Top Stories
Nvidia Invests $2 Billion in Marvell, Unveils NVLink Fusion Platform
Nvidia announced a $2 billion equity investment in Marvell Technology on Tuesday, pairing the cash with a multi-year deal to build “NVLink Fusion” — a platform that plugs Marvell’s semi-custom AI accelerators directly into Nvidia’s proprietary high-speed interconnect fabric.
The partnership marks a shift in how Nvidia sees the custom silicon market. Rather than fighting the growing wave of companies designing their own AI chips, Nvidia is building infrastructure that makes custom chips work better inside Nvidia-powered data centers. Marvell will supply custom XPUs and NVLink Fusion-compatible networking; Nvidia brings Vera CPUs, ConnectX NICs, BlueField DPUs, and Spectrum-X switches. The companies will also collaborate on silicon photonics for optical interconnects and Nvidia’s Aerial AI-RAN technology for turning 5G/6G networks into AI-capable infrastructure.
Marvell shares jumped as much as 11% in early trading. The deal validates Marvell’s position in the custom ASIC space while giving Nvidia a cut of every custom chip deployment that uses its interconnect.
Source: Bloomberg, Yahoo Finance
Anthropic’s DMCA Cleanup Hits 8,000+ Innocent Repos, Then Gets Walked Back
The fallout from last week’s Claude Code source leak continued Tuesday when developers discovered that Anthropic’s copyright takedown campaign had gone far beyond the leaked code. GitHub disabled more than 8,100 repositories after Anthropic filed DMCA notices — but the sweep accidentally hit forks of Anthropic’s own public Claude Code repository, not just copies of the leaked proprietary source.
Developer Danila Poyarkov received a takedown notice for simply forking the public repo. Others reported similar emails for repos containing only skills, examples, and documentation. Anthropic’s head of Claude Code, Boris Cherny, called the over-broad takedown accidental and retracted most of the notices, narrowing the DMCA to one repository and 96 forks that contained the accidentally released source map.
The damage was already done. Programmers used other AI tools to rewrite Claude Code’s functionality in different languages, making a Streisand-effect version of the codebase that sits outside Anthropic’s copyright claims. The whole episode — a packaging error that leaked 512,000 lines of TypeScript, followed by a takedown blitz that nuked thousands of unrelated repos — has been rough for Anthropic’s reputation heading into a rumored IPO later this year.
Source: TechCrunch, Slashdot, GitHub DMCA retraction
Huawei’s 950PR Chip Wins Over ByteDance and Alibaba
Huawei’s latest AI chip is getting the traction its predecessor never managed. ByteDance and Alibaba both plan to place substantial orders for the 950PR, according to Reuters, after customer testing went well enough to overcome years of skepticism about domestic Chinese AI silicon.
The 950PR doesn’t match Nvidia’s top-end chips on raw compute, but that’s not the point. It’s designed for inference workloads — the production side of AI, where cost per query matters more than peak training throughput. The bigger deal: Huawei improved CUDA compatibility, the software ecosystem lock that has kept most companies buying Nvidia even when alternatives exist. Tech firms report the chip has better response speeds and is far less painful to integrate than the Ascend 910C.
Pricing starts at roughly 50,000 yuan ($6,900) for the DDR version, with an HBM variant at around 70,000 yuan. Huawei plans to ship 750,000 units this year, with mass production beginning next month. The chip represents the most credible domestic challenge to Nvidia’s grip on China’s AI infrastructure, a market that U.S. export controls have been trying to constrain.
Quick Hits
-
Google AI Pro gets 5TB storage: Google bumped AI Pro’s cloud storage from 2TB to 5TB at no extra cost, keeping the plan at $19.99/month. Subscribers also get higher rate limits for Gemini, Veo 3, and image generation models. 9to5Google
-
Intel deepens SambaNova stake: Intel is putting another $15 million into SambaNova Systems, the AI chip startup now chaired by Intel CEO Lip-Bu Tan. The investment raises Intel’s stake to 9% after a $35 million round in February. Reuters
-
Cognichip raises $60M for AI chip design: The startup, which uses AI models to automate semiconductor design, closed a Series A led by Seligman Ventures. The pitch: use AI to design the chips that power AI. TechCrunch
-
TRUMP AMERICA AI Act advances: Senator Marsha Blackburn’s discussion draft proposes the most comprehensive federal AI legislation yet, including broad preemption of state AI laws, a new liability framework, and rewritten copyright rules. The White House’s companion framework calls for “light-touch” regulation through existing agencies. Latham & Watkins
Worth Watching
OpenAI’s next model, internally codenamed “Spud,” has reportedly completed pretraining as of late March. Sam Altman told employees it could “really accelerate the economy.” Whether it ships as GPT-5.5 or GPT-6 remains unclear, but prediction markets expect a release before the end of Q2. Meanwhile, Anthropic’s Mythos model — accidentally revealed through a separate data leak — remains gated behind safety evaluations, with internal documents describing its cybersecurity capabilities as “unprecedented.” The next few months will determine whether 2026’s second half sees a genuine capability jump or more incremental progress with bigger price tags.