> ## Content Index > Fetch the complete content index at: https://www.james-corcoran.com/llms.txt > Use this file to discover other available public pages before exploring further. # Banks confront the economics of AI agents. Inference performance a systems issue. Specialist models for cyber and hardware engineering. - URL: https://www.james-corcoran.com/banks-confront-the-economics-of-ai-agents-inference-performance-a-systems-issue-specialist-models-for-cyber-and-hardware-engineering/ - Published: 2026-11-11T07:04:18.000Z - Updated: 2026-11-11T09:08:42.000Z - Description: The economics of AI are a financial-market story, from banks' agent costs to the financing of data centres and the Anthropic IPO. Memory, storage and specialised accelerators are constraints on scaling AI inference. AI agents for formal software verification and physical hardware design. - Author: James Corcoran - Tags: Weekly 3x3 ## MARKETS ### Banks confront the economics of AI agents Banks are beginning to grapple with the operating costs of deploying AI agents at scale. McKinsey's latest Global Banking Annual Review estimates that token spending could eventually account for as much as 20% of total labour costs. Agent-to-agent communication can consume tokens rapidly, making simply routing requests to cheaper models less effective. Some banks are establishing dedicated AI FinOps teams to monitor usage and focus investment on deployments with measurable returns. [Read more — Reuters](https://www.reuters.com/business/finance/banks-seek-rein-soaring-ai-agent-costs-mckinsey-says-2026-11-10/?ref=james-corcoran.com) ### Banks turn to insurance to manage AI infrastructure exposure The financing of AI infrastructure is creating new challenges for bank balance sheets. Barclays has insured some loans to AI data-centre projects, while HSBC has explored similar arrangements. Credit insurance can transfer 20–50% of a project's loan risk to insurers, allowing banks to increase their lending capacity. More than 70 insurers now offer project-finance credit cover. The economics are not trivial: premiums can absorb 50–70% of the income banks earn on the insured portion of a loan. The development illustrates how the financing of AI infrastructure is creating opportunities well beyond technology companies, while introducing new layers of risk transfer into the financial system. [Read more — Financial Times](https://www.ft.com/content/599932b7-a548-4d45-8ba4-70bdb8d14565?ref=james-corcoran.com) ### Anthropic's proposed $1.4 trillion IPO tests AI market appetite Anthropic is preparing an IPO targeting a valuation of up to $1.4 trillion, according to Reuters. The proposed offering could raise as much as $64.4 billion, with a further $26 billion private placement involving major technology investors including Google, Nvidia and AMD. The scale of the proposed listing reflects the extraordinary capital requirements of frontier AI development. Anthropic's prospectus also highlights the other side of the equation: substantial future computing commitments, competition and the challenge of converting rapid revenue growth into sustainable returns. [Read more — Reuters](https://www.reuters.com/business/anthropic-ipo-raise-much-644-billion-latest-draft-prospectus-shows-2026-11-10/?ref=james-corcoran.com) --- ## MACHINES ### Google expands its TPU portfolio for secure and long-context AI Google has introduced two additional systems in its eighth-generation TPU portfolio: TPU 8tx, designed for confidential inference, and TPU 8i Max, targeting workloads with very large context windows. The 8tx combines Google's TPU technology with AMD host processors and hardware-level confidential computing to protect data while it is being processed. The 8i Max combines two TPU chips in a single package, doubling compute, SRAM and HBM bandwidth relative to the standard 8i. The announcement is notable for its emphasis on system design rather than just chip performance. [Read more — Google Cloud](https://cloud.google.com/blog/topics/ai-infrastructure/tpu-8tx-and-tpu-8i-max?ref=james-corcoran.com) ### KV cache moves beyond GPU memory Supermicro, IBM and NVIDIA have announced a joint infrastructure solution that moves persistent LLM key-value (KV) cache into shared, high-performance storage. KV caches retain intermediate attention states so that models do not need to repeatedly process the same context. As agentic applications and long-context workloads generate increasingly large caches, keeping all that state in GPU memory becomes expensive and restrictive. The solution combines Supermicro all-flash storage, IBM Storage Scale, NVIDIA Dynamo and Spectrum-X networking. In joint testing with a 120-billion-parameter model, the companies report up to 56 times faster time-to-first-token for prompts of 130,000 tokens compared with a system without persistent KV caching. The development reflects a broader shift in AI infrastructure: inference performance increasingly depends on how efficiently memory and model state are distributed across the system, not just on the accelerators themselves. [Read more — Supermicro](https://ir.supermicro.com/news/news-details/2026/Supermicro-Launches-KV-Cache-Solution-with-IBM-and-NVIDIA-Demonstrating-Significant-Inference-Performance-Benefits/default.aspx?ref=james-corcoran.com) ### AI demand drives a dramatic expansion in DRAM revenues Counterpoint Research's Q3 2026 memory tracker reports that global DRAM revenue has increased almost fivefold year-on-year, driven by demand from AI accelerators and data centres, alongside higher pricing. Samsung retained its leading position with 38% market share, while SK hynix recorded sequential revenue growth of more than 40%, narrowing the gap. Micron also benefited from demand for data-centre memory. The figures underline how the AI infrastructure buildout is extending beyond GPUs into the wider memory supply chain. High-bandwidth memory is particularly important to accelerator performance, while growing server-memory requirements add further pressure to supply. [Read more — Counterpoint Research](https://counterpointresearch.com/en/insights/dram-super-cycle-nears-dollar-200b?ref=james-corcoran.com) --- ## MODELS ### OpenAI introduces a specialised model for cybersecurity research OpenAI has introduced GPT-6 Cyber, a model designed to identify and validate software vulnerabilities while remaining within the scope of authorised security research. The model scored 48.5% on OpenAI's ExploitGym evaluation, compared with 32.1% for its predecessor, GPT-5.6 Cyber. In a separate research campaign, it identified 81 previously unknown high-severity Firefox vulnerabilities and developed working exploits for 948 Linux kernel vulnerabilities. The release illustrates the increasingly specialised capabilities of frontier models. It also highlights the importance of alignment: OpenAI reports that GPT-6 Cyber avoided exploiting out-of-scope services in 99.93% of tests conducted without deployment safety prompts. [Read more — OpenAI](https://openai.com/index/introducing-gpt-6-cyber/?ref=james-corcoran.com) ### Google DeepMind uses AI agents to formally verify Signal's cryptography Google DeepMind has introduced AlphaProof Code, an agentic framework used to formally verify key security properties of the code underlying the Signal messaging protocol. The system combines human expertise with multiple AI agents. Researchers first translated Signal's security requirements into mathematical specifications, then agents generated more than six million lines of Lean code to prove that the implementation satisfied those requirements. The work took four weeks, with human experts reviewing the specifications and areas flagged by the system. The significance is the move from using AI to inspect code or suggest potential vulnerabilities towards using it to generate machine-checked mathematical proofs of software correctness. Google DeepMind says the work demonstrates a new approach to verifying widely deployed cryptographic software, although the proofs cover the specific properties and implementation examined rather than every aspect of the system. [Read more — Google DeepMind](https://deepmind.google/blog/introducing-alphaproof-code/?ref=james-corcoran.com) ### Claude designs and builds working electronic devices Anthropic has demonstrated a further extension of Claude's capabilities beyond software engineering. Working with electronics company Diode Computers, it used Claude Fable 5.2 to design three devices: a speakerphone, a voice-assistant hub and a handheld virtual pet. Starting from relatively simple prompts, Claude produced specifications, circuit schematics, PCB component layouts, routed connections, 3D-printable enclosures, firmware and parts lists. Diode manufactured the prototypes from the resulting design files, with all three devices working without electrical modifications. The work is an interesting demonstration of AI agents operating across software and physical engineering workflows. There are still important limitations: the devices remain prototypes and have not undergone full electrical validation, thermal testing or commercial certification. Nevertheless, the ability to move from a high-level description to a functioning hardware prototype points towards a broader role for AI in engineering design. [Read more — Anthropic](https://www.anthropic.com/research/claude-electronics-design?ref=james-corcoran.com)