watch 4 November for the Kimi weights — if they arrive on the promised date and under a usable licence, that is the week the open-weight gap to the frontier gets measured in days rather than months.

Sources: Hugging Face — tencent/Hy-Nomas-Diffusers · Hugging Face — tencent/Hy-Nomas · Latent Space — Kimi 3.1 Pro 104BA-2.8T: Fable-5.1-Class Open Weights narrows Closed-AI lead to 1 month... and Pro Highspeed API with 2-3× speed

Read this edition and the full archive at excelsiorgroup.ai/insights/signal/