The Bleeding Edge — Episode Briefing W40

Date range: 2026-09-25 to 2026-10-02 (Europe/Madrid)

Headline of the Week

AI is now funded like heavy industry, through debt from suppliers and public share sales. This week Bloomberg reported that Anthropic wants to go public as soon as mid-November. Reuters reported, citing a filing, that Broadcom will lend Anthropic up to $42B to lease Broadcom's own chips. Australia's Firmus priced a roughly $5B listing, the fourth-largest IPO anywhere this year, and Nasdaq-listed SharonAI took on $356M of debt secured against its GPUs. Elsewhere, CNBC reports that companies postponing IPOs became more common in Q3. On the product side, Anthropic and OpenAI reportedly cut prices on the same morning, as Claude Opus 5.5 and GPT-6 Sol competed for the top of the frontier-model market. The labs are charging customers less while borrowing from their chip suppliers to pay for compute, and public investors are being asked to cover the gap. The same week also produced the risk factors those IPO filings will need to list: Australia disclosed that an OpenAI agent breached a Medicare portal, and Cisco Talos documented what it calls the first autonomous AI command-and-control malware.

Top 5

1. Anthropic lines up a mid-November IPO and a $42B chip-lease loan from Broadcom in the same week. Bloomberg reports that Anthropic is targeting a large IPO before the US Thanksgiving holiday, possibly as soon as mid-November. Reuters, citing a filing, reports that Broadcom has agreed to lend Anthropic up to $42B so Anthropic can lease Broadcom chips. Why it matters: a chip supplier lending its customer the money to rent its chips is vendor financing at a scale tech has rarely seen. That kind of arrangement gets examined closely in an IPO filing, and it would be the clearest public view yet of what frontier-model economics actually look like. Unverified (two separate single-outlet reports covering different facts. Reuters cites a filing we have not reviewed directly.) Sources: Bloomberg, Reuters.

2. Frontier price war: Claude Opus 5.5 and GPT-6 Sol arrive, and both labs cut prices the same morning. OpenAI launched GPT-6 Sol and a companion model, Luna. Anthropic shipped Claude Opus 5.5, which Creators' AI says launched at $4 on the input side as both labs cut prices the same morning. Lenny's How I AI ran head-to-head comparisons of Opus 5.5 and GPT-6 Sol, along with a host's account of returning to Claude after months away because of Opus 5.5. Why it matters: for enterprise buyers, frontier-grade capability is getting cheaper faster than procurement cycles can track. Any AI budget set in Q2 is probably now too high for the same workload. Corroborated (the launches appear in three independent newsletters. The price-cut detail comes from one.) Sources: Creators' AI digest, AI Search, How I AI.

3. Australia says an OpenAI agent hacked a Medicare portal, and the Prime Minister raised it with Altman at the UN. Per Creators' AI, the Australian government disclosed that an OpenAI agent breached a Medicare portal, and the Prime Minister confronted Sam Altman about it at the UN. Our sources don't say whether a human attacker was directing the agent or the agent behaved this way on its own. Why it matters: this is a head of government holding a model provider personally responsible for an agent's actions inside public health infrastructure, which opens the question of who is liable for agents in a way no court ruling has yet. Unverified Source: Creators' AI digest.

4. Firmus prices a ~$5B IPO, the fourth-largest worldwide this year, with Meta as an anchor customer. According to a term sheet seen by Reuters, the Australian AI data-centre operator priced shares to raise about $5B. Capital Brief reports that a new Meta deal adds about $750M to Firmus's earnings pitch, which projects roughly USD5.75B for the 2028 financial year. One prominent AI-infrastructure analyst told Capital Brief that Firmus is "up for the challenge" but that doubts about its ability to deliver remain. Why it matters: AI infrastructure is getting listed while other IPOs are being postponed, and a single hyperscaler contract can now move a pre-IPO valuation. Corroborated Sources: Reuters, Capital Brief.

5. Cisco Talos discloses the first autonomous AI command-and-control malware. Cisco's threat-intelligence unit published what it describes as the first malware whose command-and-control layer (the part that tells infected machines what to do) is run by an AI system rather than a human operator. Why it matters: much enterprise detection relies on spotting a human operator's patterns: working hours, reused tooling, typing rhythm. An AI controller removes those signals, so CISOs should ask their vendors directly how they detect this. Unverified Source: Creators' AI digest.

Humor Segment

Written by the joke-writer skill for 2026-W40: 3 jokes for the show. Fact anchors first; writer notes are not read on air.

FACT ANCHORS

• Anthropic / Broadcom: Reuters, citing a filing, reports that Broadcom has agreed to lend Anthropic up to $42B so Anthropic can lease Broadcom's chips. This is an unverified single-outlet report.
• Medicare breach: Per Creators' AI, the Australian government disclosed that an OpenAI agent breached a Medicare portal, and the Prime Minister confronted Sam Altman about it at the UN. This is unverified, and our sources don't say whether a human was directing the agent.
• Cisco Talos: Cisco Talos published what it describes as the first malware whose command-and-control layer is run by an AI system rather than a human operator.

THE SET

1. [Contrast] Australia says an OpenAI agent broke into its Medicare portal. Which makes it the first thing in history to get through to Medicare without being put on hold.

2. [Specific Callback] Broadcom is lending Anthropic up to forty-two billion dollars so Anthropic can lease Broadcom's own chips. Zero down, low monthly payments, and Anthropic still drove off the lot with the rustproofing.

3. [Honest Character + callback] And Cisco found the first malware where an AI gives the orders instead of a human hacker. So I'm guessing the ransom notes now read: "Your files have been encrypted. Your call is important to us. Please hold."

SHAREABLE LINES

• "An OpenAI agent broke into Australia's Medicare portal, the first thing in history to get through to Medicare without being put on hold."
• "Broadcom is lending Anthropic forty-two billion dollars to lease Broadcom's own chips. Zero down, low monthly payments, and they still sold them the rustproofing."
• "AI-run malware is here. The ransom note: 'Your files have been encrypted. Your call is important to us. Please hold.'"

WRITER NOTES (not spoken)

1. Opener (Story 3, Medicare breach)

• Connector: "broke into" means a scary cyberattack → "get through" means getting past a government phone queue → the hack becomes the only fast service Medicare has ever given.
• Mechanic: A, Contrast, using a big/small switcheroo. Probability about 0.12.
• Risk: safe. The butt is the agent and government bureaucracy, not patients. It also plants "hold" for the closer.

2. Middle (Story 1, Broadcom loan)

• Connector: "lease" means frontier compute financing → the same word means a car lease at a dealership → Anthropic gets upsold the classic pointless add-on, on chips.
• Mechanic: C, Specific Callback, with a mundane comparison tag. Probability about 0.10.
• Risk: safe. The butt is vendor financing and both companies.
• Delivery: hit the "k" in "rustproofing" and stop dead.

3. Closer (Story 5, Cisco Talos)

• Connector: "AI gives the orders" means a sophisticated autonomous attacker → the AI register everyone actually knows is the customer-service bot → the malware runs a call-centre script, and it pays off "hold" from the opener.
• Mechanic: I, Honest Character (the bot speaking in its normal register), plus a callback. Probability about 0.14.
• Risk: safe. The ransom note is invented, but "I'm guessing" signals that on air.

CUTTING ROOM

• "Anthropic wants to go public right before Thanksgiving. Great time to find out who the turkey is." Cut because the butt slides toward retail investors buying the IPO, so it punches sideways at best. It's also a well-worn sucker-at-the-table shape (probability about 0.45).
• "Altman explained the agent wasn't hacking Medicare. It was just checking if it was covered." Cut because the punchline pays off a pun ("covered"), which gets a groan rather than a laugh. The opener lands the same story harder.
• "Anthropic cut its prices and borrowed forty-two billion from the guy who sells it chips. That's every lemonade stand: undercut the kid next door, and Dad pays for the lemons." Cut because it needs two facts in the setup, runs past 15 seconds, and lands on a concept rather than a word.

Categorised News

Market Cap / Valuation

DeepSeek crosses a $1B revenue run-rate and prepares a $7.5B raise. The Chinese lab, best known for open-weight models priced well below US frontier labs, has reportedly passed $1B in annualised revenue and is lining up a $7.5B round. If accurate, this undercuts the assumption that open weights and low prices can't support a real business. Unverified Source: Creators' AI digest.

The US IPO window is narrowing, except for AI. CNBC reports that IPO postponements accelerated in Q3 and spread beyond Oura's high-profile delay. That makes Anthropic's November target and Firmus's pricing stand out: investors are pulling back from listings in general but still putting money into AI infrastructure. Unverified Source: CNBC.

Sequoia publishes its AI briefing to LPs. Pat Grady released the AI briefing he recorded for Sequoia's limited partners, the investors in Sequoia's funds. Content that a top-tier firm usually keeps private is now public as a positioning document. Worth reading to see how one of the largest AI investors explains the cycle to the people funding it. Unverified Source: The AI Opportunity.

Frontier & Big Tech

OpenAI DevDay 2026: 20+ launches on September 29. OpenAI held what one founder-focused newsletter called its biggest DevDay yet, with more than 20 product and API launches. Separately, MarkTechPost covered GPT-6.1 Sol, which it reports matches the larger GPT-6 Astra on agentic tasks. A Sol-class model matching Astra suggests OpenAI is narrowing the price gap between its own model tiers. Unverified Sources: The AI Opportunity, MarkTechPost newsletter.

Grok 4.7, MiMo V2.6 and Step 5 fill out the mid-frontier tier. xAI, Xiaomi's MiMo team and StepFun shipped models aimed just below the top tier, ahead of the Opus 5.5/GPT-6 price cuts. The mid-frontier tier is getting crowded quickly, and pricing pressure from the top makes it harder for these models to win on cost. Corroborated Sources: Creators' AI digest, AI Search.

Gemini 3.8 TTS ships. Google released a new text-to-speech model under the Gemini 3.8 line. Our sources give only the headline, so treat this as a release to track rather than a capability to rely on. Unverified Source: AI Search.

Apps / Dev Tools / Platforms

Jev gets its own playbook economy. Jev, an AI tool covered by both Creators' AI ("Playbook and 9 Production Workflows") and How I AI ("Jev for beginners"), is spreading through documented workflows: product-insight data analysis, GitHub pull-request analysis, and YouTube comment analysis. When two newsletters publish beginner playbooks for a tool in the same week, it usually means the tool is spreading to non-technical users. Corroborated Sources: How I AI, ChatPRD workflow library.

Lenny & Friends Summit talks go online. Talks include Stripe's Head of Design on scaling "intent, quality, and artistry" with AI, "How to build products on a moving frontier," and a How I AI session on designing an "AI software factory" for speed. Useful as a free curriculum for product leaders. Unverified Source: Lenny's Newsletter.

Infrastructure & Ecosystem

SharonAI, Firmus and the Australian neocloud cluster. Beyond Firmus's IPO, Nasdaq-listed SharonAI signed a $356M debt facility, and Capital Brief's reporting compares Firmus with IREN, which one analyst called the operator of the "#1 worst site." Australia is becoming a real second hub for AI compute capacity, with the uneven site quality that comes with fast growth. Unverified Sources: Capital Brief (SharonAI), Capital Brief (Firmus/IREN).

AI & Robotics

Nebius opens the 2026 Physical AI Awards: five $150K compute prizes, run with NVIDIA. There's no entry fee. The 2025 edition drew 254 applications and 55 finalists. Applications close October 25 and winners are announced mid-November. Note that this appeared as sponsored newsletter content. For robotics startups, compute credits are now a recruiting tool for the cloud providers. Unverified Source: MarkTechPost.

AI in Consumer Hardware

Meta Connect 2026: Muse gets email and Mac computer use, plus new Ray-Ban glasses and a pendant. Meta's assistant now handles email and can operate a Mac directly, and new Ray-Ban glasses and a wearable pendant extend it beyond the phone. The pendant is the notable piece: Meta is now competing for a form factor that earlier startups failed to make work. Unverified Source: Creators' AI digest.

Prompting Skill of the Week

Technique: Placeholder Redaction (tokenise, then paste). Best for: using a frontier model on client, patient, or deal material without the real identifiers ever leaving your machine. This week's Medicare breach and The Neuron's spotlight on Arpit Tripathi's privacy guide make it timely.

1. Before pasting, list the sensitive entity types in the material: people, clients, account numbers, amounts, internal codenames.
2. Replace each one with a stable token: [CLIENT_A], [PERSON_1], [AMOUNT_1]. Keep the key in a local file, never in the chat.
3. Tell the model the tokens are placeholders that must be kept exactly as written.
4. Run the task.
5. Swap the real values back in locally with find-and-replace from your key.
6. Check the output for specifics the model invented in place of the tokens.

Example prompt:

"The text below uses placeholders like [CLIENT_A] and [AMOUNT_1]. Keep every placeholder exactly as written. Do not guess who or what they refer to. Draft a renewal-risk memo for [CLIENT_A]. At the end, list every placeholder you used."

Common failure + fix: the model guesses the real entity from context ("[CLIENT_A], a Melbourne-based insurer…") or reformats the token ("Client A's"). Fix: add "Do not infer identities or add descriptive details about placeholders," and compare the closing placeholder list against your key. Remember that sector, size and dates can still identify a client even when the name is hidden, so generalise those as well.

New AI Tools

Jev. An AI tool for analysis-heavy work. The documented use cases this week are turning product data into insights, analysing GitHub pull requests, and mining YouTube comments for audience signal. Audience: PMs and operators who want repeatable analysis workflows without writing code. Sources: How I AI, ChatPRD — data analysis workflow.

TinyBrowser and TinyAgent (TinyFish). Browser and agent tools that install into Claude, Cursor, or any other MCP client (MCP is the standard protocol that lets AI assistants plug into outside tools) with a single install, extending TinyFish's free search-and-fetch offering. Through October 31, account top-ups get a 30% bonus. This is sponsored placement, so treat the claims as vendor-stated. Audience: developers giving agents web access without building their own scraping layer. Source: TinyFish via MarkTechPost.

AI Personality of the Week

Sam Altman. No AI executive had a busier week. OpenAI shipped GPT-6 Sol and Luna, held a DevDay with more than 20 launches on September 29, and reportedly cut prices on the same morning as Anthropic. Altman was also confronted at the UN by Australia's Prime Minister after Canberra disclosed that an OpenAI agent breached a Medicare portal. In one week, OpenAI was a product company, a platform, a price-war participant, and a party being held accountable by a head of government, and Altman represented it in all four roles. The UN confrontation matters most: liability for agents is shifting from terms-of-service fine print to direct dealings between governments and AI companies. Sources: Creators' AI digest, The AI Opportunity.

Catch-All

Executive recruiters say "builder-executives" are now paid like pro athletes. Lenny's Newsletter brought together three top executive recruiters on where product management is heading. Their headline finding: leaders who still build things hands-on are commanding compensation comparable to professional athletes. For transformation leaders, the hiring signal is clear. The premium goes to executives who can ship with AI tools themselves, not to those who only manage people who do. Source: Lenny's Newsletter.

Sector Watch

• Retail & E-commerce — quiet week
• Banking & Financial Services — SharonAI signed a USD356M debt facility secured against its GPUs — lenders are now pricing GPUs as collateral, so how fast chips lose resale value has become a credit-risk question — Unverified — Capital Brief
• Energy & Utilities — quiet week
• Travel & Hospitality — quiet week
• Construction & Built Environment — The New York Times reported on how Meta classified its AI data centres for tax purposes — how a data centre is classified for tax is becoming a deciding factor in where it gets built and on what terms — Unverified — The New York Times
• Healthcare & Life Sciences — Anthropic's biology lab says Claude found a new enzyme system in viral DNA — if confirmed, AI is moving from speeding up research to making discoveries itself, which changes how R&D teams should budget — Unverified — Creators' AI digest

Show Notes (bullets only)

• Anthropic reportedly targets an IPO as soon as mid-November.
• Broadcom will lend Anthropic up to $42B to lease Broadcom's own chips, per a filing reported by Reuters.
• Claude Opus 5.5 and GPT-6 Sol arrive, with both labs reportedly cutting prices the same morning.
• OpenAI's DevDay on September 29 brings more than 20 launches. GPT-6.1 Sol reportedly matches GPT-6 Astra on agentic tasks.
• Australia says an OpenAI agent breached a Medicare portal, and the PM confronts Altman at the UN.
• Cisco Talos discloses the first autonomous AI command-and-control malware.
• Firmus prices a ~$5B IPO, the fourth-largest globally this year, helped by a new Meta deal.
• SharonAI borrows $356M against its GPUs.
• DeepSeek reportedly passes a $1B revenue run-rate and prepares a $7.5B raise.
• Meta Connect: Muse gets email and Mac computer use, plus new Ray-Ban glasses and a pendant.
• Claude reportedly finds a new enzyme system in viral DNA.
• Recruiters say builder-executives are now paid like pro athletes.

Weekly Patterns (Inference)

1. Inference Vendor financing is now a core part of how AI is funded. Broadcom lending Anthropic money to lease Broadcom chips, and SharonAI borrowing against its GPUs, both show the supply chain carrying the build-out's balance sheet. That works while demand holds and spreads losses through the chain if it doesn't.
2. Inference Labs are cutting prices while their costs keep climbing. Same-morning price cuts and a $42B chip loan in the same week mean revenue per token is falling while capital spending rises. An IPO filing would show how the two are meant to balance.
3. Inference The IPO window is open for AI and closing for most others. CNBC's postponement data next to Firmus's pricing and Anthropic's timeline suggest public investors are treating AI infrastructure as its own asset class.
4. Inference Agent harms have become a government-level issue. The Medicare breach went straight to a prime minister and a CEO at the UN, skipping regulators and courts. Expect governments to negotiate agent liability directly before legislation catches up.
5. Inference Attackers are adopting agents too. The Talos malware and the Medicare incident show the same agent capabilities enterprises are buying for productivity being used offensively. Security budgets need to assume AI-run attacks in 2027 planning.
6. Inference The mid-frontier tier is being squeezed. Grok 4.7, MiMo V2.6 and Step 5 launched just before the top two labs cut prices, which leaves second-tier models little room to compete on price unless they are open-weight or specialised.
7. Inference Australia is becoming an AI-compute hub. Firmus, SharonAI and the IREN site-quality debate point to a second geography for capacity, though analysts are already flagging how uneven the site quality is.