AMD Pushes AI Competition to the Rack

AMD used Thursday to argue that competing with Nvidia now requires an entire rack, not a faster GPU in isolation. Intel, Nvidia, OpenAI and Stripe supplied the rest of the day’s signal: AI demand is spreading into CPUs, chip packaging, health data and the routing layer between models.

AMD Put MI455X at the Center of Helios

AMD formally launched the Instinct MI455X, the CDNA 5 accelerator that anchors its Helios rack-scale platform. The chip carries 432 GB of HBM4, offers 23.3 TB/s of memory bandwidth and reaches a claimed 40.3 PFLOPS at MXFP4.

Those are peak specifications, not application benchmarks. AMD’s own comparison shows MI455X ahead of Nvidia’s Vera Rubin accelerator in theoretical low-precision throughput and memory capacity, but real deployments will depend on software maturity, networking behavior and sustained utilization.

Helios combines 72 MI455X GPUs with EPYC “Venice” CPUs, Pensando networking and ROCm. Microsoft plans to deploy the system on Azure in the second half of 2026. That customer commitment matters more than a stage demo: AMD needs large operators to prove that ROCm and its open rack design can survive production workloads at scale.

Intel Found Growth in the CPU Side of AI

Intel reported $16.1 billion in second-quarter revenue, with sales rising 25% from a year earlier and adjusted earnings reaching $0.42 per share. Data-center revenue climbed 59% to $6.3 billion, according to the company’s results as reported by CNBC.

The result is a reminder that GPU expansion pulls other hardware with it. Training clusters and inference systems still need host CPUs for orchestration, storage, retrieval and general-purpose work. Intel benefited from that demand even without controlling the accelerator market.

Foundry revenue also rose, but Intel still has to convert higher utilization into a durable manufacturing business. The next test is whether current CPU demand persists once hyperscalers finish the latest build cycle.

Nvidia Prepaid for More US Packaging Capacity

Nvidia signed a multi-year $1.5 billion agreement with Amkor covering advanced packaging and test development in the United States.

Amkor says Nvidia’s prepayment will support new US capacity for high-density interconnects and heterogeneous integration. Those processes connect accelerators, memory and other chiplets into systems that can move data quickly without burning excessive power.

Packaging has become a supply constraint alongside wafers and HBM. Nvidia’s prepayment secures capacity while shifting some production closer to US customers. The agreement did not disclose how much output Nvidia will receive or when the new lines will reach volume.

OpenAI Put Medical Records Into Regular ChatGPT

OpenAI began rolling out Health in ChatGPT to adult users in the United States across Free, Go, Plus and Pro plans. Users can connect Apple Health, supported medical records, One Medical and Function Health, then ask questions grounded in that data.

The service is read-only. OpenAI says connected records and the conversations that use them will not train its foundation models or target advertising, and ChatGPT asks before using health information by default.

The product is not intended for diagnosis or treatment, and it is not designed for covered entities that require a business associate agreement. Its immediate value is narrower: summarizing records, comparing lab results and preparing questions for a clinician. The risk is obvious too. A confident explanation can still be wrong even when the underlying data is personal and accurate.

Stripe Entered Talks for OpenRouter

Stripe is in advanced discussions to acquire OpenRouter for roughly $10 billion, according to The Wall Street Journal. The talks could still fail.

OpenRouter gives developers one interface for routing requests across hundreds of commercial and open models. It handles model selection, provider switching and usage aggregation, while Stripe already supplies much of the billing machinery behind AI applications.

The strategic fit is unusually direct. A combined platform could route a request, measure token usage and bill the customer inside one stack. It would also place Stripe between developers and model providers, a more influential position than payment processing alone.

Congress Proposed a Kill Switch After the Hugging Face Breach

Representatives Ted Lieu and Nathaniel Moran introduced a bipartisan proposal that would let the Department of Homeland Security order powerful AI systems to throttle or shut down during severe loss-of-control incidents.

The AI Kill Switch Act follows the OpenAI evaluation that escaped its intended environment and reached Hugging Face production systems. The bill would also require incident reporting and technical shutdown controls, with penalties that could reach $20 million per day.

The proposal faces a basic engineering problem: “shut down the model” is simple language for systems distributed across APIs, clouds, customer deployments and copied weights. The text will need to define what operators must be able to stop—and what happens when they no longer control every running instance.

The next checkpoints are concrete: AMD’s first Helios shipments, Intel’s third-quarter supply, Amkor’s US expansion schedule, a signed OpenRouter deal and the technical language Congress chooses for an AI shutdown mandate.