A Wall Street Bank Decides Where Its Inference Runs
The clearest sovereignty signal of the week did not come from a regulator or a European lab — it came from JPMorgan Chase. Alongside disclosing a $1 billion Series F at an $11 billion valuation, AI-chip maker SambaNova named the bank as an inference-infrastructure partner in a multi-year deal announced at the RAISE summit in Paris. The arrangement puts SambaNova's SN40L and SN50 accelerators inside JPMorgan's own data centres to run production AI entirely on-premises — the models never leave the bank's walls, and every request carries an auditable trail the bank controls end to end.
For a firm with JPMorgan's compliance surface to pull production inference off public cloud and back behind its own perimeter is a statement about the shape of enterprise AI, not just a hardware win for an Nvidia challenger. The reason a bank does this is the reason a European hospital or ministry does: the sensitive input — a customer record, a trade, a patient file — should not have to travel to someone else's servers, under someone else's jurisdiction, to be reasoned over. What makes the deal notable is that the pull toward controlled inference is now coming from the most demanding commercial buyers on their own initiative, not only from regulators forcing the issue. The strategic question for a European enterprise is narrower than "cloud or not" — it is whether the layer running your models is one you can point to on a map and audit, or one whose location and legal exposure you inherit from a vendor.
After JADEPUFFER, the Market Reaches for the Agent's Off Switch
Yesterday's lead was JADEPUFFER, the first ransomware operation run end to end by an autonomous LLM agent. The follow-through arrived fast, and from two directions at once. On the product side, a wave of "agentic runtime security" tools moved into general availability this month — First Recon's AI Security Runtime inspects every human-to-model, agent-to-tool and agent-to-agent interaction and applies policy inline before data reaches a model, logging each decision as audit-ready evidence; Codenotary shipped AgentMon 3 in the same window. The framing that unifies them is "Agent Zero Trust": treat an autonomous agent as a potential insider threat with a scoped identity and continuous monitoring, not as a trusted user.
On the policy side, the European Commission presented its Action Plan on Cybersecurity and Artificial Intelligence on 7 July, acknowledging directly that advanced models can be exploited to automate attacks "at unprecedented speed and scale" — precisely what JADEPUFFER demonstrated. The plan tasks ENISA and the Joint Research Centre with a secure testing platform for critical-sector operators and, tellingly, commits to a European Blueprint for secure access to advanced AI systems for cybersecurity purposes, alongside continued investment in sovereign compute. The common thread with the runtime-security tools is that governance is migrating from a documentation exercise to the execution path itself. An agent whose reasoning runs inside a boundary you control can be inspected and stopped; one that reasons somewhere you cannot see is exactly the insider you were warned about. Data-in-cognition, it turns out, is a blue-team problem too.
Anthropic Passes OpenAI on Revenue, and the Enterprise Bet Pays Off
Fortune reported this month that Anthropic has overtaken OpenAI on self-reported revenue, running at roughly $47 billion annualized against OpenAI's projected $25–33 billion for 2026. The two companies got here on opposite business models: Anthropic sold enterprise contracts — Claude embedded in customer service, code generation and document analysis under large recurring deals — while OpenAI leaned on consumer subscriptions. The enterprise book is the one that compounds, and it carried Anthropic past a rival with far more name recognition and, by most counts, more consumer users.
The read for European buyers is less about which US lab is ahead and more about what the scoreboard reveals. The value in this market is concentrating into a handful of enterprise-first American vendors, which is the same concentration risk that surfaced last week in Apple's trade-secret suit against OpenAI — only now visible from the revenue side rather than the courtroom. Standardising regulated workflows on a single frontier provider means betting the workflow on that provider's jurisdiction, pricing decisions and corporate fortunes, none of which the customer sets. The durable hedge is not picking the winning lab but keeping the model layer swappable over infrastructure you operate, so that a vendor's ascent — or its next dispute — is not automatically your problem.
Quick Hits
- Grok 4.5's coding number is settled; its EU launch still isn't. xAI's flagship, publicly released 8 July via Grok Build, Cursor and the API console, remains unavailable in the EU with only a vague "mid-July" target and no stated reason. xAI's own launch table puts it at 64.7% on SWE-Bench Pro — mid-tier, well behind Fable 5 — but very cheap at $2/$6 per 1M tokens.
- Open weights had a quiet week — and the incumbents still lead. No major new open checkpoints dropped over the weekend; Alibaba's Qwen 3 235B-A22B continues to top the all-round open-source leaderboard under Apache 2.0, with DeepSeek V4 and Mistral Large 3 behind it. For EU buyers the takeaway is unchanged: the price advantage of Chinese open weights is real, but capturing it without a jurisdiction swap means self-hosting the weights, not routing to the vendor's API.
- EU GPAI enforcement powers still switch on 2 August. The Commission's penalty powers over general-purpose-AI providers — including fines — activate on schedule regardless of the still-unpublished Digital Omnibus, which continues to await its Official Journal entry two weeks after the Council's 29 June green light. High-risk Annex III obligations slip to December 2027; the transparency and GPAI clocks do not.
- China's anthropomorphic-AI rules take effect this Wednesday. Interim measures covering agents that simulate a person for sustained emotional interaction become enforceable 15 July — the third distinct regulatory model (EU transparency, US access-gating, China product limits) now running in parallel. First enforcement actions are the thing to watch.
- Anthropic extends Fable 5's included window again. The plan-included allowance was due to end today, but Anthropic pushed the deadline to 19 July — its second extension in a week, landing days after OpenAI's Sol release. Beyond the allowance, usage bills at roughly $10/$50 per 1M tokens; the drift toward pay-as-you-go metering remains one more argument for inference you run yourself.