Three threads converged this week to trace the emerging shape of AI competition: who controls the compute, who owns the training data, and who sets the rules. Today AMD fires the starting gun on a new hardware cycle, Washington escalates its IP-theft dispute with Chinese labs, and a federal court signs off on the settlement that clears Anthropic's path to its IPO.
AMD Launches EPYC Venice and the Helios AI Rack
AMD formally launched the EPYC “Venice” processor today at its Advancing AI 2026 conference in San Francisco, the first high-performance server CPU to enter production on TSMC's 2 nm node. Venice scales to 256 Zen 6 cores, a 33 per cent jump over the 192-core Turin it replaces, with roughly 1.7 times the performance and 1.6 TB/s of memory bandwidth.
Alongside Venice, AMD introduced the Instinct MI455X GPU and the Helios rack-scale platform. A single Helios rack holds 72 MI455X accelerators and 31 TB of HBM4 memory, delivering 3 AI exaflops of peak performance. AMD confirmed a 6 GW supply deal with Meta for Helios rigs, with initial shipments in the second half of 2026. Lisa Su's keynote follows tomorrow, with OpenAI, xAI, Oracle, and Microsoft among the partner organisations on stage.
Venice resets the cost-per-token calculus for on-premises and co-location deployments. If ROCm software maturity keeps pace with the hardware, Helios gives hyperscalers a credible alternative to NVIDIA's GB300 rack for inference at scale. The critical unknown remains software ecosystem depth: AMD has closed the gap on silicon; the compiler and library stack is the next test.
Bessent Threatens Sanctions on Chinese AI Models
US Treasury Secretary Scott Bessent said on Tuesday that Washington would scrutinise Chinese open-weight AI models for intellectual property theft and would use sanctions authority against any lab found stealing from US companies. Bloomberg and CNBC reported the statement as the first time a senior administration official has directly threatened sanctions against AI model makers rather than chip exporters.
The trigger is the intersection of two live disputes: Moonshot AI's Kimi K3 outperforming US frontier models in several coding benchmarks, and Anthropic's June lawsuit accusing operators connected to Alibaba of distilling Claude through roughly 25,000 fraudulent accounts and 28.8 million interactions. The enforcement tools under discussion include export blacklists, federal procurement bars, and security advisories. All travel through cloud providers.
The practical constraint is severe. Washington cannot un-publish a Chinese open-weight model. Once Kimi K3's weights land on 27 July, any organisation anywhere can download and run them on its own hardware. The real policy lever is upstream: whether Washington can raise the cost of Chinese lab training before the next parameter-scaling cycle begins. A sanctions framework that arrives after the weights are published is, at best, a deterrent for the round after next.
Anthropic's $1.5 Billion Copyright Settlement Approved
A federal judge approved Anthropic's $1.5 billion class-action settlement with authors and publishers on Monday, closing the first major AI training copyright case in the United States. Judge Araceli Martinez-Olguin signed the order despite objections from a subset of authors who argued the payout was insufficient.
The settlement pays approximately $3,000 per work across an estimated 500,000 covered titles, the largest known US copyright settlement by dollar amount. The underlying ruling, that training Claude on pirated books constituted fair use but that storing 7 million pirated books in a centralised library did infringe copyright, now functions as the working legal foundation for the US AI industry until a higher court rules otherwise.
For Anthropic, with a confidential S-1 filed in June and an IPO expected before year end, clearing this liability removes one of the most visible risk factors from its prospectus. For every other lab still facing copyright suits, the settlement establishes a pricing floor of roughly $3,000 per text work that will shape every subsequent negotiation in the space.
Kimi K3 Hits a Compute Ceiling; Open Weights Remain on Track
Moonshot AI paused new Kimi K3 subscriptions on 19 July, within 48 hours of the model's launch, after user demand pushed its GPU cluster to capacity. The company is restoring access in controlled batches while scaling its infrastructure. The open-weight release remains scheduled for 27 July under a Modified MIT licence.
MXFP4 quantised weights require approximately 1.4 TB of storage, a 4x reduction from the 5.6 TB FP16 equivalent, making self-hosting feasible for operators running A100 or H100 clusters. TrendForce reports that Microsoft is evaluating Kimi K3 for internal deployment, which would be a significant endorsement from the company that also backs OpenAI.
The subscription pause exposes a structural tension in the open-weight playbook. A lab that publishes its weights removes the monetisation moat it needs to fund the next training run. Moonshot's situation is not a crisis, but it shows that even a frontier-class open-weight model needs a revenue model that can survive the compute cycle it sets in motion.
The through-line across today's stories is constraint: hardware constraint in the 2 nm cycle AMD is opening, legal constraint on training data from Anthropic's settlement establishing a $3,000-per-work floor, enforcement constraint from Washington's inability to close the barn door once weights are published, and capacity constraint from the GPU ceiling Kimi K3 hit within two days of launch. For a senior operator, the question is which constraint binds first in your deployment scenario, and whether your AI vendor has the capital structure to absorb the next round.