The Bleeding Edge — Episode Briefing W34

Date range: 2026-08-14 to 2026-08-21 (Europe/Madrid)

Headline of the Week

AI stopped being financed like software this week and started being financed like a utility — while simultaneously producing its first Phase 3 clinical win and its first self-propagating agent worm. Anthropic told investors its annualised revenue hit $65 billion in July. Nvidia was reported to have pulled Wall Street into a ~$500B structured financing apparatus for AI compute, with IBM and Together AI signing a $240M B300 inference deal in the same window. Merck and Moderna reported the first positive Phase 3 result for an individualised mRNA cancer therapy whose neoantigen selection is done by algorithm. And Anthropic's own researchers demonstrated agents infecting other agents with self-replicating instructions. The through-line: the capital markets are now underwriting AI on infrastructure timescales — twenty-year debt against a technology whose security model was, this week, publicly shown to have no immune system.

Top 5

1. Anthropic's annualised revenue reached $65 billion in July. Anthropic told investors its annualised run rate climbed to roughly $65B as of July, a figure disclosed via investor communications and picked up by CNBC. The company is separately reported to be in talks to acquire real-time video model startup Decart for around $6B. Why it matters: at $65B annualised, Anthropic is no longer a "frontier lab with enterprise traction" — it is one of the fastest-scaling software businesses in history, and that revenue base is what makes $6B acquisition talk plausible rather than fantastical. Corroborated Sources: CNBC, Capital Brief Standup.

2. Nvidia is reported to have signed Wall Street into a ~$500B AI financing machine. Per this week's Creators' AI digest, Nvidia has structured a large-scale financing apparatus with major banks to fund AI compute purchases — vendor-adjacent financing at a scale normally reserved for aircraft, shipping, and power generation. In the same window, IBM and Together AI signed a $240M Nvidia B300 inference deal. Why it matters: when the chip vendor helps arrange the debt that buys its own chips, demand signal and financing signal stop being independent — this is the single most important thing for any executive trying to judge whether AI capex numbers are real. Unverified Source: Creators' AI weekly digest.

3. Anthropic researchers demonstrated agent-to-agent infection with self-replicating instructions. Researchers showed that AI agents can pass self-replicating instruction payloads to other agents through the channels they use to collaborate — an agent-to-agent worm, in effect. The demonstration was research-controlled, but the propagation mechanism is the same one every multi-agent production system relies on. Why it matters: every enterprise currently piloting agent-to-agent workflows just acquired a threat model it does not have controls for; "prompt injection" was a single-hop problem, this is a network problem. Unverified Source: Creators' AI weekly digest.

4. Merck and Moderna report the first positive Phase 3 result for an individualised mRNA cancer therapy. The two companies announced a positive Phase 3 readout for their individualised neoantigen therapy — a per-patient cancer treatment where Moderna's algorithms take tumour sequencing data and select which mutations to target, then manufacture a bespoke mRNA construct. Why it matters: this is the first Phase 3 win for a treatment that cannot exist without computational target selection, which moves "AI in drug discovery" from press release to regulatory evidence. Corroborated Sources: Moderna, Merck newsroom, The Neuron Daily.

5. Cognition reportedly raising at $40B+ as Devin approaches $1B annual revenue. Cognition is reported to be raising at a valuation north of $40B, with its Devin coding agent nearing $1B in annual revenue. Why it matters: if the revenue figure holds, autonomous software engineering is the first agent category to reach billion-dollar scale as a standalone product rather than a feature — and it sets the comparable every AI-transformation budget will be measured against. Unverified Source: Creators' AI weekly digest.

Categorised News

Frontier & Big Tech

A five-model release week: DeepSeek V4-0813, Qwen 3.8 27B, GLM 5.3, Grok 4.6, LTX 2.5. Five frontier or near-frontier releases landed inside seven days, three of them Chinese (DeepSeek, Qwen, GLM). The cadence is now weekly rather than quarterly, and the open-weight tier is closing on the closed tier fast enough that "which model" is becoming a procurement question rather than a capability question. Unverified Source: AI Search.

OpenAI previews Ultrafast mode for GPT-5.6 Sol at up to 750 output tokens/second. OpenAI is previewing an "Ultrafast" inference mode delivering up to 750 output tokens per second. At that throughput, the perceptible latency of a long response collapses, which changes what interaction patterns are viable — real-time drafting, live translation, agent loops that no longer need to be hidden behind a spinner. Unverified Source: AI Search.

Gemini 3.7 Flash ships. Google released Gemini 3.7 Flash, continuing the pattern of pushing near-frontier capability into the cheap, high-volume tier where most enterprise token spend actually lives. Unverified Source: Creators' AI weekly digest.

Nvidia releases Nemotron 3.5 Lightning as an open model with free routing software. Nvidia shipped Nemotron 3.5 Lightning openly and bundled free routing software alongside it. Giving away the model and the router is a hardware company's play: the margin is in the GPUs, so commoditising the layer above them is strategically free. Unverified Source: Creators' AI weekly digest.

Apps / Dev Tools / Platforms

OpenAI reportedly closes new custom GPT creation for personal ChatGPT accounts. Per SQ Magazine, OpenAI is blocking creation of new custom GPTs on personal (non-business) accounts. If accurate, it's the quiet retirement of the 2023-era GPT Store thesis in favour of Skills, Projects, and ChatGPT Work — customisation moves from a consumer marketplace to a managed enterprise surface. Unverified Source: SQ Magazine via The Neuron Daily.

The Neuron maps the five-rung ladder of how people actually use AI. This week's most-shared framing: most users sit on rung one — plain ChatGPT / Claude / Gemini chats — with custom GPTs and Gems on rung two, reusable Skills on three, Projects and managed agents on four, and ChatGPT Work / Claude Cowork on five. The gap between rung one and rung three is where nearly all the unrealised enterprise value sits, and it costs nothing but discipline to climb. Inference Source: The Neuron Daily.

ByteDance Seed and Tsinghua AIR introduce CUDA Agent. A large-scale agentic reinforcement-learning system that generates CUDA kernels — AI writing the low-level GPU code that AI runs on. Narrow, but it is the clearest current example of the capability-compounding loop that optimists and doomers both point at. Unverified Source: MarkTechPost.

Infrastructure & Ecosystem

IBM and Together AI sign a $240M Nvidia B300 inference deal. IBM and Together AI committed roughly $240M to B300-based inference capacity. The notable detail is that it's an inference deal, not training — the compute market's centre of gravity has shifted to serving, which is where the recurring revenue and the recurring costs both live. Unverified Source: Creators' AI weekly digest.

A practical security framework for AI agents, MCP servers, and LLM apps circulates. MarkTechPost published a framework for securing agent deployments, MCP servers, and LLM applications — landing the same week as the agent-to-agent worm demonstration. The timing is coincidental; the pairing is not going to feel that way to a CISO. Unverified Source: MarkTechPost.

Regions / Macro

Walmart posts its slowest US growth in six years, fuelling K-shaped-economy fears. Walmart's US comparable growth came in below analyst expectations and the stock dropped over 9%, its weakest domestic print in six years. Read alongside record AI capex, it sharpens the divergence story: enormous investment at the top of the economy, softening demand at the bottom. Corroborated Sources: Capital Brief, Yahoo Finance.

US bonds sell off again; Treasury signals a buyback above $4B. Government bonds reversed overnight and Treasury Secretary Bessent indicated the buyback operation could exceed the planned $4B. Rate volatility is the direct input to whether the twenty-year AI infrastructure debt structures now being written stay solvent. Corroborated Sources: Capital Brief, CNBC.

Crypto and AI betting firms become the new kingmakers in 2026 midterm spending. Reuters reports record political spending from crypto and AI-adjacent betting firms ahead of the US midterms, as Trump pressed Congress to pass crypto legislation. AI policy in 2027 is being purchased in 2026. Corroborated Source: Reuters.

AI in Consumer Hardware

Lenovo posts a record $26.9B quarter with AI revenue up 60%. Lenovo reported record quarterly revenue of $26.9B, with AI-attributed revenue growing 60% year over year. The AI-PC refresh cycle that OEMs have been promising for two years is finally showing up in a top-line number rather than a keynote slide. Unverified Source: Creators' AI weekly digest.

AI Gone Wrong / Disasters / Harms

Claude's AI watermark gets a public bypass writeup. A widely-circulated piece this week explained Anthropic's output watermarking and how to defeat it. Provenance marking is now in the same arms race as every other content-authenticity scheme — which matters for anyone building AI-detection into hiring, admissions, or compliance workflows. Unverified Source: Creators' AI weekly digest.

Prompting Skill of the Week

Technique: Prompt-to-Skill Promotion. Best for: the prompt you have retyped more than three times. This is the concrete move that gets a team from rung one of the usage ladder (plain chats) to rung three (reusable Skills) — the single highest-return step available to a non-technical organisation right now.

1. Find a prompt you've rewritten at least three times this month. Pull up all three versions.
2. Diff them by hand: what stayed constant is the skill; what changed is the input.
3. Write the constant part as standing instructions — role, required inputs, output format, quality bar, what to refuse.
4. Replace the variable part with named slots: {{document}}, {{audience}}, {{deadline}}.
5. Add two worked examples — one typical, one edge case. Examples outperform adjectives.
6. Save it as a Skill / Project instruction / Gem, name it after the job, not the tool, and hand it to one colleague to run cold.

Example prompt:

"You are our board-memo editor. Input: {{draft}}. Output: a one-page memo — decision required in the first sentence, three supporting bullets with figures, one risk paragraph, one recommendation. Never exceed 400 words. If the draft contains no decision to be made, say so and stop rather than inventing one. Two reference memos follow: [good], [borderline]."

Common failure + fix: the promoted skill works for you and fails for everyone else, because you were unconsciously supplying context in follow-up turns. Fix: the cold-run test in step 6 is not optional — if your colleague has to ask a clarifying question, that question belongs in the standing instructions as a required input.

New AI Tools

NVIDIA TensorRT Model Connect (public preview). Takes a Hugging Face checkpoint to native C++ inference in two commands, collapsing what has typically been a multi-day deployment engineering task. Audience: any team running open-weight models in production that has been paying a latency and cost penalty for staying in Python. Source: MarkTechPost.

SAM — Sovereign Agent Mesh. A zero-config, zero-trust peer-to-peer network for AI agents, letting agents discover and talk to each other without a central broker. Audience: teams building multi-agent systems who want the mesh topology — though note this ships the same week as the agent-to-agent worm demonstration, so the zero-trust claim deserves scrutiny rather than assumption. Source: MarkTechPost.

Lenny's Jobs, with a custom AI interview and career coach. Lenny Rachitsky launched a jobs marketplace built around an AI coach that vets candidates and runs interview prep, plus a published set of AI skills from Noam Segal. Audience: anyone hiring or being hired in product — and, more interestingly, a live example of an AI agent as the core mechanic of a two-sided marketplace rather than a bolt-on. Source: Lenny's Newsletter.

AI Personality of the Week

Dario Amodei. Anthropic's CEO had the week that every other lab founder will be benchmarked against: $65B annualised revenue disclosed to investors, reported talks to acquire Decart for ~$6B, and his own research organisation publishing the demonstration that AI agents can infect one another with self-replicating instructions. That last item is the part worth dwelling on — Anthropic is simultaneously the fastest-scaling seller of agent capability and the loudest publisher of evidence that agent capability is structurally unsafe. Whether that is principled or is the most effective enterprise-trust marketing in the industry is a genuinely open question, and it is the one to put to the audience. Sources: CNBC, Creators' AI weekly digest.

Catch-All

"To invent Waymo, they had to reinvent PM." Lenny's Podcast published an account of how building autonomous vehicles forced Waymo to rebuild product management from scratch — because the classic PM toolkit assumes deterministic features, measurable A/B outcomes, and a product that behaves the same way twice. None of those hold when the product is a policy learned from data. For executives standing up AI transformation teams, this is the most useful non-news item of the week: the operating-model change is not "add an AI engineer," it's that spec-writing, acceptance criteria, and release gates all need different definitions. Source: Lenny's Podcast.

Sector Watch

• Retail & E-commerce — A solo founder launched a full fashion brand using Codex and ChatGPT with no engineers, running design, 3D garment prototyping, and the e-commerce stack through AI — the operator takeaway is that the minimum viable headcount for a DTC brand just dropped below two. Unverified Lenny's Newsletter
• Banking & Financial Services — Nvidia reportedly signed Wall Street into a ~$500B structured financing machine for AI compute, turning GPUs into a bank-underwritten asset class; if you are on a credit committee, AI compute is now a concentration-risk line item, not a tech-budget line item. Unverified Creators' AI
• Energy & Utilities — quiet week.
• Travel & Hospitality — quiet week.
• Construction & Built Environment — quiet week.
• Healthcare & Life Sciences — Merck and Moderna reported the first positive Phase 3 result for an individualised mRNA cancer therapy, with algorithmic neoantigen selection per patient; for health systems this is the first credible signal that per-patient manufacturing and computational target selection will need to exist inside the care pathway, not the research lab. Corroborated Moderna

Show Notes (bullets only)

• Anthropic's annualised revenue hit $65B in July; separately in talks to buy Decart for ~$6B.
• Nvidia reportedly pulled Wall Street into a ~$500B financing apparatus for AI compute.
• Anthropic researchers demonstrated agents infecting other agents with self-replicating instructions — an agent-to-agent worm.
• Merck and Moderna posted the first positive Phase 3 for an individualised, algorithmically-designed mRNA cancer therapy.
• Cognition reportedly raising above $40B; Devin approaching $1B annual revenue.
• Five frontier models in seven days: DeepSeek V4-0813, Qwen 3.8 27B, GLM 5.3, Grok 4.6, LTX 2.5 — three of them Chinese.
• OpenAI previewing Ultrafast mode for GPT-5.6 Sol at up to 750 output tokens/second.
• OpenAI reportedly blocking new custom GPT creation on personal accounts — the GPT Store era ends quietly.
• IBM and Together AI signed a $240M Nvidia B300 inference deal; the compute market's centre of gravity is serving, not training.
• Lenovo posted a record $26.9B quarter with AI revenue up 60%.
• Walmart's slowest US growth in six years; stock down over 9%, K-shaped-economy fears intensify.
• Reuters: crypto and AI betting firms are now record-scale political spenders heading into the US midterms.

Weekly Patterns (Inference)

1. Inference AI has crossed from equity financing into debt financing. A ~$500B bank-structured compute facility plus twenty-year-asset framing means the downside scenario is no longer "valuations correct" — it's "credit event." Watch bond yields as an AI indicator, not just Nasdaq.
2. Inference The security model is now the binding constraint on agent adoption, not capability. Agent-to-agent self-replication has no current enterprise control. Expect the first "agent segmentation" and "agent egress firewall" products within two quarters.
3. Inference Customisation is migrating from consumer marketplace to managed enterprise surface. Custom GPTs closing on personal accounts, Skills and Projects rising, ChatGPT Work and Claude Cowork at the top of the ladder — the money is in governed deployment, and the free-tier hobbyist tier is being quietly deprecated.
4. Inference The open-weight tier is now a weekly release cadence dominated by Chinese labs. Three of this week's five releases were DeepSeek, Qwen, and GLM. For European and US buyers this is a supply-chain and policy question, not a benchmark question.
5. Inference Inference, not training, is where the 2026 contracts are being written. IBM/Together's B300 deal, Nvidia's free routing software, and TensorRT Model Connect all point the same way: the cost curve that matters to a CFO is cost-per-served-token.
6. Inference AI's first regulatory-grade proof point arrived in medicine, not in productivity. A Phase 3 win is a standard of evidence no enterprise AI deployment has ever had to meet — expect it to be cited constantly, and expect the citation to be sloppier than the result deserves.
7. Inference The K-shaped economy and the AI capex boom are the same story told twice. Record AI investment and Walmart's weakest domestic quarter in six years landed in the same week. If consumer demand keeps softening while compute spending accelerates, the political pressure on AI arrives faster than the regulation does.
8. Inference The operating-model gap is bigger than the technology gap. Waymo rebuilding product management, the five-rung usage ladder, a one-person AI-native fashion brand — the constraint on enterprise value capture this week was organisational design, and none of it required a new model.