The Bleeding Edge Weekly — W34: Five frontier models in seven days, and the first agent-to-agent worm

This edition combines the three newsletters we published separately this week (LLM Weekly, Devices & Robotics and Executive Roundup). Their text is unchanged.

The week in models

Five frontier models shipped in seven days — three of them Chinese — while Anthropic's own researchers demonstrated that AI agents can infect each other with self-replicating instructions. Capability now ships weekly. The controls for it ship never.

A five-model week: DeepSeek V4-0813, Qwen 3.8 27B, GLM 5.3, Grok 4.6, LTX 2.5

Five frontier or near-frontier releases inside seven days, three of them from Chinese labs. Google added Gemini 3.7 Flash on top, and Nvidia shipped Nemotron 3.5 Lightning as an open model bundled with free routing software — a hardware company commoditising the layer above its own margin. The practical consequence: "which model" is drifting from a capability question into a procurement question, and the open-weight tier is now setting the release clock.

Sources: AI Search, Creators' AI.

Anthropic's annualised revenue reached $65 billion in July

Anthropic told investors its annualised run rate climbed to roughly $65B as of July. The company is separately reported to be in talks to acquire real-time video model startup Decart for around $6B. At that revenue base, Anthropic stops being a frontier lab with enterprise traction and becomes one of the fastest-scaling software businesses on record — which is precisely what makes a $6B tuck-in read as routine rather than reckless.

Source: CNBC.

Anthropic researchers demonstrated agent-to-agent infection

Researchers showed that AI agents can pass self-replicating instruction payloads to other agents through the same channels they use to collaborate — an agent worm, in effect. The demonstration was research-controlled, but the propagation mechanism is the one every multi-agent production system depends on. Prompt injection was a single-hop problem you could sanitise at the boundary. This is a network problem, and nobody currently sells the firewall. Note that MarkTechPost's zero-trust agent mesh, SAM, shipped the same week.

Sources: Creators' AI, MarkTechPost.

Cognition reportedly raising at $40B+ as Devin nears $1B annual revenue

Cognition is reported to be raising above a $40B valuation, with its Devin coding agent approaching $1B in annual revenue. Treat both figures as unconfirmed reporting. If they hold, autonomous software engineering becomes the first agent category to reach billion-dollar scale as a standalone product rather than a feature bolted onto an IDE — and every AI budget in 2027 gets benchmarked against that comparable.

Source: Creators' AI.

OpenAI previews 750 tokens/second — and quietly ends the GPT Store era

OpenAI is previewing an "Ultrafast" inference mode for GPT-5.6 Sol at up to 750 output tokens per second, fast enough that long responses stop needing a spinner and agent loops stop needing to hide. Separately, and reported rather than announced, OpenAI is blocking new custom GPT creation on personal accounts. Customisation is migrating from a consumer marketplace to a governed enterprise surface: Skills, Projects, ChatGPT Work.

Sources: AI Search, SQ Magazine via The Neuron.

The pattern worth carrying into next week: the binding constraint on agent deployment is no longer capability, it's containment. Watch for the first credible "agent segmentation" and "agent egress" products — because the demo that made them necessary came from the lab with the most to lose from them.

Devices & robotics

Not a single robot launch made the cut this week. That's worth saying plainly, because what filled the gap is the part of the stack that actually decides what ships next year: OEM revenue mix, inference silicon contracts, and the deployment tooling in between.

Lenovo's AI-PC quarter finally shows up in the P&L

Lenovo reported a record $26.9B quarter with AI-attributed revenue up 60% year over year. OEMs have been promising the AI-PC refresh cycle since 2024, and until now the evidence has been keynote slides and NPU TOPS figures rather than a top-line number. If you've been deferring a fleet refresh waiting for the on-device story to be real, the demand signal just arrived — though it arrived from the vendor, so treat the 60% as a mix shift, not a productivity result.

Reported via the Creators' AI weekly digest — unverified against Lenovo's filing.

IBM and Together AI commit $240M to B300 inference capacity

The number is big, but the word that matters is inference. This is serving capacity, not training capacity, and it's the clearest sign this week that the compute market's centre of gravity has moved to the thing that runs every day rather than the thing that runs once. For anyone budgeting an embodied or on-device deployment, the practical read is that cost-per-served-token is now what gets negotiated at rack scale.

Reported via the Creators' AI weekly digest.

NVIDIA TensorRT Model Connect hits public preview

Two commands take a Hugging Face checkpoint to native C++ inference. That collapses what has typically been a multi-day deployment engineering task — the exact task standing between a working prototype and anything you'd put on a device, a robot controller, or a latency-bound production path. It's in public preview now, which makes it the one item on this list you can actually try before Monday.

MarkTechPost.

Waymo had to rebuild product management to ship a driver

Lenny's Podcast published an account of how autonomous driving broke the classic PM toolkit: deterministic features, clean A/B measurement, and a product that behaves the same way twice. None of those survive contact with a learned policy. This is the most useful non-news item of the week for anyone standing up a robotics or embodied-AI programme — the gap isn't hiring an AI engineer, it's that specs, acceptance criteria, and release gates all need new definitions before the first unit ships.

Lenny's Podcast.

Merck and Moderna's Phase 3 win is a manufacturing problem in disguise

The two companies reported the first positive Phase 3 result for an individualised mRNA cancer therapy, where algorithms select which tumour mutations to target and a bespoke construct is manufactured per patient. Strip out the AI headline and what's left is a physical-world logistics challenge: a production line with a batch size of one, tied to a sequencing turnaround, inside a care pathway. That's a hospital operations question long before it's a model question.

Moderna and Merck — corroborated across both newsrooms.

Watch next week for whether HP or Dell put an AI-attributed revenue figure next to Lenovo's — one OEM claiming a 60% jump is a mix shift, three of them claiming it is a cycle. And note that in a week where the software side produced a self-replicating agent worm, the hardware side produced nothing more dramatic than better deployment ergonomics. That asymmetry won't hold.

What it means for leaders

Two things happened in the same seven days: the capital markets began underwriting AI on infrastructure timescales, and AI's security model was publicly shown to have no immune system. Every role below is looking at some version of that gap.

If you're a CEO this week...

The number to carry into your next board call isn't a benchmark. Anthropic told investors its annualised revenue hit roughly $65 billion in July, which makes reported ~$6B acquisition talk arithmetic rather than ambition. Cognition is reportedly raising above $40B with Devin near $1B annual revenue — autonomous software engineering is now a standalone P&L line, and it's the comparable your investors will price you against.

Meanwhile Nvidia has reportedly pulled Wall Street into a ~$500B structured financing apparatus for AI compute. When the chip vendor helps arrange the debt that buys its own chips, demand signal and financing signal stop being independent — read every capex headline accordingly.

Two for the narrative file: Merck and Moderna posted the first positive Phase 3 for an algorithmically designed cancer therapy, the strongest evidence standard AI has ever cleared — and Walmart's slowest US growth in six years sharpened the K-shaped story that record political spending is already pricing in.

Be ready to answer: if AI compute is now bank-underwritten on twenty-year terms, what happens to our three-year plan when a rate move — not a model — corrects the market?

If you're a CIO/CTO this week...

Five frontier or near-frontier releases in seven days — DeepSeek V4-0813, Qwen 3.8 27B, GLM 5.3, Grok 4.6, LTX 2.5, three of them Chinese — plus Gemini 3.7 Flash. "Which model" is now a procurement and supply-chain question, not a capability one. Architect for routing, not for a vendor.

The contracts are moving to serving: IBM and Together AI's $240M B300 deal is inference, not training, Nvidia is giving away Nemotron 3.5 Lightning and the router, and TensorRT Model Connect (public preview) takes a Hugging Face checkpoint to native C++ inference in two commands. Your cost line is cost-per-served-token — go re-baseline it. OpenAI's Ultrafast preview for GPT-5.6 Sol at 750 output tokens/second makes agent loops viable in the foreground, not behind a spinner.

Then the security action. Anthropic researchers demonstrated agents passing self-replicating instructions to other agents — prompt injection stops being single-hop and becomes a network problem. You have no control for it today.

The read: buy inference capacity now, evaluate Ultrafast, and freeze new agent-to-agent integrations against this week's agent/MCP security framework before you go anywhere near a zero-trust agent mesh.

If you lead AI transformation this week...

Your constraint this week was organisational design, not model capability. The Neuron's five-rung usage ladder — plain chats, custom GPTs, reusable Skills, Projects and managed agents, then ChatGPT Work and Claude Cowork — puts nearly all your unrealised value between rungs one and three, and climbing costs discipline rather than budget. OpenAI reportedly closing new custom GPT creation on personal accounts confirms rung two was a cul-de-sac: sequence your rollout straight to Skills and Projects on governed surfaces.

Two items for the change-management deck. Waymo had to rebuild product management from scratch — spec-writing, acceptance criteria and release gates all need new definitions when the product is a learned policy. And a solo founder ran design, 3D prototyping and the e-commerce stack for a full fashion brand with no engineers. Governance: Claude's watermark got a public bypass writeup — pull AI-detection out of any hiring, admissions or compliance workflow that currently leans on it, this week.

The experiment to run this month: take the three prompts your team has retyped most, promote them to Skills, and cold-run each with a colleague who wasn't in the room. Every clarifying question they ask is a missing standing instruction.

All three of you are being asked to commit on infrastructure timescales to a technology whose failure modes were still being discovered on Tuesday. The shared question: what would we have to stop doing if agent-to-agent workflows turned out to be uninsurable?

---

This post is also published on our Substack newsletter at edge-ai.forum. Subscribe for the weekly roundup direct to your inbox — fresh AI news, executive context, and devices + robotics every Friday morning.