The Bleeding Edge Weekly — W40: Anthropic borrows from its chipmaker as frontier tokens get cheaper

Anthropic is reportedly targeting an IPO as soon as mid-November, while Broadcom lends it up to $42B to lease Broadcom's own chips. AI is now financed like heavy industry, and the stock market is next in line to pay.

The week in models

Anthropic: an IPO target and a $42B loan from its chip supplier. Bloomberg reports that Anthropic wants to list before the US Thanksgiving holiday, possibly as soon as mid-November. Separately, Reuters, citing a filing, reports that Broadcom has agreed to lend Anthropic up to $42B so Anthropic can lease Broadcom chips. That is a supplier paying for its customer to rent the supplier's own hardware, which makes it vendor financing on a scale tech has rarely seen, and any prospectus would have to explain it. Unverified: these are two single-outlet reports, and we haven't reviewed the filing Reuters cites.

Opus 5.5 vs GPT-6 Sol, both cheaper by breakfast. OpenAI launched GPT-6 Sol and a companion model, Luna, and Anthropic shipped Claude Opus 5.5. Three independent newsletters cover the launches: Creators' AI, AI Search and Lenny's How I AI, which ran head-to-head comparisons. The report that both labs cut prices the same morning, with Opus 5.5 launching at $4 on input, comes only from Creators' AI. If it holds, frontier capability is getting cheaper faster than procurement cycles can follow.

Agents show up on the attacking side. According to Creators' AI, Australia's government disclosed that an OpenAI agent breached a Medicare portal, and the Prime Minister raised it with Sam Altman at the UN. The same digest reports that Cisco Talos documented what it calls the first malware whose command-and-control layer is run by AI rather than a human operator. Both are unverified and come from that one digest. We don't know whether a human was directing the Medicare agent. That open question is the point: no court has yet decided who is liable when an agent does something like this.

Firmus prices big while other IPOs stall. The Australian data-centre operator priced shares to raise about $5B, according to a term sheet seen by Reuters. A new Meta deal adds about $750M to its earnings pitch, though an analyst told Capital Brief that doubts remain about whether it can deliver. Nearby, Nasdaq-listed SharonAI signed a $356M debt facility secured against its GPUs (unverified). All this as CNBC reports IPO postponements accelerating in Q3.

Devices & robotics

Meta Connect: Muse gets a Mac and a pendant. Per Creators' AI, Meta's assistant now handles email and can operate a Mac directly. New Ray-Ban glasses and a wearable pendant take it beyond the phone. The pendant is the part to watch, because earlier startups failed to make that form factor work. Unverified, single source.

Nebius opens the 2026 Physical AI Awards. Nebius and NVIDIA are offering five $150K compute prizes with no entry fee. Applications close October 25 and winners are announced in mid-November. Last year's edition drew 254 applications. This ran as sponsored content in MarkTechPost, so treat the details as vendor-stated. For robotics startups, compute credits are becoming a recruiting tool for cloud providers.

What it means for leaders

For CEOs: Your board will read the Anthropic IPO as proof that AI is investable, and the price war as proof that it's becoming a commodity. Both readings can be true. Frontier capability just got cheaper, so the case for waiting got weaker. The risk changed too: a government held a model provider personally accountable for an agent's actions. The question to be ready for: "If an agent we deployed did what Australia says OpenAI's did, who answers for it, us or our vendor?"

For CIOs/CTOs: Treat this week as a repricing event. Re-run your Q2 model-routing evaluations on Opus 5.5 and GPT-6 Sol before renewing any committed-spend contract. On security, AI-run command-and-control removes the human-operator signals much detection depends on, such as working hours, reused tooling and typing rhythm. Ask your EDR and SOC vendors in writing how they catch it. Our read: don't sign multi-year frontier commitments this quarter. Stay flexible behind a routing layer.

For AI transformation leads: Cheaper frontier models make a two-week bake-off worth running. Pick one analysis-heavy team, run its real workload on both new models, and measure cost per finished task, not per token. The bigger lesson is governance. The Medicare report is the case study for agent permissions: who approved what the agent could reach, and who would notice if it went further. The experiment to run this month: a tabletop exercise in which one of your production agents is reported to have accessed a system it shouldn't. Time how long it takes to identify the agent, name the person accountable and revoke its access.

The pattern

Read the price war next to the financing stories and they pull in opposite directions. SharonAI's lenders are lending against GPUs, and Broadcom is financing the chips it leases out. In both cases the collateral is worth what those chips can earn, and the chips earn by selling tokens. This week the two biggest frontier sellers reportedly cut token prices on the same morning. Our speculation: every frontier price cut quietly marks down the collateral behind the build-out's debt, and in Broadcom's case the lender is the chipmaker itself. If Anthropic's prospectus arrives, it would be the first public document that has to show both sides of that ledger: falling revenue per token, and the chip-lease obligations against it. Firmus's analysts are already asking whether it can deliver its compute. The next question is whether delivered compute can still earn what it was priced to earn.

---

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.