Racks built for agents, AI code breaking production, and a quantum firm buys agent security

AMD unveiled Helion on October 12 at the Open Compute Project Global Summit in San Jose. CEO Lisa Su presented it as a rack reference design for agentic AI that extends Helios, AMD's GPU-heavy rack-scale platform launched earlier this year. AMD plans to contribute the Helion specification to the Open Compute Project.

Helion comes in two configurations. The hyperscale version is CPU-only, with up to 32,768 Zen 6 cores from sixth-generation EPYC 'Venice' processors, and hands model inference off to separate accelerator racks. The enterprise version mixes EPYC CPU blades with up to 16 Instinct MI350P GPUs in a single rack, adding GPU capacity in pairs of 8-GPU servers.

Both use a 21-inch OCP ORv3 rack with standard modular compute boards, liquid or air cooling, up to 102Tbps of programmable networking and an open-source software stack. AMD described server types tuned for different agent workloads: dense 256-core servers for personal agents, balanced servers for enterprise agents and high-frequency 96-core servers for technical agents. Configurations from Cisco, Dell, HPE, Lenovo, Supermicro and others are expected from November.

Why CPUs? AI agents spend much of their time not running the model but doing everything around it: calling tools, running code, querying databases, browsing, waiting on APIs and orchestrating sub-tasks. That work runs on general-purpose cores. A rack designed around thousands of cores for agent 'busywork', with GPUs or separate racks for the model itself, reflects how agent workloads actually behave.

No independent benchmarks or detailed availability dates were published at launch. But Helion signals that hardware vendors now see agents as a distinct workload class, which will shape data-centre purchases, cloud instance types and software design over the next few years.

Why it mattersHardware is now being designed around how agents behave, with lots of tool calls and orchestration, which will change what AI infrastructure costs and how it's bought.

AI coding tools are speeding up how fast code gets written, but not how fast it gets safely shipped. That's the conclusion of a survey published by Undo, a company that makes debugging tools, of 300 senior engineering leaders at large US and UK companies building mission-critical software. Questions were fielded in July and August 2026, and LeadDev reported the results.

The headline numbers are stark. 81% of leaders said their organisation had a production incident or outage linked to AI coding tools in the past six months. More than nine in ten reported serious defects, poorly optimised code or test escapes reaching production, and 93% said an AI hallucination had led their team to misdiagnose a root cause.

The productivity picture is mixed at best. 79% said their release cycles are no faster than before AI. Engineers now spend an average of 16.9 hours a week debugging versus 9.8 hours writing code, and 35% of AI-generated code reaches production before teams fully understand it. Undo's CTO Mark Williamson summed it up: AI has helped the input end of the pipeline a lot, the output end not so much, at times acting like a 'code machine gun'.

The survey comes from a vendor that sells debugging tools, so it has an interest in the problem, and it reports aggregate results rather than specific incidents. But the findings match a broader pattern: AI has moved the bottleneck from writing code to verifying, understanding and maintaining it, especially in long-lived systems full of undocumented assumptions.

The recommendation is to measure what matters, defect rates and time to resolution rather than lines of code produced. For businesses, it's a signal that review, testing and observability skills are rising in value just as code generation gets cheaper.

Why it mattersAs AI writes more code, the value shifts to the people and tools that verify it, which is where demand and budgets will follow.

IonQ, the quantum computing company, announced on October 12 that it will acquire Elisity, a zero-trust network security company, in an all-stock deal worth up to $1.3 billion, including up to $150 million in performance-based earnouts. The share count will be set at closing from a 20-day volume-weighted average price within a fixed range, and the deal is expected to close in the fourth quarter of 2026, subject to regulatory approval.

Elisity's platform applies least-privilege rules to control which users, devices, workloads and, increasingly, AI agents and other non-human identities can communicate on an enterprise network. Founded in 2019, Elisity says it has more than 100 customers, including a growing number of Fortune 500 companies. It will keep its leadership and brand after closing.

IonQ's rationale combines two security shifts. One is post-quantum cryptography: IonQ plans to pair its quantum-safe encryption and quantum key distribution work with Elisity's platform, extending Elisity's IdentityGraph with an inventory of cryptographic risk. The other is the explosion of non-human identities, as AI agents, bots and service accounts multiply inside corporate networks and each needs its own access rules.

IonQ also points to access to the microsegmentation market, which it sizes at more than $10 billion. Some secondary coverage reports rapid revenue growth at Elisity, but that comes from less authoritative sources than IonQ's own announcement and filing.

The deal is another sign that 'agent identity', meaning knowing what each AI agent is, what it may access and who it may talk to, is becoming a core security category. Expect more acquisitions as large vendors race to own the control layer for non-human identities.

Why it mattersAI agents are becoming network identities that need their own access rules, and security vendors are paying billion-dollar prices for that control layer.