The Bleeding Edge — Episode Briefing W44

Date range: 2026-10-23 to 2026-10-30 (Europe/Madrid)

Headline of the Week

This week the AI build-out started being paid for with borrowed money, and the market noticed. Microsoft opened its first US high-grade bond sale since 2017: a $25B offering that drew peak orders of $115B. Bloomberg reported that global bond sales hit a record, driven by an "AI borrowing binge." Over the same days, Alphabet posted revenue and income that disappointed, and Nasdaq futures slid as mixed Big Tech results tested the AI trade. Further up the stack, the big players are moving into services and the wider ecosystem. The FT reports that Anthropic is getting into consulting. Nvidia's reported $12.93B Hugging Face deal is drawing questions. Anthropic's open-source security scanner says it has logged 29,000 findings. The labs and chipmakers are borrowing to build capacity and then selling services to make sure that capacity gets used.

Top 5

1. Microsoft taps the bond market for the first time since 2017, and AI borrowing pushes global bond sales to a record. Microsoft started a $25B US high-grade bond sale. Peak orders reached $115B, and Bloomberg reports that AI-driven borrowing pushed global bond sales to a record. On the same day, Reuters reported Nasdaq futures sliding as mixed Big Tech earnings tested the AI trade, after Alphabet's disappointing revenue and income sent its shares lower. Why it matters: the biggest balance sheets in tech are now financing AI capex with debt as well as cash flow. For executives, a falling cost of AI capacity now partly depends on credit markets staying open. Unverified Sources: Bloomberg: Microsoft bond sale, Bloomberg: global bond sales record, Reuters: Nasdaq futures slide.

2. Nvidia's reported $12.93B Hugging Face deal draws scrutiny. The Neuron reported that questions are being raised about Nvidia's $12.93B deal involving Hugging Face, the largest public hub for open-weight models and datasets. The newsletter excerpt does not give the deal terms or say who is asking the questions. Why it matters: if it holds, the dominant chip supplier gains influence over the main distribution point for open models, which is the part of AI that has been most vendor-neutral so far. Expect neutrality and competition to dominate the conversation. Unverified Source: The Neuron, Oct 29.

3. Anthropic moves into consulting. On October 23 the Financial Times reported that Anthropic is entering the consulting business, according to The AI Opportunity's write-up of the story. Why it matters: a frontier lab selling implementation work competes head-on with the systems integrators that resell its models. Enterprise buyers now have to decide whether they want the model vendor also designing their transformation programme. Unverified Source: The AI Opportunity: The Anthropic AI Consulting Playbook.

4. AI bug-hunting at industrial scale: Anthropic's OSS scanner logs 29,000 findings, 516 patched. AI Dev Signals (MarkTechPost) reports that Anthropic's open-source vulnerability scanner has produced 29,000 findings, and 516 of them have been patched. Separately, The Neuron reported that 62 critical or high-severity Chrome bugs were found this week. The excerpt does not name the tool or team. Why it matters: the gap between 29,000 and 516 is the story. AI can now find bugs much faster than maintainers can fix them, so the bottleneck has moved to triage and remediation. Unverified Sources: AI Dev Signals, The Neuron, Oct 29.

5. Claude Duet gets its first hands-on reviews. Two separate newsletters covered Claude Duet this week. AI Search listed it among the week's launches. Lenny's "How I AI" published an honest review with worked workflows for building a sales deck and a sales pipeline in Duet. Why it matters: the examples are sales and GTM work, not code. That suggests Anthropic's newest surface is aimed at the business user, the same buyer its reported consulting push would serve. Corroborated Sources: Lenny's Newsletter: How I AI, Claude Duet review, ChatPRD: Claude Duet review, AI Search.

Humor Segment

Written by the joke-writer skill for 2026-W44: 3 jokes for the show. Fact anchors first; writer notes are not read on air.

FACT ANCHORS

• Story 1: Microsoft opened a $25B US high-grade bond sale, its first since 2017, and peak orders reached $115B.
• Story 3: The Financial Times reported on October 23 that Anthropic is entering the consulting business.
• Story 4: Anthropic's vulnerability scanner has produced 29,000 findings, and 516 of them have been patched.
• Story 5: Lenny's "How I AI" reviewed Claude Duet with workflows for building a sales deck and a sales pipeline.

THE SET

1. [A: Contrast] Anthropic's bug-hunting AI has flagged twenty-nine thousand security problems. Five hundred and sixteen have been fixed. Also this month, Anthropic's getting into consulting. So I think we know who to call about the other twenty-eight thousand.

2. [J: Literal Quote] Anthropic launched Claude Duet. A duet: two voices, in perfect harmony. The first big review shows you how to use it to build a sales pipeline. Which is technically still two voices. So is a cold call.

3. [C: Specific Callback] Microsoft went to the bond market for the first time since 2017. It asked for twenty-five billion dollars. Lenders lined up with a hundred and fifteen. First time in history anyone's asked for more Microsoft. Normally it just shows up in an update. With twenty-eight thousand bugs.

SHAREABLE LINES

• "Microsoft asked to borrow $25 billion and lenders lined up with $115 billion. First time in history anyone's asked for more Microsoft."
• "Claude Duet: two voices in perfect harmony, used to build a sales pipeline. Technically still a duet. So is a cold call."
• "Anthropic's AI flagged 29,000 security problems and 516 got fixed. The same month, Anthropic got into consulting. I think we know who to call about the other 28,000."

WRITER NOTES (not spoken)

• Joke 1 (opener)
  • Stories: 4 (29,000 findings, 516 patched) and 3 (Anthropic consulting).
  • Connector: the listener assumes the AI finds bugs so they get fixed. The connector is "consulting." The second reading is that the bug list doubles as a sales pipeline: the finding is free and the fixing is billable.
  • Mechanic: A, Contrast. Candidate p ≈ 0.12.
  • Delivery: the number is the reveal, so say "the other twenty-eight thousand" slowly. It also plants the callback for the closer.
  • Risk: edgy-but-defensible. It hints at a conflict of interest, but it's framed as a guess ("I think we know"), not as a claim of fact.
• Joke 2 (middle)
  • Story: 5 (Claude Duet reviews).
  • Connector: "duet" sets up two voices cooperating in harmony. The connector is "two voices." The second reading is a sales call, where there are two voices and one of them wants to hang up.
  • Mechanic: J, Literal Quote (the product name read literally). Candidate p ≈ 0.18.
  • Delivery: pause before "So is a cold call." The hard k's carry it.
  • Risk: safe.
• Joke 3 (closer)
  • Story: 1 (Microsoft bond sale, $115B peak orders).
  • Connector: "more Microsoft" first reads as investor demand. The second reading is the Microsoft everyone gets without asking: forced updates and bundled software.
  • Mechanic: C, Specific Callback, with two laughs. The first is "more Microsoft." The tag "twenty-eight thousand bugs" calls back the opener's number. Candidate p ≈ 0.14.
  • Risk: safe. The target is Microsoft's software reputation.

CUTTING ROOM

• Nvidia's $12.93B deal for Hugging Face, a company named after a hug emoji, is drawing questions. Mostly: is it a hug, or a headlock? The connector is real, but the hug/headlock pun is the obvious joke (p ≈ 0.4), and every other show will make it. Good material if Ralph or Emile want to riff on it.
• Asked Copilot how Microsoft plans to pay back twenty-five billion dollars. It offered to summarize the meeting. A good Bad Answer shape, but the laugh depends on knowing Copilot's in-meeting summary prompts. It fails the non-insider test.
• An AI flagged 29,000 bugs and 516 got fixed. It's not superintelligence yet, but it's a world-class backseat driver. A clean image, but it's mostly true as an observation and earns a nod, not a laugh. It also loses to the consulting reveal on the same numbers.

Categorised News

Market Cap / Valuation

Legal AI: Harvey and Legora raised almost $1.9B between them in 2026. The AI Opportunity calls legal AI "one of the biggest opportunities ever," based on combined 2026 fundraising of nearly $1.9B for the two category leaders. The Neuron also reported a launch of ready-made artifacts for legal teams this week, but the excerpt does not name the vendor. Unverified Sources: The AI Opportunity, The Neuron, Oct 29.

Instinct attaches a $10B figure to the "text your AI assistant" category. Creators' AI reports that texting an AI assistant has become mainstream behaviour and that Instinct "got $10B" for it. The excerpt does not say whether that is a valuation or capital raised, so check before quoting on air. The same piece points back to the earlier Clawdbot and OpenClaw wave of messenger-native agents. Unverified Source: Creators' AI.

Frontier & Big Tech

Alphabet disappoints and the AI trade wobbles. Bloomberg reported that Alphabet's revenue and income disappointed and its shares fell. Reuters linked the broader Nasdaq futures slide to mixed Big Tech results that are "testing the AI trade." Corroborated Sources: Bloomberg, Reuters.

Moonshot ships Kimi K3.1 Pro; MiniMax M3.1 lands the same week. Moonshot AI introduced Kimi K3.1 Pro for coding, research and 3D design work. MiniMax M3.1, Carbon X and Deep Infer appeared in the same AI Search roundup. Chinese labs are still releasing on a near-weekly cadence. Unverified Source: AI Search.

Inside PM at Anthropic: who makes it, and who doesn't. Lenny's Newsletter cross-posted a look at how product management works inside Anthropic. It is useful context for anyone hiring PMs into AI-native teams. Unverified Source: Lenny's Newsletter.

Apps / Dev Tools / Platforms

Nace AI's Drex for Search scores 92.05% on BrowseComp-Plus with an average of 12.6 search calls. Drex is a decision model for agentic search. Instead of producing free-text reasoning, it returns a probability for each option, which lets an agent decide when to stop searching. The headline result is accuracy and efficiency together: a search agent that knows when to stop. Unverified Source: AI Dev Signals.

OmniParseBench lets teams rebuild the document-parser leaderboard themselves. It is a reproducible benchmark for document parsers, so teams can test vendor claims on their own documents instead of relying on published rankings. The same AI Dev Signals issue also covered a "drop the KV cache, not the context" technique for long-context inference. Unverified Source: AI Dev Signals.

"Rent the model, own the context." Daniil at Creators' AI says the "best model" for his work changed hands at least four times this year. His conclusion is to treat models as rented infrastructure and invest in owned context: prompts, memory, documents. A companion piece describes "hiring four AI agents and becoming their middle manager." Unverified Source: Creators' AI.

Infrastructure & Ecosystem

Marvell reportedly in talks to buy data-centre optics maker Lumentum. Lumentum stock rose on a report of a possible merger with Marvell. Optical interconnect is a choke point in AI data centres, and consolidation there affects networking costs across the build-out. Unverified Source: Yahoo Finance.

AI & Robotics

Applied Intuition, the "$15B AI company nobody knows." The AI Opportunity profiles Applied Intuition, which builds the software layer for autonomous vehicles and machines. The piece argues the company could eventually be worth $100B, drawing on a 52-minute conversation with CEO Qasar Younis. Palantir's Alex Karp is cited as offering a framework for which AI companies survive. Unverified Source: The AI Opportunity.

Prompting Skill of the Week

Technique: the Portable Context Pack. Best for teams that switch between models as leadership changes and don't want to rebuild their workflows each time. This is the practical version of "rent the model, own the context."

1. Choose one recurring task, such as a weekly board summary or a sales-call recap.
2. Write a model-agnostic context file covering role, audience, house style, definitions, three "good output" examples, and hard constraints.
3. Keep it in a plain-text file you own, not in a vendor's memory feature.
4. Run the same pack plus the same input through two or three models.
5. Score the outputs against a fixed five-point rubric you write once.
6. Revise the pack, not the model-specific prompt, whenever an output fails.

Example prompt:

"You will receive a CONTEXT PACK and an INPUT. Treat the CONTEXT PACK as binding: follow its definitions, style rules and constraints exactly. If the INPUT conflicts with the pack, flag the conflict before answering. CONTEXT PACK: [paste]. INPUT: [paste]. Output in the format specified in section 4 of the pack."

Common failure + fix: the pack grows into a 6,000-word wall and models start ignoring parts of it. Fix: open with a "Top 5 rules" section, move reference material to an appendix, and test by deleting a section. If the output doesn't change, that section wasn't doing anything.

New AI Tools

Drex for Search (Nace AI). A decision model for agentic search that returns per-option probabilities and reports 92.05% on BrowseComp-Plus with an average of 12.6 search calls. For teams building research or procurement agents where search cost and runaway loops are real problems. Source: AI Dev Signals.

OmniParseBench. A benchmark for checking document-parsing vendors (PDFs, scans, tables) on your own corpus instead of trusting their marketing. For operations and data teams choosing a parser for invoice, contract or claims pipelines. Source: AI Dev Signals.

TinyFish Search + Fetch. Free web search and fetch APIs for agents, with no credit card required. This was a sponsored placement in MarkTechPost's newsletter, so treat the claims as marketing until tested. For indie developers and small teams prototyping agents that need web access. Source: MarkTechPost (sponsored).

AI Personality of the Week

Qasar Younis. The co-founder and CEO of Applied Intuition got an unusually long hearing this week: a 52-minute conversation with The AI Opportunity, framed around a $15B company "almost nobody knows about" that the newsletter argues could be worth $100B. Younis matters because Applied Intuition is a counter-example to the chatbot-centred view of AI. It sells the simulation and software stack behind autonomous vehicles and machines into defence and automotive, a business that doesn't depend on which frontier model leads this month. In a week when the frontier labs are borrowing money and moving into consulting, a quiet, vertical, B2B autonomy company at $15B is a useful reminder of where else value is building up. Source: The AI Opportunity.

Catch-All

Reddit built its own AI prototyping tool instead of buying one. In Lenny's "How I AI," Reddit's Blair Metcalf explains why the company built an in-house AI prototyping tool. The published workflows cover prototypes that run on live Reddit data and a path from a Figma design to a draft pull request. For executives this is the build-versus-buy question in concrete form. Off-the-shelf "vibe-coding" tools can't reach your production data or your design system, so large product organisations are starting to build thin internal layers on top of the models. Sources: ChatPRD: Blair Metcalf / Reddit, Reddit live-data prototype workflow.

Sector Watch

• Retail & E-commerce — DoorDash agreed to buy delivery rival Instacart. This is not an AI deal, but it consolidates the order-and-delivery layer that shopping agents will send purchases through. Unverified — Capital Brief
• Banking & Financial Services — Global bond sales hit a record, driven by AI-related borrowing. Debt desks and credit investors are now directly underwriting the AI build-out and carry concentration risk if the AI trade sours. Corroborated — Bloomberg
• Energy & Utilities — quiet week
• Travel & Hospitality — quiet week
• Construction & Built Environment — quiet week
• Healthcare & Life Sciences — quiet week

Show Notes (bullets only)

• Microsoft launches its first US high-grade bond sale since 2017: $25B offered, $115B in peak orders.
• AI borrowing pushes global bond sales to a record, per Bloomberg.
• Alphabet disappoints on revenue and income; Nasdaq futures slide as mixed Big Tech results test the AI trade.
• Nvidia's reported $12.93B Hugging Face deal draws questions.
• The FT reports Anthropic is moving into consulting.
• Anthropic's open-source scanner: 29,000 findings, 516 patched. Separately, 62 critical or high-severity Chrome bugs found.
• Claude Duet gets hands-on reviews, with sales-deck and sales-pipeline workflows.
• Harvey and Legora have raised almost $1.9B between them in 2026.
• Moonshot ships Kimi K3.1 Pro; MiniMax M3.1 lands the same week.
• Nace AI's Drex for Search: 92.05% on BrowseComp-Plus with 12.6 search calls on average.
• Marvell reportedly eyeing data-centre optics maker Lumentum.
• Reddit explains why it built its own AI prototyping tool.

Weekly Patterns (Inference)

1. Inference AI capex is moving from cash-funded to debt-funded. When Microsoft borrows for the first time in nine years and global issuance hits a record, the cost of AI capacity becomes linked to interest rates and credit appetite. That is a new risk for anyone budgeting around falling inference prices.
2. Inference Labs and chipmakers are moving into services and distribution to secure demand. Anthropic in consulting and Nvidia reportedly at Hugging Face are the same strategy: own the route by which customers adopt your product.
3. Inference Partners are becoming competitors. A lab that does consulting competes with its integrators, and a chipmaker tied to the main open-model hub weakens that hub's neutrality. Expect channel partners to start hedging across vendors.
4. Inference Security is now limited by remediation capacity, not detection. 29,000 findings against 516 patches shows AI produces vulnerabilities faster than humans can fix them. The next budget line is AI-assisted patching and triage, not more scanners.
5. Inference Value is building up in less visible verticals. Applied Intuition at $15B and legal AI at $1.9B raised show that vertical, workflow-heavy businesses are being priced as durable while the frontier-model race keeps changing leaders.
6. Inference Model-agnosticism is becoming standard practice. "Rent the model, own the context," Reddit building its own tooling, and weekly Chinese model releases all point the same way: the lasting asset is your context and workflow layer, not your choice of model vendor.
7. Inference Agents are being judged on when they stop, not only on whether they're right. Drex's "12.6 search calls" headline suggests agent benchmarks are starting to price in cost per answer, which is the metric operations leaders actually pay for.